Erneuerbare Energien und Energieeffi zienz Renewable Energies and Energy Effi ciency. Ümit Cali

Transcription

1 Erneuerbare Energien und Energieeffi zienz Renewable Energies and Energy Effi ciency Ümit Cali

2 Erneuerbare Energien und Energieeffizienz Renewable Energies and Energy Efficiency Band 17 / Vol. 17 Herausgegeben von / Edited by Prof. Dr.-Ing. Jürgen Schmid, Universität Kassel

3

4 Ümit Cali Grid and Market Integration of Large-Scale Wind Farms using Advanced Wind Power Forecasting: Technical and Energy Economic Aspects kassel university press

5 This work has been accepted by the faculty of Electrical Engineering and Computer Science of the University of Kassel as a thesis for acquiring the academic degree of Doktor der Ingenieurwissenschaften (Dr.-Ing.). Supervisor: Prof. Dr.-Ing. Jürgen Schmid Co-Supervisor: Prof. Dr. rer. nat. Heinrich Werner Defense day 12 th July 2010 Bibliographic information published by Deutsche Nationalbibliothek The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data is available in the Internet at Zugl.: Kassel, Univ., Diss ISBN print: ISBN online: URN: , kassel university press GmbH, Kassel Cover design: Grafik Design Jörg Batschi, Kassel Printed in Germany

6 Abstract Abstract According to the forecast of future energy scenarios, wind power is one of the fastest growing energy sources in Germany and many other countries. Therefore grid and market integration of wind power is becoming an increasingly critical issue. The variability and limited predictability of the intermittent Renewable Energy Sources ( RES ) such as wind and solar power are two major factors that have a major influence on power system operation as well as power market prices. Thus, it is essential to investigate the different impacts of large scale RES on the power market and grid by developing strategies for an effective integration of RES. Large-scale deployment of renewable electricity supplies, especially wind power, requires a well-structured innovative network and market integration concepts such as innovative energy management concepts, utilization of energy storage facilities, active network management, wind power forecasting methods and utilization of optimal intraday trading. Among these concepts, wind power prediction plays a very important role to improve the economic and technical integration capability of wind energy. The aim of this study is to develop advanced wind power forecasting techniques to improve the accuracy of prediction systems using a variety of methods and to investigate the role of advanced wind power forecasting techniques in the grid and market integration of onshore and offshore wind farms. A Multi-Model approach using ensemble Numerical Weather Predictions ( NWPs ) and Multi NWP approach are developed in order to optimize the Artificial Neural Network ( ANN )-based wind power forecasting models. A new approach for an offshore-specific wind power prediction system based on NWP data and oceanographic parameters is presented. In conclusion, energy economical and technical benefits of wind power forecasting are analyzed. I

7 Abstract Keywords: Grid and market integration, wind power, wind power forecasting, offshore, ensemble forecasting, intraday trading, artificial neural networks, multi and single numerical weather prediction, energy meteorology, power markets II

8 Contents Contents Abstract... I Contents... III Figure List... VI Table List... IX Abbreviations... X Acknowledgement... XI 1 Introduction and Summary Motivation Problem statement Purpose of the PhD Work Scope and the Limitations of the Study Development of Advanced Wind Power Forecasting Models The Necessity of Wind Power Predictions State of the Art in Wind Power Forecasting Existing Wind Power Forecasting Methods The Physical Approaches The Statistical Approaches Artificial Intelligence Based Models Combined Wind Power Forecasting Models Comparison of Existing Wind Power Forecasting Models Energy Meteorology and Numerical Weather Prediction Systems Statistical Methods to Measure the Accuracy of Wind Power Forecasting Arithmetic Mean Value Standard Deviation Correlation Root Mean Squared Error Mean Absolute Error Improvement on a Reference Prediction Selected Methodology: ANN based Wind Power Forecasting ANN Meets with Wind Power Forecasting History of ANN Types of Training (Learning Methods in ANN) Neural Network Topologies III

9 Contents Training of Artificial Neural Networks Testing Phase ANN Theory and Mathematical Foundations of ANN Mathematical Modeling of an Artificial Neuron Activation Functions Multi-Layer Network Supervised Learning Multilayer Perceptrons (MLP) and Back- Propagation Rule Local Minima and Momentum Terms Iset s Wind Power Prediction Approach Regional Up-scaling using Online Model Day Ahead and Very Short-term Wind Power Forecasting Smoothing Effect Software Development for Modeling an ANN Based WPF System Comparison of Three ANN Based Wind Power Forecasting Models Conclusion & Future Challenges Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Introduction The Multi-Model Approach The Multi-NWP Approach for Germany The Multi-NWP Approach for Single Onshore Wind Farms The Multi-Scheme Ensemble Weather Prediction Approach The WEPROG Ensemble Weather Prediction Model Evaluation of the Multi Scheme Weather Ensemble Prediction Approach Conclusion Development of New Offshore-Specific Wind Power Forecasting Models Introduction State of the Art in Offshore Wind Power Forecasting Wind Wave Interaction Wave Spectra Spectral Characteristics Site Description of the Horns Rev Wind Farm Wind Measurement Masts Wind Conditions in Horns Rev Wave Measurements in Horns Rev Evaluation of the Offshore-Specific WPF Models Day Ahead Wind Power Forecasting in the Offshore Environment Influence of Several NWP Parameters on the Accuracy of the WPF Integration of Wave Forecasts into the Offshore WPF Model Results of the Day Ahead Offshore-Specific Wind Power Forecasting Model IV

10 Contents Evaluation of the MS EPS Approach Combination Models to Improve the Accuracy of Wind Power Forecasting Comparison of the Day Ahead Wind Power Forecasting Results Evaluation of the Two Hours Ahead Very Short-term Offshore WPF Model Descriptions of the Very Short-term Wind Power Forecasting Experiments Results of the Offshore Very Short-term Wind Power Forecasting Models Comparison of Day Ahead and Very Short-term Offshore WPF Model Conclusion and Discussion Energy Economic and Technical Benefits of Wind Power Forecasting Introduction The Influence of Wind Power Generation on the Electricity Prices Theoretical Approach Practical Approach Overview of the European Electricity Markets SoA in Energy Economic Benefits of WPF and Utilization of Intraday Trading Wind Power Trading Evaluation and Utilization of Intraday Trading Description of the Possible Trading Options Cases studied Comparison of the Results Conclusion Conclusion and Outlook Conclusion Outlook References Acknowledgement ANNEX: Very short-term offshore wind power forecasting experiments V

11 Figure List Figure List Figure 1 : Overview of the existing wind power forecasting models Figure 2: Horizontal grid of a global numerical weather prediction model and more detailed area covered by a local area model Figure 3: A sample power curve of a wind turbine [18] Figure 4: The sketch of an artificial neural network (ANN) used for wind power forecasting Figure 5: Overview of ANN learning methods Figure 6: Supervised learning Figure 7: The structure of a biological neuron Figure 8: The structure of an artificial neuron Figure 9: The structure of an artificial neuron (mathematical model) Figure 10: The model of a multilayer artificial neuron Figure 11 : Illustration of the error vector in the supervised learning method Figure 12 : The process of the adjustment of the weights in the back propagation algorithm Figure 13 : The illustration of the absolute minimum, local minimum and constant weights Figure 14: The graphical user interface of the Wind Power management System ( WPMS ), Figure 15: The structure of a day ahead and very short-term wind power forecasting model using AI and statistical methods Figure 16: Cell squares used by WPMS for regional up-scaling Figure 17: Real and predicted course of wind generation Figure 18: Comparison of day ahead and short-term (2 and 4 hours ahead) wind power forecasting errors (in terms of nrmse) for Germany Figure 19: Smoothing effect for a single wind farm, control zone and Germany Figure 20: Comparison of the normalized RMSE of a wind power forecast for Germany which is obtained by means of the WPMS based on ANN with input data from three different NWP models and with a combination of these models Figure 21: The representative wind farms Meerhof (WF 1) and Meppen (WF 5) in RWE region Figure 22: Comparison of day ahead wind power forecasting error for the WF 1 (2005) Figure 23: Comparison of day ahead wind power forecasting error for the WF 1 (2006) Figure 24: Comparison of 2 hours ahead short-term wind power forecasting error for wind farm 1 (2006) Figure 25: Structure of the 75 member MS-EPS model Figure 26: Sketch of ANN based prediction model for each MS EPS member Figure 27: nrmse of the 75 members for the Wind Farm 1 (Meerhof) and Wind Farm 5 (Meppen) Figure 28: Sketch of the combination model using simple averaging VI

12 Figure List Figure 29: Sketch of the two step ANN combination model Figure 30: Comparison of the nrmse values of wind power forecasts with a single NWP model (from different meteorology service provider), the EPS Mean and the combination models for the wind farm 1 (Meerhof) Figure 31: Development of offshore wind power in the EU Figure 32: Total offshore wind power installed by the end of 2007 according to EWEA`s records Figure 33: Sketch of influences on the wind field over coastal waters Figure 34: Representation of wave spectra Figure 35: Location of offshore wind farm Horns Rev Figure 36: The location of the Horns Rev wind farm and the surrounding measurement platforms Figure 37: Wind Speed Distribution of the M2 at 62m Figure 38: Wind distribution level 30 m at Horns Rev Wind Farm Figure 39: Wind speed versus Hmax in Horns Rev Figure 40: The sketch of all the combinations of the new offshore wind power forecasting models Figure 41: The sketch of the day ahead offshore wind power forecasting model using only NWP variables Figure 42: The structure of the day ahead offshore wind power forecasting model using NWP and forecasted wave variables Figure 43: The general structure of the MS EPS approach Figure 44: Wind power forecasting error (nrmse) for forecasts based on input from 75 the MSEPS ensemble members for the Horns Rev wind farm Figure 45: Results of the combination models by employing the simple averaging and the double ANN (2ANN) methods Figure 46: Comparison of the day ahead wind power forecasting results for the wind farm Horns Rev Figure 47: Prediction error values (in terms of nrmse) vs. improvements of 2 hours ahead wind power forecasting for the Horns Rev wind fa Figure 48: Comparison of probability density function values (ST2_1, ST2_6W vs. ST2_Pers.) Figure 49: Comparison of the prediction error values of the day ahead and 2 hours ahead wind power forecasting models Figure 50: A sample for the operational unit order of some power plants Figure 51: Dependency between the day ahead wind power forecast results and the Phelix-based prices in 1000 MW classes (2006) Figure 52: Overview of the wholesale market Figure 53: General structure of power market and utilization of intra-day market Figure 54: Wind power trading on the day ahead market Figure 55: Wind power trading on the hour ahead market (adjustment by using shortterm wind power forecasting model) Figure 56 : Wind power forecasting results of the various experiments Figure 57: Average Energy revenue Euro per MWh (for the 7150 hours period) Figure 58: Total revenue values according to the various case scenarios Figure 59: The structure of the very short-term offshore wind power forecasting model using the NWP and measured wind power data VII

13 Figure List Figure 60: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data and forecasted wave parameters Figure 61: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data and measured wave parameters Figure 62: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data and measured wind parameters Figure 63: The structure of the very short-term offshore wind power forecasting model using the NWP, forecasted wave, measured wind power and measured wind data Figure 64: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data, measured wave data, measured wind parameters and forecasted wave variables Figure 65: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data, measured wave parameters and forecasted wave variables Figure 66: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data, measured wind parameters and measured wave variables VIII

14 Table List Table List Tab. 1: Comparison of wind power forecasting models Tab. 2: Overview of some existing NWP models Tab. 3: Overview of three wind power forecasting tool features Tab. 4: Comparison of the results from the three wind power forecasting tools Tab. 5 : Main Characteristics of the selected NWP models Tab. 6 : Available Input Parameters from MS EPS Tab. 7: The marine statistics of the wind farm Horn Rev using Hmax, Hmo, Tp and T02 parameters Tab. 8: The description of the day ahead offshore wind power forecasting experiments (without wave forecasts) Tab. 9: The description of the day ahead offshore wind power forecasting experiment DA_10 (with wave forecasts) Tab. 10: The results of the day ahead offshore wind power forecasting experiments for the Horns Rev Wind farm Tab. 11: Overview of the short-term offshore wind power forecasting models Tab. 12: Overview table of the very short-term offshore wind power forecasting results of the Horns Rev wind farm Tab. 13: Overview of the different simulated scenarios Tab. 14: Needed Up and Down regulation values according to the different cases (for the 7150 hours) IX

15 Abbreviations Abbreviations ANN DA DRP DSO DWD ECMWF HA HIRLAM MAE MLP MOS MS EPS nrmse NWP OM OTC PSO RES RES SCADA SNNS SP TSO URP WAM WEPROG WF WPF WT Artificial Neural Networks Day Ahead Down-Regulation Price Distribution System Operator Deutsche Wetter Dienst (German Weather Service) European Center for Medium/Range Weather Forecasting Hour Ahead High Resolution Limited Area Model Mean Absolute Error Multi-Layer Perceptron Model Output Statistics Multi Scheme Ensemble Weather Prediction System normalized Root Mean Square Error Numerical Weather Prediction Online Model Over the Counter Public Service Obligation Renewable Energy Sources Renewable Energy Source Supervisory Control and Data Acquisition Stuttgarter Neural Network Simulator Spot Price Transmission System Operator Up-Regulation Price Wave Model Weather and Energy Prognosis Wind Farm Wind Power Forecasting Wind Turbine X

16 Acknowledgement Acknowledgement It is my pleasure to thank those who have made this thesis possible. First and foremost I offer my sincerest gratitude to my supervisors Prof Dr.- Ing. Jürgen Schmid and Prof. Dr. rer. Nat. Heinrich Werner for giving me the opportunity to work on this interesting topic. Their continuous support, guidance, patience and knowledge while allowing me the room to work my own way have been so crucial for this thesis. I attribute the level of my PhD to their encouragement and effort. I owe my deepest gratitude to Dr. Kurt Rohrig and Dr. Bernhard Lange for their support and beneficial suggestions throughout my work and for giving me the opportunity to work on my PhD at the Fraunhofer Institute IWES (Former ISET e.v.) in Kassel. I am thankful to Prof. Dr. Siegfried Heier for his advice and support as my thesis committee member. I am indebted to many of my former and current colleagues at Faunhofer Institute IWES who supported me. All of you made the working atmosphere so fun. I will miss our many talks and discussions. Your inputs have greatly impacted the way this thesis turned out. So thank you very much! Special thanks go out to Melih Kurt, Nilgün Kablan, Ümit Ciplak, Ali Cebe and Zouhair Khadiri Yazami for your excellent assistance during numerous experiments and Oya Bengi for the final proof reading of my thesis! Furthermore I want to thank Dr. Corinna Möhren, who always lent me her ear whenever I encountered problems during my experiments as well as for doing everything with excellence. Last but not least, I am utterly thankful to my mother (Salime Cali) and father (Zeki Cali), who supported me morally during my studies and always believed the best in me. XI

17

18 Introduction and Summary 1 Introduction and Summary The EU commission has defined various targets in order to decrease carbon dioxide ( CO 2 ) emissions in future while increasing the proportion of electricity generated by renewable energy. In addition to the development of onshore wind power capacity, the electricity production from offshore wind farms is becoming an increasingly promising option to meet targets determined by the EU commission and other several countries. Therefore grid and market integration of wind power is becoming an increasingly critical issue. Integration of great amount of wind power leads to some technical and economic challenges. Thus, it is necessary to overcome these kinds of barriers by employing some innovative techniques such as new generation energy management concepts, utilization of energy storage facilities, active network management, wind power forecasting methods and utilization of optimal intraday trading. Among these concepts, wind power prediction plays a very important role to improve the economic and technical integration capability of wind energy. Accordingly, this study focuses mainly on the importance of advanced wind power forecasting systems and optimal intraday trading of wind power in order to improve the integration potential of onshore and offshore wind farms. This PhD work deals with how different combinations of Numerical Weather Prediction ( NWP ) models influence wind power forecasting as well as the development of an advanced offshore-specific wind power forecasting model using measured wind farm power output, meteorological and oceanographic parameters. During this study, the development of new approaches will be investigated in order to increase the reliability of wind power prediction models. Energy economic benefits of wind power forecasting will be analyzed on the power market in the final part of the study. Beside the direct economic usage 1

19 Introduction and Summary of wind power forecasting, the reduction in the needed balancing power via utilization of the intra-day trading option will also be explored. 1.1 Motivation The importance of a more accurate output of the advanced wind power forecasting models from the perspective of major market players can be summarized very briefly; The TSOs have the responsibility for the secure operation of the electricity system. Indeed, the usage of developed wind power prediction models can be used to optimize power plant scheduling, power balancing and determine the reserve power needs. The application of regional wind power forecasting approaches can be used for the optimal grid operation and congestion management purposes. The DSOs are responsible for maintaining the distribution systems and covering the energy losses. Wind power forecasting tools can be used to accommodate the sudden changes caused by wind farm output in their low-voltage power grids. Wind farm operators are able to sell their power output in countries where the direct marketing is allowed. Therefore, wind power forecasting tools can be used to perform optimal trading operations. 1.2 Problem statement The following issues were the starting point of this PhD work: 1. Selection of methodology: Which methodology can be selected to develop a realistic and effective wind power forecasting model? 2. Optimization of the existing model: What are the possible optimization techniques to improve the accuracy of the existing wind power forecasting models? 2

20 Introduction and Summary 3. Transferring the knowhow to a new environment: Is it possible to transfer the existing onshore specific wind power forecasting experiences to the offshore environment and what kind of additional input parameters can be used in order to develop next generation offshore-specific wind power forecasting models? 4. Technical and economic validation: How can these technical investigations be demonstrated from the energy economic point of view? 1.3 Purpose of the PhD Work The purpose of this study is to develop advanced wind power forecasting models in order to improve the accuracy of prediction systems using a variety of methods and to investigate the role of wind power forecasting in the grid and market integration of onshore and offshore wind farms. 1.4 Scope and the Limitations of the Study In order to achieve this, the following objectives will be studied: I. Development of the advanced wind power forecasting models The aim of a wind power forecast is to estimate the expected generation of wind power from one or more wind turbines. Forecasting of the expected wind power generation plays a key role in integrating the wind power efficiently into the existing electricity system. The first chapter is focused on explaining the importance of wind power forecasting. The overview of the existing wind power forecasting methods will be also stated in this chapter. Especially, the description of the Artificial Neural Network ( ANN )-based wind power forecasting approach is given. ANNs are one of the most popular artificial intelligence techniques which are used for forecasting applications. For the purpose of learning the relationship between meteorological data and wind farm power output, the ANN needs to 3

21 Introduction and Summary be trained with numerical weather forecasts and measured power values from the past. NWP variables are used as input variables for ANN in order to predict the wind power output of the wind farm or cluster for the next minutes, hours or days. ANN represents the multidimensional patterns of the power curve of the wind farm or cluster by using the historical NWP and measured wind power information. The effect of the atmospheric parameters which are used as major input variables of the wind power prediction system will be analyzed. The importance and usage of wind power forecasting models will be outlined in the final part of the chapter. II. Improvement of wind power forecasting accuracy using the multimodel approach The second chapter focuses on the improvement of the existing ANN based wind power forecasting model using several optimization methods. The investigations in this chapter are mainly performed for the onshore wind farms. A similar method will be applied to the offshore wind farm environment in the following chapter. The multi model approach for wind power forecasting is an effective way to reduce the forecasting error. This approach, which is covered in this chapter, consists of two main methods: The multi numerical weather prediction ( NWP ) approach. The Multi-scheme ensemble weather prediction approach. The aim of the Multi NWP approach is to investigate the effect of the different NWP models on the accuracy of the wind power forecast. The use of the Multi- Scheme Ensemble Weather Prediction System ( MS EPS ) was investigated as an alternative to the costly combination of different Numerical Weather Prediction ( NWP ) models in order to improve the accuracy of wind power forecasting. Afterwards, combination methods using statistical methods (e.g. the simple averaging approach) and artificial intelligence based approaches (e.g. the two step ANN approach) were evaluated in order to optimize the MS EPS-based wind power forecasting output. The Multi NWP approach was 4

22 Introduction and Summary performed both for whole of Germany and 2 single wind farms located in Germany. The multi-scheme ensemble weather prediction approach was only performed for two single wind farms. III. Development of new offshore-specific wind power forecasting models Advanced wind power prediction based on Numerical Weather Prediction ( NWP ) is supposed to be a key technology for dealing with the network problems caused by power fluctuations of wind farms. In this chapter, a new offshore-specific wind power prediction approach will be presented. The work was carried out with the measured data from the offshore wind farm Horns Rev in Denmark according to the project High Resolution Ensemble for Horns Rev (HRensembleHR) funded by the Danish PSO Programme The objectives of the study are strongly matching with the pathways determined by the European Wind Energy Technology Platform (TPWind) 1. TPWind is the Europe-wide platform for the crystallization of the wind energy related policy and technology research. The following research issues concerning the offshore wind power meteorology are stated in the Strategic Research Agenda [1] that were published by TPWind: Development of dedicated offshore wind power forecasting including specific offshore meteorology, online measurements and satellite remote sensing data, Use of real-time or near-real time measurements data (e.g. wind farm power measurements, mast measurement, LIDAR or satellite measurements) for very short-term forecasting, system, integration and safe grid operation, 1 5

23 Introduction and Summary Integration of online wind measurements from masts, remote sensing (offshore) and the wind turbines themselves to improve accuracy, Development of Multi-model combined forecasting approaches should minimize prediction errors. The optimization of the offshore wind power forecasting system by using the Multi-Scheme Ensemble Weather Prediction System ( MS EPS ) was developed in two phases. In the first step, wind power forecasts for each of the 75 ensemble members were generated individually with neural networks. In the second step, all predictions from each ensemble members were combined using different statistical methods (e.g. the simple averaging approach) and artificial intelligence based approaches (e.g. the two step ANN approach). In order to further improve this wind power prediction system, the influence of additional oceanographic and atmospheric parameters on the forecast accuracy was investigated. The most influent parameters were then integrated in the prediction system to improve the prediction accuracy. The results of first trials to include these new oceanographic and atmospheric parameters into the ANN-based wind power forecasting system in order to improve offshore wind power will be discussed in this chapter. IV. Energy economic and technical benefits of wind power forecasting In the final chapter, an improved market integration of wind power by employing the intraday markets is represented in order to reduce the need of reserve power and increase the value of wind power on the power market. Energy economic and technical benefits of wind power forecasting will be analyzed in final part of the study. 6

24 Development of Advanced Wind Power Forecasting Models 2 Development of Advanced Wind Power Forecasting Models Abstract This chapter is dedicated to explain how wind power forecasting systems can be developed and utilized to improve the technical and economic integration capability of wind power into the power system. The summary of the existing wind power forecasting models is given in this section. Additionally, in this chapter the description of the Artificial Neural Networks ( ANN ) - based wind power forecasting approach is described. Furthermore, the theory of the ANN is summarized. 2.1 The Necessity of Wind Power Predictions According to the forecast of the future energy scenarios, wind is the fastest growing distributed energy source in many countries. By the end of 2009, approximately Wind Turbines ( WTs ) with an installed capacity of more than 24.8 GW were in operation in Germany. In addition to the installed onshore wind power capacity, the Federal German Government plans to develop the offshore wind power production in the North and the Baltic Sea up to 25 GW in the foreseeable future. This planned capacity will amount to around 15 % of the German s electricity consumption by Intermittent and distributed energy sources like wind and solar energy generators require wellorganized forecasting concepts in contrast to dispatchable distributed energy sources such as fuel cells and microturbines or other conventional power generation technologies. The fluctuating and intermittent behavior of distributed power generation leads to some challenges in power system operation. Integration of this great amount of wind power into the conventional electricity grid will require some additional control techniques to eliminate the technical barriers. Balancing the grid load is one of the main tasks of the TSOs. Therefore, TSOs have to predict potential unwanted effects of their grid 7

25 Development of Advanced Wind Power Forecasting Models area very precisely. For the conventional power generation this is a relatively easier process. However, predicting the wind power generation is a complex issue which has to be done by TSOs with great accuracy to minimize the cost of balancing power. Currently, TSOs have been using an online monitoring system to determine the amount of generated wind power in each control zone which is operated in 15 minutes intervals in Germany. Day ahead and shortterm wind power prediction tools are used by TSOs in order to integrate the generated wind power into the power system effectively. The necessity of wind power forecasting models is not restricted only by TSO but also, these systems are used by other power market actors such as DSOs and the wind power operators. This chapter focuses mainly on wind power prediction. In addition to the prediction of wind power generation, it is also essential to predict photo-voltaic energy and other distributed energy generation technologies using similar forecasting techniques. Please note that, this study mainly deals with wind power forecasting. Each of four German TSOs has a national quota of wind power which they have to balance system according to the German Renewable Energy Act ( EEG ) mechanism. German TSOs are using additional control power that is generally applied in order to compensate the sudden deviations between the load and generation [2]. The TSOs use day ahead predictions in order to determine the amount of electricity to be exchanged between their grid control areas on a daily band delivery basis. In order to overcome the challenges caused by high amount of wind power, wind power forecasting has become an integral part of the power system in Germany and many other countries. For this purpose, like ISET e.v. (now Fraunhofer IWES) several institutes and companies have developed wind power management and prediction concepts. ISET s wind power forecasting model is one of the most successful models in this category. WPMS (ISET s wind power forecasting system) has been used by three German transmission system operators ( TSOs ) for several years. The system is based on several representative online measurements - which are distributed all across the country and the numerical weather prediction data for the same grid points in order to determine the current and expected wind power value. 8

26 Development of Advanced Wind Power Forecasting Models The recent advances in short-term wind power prediction are discussed in this chapter. The influence of the combination of different Numerical Weather Prediction ( NWP ) models on the wind power forecast and advanced offshore-specific wind power forecasting using meteorological and oceanographic parameters will be investigated in the following chapters. 2.2 State of the Art in Wind Power Forecasting Several institutes and companies have developed a variety of wind power forecasting systems in the last decade. The existing models are based on a three dimensional weather models and numerical weather prediction (NWP) models such as German Meteorology Service ( DWD ), High Resolution Limited Area Model ( HIRLAM ) and the UK Met. Office Meso-scale Model ( UK MESO ). The actual wind power forecasting methods and systems that are currently under research and in operation are summarized by Giebel [3]. Short-term wind power forecast horizons can be several minutes (e.g. 15 minutes) and up to several days. NWP parameters, such as wind speed, wind direction, air pressure and humidity are the main input variables for the classical day ahead wind power forecasting models. More advanced wind power forecasting systems based on nested or multi NWP and ensemble forecasts are able to provide more accurate predictions. The Ensemble multi NWP model approach is applied on onshore and offshore wind farms in this work. New generation offshore-specific wind power forecasts will use additional oceanographic variables like wave height and wave periods. The application using oceanographic variables in addition to the meteorological variables as input parameter for offshore wind power forecasting system will be conducted in this study. 9

27 Development of Advanced Wind Power Forecasting Models 2.3 Existing Wind Power Forecasting Methods The main goal of a wind power prediction model is to develop a relationship between the numerical weather prediction information (especially wind speed and wind direction) and the power output of the wind turbine to generate the wind power forecasting time series for the very near future. There are mainly three types of wind power forecasting model: Physical models Statistical models Artificial intelligence based models The Physical Approaches NWPs provide the weather forecasts at specific grid nodes in a certain area. Therefore, it is necessary to adapt the weather forecasts at the desired location and at the wind turbine hub height. Physical models use detailed physical equations including the description of the lower atmosphere in order to transform the wind speed given by the wind prediction of NWP model over a numerical grid to the local (on-site) conditions of the wind farm. That means that the main goal of the physical wind power forecasting models is to transform the NWPs to the wind turbine hub height and the location of the wind farm by using physical considerations concerning the terrain information such as roughness and orography and by modeling the shadowing effects of the wind farm. One of the first physical models was developed by Lanberg [4], [5] at the National Laboratory in RISOE in the early 1990s The Statistical Approaches Statistical models are usually based on measured information and NWP data. These types of models use statistical methods in effort to establish the relationship between the weather forecasts and the wind power production data from the history and generate the wind power forecasting time series for 10

28 Development of Advanced Wind Power Forecasting Models the future. Statistical wind power forecasting models include linear and nonlinear statistical methods. Statistical models usually consist of autoregressive statistical methods to gauge the persistent behavior of the wind. On the other hand, stochastic behavior of wind uses non-linear methods to transform the meteorological variables. SIPEOLICO and WPPT are two examples of statistical wind power forecasting models Artificial Intelligence Based Models Similar to the statistical models, Artificial Intelligence ( AI ) - based models also derive a relation between the NWP data and measured wind power information from past to predict wind power time series for the future. However, AI-based models use some AI learning methods instead of clearly defined statistical formulas to describe non-linear and complex problems. Artificial intelligence based models can be also defined as black-box models which are able to predict the wind power output without knowing the physical conditions of the wind farm. A variety of artificial intelligence methods are used to predict wind power output such as: Artificial neural networks ( ANN ) Fuzzy- ANN Support vector machines ( SVM ) Nearest neighbour search ( NNS ) ISET s Wind Power Management System ( WPMS ) model is an example of AI based wind power forecasting model that uses artificial neural networks for the prediction. 11

29 Development of Advanced Wind Power Forecasting Models Combined Wind Power Forecasting Models These models are the combination of the statistical, physical and AI based WPF models. The following figure summarizes all types of WPF models. Figure 1 : Overview of the existing wind power forecasting models Comparison of Existing Wind Power Forecasting Models A sufficient amount of measured wind power output and numerical weather prediction data is essential for both statistical and AI-based models in order to generate a reliable wind power forecasting time-series. As discussed in [6] the systematical errors are reduced in both statistical and AI-based models in comparison to the physical models if the statistical and AI based predictions are adapted according to the location of the wind farm. The following table summarizes some of the important wind power forecasting models: 12

30 Development of Advanced Wind Power Forecasting Models Tab. 1: Comparison of wind power forecasting models Model Name Provider Method Country WPMS ISET ANN based Germany Previento Uni. of Oldenburg, En. & Met. Systems Physical Germany Prediktor RISOE, DTU and Eltra Physical (WaSP) Denmark WPPT IMM, University of Copenhagen Statistical Denmark Zephyr RISOE Physical & Statistical Denmark SIPREOLICO University of Carlos III, Red Electrica Statistical Spain ANEMOS RISOE, ARMINES, RAL Physical & Statistical EU AWPPS Armines, Ecole des Mines de Paris Fuzzy ANN based France LocalPred CENER Adaptive Regression, FL Spain Casandra MOMAC Group Physical Spain 3 Tier 3 Tier Statistical USA WEPROG WEPROG Ensemble Ireland and DK GH Forecaster WEPROG Hybrid UK Scirocco Aeolis Forecasting Services Hybrid The Netherlands OWPMS ISET, University of Kassel ANN, Offshore DE and DK

31 Development of Advanced Wind Power Forecasting Models WPMS is a prediction model that combines measured wind power data and NWP from DWD and other national NWP providers using ANN. WPMS is developed by ISET and has been in operation in 3 German TSOs for several years. The predicted wind power output of WPMS is generated on an hourly basis and the forecasting horizon is between 1 hour and several days. [7], [8] Previento is a short-term wind power forecasting model based on meteorological wind predictions and was developed at the University of Oldenburg. The predicted wind power output of Previenato is in hourly resolution and the forecasting horizon is between 1 hour and several days. The accuracy of this model can be improved by using measurement values from wind farms. [9], [10], [11] Prediktor is a physical short-term prediction model based on meteorological wind predictions where the wind is transported to the surface using the geotropic drag law and logarithmic wind profile. Prediktor uses the WaSP program that takes local conditions into account. This model was developed by RISOE National Laboratory in Denmark. Shadowing effects are also taken into account by using RISOE PARK program. Finally, Model Output Statistics ( MOS ) methods are applied to reduce general errors. [12], [13] WPPT is a short-term forecasting model based on advanced time series analysis which is applied for forecasting the expected wind power output in larger areas, using online wind farm data of representative wind farms among the total population of wind turbines in the selected area. WPPT was developed by IMM at the Technical University of Denmark in co-operation with ELSAM and ELTRA. The predicted wind power output is generated every 30 minutes and the forecasting horizon is between 1/2 hour and several days (36 hours). [14], [15] Zephyr is a short-term forecasting model using WPPT in order to combine wind power data and NWP predictions. The main objective of the Zephyr model is to merge statistical and physical-based wind power forecasting models. The Zephyr project was developed to combine the two existing prediction systems (WPPT 14

32 Development of Advanced Wind Power Forecasting Models and Prediktor) and to take advantage of JAVA client implementation. This project was developed by RISOE, IMM and all Danish utilities. [16], [17] SIPREOLICO is a short-term wind power prediction model based on statistical and combined forecasting methods. The forecasting horizon is between 1/2 hour and several days. This model was developed at the University of Carlos III and Red Electrica in Spain and has been in operation since [18] ANEMOS is an EU-wide research project which aims to develop more accurate and advanced wind power forecasting models for onshore and offshore wind farms by combining existing models. The 22 partners participated in this project from seven countries. 2 AWPPS (ARMINES Wind Power Prediction System) is a short-term prediction model based on fuzzy adaptive neural networks and developed at Armines / Ecole des Mines de Paris. This model has been in operation in Crete, Ireland and Azores since The predicted wind power output is generated every minutes and 1 hour and the forecasting horizon is between several hours (4-6 hours) and several days (72 hours). 3 LocalPred is a statistical wind power forecasting model which is based on meteorological wind predictions and uses adaptive regression and fuzzy logic methods. This model was developed by CENER and has been in operation in a number of wind farms in Spain since [19] Casandra is a physical wind power prediction model using a mesoscale model and has been in operation in Spain since This model was developed by MOMAC Group. The forecasting horizon is up to several days (72 hours). [20]

33 Development of Advanced Wind Power Forecasting Models 3 Tier is an US company that has developed a wind power forecasting model based on neural networks and support vector machines using off-site wind measurements. This model provides hour, day and week ahead forecasts. 4 WEPROG is a hybrid model which consists of two models - A wind power prediction model and a weather prediction model based on multi-scheme ensemble prediction ( MS EPS ) technique. MS EPS is a short-term ensemble wind power forecasting model based on a limited area NWP. This model was developed at the University College Cork. GH Forecaster is a hybrid WPF model which uses multi-parameter statistical regression algorithms to transform global NWP with adequate geographical resolution and site data to the site-specific models. This model was developed by Garrad Hassan in the UK. Scirocco is a hybrid WPF model. The wind power forecast is an output of a model chain with consecutive steps from statistical and physical models. This model was developed by Aeolis Forecasting Services. Scirocco adapts itself to the local wind farm conditions and characteristics within the first months of operation. OWPMS (Offshore Wind Power Management System) is an extended version of WPMS. OWPMS was developed for predicting the wind power output of offshore wind farms using the ANN method. In order to increase the accuracy of the system ensemble prediction methods were used. Besides the classical inputs such as NWP prediction data, wave predictions, wind and wave measurements are used as additional input variables More detailed description of the model will be given in the following chapters. Please note that this is not the official name of this model. This model is developed by the author and named as OWPMS. 16

34 Development of Advanced Wind Power Forecasting Models 2.4 Energy Meteorology and Numerical Weather Prediction Systems Weather predictions from NWP models are the most fundamental inputs for all wind power forecasting models. NWP uses mainly current weather conditions as input into mathematical equations such as fluid dynamics and thermodynamics in order to predict the state of the atmosphere at a certain time in future. All available data such as meteorological measurements, observations and satellite information are used as input for a global NWP model. In essence, the global models are cruder in comparison to the local NWP models. The following figure shows the horizontal grid of a global numerical weather prediction model and a more detailed area covered by a local area model. Figure 2: Horizontal grid of a global numerical weather prediction model and more detailed area covered by a local area model 6 Local area models ( LAM ) are used in order to provide more accurate weather forecasts, which focuses on a certain part of the globe with higher resolution. Due to long computational times, the global NWP models have less spatial resolution. Lokal Modell ( LM ) is the local NWP model performed by DWD (German Meteorology Service) that covers central Europe with a 7 km horizontal resolution. The recent models by DWD have 2.8 km and 1.1 km resolution by one-way interactive nesting without introducing urbanization of physiographic parameters. [21] The LM has a time resolution of one hour and 48 hours forecast horizon. Every local NWP model (such as Met Office, WEPROG, etc.) has its own structure of spatial and time resolution. Some models such as LM run twice a day at 00 UTC and 12 UTC. On the other hand, the other models may run 6 Source: DWD 17

35 Development of Advanced Wind Power Forecasting Models more than twice a day. Covered area, accuracy, forecast horizon, spatial and temporal resolutions are important criteria for the wind power forecast providers in order to have more accurate and reliable predictions. The following table gives an overview of different NWP models: Tab. 2: Overview of some existing NWP models Model Name Resolution Model Horizon Frequency DWD 1.1, 7 & 6 km 48 hrs.to 174 hrs. 12 Hourly MS EPS 6 & 45km Up to 72 hrs. 6 Hourly ECMWF 40 & 80km 3 days to 10 days 6 Hourly UK Met Office Approx. 60km 120 to 144 hrs. 12 Hourly Environment Canada GEM 15 & 100km 10 to 48 hrs. Daily (Global), 12 Hourly (Regional) Beside single NWP models, there are also ensemble weather prediction models in operation. In this study, the Multi Scheme Ensemble Prediction System ( WEPROG s MS-EPS ) and DWD models are used for onshore and offshore wind farms. The detailed investigations will be discussed in the following chapters. 2.5 Statistical Methods to Measure the Accuracy of Wind Power Forecasting In this section, the statistical techniques that are used for indicating the quality of the forecast accuracy will be explained. Accuracy is the most important criteria for the quality of the prediction. A variety of statistical methods were used in this study to evaluate these criteria such as: MAE nrmse Arithmetic mean value Standard deviation Correlation coefficient 18

36 Development of Advanced Wind Power Forecasting Models Arithmetic Mean Value In statistics, the arithmetical mean value of a sequence of numbers is the sum of the entire sequence divided by the number of terms in the sequence. If we denote a sequence of data by x= x1 + x xn, then the sample mean can be denoted by x arith value.. The following equation describes the arithmetical mean x arith = x x x n n xi = (2.1) n n i= 1 Fp is to be in accordance with the arithmetical mean value of the prediction error where P m : Measured wind power P p : Predicted wind power F p : Prediction error The following equations give the arithmetical mean of prediction error: F = p P m - P p (2.2) F 1 n p = ( F p) n (2.3) i = 1 19

37 Development of Advanced Wind Power Forecasting Models Standard Deviation In statistics, standard deviation is a measure of the variability of a data set. Standard deviation is the square root of the variance and described by (σ ). The following formula demonstrates the standard deviation of the wind power forecasting error: n 2 2 n F p F p i= 1 1 σ = (2.4) n n - i= ( n -1) Correlation Correlation is one of the most common statistical techniques which show the strength and the direction of a linear relationship between pairs of variables. Correlation coefficient (ρ) is the component for the correlation calculations. ρ: Correlation coefficient σ: Standard deviation Cov: P m : P p : Covariance Measured wind power Predicted wind power P m : Mean value of measured wind power P p : Mean value of predicted wind power Cov(meas., < pred.) ρ meas.,pred. = (2.5) σmeas.,σ,σpr. 20

38 Development of Advanced Wind Power Forecasting Models Where 1 ρmeas.,pred. 1 and Cov( meas., pred.) n = 1 ( Pm - Pm) ( Pp - Pp) n (2.6) i= 1 ρ 1 n n ( Pm Pm) ( Pp Pp) i= 1 meas., pred. = (2.7) meas. pred. σ, σ Root Mean Squared Error Root mean squared error (RMSE) is a statistical measure where the individual errors that are squared, added together, divided by the number of individual errors and finally square rooted. RMSE is evaluated by the equation below: RMSE 1 n = 2 ( Fp ) n (2.8) i= Mean Absolute Error The mean absolute error measures the average magnitude of the absolute value of errors on a set of forecasts. The mean absolute error (MAE) is described by formula: MAE = 1 n n i= 1 Fp (2.9) 21

39 Development of Advanced Wind Power Forecasting Models Like RMSE, the MAE is a common measure of forecast error in time series analysis Improvement on a Reference Prediction Improvement on a reference prediction represents the improvement of the prediction accuracy and can be calculated by using any EC (Evaluation Criteria) such as MAE or RMSE. IMP EC EC ref (k)- (k) ref,ec (k) = (2.10) ECref (k) 2.6 Selected Methodology: ANN based Wind Power Forecasting ANN Meets with Wind Power Forecasting Artificial Neural Networks ( ANNs ) are one of the most popular artificial intelligence techniques that are used in forecasting applications. NWP variables are used as input variables for ANNs in order to predict the wind power output of the wind farm or cluster for the next minutes, hours or days. ANN represents the multidimensional patterns of the power curve of the wind farm or cluster by using the historical NWP and measured wind power information. During this study, different ANN based applications are investigated in order to minimize the prediction error. The complex wind curve characteristics of the wind farm or cluster can easily be learnt by ANN based wind power forecasting techniques. A power curve demonstrates the relationship between the wind power and the wind speed of a wind turbine. The following figure demonstrates the manufacturer power curve of a wind turbine. 22

40 Development of Advanced Wind Power Forecasting Models Figure 3: A sample power curve of a wind turbine [18] Every wind turbine has typical characteristics that are certified by an authorized institute. However in practice, the production output of the wind turbines may be slightly different than the published power curves by the wind turbine manufacturers due to other environmental, meteorological, mechanical and geographical influences. These kinds of effects make it difficult to predict the power output of the wind farms using physical wind power. If there is enough historical data (measured wind power and NWP data), ANN based wind power forecasting can solve this stochastic problem. In order to find a relationship between input data and wind farm power output, an ANN based prediction system is deployed. The main advantage of ANN over other prediction methods is that they learn from experience and interpolate results, even when their inputs are contradictory or incomplete. Its performance is great if there is enough amount of data set available. Various ANN modules (more than 200) are trained to learn the relationship between variations in the meteorological data and the wind power output using past wind and power data during this study. By comparing the results with measured power data, the optimal configuration of ANN modules is determined. The following sketch 23

41 Development of Advanced Wind Power Forecasting Models illustrates the general structure of the ANN-based wind power forecasting model. Supervised learning methods (Multi Layer Perceptrons ( MLP ) and Back Propagation) were selected for our applications. Supervised learning algorithms are one of the most suitable methods in order to forecast wind power output. That means that, the measured wind power values are used as supervisor or teacher while the ANNs are trained. The ANN models with MLP architecture consist of three different types of layers (input, hidden and output layers). Meteorological parameters (such as wind speed, wind direction, temperature, pressure and humidity) are used as input variables. The equation that is illustrated in the following figure shows the general mathematical form of a back propagation ANN. The number of the hidden layer differs from application to application. The mathematical modeling of MLP and back prorogation ANNs will be discussed in the following section. Figure 4: The sketch of an artificial neural network (ANN) used for wind power forecasting For the purpose of learning the relationship between meteorological data and wind farm power output, the ANN needs to be trained with numerical weather forecasts and measured power values from the past. It is also possible to use 24

42 Development of Advanced Wind Power Forecasting Models shorter intervals like 15 minutes according to this application. After this initial training phase, the model can be implemented and operated. During operation, the ANNs are trained again in regular intervals with the increased amount of data available. As it is mentioned above, classical ANN-based prediction model has two main kinds of input data sets, Numerical Weather Prediction ( NWP ) values received from DWD or other MET offices and measured power data. The accuracy of the prediction model is mainly dependent on the reliability and accuracy of NWP data. Another important issue to build a model which calculates the time series is the representative wind farms on the total feed-in area of all Wind Turbines ( WTs ) in the control zone History of ANN The research activities concerning ANN initiated in the 1940s with Mc Culloch and Pitts. They were the first researchers who tried to model the human neuron cells with mathematical models. In the 1950s, Rodenblatt s investigation resulted in a two-layer neural network, which was able to learn certain classifications by adjusting the connection weights. This model was called the perceptron. After almost 30 years, researchers started investigating the neural networks in the early 1980s. Recent investigations in ANN field are focused mainly on competitive learning models, Boltzmann machines, Hopfield nets, multi-layer networks and hybrid models such as fuzzy-neural network models. [22] There are some advantages and disadvantages of ANNs. Some of the important points are listed below: Advantages: Capability to determine a solution for the problem in which relationships are dynamic and nonlinear. Parallel organization structure of ANN allows solutions to be developed for which multiple limitations have to be satisfied. 25

43 Development of Advanced Wind Power Forecasting Models ANNs simplify the model phenomena, where the mathematical and physical models are insufficient. If an element of ANN fails, it is able to continue without any problem due to its parallel programming structure. An ANN model can be implemented in any application easily. Disadvantages: Relatively big data set is necessary to train the ANNs. The computational time may be higher for larger data sets. An ANN needs to be operated during the training phases. 2.7 Types of Training (Learning Methods in ANN) Neural Network Topologies In this section, the pattern of connections between the units and the propagation data is discussed. Two types of neural network topologies are described in this part: Feed-forward neural networks, are one of the most popular and widely used ANN models which there is no direct connection between the units. In feed-forward ANNs, there is no feedback connection and the data processing can be extended over multiple units. In other words, there are no cycles or loops in the network structure and the data flows only in one direction between the nodes and up to the output node. [23] Recurrent neural networks: In contrast to feed-forward ANN models, recurrent neural networks have direct connections between the nodes. A dynamic system theory is usually used to model this feature of the recurrent neural networks. In such neural networks, each neuron consists of a weight associated with each input. A function of the weights and inputs is then yielded as output. [24] 26

44 Development of Advanced Wind Power Forecasting Models Training of Artificial Neural Networks The configuration of a neural network has to be organized such a way that a set of inputs should be able to generate the desired set of outputs by using learning methods. There are several methods to determine the weights of the connections between the artificial neurons. 7 There are mainly three types of ANN learning models as shown in the following figure: Figure 5: Overview of ANN learning methods 8 Supervised learning models can be also called associated learning models, refer to models in which training occurs by using matching output patterns and input. Error function is minimized by the supervised learning algorithm in which one specifies the value of the desired output for each input pattern. The input vector can be provided by an external teacher. 7 [22] 8 Multi Layer Perceptron ( MLP ), Adaptive Resonance Theory ( ART ), Self Organizing Maps ( SOM ) 27

45 Development of Advanced Wind Power Forecasting Models Figure 6: Supervised learning Unsupervised learning models have some advantages in solving complex problems. These types of models are trained to respond to clusters of pattern within the input. Unsupervised learning models learn on their own way through a self-learning mechanism. These types of models only have input training patterns without any teacher or supervisor. Learning occurs by storing up an archive based on previous patterns. Reinforcement Learning is an intermediate form of supervised and unsupervised learning models Testing Phase During the testing phase, checks are carried out to ensure that the weights have been modified correctly during the training phase. Unknown input data are used to perform the testing. 28

46 Development of Advanced Wind Power Forecasting Models 2.8 ANN Theory and Mathematical Foundations of ANN Artificial neural networks are inspired by human brain and applied in a variety of applications such as forecasting, intelligence control and pattern recognition. The human brain is a highly complex non-linear and parallel computer which has the capability to organize its structural components by using approximately 10 billion neurons. In other words, ANNs are the programs that imitate the biological neurons of humans. Biological neural networks are made up of real neurons which are able to perform certain functionalities and computations such as motor control of muscles and pattern recognition of eyes. These kinds of functionalities are related to the peripheral nervous system or the central nervous system in human body. [22] Similarly, ANNs are made up of interconnected artificial neurons that are supposed to imitate the nature of biological neural networks. The following figure shows the structure of a biological neuron. Figure 7: The structure of a biological neuron 9 Each biological neuron cell has approximately neighboring connections. It remains a phenomenon of how brain trains itself to process information. The cell body (soma) is the central part of neuron that has two main parts: the dendrites and axons. A cell body consists of a nucleus which is responsible for containing 9 Source: 29

47 Development of Advanced Wind Power Forecasting Models proteins. Dendrites are the branch-like structures in a neuron that act to conduct the electrochemical signal collected from other neighboring cells to other parts of the body (cell body).the electrical signal is transmitted into dendrites via synapses. Dendrites are responsible to integrate these synaptic inputs. Axon conducts electrical signal generated at axon hillock thorough it s thin and long structure. These signals are called action potentials. When a neuron receives a signal input that is sufficiently larger than a certain threshold value, it sends a signal of electrical activity through its axon. Learning occurs by exchanging the effectiveness of the synaptic signals Mathematical Modeling of an Artificial Neuron There are number of similarities between biological and artificial neurons. Typically, a neuron is a model with many inputs (dendrite) and an output (axon). An artificial neuron can have multiple inputs and an output. The similarities between a biological and an artificial neuron can easily be seen in the following figure: Figure 8: The structure of an artificial neuron The artificial neuronal model is illustrated in the following figure that includes; externally applied bias ( b i ) and activation function ( f (.) ).[22] 30

48 Development of Advanced Wind Power Forecasting Models Figure 9: The structure of an artificial neuron (mathematical model) Each input has an associated weight (w) which can be adjusted in order to perform synaptic learning, where the input signals are i1, W i 2,..., W im X,... 1, X 2 X m and W are the synaptic weights of the neuron i. We can demonstrate the neuron i by writing the following equation: i m = ij x j (2.11) j=1 U w The usage of bias has an influence on the output of the linear combiner [ V i ]. V + i = Ui bi (2.12) In mathematical terms, we may show the output of the neuron by using the following equation, where the activation function is denoted by f (.) : X = f U + b ) (2.13) i ( i i 31

49 Development of Advanced Wind Power Forecasting Models The bias ( b i ) is an external parameter of an artificial neuron i and can be used as a new synaptic weight equal to W i0 ( i W i 0 b = ). Then, i V can be given as: i m = ij x j (2.14) j=0 V w The output of the neuron can be shown as: X = f V ) (2.15) i ( i And the same equation can be driven by using (2.14) as follows: X i = f m ( j= 0 w ij x j ) (2.16) Activation Functions The activation function acts as a squashing function and denoted by f(v) that defines the output of a neuron in terms of induced local field v. In general there are three types of activation functions. 32

50 Development of Advanced Wind Power Forecasting Models 1. Threshold Function: 1... ifv 0 ( v) = 0... ifv < 0 f (2.17) 2. Piecewise Linear Function f ( v) = 1... ifv > ifv = ifv < 0 (2.18) 3. Sigmoid Function f (v ) 1 = 1 + exp ( αv ) (2.19) Multi-Layer Network A multilayer network (where the input ( can be represented as in the following figure. X i ) is transformed to an output ( Y k )) 33

51 Development of Advanced Wind Power Forecasting Models Figure 10: The model of a multilayer artificial neuron The following initial descriptions for the mathematical modeling of the multilayer network will be used in this study [22]: W ji : Weights between input layer i and hidden layer j, W kj : Weights between hidden layer j and output layer k H w j x = ji i : The net input of the hidden layer j (2.20) i I w k y = kj j : The net input of output layer k (2.21) j Y = f H ) : The output produced in the hidden layer j (2.22) j ( j 34

52 Development of Advanced Wind Power Forecasting Models Y = f I ) : The output produced in the output layer k (2.23) k ( k A multilayer network (where the input ( X i ) is transformed to an output ( Y k )) can be represented by deploying the following equations: Y k = f ( I k ) = f ( w kj y j ) (2.24) j If we drive the equation (2.22) into the equation (2.24), we denote the following expression: [ j Y k = f wkj f ( H )] (2.25) j Finally, the output ( Y k ) can be given as: Y k = f [ w kj f ( w ji xi )] (2.26) j i Supervised Learning In this case, the supervisor knows the correct output value (target value) where the error value (E) is inserted again to the supervised learning algorithm. 35

53 Development of Advanced Wind Power Forecasting Models Figure 11 : Illustration of the error vector in the supervised learning method The error vector can be shown as: E = (2.27) V i - Z i Where; Z i : The output of the neuron i V i : The target value that is expected from the neuron i Multilayer Perceptrons (MLP) and Back- Propagation Rule In comparison to simple perceptrons, MLPs are more flexible. The hidden layers are located between input and output nodes. The backpropagation is one of the most important learning rules for network training. One of the common ways to apply back propagation is the delta rule method. The Back-propagation algorithm learns the weights for a multilayer network using the input variables and interconnections. We can define the error function of the neuron as the average squared error by using the following formula [25], [22]: 36

54 Development of Advanced Wind Power Forecasting Models 2 E = 1 [ Vi - Z ] 2 i (2.28) i Where w is the weight vector. Adjusting the weights may increase or decrease the error value. Gradient descent is a basic ANN method which is used to reduce the error value. The weights are adjusted in the negative gradient direction: In this step, minimization and weight correction of the error value is achieved by using the gradient descent method. According to the gradient descent method, weights are adjusted proportionally to the negative of the error with respect to each weight. The gradient descent method is employed in order to minimize the squared error between the network output and the target values for the outputs. Here, we apply the gradient descents rule, where the learning rate factor of the network is denoted by η and δ k is the local gradient. de w ji = η = ηδky j (2.29) dwji We now summarize the relations that we have derived previously for the back propagation algorithm. The weight correction function can be applied to the synaptic weight connecting neuron i to neuron j by deploying the delta rule: ηδ w ji = ky j (2.30) 37

55 Development of Advanced Wind Power Forecasting Models Figure 12 : The process of the adjustment of the weights in the back propagation The local gradient layer. algorithm δ k depends on whether neuron is an output or a hidden 1. If the neuron is an output node, δ k equals to the product of the derivative ' f ( ) and the error signal - Z ) I k ( V k k ; ' δ k = ( Vk - Z k ) f ( I k ) (2.31) If the neuron is a hidden node, δ j equals to the product of the derivative ' f ( ) and the weighted sum of the δ kj ; H j δ j = ' f ( H j ) δ k wkj (2.32) j j k w In general, we have used the sigmoid function as an activation function. We may apply the same learning rule by using the sigmoid function 38

56 Development of Advanced Wind Power Forecasting Models ( v) 1 ( f = ( αv) ). During this operation we take the derivative value of 1+ exp the sigmoid function. The equations below demonstrate the local gradient values for the output and hidden layers respectively; δ k Zk (1 - Zk )(Vk - Zk ) = (2.33) δ j = Y j (1- Y j ) wkjδ k (2.34) j Local Minima and Momentum Terms As mentioned previously, in order to reduce the error to an acceptable minimum level, the direction of the back propagation rule weight vector should be shifted in the negative direction. In the figure 13, the demonstration of the absolute minimum, local minimum and the constant weight values can be seen. Figure 13 : The illustration of the absolute minimum, local minimum and constant weights 39

57 Development of Advanced Wind Power Forecasting Models One of the standard representations of the gradient descent is the addition of momentum term. Momentum term is a constant value ( a ), which is used to keep the weight value a little bit of the direction of the previous weight change in each iteration and is used to determine the direction of the next step by using the previous information. [25], [22] w ( t + 1) = ηδ y + a w ( t) (2.35) ji k j j 2.9 Iset s Wind Power Prediction Approach In this part of the study, the structure of ISET S Wind Power Forecasting Model ( WPMS ) was used and improved further by using additional techniques. The Wind Power Management System ( WPMS ) was developed by ISET. This model is currently in operation in 3 German TSOs. The system consists of the following parts: The online monitoring, which is responsible for producing the generated wind power time series that belong to the representative wind farms on the total feed-in area of all Wind Turbines ( WTs ) in the control zone (Upscaling). The day-ahead forecasting module produces 24 hourly mid-term wind power prediction values based on numerical weather prediction data. The very short-term forecast is a model, which generates shorter range predictions (between 15 minutes to 8 hours) based on NWP and online wind power measurements in order to produce more accurate prediction results according to the day ahead forecast. 40

58 Development of Advanced Wind Power Forecasting Models Figure 14: The graphical user interface of the Wind Power management System ( WPMS ), Regional Up-scaling using Online Model Germany has a current wind capacity of more than 24.8 GW. There are more than installed wind turbines in Germany. Therefore, it is very difficult to measure and predict all wind farms individually. In order to overcome this problem, a kind of so-called upscaling algorithm is necessary to estimate the cumulative amount of generated and predicted wind power values for the whole region or country. In the first step, the wind power predictions for most representative wind farms are evaluated. A special transformation function called Online Model is then used to upscale the representative wind farm results for generating larger area such as a region or county-wide predictions. 10 The following sketch demonstrates the structure of a day head and very short-term wind power forecasting model using AI or statistical methods. For both day 10 [27], [26] 41

59 Development of Advanced Wind Power Forecasting Models ahead and very short-term wind power forecasting modules, it is possible to evaluate an upscaling algorithm in order to generate regional forecasts. Figure 15: The structure of a day ahead and very short-term wind power forecasting model using AI and statistical methods This method reduces the amount of needed numerical weather prediction data. The Online Model, which was developed by ISET e.v. in Germany, has the following mechanism. The target area is subdivided into the individual cell squares or matrixes. For each cell, the installed capacity of wind farms coordinates the roughness factor of the terrain and hub heights are inserted into the system. [26] 42

60 Development of Advanced Wind Power Forecasting Models Figure 16: Cell squares used by WPMS for regional up-scaling 11 The figure above illustrates the cell squares for Germany. In this sample, each square shows the installed capacities of all the wind farms in the selected area where the power output of one cell square is computed from the weighted power outputs of all representative wind farms. If the respective wind farms are closer to the cell square, the influence of these wind farms is greater than that selected cell square.[27], [26] 11 Source: ISET 43

61 Development of Advanced Wind Power Forecasting Models If we assume I as cell squares and j representative wind farms, the power output of the total region, P total is the summation of the power output and P i of all cell squares: P total = P i (2.36) i The power output of one cell square outputs of all representative wind farms: P i is computed by weighting power P i = ki A ij Pj (2.37) j Where Pj is the power output of representative wind farm j, and k i normalization factor, and A ij is the weighting factor. S ij A ij = exp PIP, i S (2.38) 0 Here, P IP, i is the installed power in cell square I, S ij is the distance between the cell square and the representative wind farm and S 0 is a spatial correlation parameter. The normalization factor ( k i ) denoted by the following equation where the sum of all weighing factors equals one. 44

62 Development of Advanced Wind Power Forecasting Models k = i 1 s * A j j ij (2.39) Day Ahead and Very Short-term Wind Power Forecasting Besides the prediction of the entire power output of the WTs for the next few days (up to 72 hours), very short-term high-resolution forecasts of the wind power generation are essential to manage the power system management more accurately. Wind power data measured online is used as an additional input for the ANN-based prediction model, in order to optimize a very short-term prognosis from 15 minutes to 8 hours. The shorter the forecast horizon, the more accurate the wind power forecast. Transformation Model is used to upscale the produced / forecasted wind power values based on the total feed-in area. Furthermore, sudden changes in the weather conditions cannot be taken into account by models that are based purely on numerical weather forecasts. Therefore, it is useful to insert the additional data from the very near past such as measured wind power output (online metering value) and meteorological from the site, if it is available. [27], [26] As the local weather conditions in the near past (and present) are indirectly recorded over the measured power of the wind farm. The prognosis material can be significantly improved for short time horizons through the inclusion of this information. Highly resolved and precise wind farm predictions, for the next 15 minutes up to 4 hours, are especially important for the operational control strategies mentioned. [28], [29] 45

63 Development of Advanced Wind Power Forecasting Models Figure 17: Real and predicted course of wind generation The figure 17 illustrates the real and predicted course of the wind generation. The deviation of the day ahead prediction is corrected by the very short-term prediction (2 hours ahead). The improvement of the prognosis can be observed by the inclusion of actual measurement data from the near past. As it can be seen in the figure 18, the forecast error (RMSE in % of installed capacity) values from very short-term prediction (2 and 4-hour ahead forecasts) of wind generation are dramatically lower than the day ahead prediction value. For this investigation NWP data from DWD are used. The forecasting results of total Germany are illustrated in the following graph where the frequency distribution of the prediction error of the day ahead and very short-term (2 and 4 hours ahead) prediction results are given. The frequency distribution of the prediction error of the day ahead is 5.7 % in terms of nrmse. The nrmse values for 4 and 2 hours ahead very short-term wind power forecasting models are 3.1 % and 2.5 % respectively. The investigation was performed for the wind power generation in Germany in the year

64 Development of Advanced Wind Power Forecasting Models Figure 18: Comparison of day ahead and short-term (2 and 4 hours ahead) wind power forecasting errors (in terms of nrmse) for Germany Smoothing Effect The accuracy of regional forecasts is usually higher than the predictions for a single wind farm or a cluster. This can be explained by the smoothing effect. The wind power prediction accuracy of a wind farm is less than the ones for a TSO control zone or a country. The more spread in the forecasted area, the lower prediction error. As illustrated in the following graph, the wind power forecast measure (nrmse) for a wind farm is 14 % for day ahead and approx. 8 % for 2 hours ahead predictions. The prediction error values calculated for TSO control zone are up to approximately 7 % for day ahead and approx. 3 % for 2 hours ahead predictions. Finally, consequence of the smoothing effect the nrmse value for Germany reduces to 5.7 % for the day ahead and approximately 3 % for 2 hours ahead predictions. 47

65 Development of Advanced Wind Power Forecasting Models Figure 19: Smoothing effect for a single wind farm, control zone and Germany Software Development Environments for Modeling an ANN Based Wind Power Forecasting System There are various wind power forecasting models currently in use and under development. In this section, the development of single ANN based wind power forecasting models are discussed. We have employed three development environments in order to develop single prediction models: Stuttgart Neural Network Simulator ( SNNS ) - based wind power forecasting model MATLAB ANN Toolbox based wind power forecasting model NeuroSolution ( NS ) - based wind power forecasting models The performances of these three models have been compared according to the following criteria: Forecast accuracy 12 The nrmse values were prepared by the author using ISET s database during his employment. 48

66 Development of Advanced Wind Power Forecasting Models Flexibility Computational time Stuttgart Neural Network Simulator SNNS Based Model SNNS is an operation system independent ANN development environment which is able to run under Windows or UNIX. SNNS was developed at the Institut für Parallele und Verteilte Systeme (IPVR) at the University of Stuttgart that gives opportunity to the users to develop their ANN models by using several ANN network topologies. The kernel of SNNS s ANN simulator is written in C. The ISET s wind power development environment runs under UNIX. In order to realize and simulate the wind power forecasting models several additional programming languages are used. ISET s development GUI was programmed by using TCK/TK and Java, the underground processes such as organizing the training, testing data set and similar tasks are programmed by using UNIX shell scripts and PL/SQL scripts. The input parameters and forecast results are stored in an Oracle data base. Matlab ANN Toolbox Based Model MATLAB is a high-level scientific programming language that enables the users to develop computationally intensive and complex applications. MATLAB is supplemented by several toolboxes which make the developers it easier to perform the desired applications. Neural network toolbox is one of the toolboxes developed by MATLAB that provides a comprehensive support for the design, application and simulation of many network topologies

67 Development of Advanced Wind Power Forecasting Models NeuroSolutions Toolbox Based Model NeuroSolutions ( NS ) for MATLAB is a programming environment that is quick and easy to use platform and provides 15 neural networks models, six learning algorithms and lots of useful utilities. It is very easy to use the NS for the developers to developing neural models, just by typing some commands. This interface feature enables the user to model neural models in few minutes without knowing the complex details of artificial neural networks methods in 3 steps. Step 1: Modeling of a neural network The standard model of the NS toolbox is multi-layer perceptron with a hidden layer. If the user wants to use another model among 15 available NS neural models it is enough to type just short commands in order to model the ANN structure. Step 2: Training of neuronal Network The input and target data has to be organized in this step before starting the training process. Input Data was the data set that is prepared to be used as input variable for the neural network model. In our application the input data was the NWP parameters such as wind speed, wind direction and other available meteorological variables. Target Data included the desired value of the production data. In our case the target data was the measured wind power information. Validation is an optional process that can be used to compare and check the data. 50

68 Development of Advanced Wind Power Forecasting Models Step 3: Testing and employing the neural networks After the training and the optional validation phases, the testing process was performed. The generation of prediction time series was the final phase of the process Comparison of Three ANN Based Wind Power Forecasting Models Three models were trained by using the same training type; input data, learning algorithm, and the testing / training data structure of all the models were also identical. The following features of the three models were common for all the models. Tab. 3: Overview of three wind power forecasting tool features KNN Type MLP Hidden Layer 1 Number of nodes 30 Transfer Function Sigmoid Three ANN models were trained using gradient descent method and back propagation algorithm. Gradient descent method minimizes the mean squared error value between the measured wind power and the forecasted wind power. For this analysis we used the NWP data from DWD for a single wind farm in Germany (with 30 MW installed capacity) in After performing the three models, the comparison of the models was performed (please see table 4). Tab. 4: Comparison of the results from the three wind power forecasting tools Correl. nrmse Bias SNNS % MATLAB % NS % As observed from the results in the table 4, correlation, nrmse and bias results of the three models are almost the same. As a conclusion, if we consider the flexibility and the short computational times of the NS and the MATLAB toolbox- 51

69 Development of Advanced Wind Power Forecasting Models based models, it is reasonable to use these two models for the next parts of the study to develop neural models Conclusion & Future Challenges In this chapter, an overview of the existing wind power forecasting systems was given. Among a variety of prediction methods, ANN-based wind power forecasting method was selected for this PhD work. Consequently, the background information about artificial neural networks and the meteorological input (NWPs) parameters were given. The issues concerning the comparison of day ahead vs. very short-term wind power forecasting models and the smoothing effects were also presented. Finally, the development of an ANN based advanced wind power forecasting system was discussed briefly. For the integration of wind power generation in Germany, the TSOs use ISET s Wind Power Management System ( WPMS ), which includes an onlinemonitoring of the current wind generation with forecast horizon between several hours to 24 hours since several years. It also builds the basis for the horizontal exchange mechanism of wind power between the control zones. The forecasting accuracy of the system has been improved continuously and very significantly during its operational use since Some features of the existing WPMS were further developed during this PhD study. As the wind power capacity grows fast in Germany and many other countries, forecast accuracy becomes increasingly important. However, it can also be expected that the improvements in forecasting accuracy should be maintained in the future. Forecast accuracy is only one of the challenges for wind power forecasting systems of the future. Additionally, the scope of systems will have to be extended to meet the future challenges: Wind power forecast in the offshore environment has the potential to become more reliable than on land, if offshore-specific forecast models are developed. Improved forecasts for shorter time horizons will be needed for grid safety. 52

70 Development of Advanced Wind Power Forecasting Models Forecasts in high spatial resolution for each grid node of the high voltage grid will be needed for high wind power penetration to challenge with the problem of congestion management. Further improvements in the forecasting methods and the improved methods for the combination of different forecasting methods may be expected to further reduce forecasting errors. Especially for very short-term wind power forecasting, additional use of online wind measurement data is very important for the accuracy of the forecast. Further extension and optimization of the set of meteorological parameters from the NWP models lead to an important improvement in the wind power forecasting. A significant reduction in the forecasting error may be obtained by combining different NWP models. Finally, the combination of different AI methods used in the wind power forecasting system has the potential to reduce the forecasting errors. 53

71 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach 3 Improvement of Wind Power Forecasting Accuracy using the Multi- Model Approach Abstract This chapter focuses on the improvement of the existing ANN based wind power forecasting model using several optimization methods. In this section, the multi model approach is investigated. This optimization approach consists of two methods; the multi NWP and the ensemble wind power prediction methods. The investigations in this chapter have mainly been performed for the onshore wind farms. The same method will be extended to the offshore wind farm environment in chapter four. 3.1 Introduction As discussed in the previous chapter, wind power forecasting plays a very important role to improve the economic and technical integration capability of large-scale wind energy. The wind power forecasting accuracy is directly related to the need for balancing energy, and system reliability of the electricity supply system and hence the cost of wind power integration. The multi model approach for wind power forecasting is an effective way to reduce the forecasting error. The approach, which is covered in this chapter, consists of two main methods: The Multi NWP ( Numerical Weather Prediction") approach The Multi-scheme ensemble weather prediction approach 54

72 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach The Multi-NWP approach was performed to investigate the influence of different NWP models on the accuracy of the wind power forecast. This approach is compared to the use of a Multi-Scheme Ensemble Prediction model ( MS EPS ) to observe the improvement wind power forecasting accuracy. The results were published in several conferences [30], [31] 3.2 The Multi-Model Approach The Multi-NWP Approach for Germany Three different NWP models were used for a day-ahead wind power forecast for Germany. The models were independent from each other, i.e. they used different global and local models and have been obtained from different providers. All three models were used as input to the WPMS which is based on the ANN method. The training of the networks was performed with data of more than one year. A concurrent data set of seven months (April October 2004) was used for the comparison. The forecast horizon was 24 hours for this investigation. Tab. 5 : Main Characteristics of the selected NWP models NWP-1 NWP-2 NWP-3 Forecast Horizon 72 Hours 48 Hours 72 Hours Model Runs 00 and 12 UTC (Universal Time) Wind Speed 00 UTC Wind Speed 00 UTC Wind Speed Wind Direction Wind Direction Wind Direction Available Parameters Temperature Temperature Temperature Air Pressure Air Pressure Air Pressure Humidity Humidity Momentum Flux The RMSE values in percent of the installed capacity of the three models are shown in the figure 20. It can be seen that the differences between the different models are small. The prediction error values are varying between 5.7 and 6.1 % in terms of nrmse. Additionally, a simple combination of the three models has been tested by averaging their forecasts. It is observed that even this simple approach improves the forecast accuracy quite significantly compared to the 55

73 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach results of the individual models. The resulting RMSE for the combined model for Germany is 4.7%. This leads to a relation reduction of approx 18 % compared to the best single NWP based wind power forecast. Figure 20: Comparison of the normalized RMSE of a wind power forecast for Germany which is obtained by means of the WPMS based on ANN with input data from three different NWP models and with a combination of these models The Multi-NWP Approach for Single Onshore Wind Farms For this case study data from the wind farms Meerhof and Meppen were used. Their geographical location in the RWE control zone is shown in the figure 21. The value of installed capacity is 97.3MW for the wind farm Meerhof and 25.6 MW for the wind farm Meppen in

74 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Figure 21: The representative wind farms Meerhof (WF 1) and Meppen (WF 5) in RWE region The multi NWP model approach was applied on an onshore wind farm (WF 1) in the RWE control zone. The use of multi model approach decreased the day ahead wind power forecasting error. A concurrent data set of two years ( ) was used for the comparison. 57

75 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Figure 22: Comparison of day ahead wind power forecasting error for the WF 1 (2005) As observed from the results graph above, the prediction error values vary from 10.38% and 11.1% in terms of nrmse in Utilization of the multi NWP method increases the accuracy of the forecasting system and the nrmse value is reduced to 9.33% by using multi NWP (Combi NWP) method. This leaded to a reduction in the nrmse value approx. 10% in comparison to the best single NWP model. Figure 23: Comparison of day ahead wind power forecasting error for the wind farm 1 (2006) 58

76 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Similar results can be observed from the results graphic above. The prediction error values vary between % and 11.88% in terms of nrmse for the same wind farm (WF1) in Utilization of multi NWP method increased the accuracy of the forecasting system and the nrmse value reduced to % using the multi NWP (Combi NWP) method. This leaded to an approx. 5.4 % reduction in the nrmse value in comparison to the best single NWP model. In addition to the application of the multi model approach on a day ahead wind power forecasting time-series, we applied the same approach on a 2 hours ahead (very short-term) wind power forecasting time series for the same wind farm. A data set of one year (2006) was used for this investigation. In this case, the prediction error value nrmse which was calculated by using single NWP models varied between 8.54 and 8.84 %. Consequently, the nrmse value for the WF1 (the wind farm Meerhof) reduced to 8.36 % by using the Multi NWP approach. The improvement of the accuracy achieved by using the multi NWP approach for 2 hours ahead wind power forecasting is less than the accuracy of the day ahead wind power prediction. This leaded to a reduction in the nrmse value approx. 2 % in comparison to the best single NWP model. Figure 24: Comparison of 2 hours ahead short-term wind power forecasting error for wind farm 1 (2006) 59

77 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach 3.3 The Multi-Scheme Ensemble Weather Prediction Approach Ensemble forecasting, i.e. by using several different forecasts instead of only one single deterministic forecast was developed in order to quantify forecast uncertainty and to improve the accuracy of deterministic forecasts. The study driven by ISET s neural network based Wind Power Management System ( WPMS ) was extended with the possibility of including and combining several members of the MS EPS operated by Weather & Wind Energy Prognosis ( WEPROG ). The new method was developed using two wind farms in the RWE control zone in Germany. [31] The WEPROG Ensemble Weather Prediction Model The ensemble system used in this study is a limited-area Multi-Scheme Ensemble Prediction System ( MS-EPS ). The multi-scheme ensemble prediction is a technique where different parameterizations or schemes are used to vary the calculation of meteorological processes (e.g. vertical diffusion, condensation) giving the potential to provide a more realistic representation of the state of the atmosphere, the uncertainty of the forecast and hence a smaller forecast error. This means that each forecast member comprises a different set of equations for certain physical or dynamical processes (called parameterization schemes ). The ensemble approach is targeted to determine the uncertainty of the weather forecast. The differences in the equations lead to different methods of solving these equations and thereby generate different end results. 14 A large number of meteorological variables are calculated in this model. A number of variables, which could potentially be useful for forecasting of wind power, have been selected. These are shown in the next table. In this study, the importance of the variables was evaluated and only the necessary ones were used as input to the wind power prediction system. [32] 14 [30] 60

78 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Tab. 6 : Available Input Parameters from MS EPS 15 Input Parameter Description Unit 1 U10m Wind Speed in x direction in 10m (exactly) m/sec 2 V10m Wind Speed in y direction in 10m (exactly) m/sec 3 U31 Level 31 wind (Wind Speed in x direction in approx.100m) m/sec 4 V31 Level 31 wind (Wind Speed in y direction in approx.100m) m/sec 5 U32 Level 32 wind (Wind Speed in x direction in approx. 35m) m/sec 6 V32 Level 32 wind (Wind Speed in y direction in approx. 35m) m/sec 7 U30 Level 30 wind (Wind Speed in x direction in approx. 170m) m/sec 8 V30 Level 30 wind (Wind Speed in y direction in approx. 170m) m/sec 9 Senf sensible heat flux W/m2 10 Momf momentum flux W/m2 11 Clc Cloud cover % 12 Pmsl Mean sea level pressure hpa 13 precip Precipitation mm/h 14 T2m Temperature at 2m in K 15 T32 Temperature at Level 32 (apprx. 35m) K 15 Provided by WEPROG

79 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach As mentioned in [32] the Multi Scheme Ensemble Prediction System ( MS EPS ) approach uses different model configurations from the traditional approaches to generate an ensemble weather forecast. It is known that, traditional meteorological approaches build ensemble of forecasts which are initiated from different initial conditions. The vertical diffusion and the condensation or cloud convection are the most important parts of the MS EPS approach. The vertical diffusion simulates the deviations from wind speed (turbulence) and also controls the mixing length of the middle and upper atmosphere in order to determine the phase speed and the amplification of atmospheric fronts. The vertical diffusion schemes are the parameterization of vertical eddy heatfluxes, momentum and water vapor in the atmosphere. [33] The parameterization of the convection and condensation processes is a challenging task in mesoscale Numerical Weather Prediction ( NWP ) models. The main task of convective schemes is to predict the rate of the sub-grid scale convective precipitation, the redistribution of heat moisture, the momentum and the latent heat flow in the atmosphere. The condensation and convection schemes are cumulus convection and condensation where the parameterization depends on the resolution of the models. [34], [35] The following 3 advection schemes are used in MS EPS approach: Eulerian advection scheme Eulerian based first-order-upstream differencing scheme Semi-Lagrangian advection scheme The general structure of MS EPS approach is demonstrated as a cubic form in the following figure. 62

80 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Figure 25: Structure of the 75 member MS-EPS model 16 MS EPS approach can be successfully used for the purpose of generation ensemble wind power forecasting time-series and their uncertainty intervals. During the PhD. study, we have tested MS EPS approach for the configurations of two onshore and an offshore wind farm Evaluation of the Multi Scheme Weather Ensemble Prediction Approach The evaluation of the Multi Scheme Ensemble Prediction System ( MSEPS ) approach included the following steps: Training for each of the 75 single members of the MSEPS model using artificial neural networks Averaging of the forecasted wind power of several ensemble members A two-step model using ANN modules for the optimized combination of the forecasted power of several members 16 This sketch is prepared according to the WEPROG s descriptions. 63

81 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Single ANN Models for Each MS EPS Members As a first step, an Artificial Neural Network (ANN)-based model was developed for each of the MS EPS 75 members. The structure of the model is illustrated in the figure 26. Figure 26: Sketch of ANN based prediction model for each MS EPS member In this study, we used an ANN module with three hidden layers for each member. The ANN was trained with gradient descent by applying the standard backpropagation algorithm. It minimizes the mean least square error between the measured power and the predicted power over a training data set. The training and testing data set had a time period of about one year (in 2005). In 64

82 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach total, 8200 hourly values were used. Half of this dataset was used for training of the ANN and the other half is for the evaluation. The following graph illustrates the nrmse values of all MSEPS members (75 single members) and the EPS Mean (mean value of 75 Ensemble members) for Meerhof wind farm (WF 1) and Meppen wind farm (WF 5). The normalized root mean square error (nrmse) values for the wind farm Meerhof (WF 1) vary between 11.1 % and 13 % of installed capacity, and the nrmse values for the wind farm Meppen (WF 5) vary between 9.9 % and 11.9 % of installed capacity. These results show that, it is necessary to determine the ranking of the members for each investigated wind farm individually. On the other hand, there are some common characteristics of the nrmse results for each wind farm. Three periodical characteristics which are caused by the calculation methods of meteorological processes (e.g. vertical diffusion, condensation and dynamics) can be observed respectively between the 1st and 25th members, between 26th and 50th members and finally between 51st and 75th members. 65

83 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Figure 27: nrmse of the 75 members for the Wind Farm 1 (Meerhof) and Wind Farm 5 (Meppen)

84 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Averaging of the Forecasted Wind Power of Several Ensemble Members A simple averaging of the output time series of all 75 members was performed as a first combination method. In the first method the arithmetical averaging was done without weights, i.e. all members had equal weights (please see the figure 28). Figure 28: Sketch of the combination model using simple averaging The Two Step Model using ANN for the Optimized Combination of Forecasted Wind Power of Several Members In this evaluation, the MS EPS variables were used again as meteorological input data for the first stage of the ANN model. The outputs of the first stage of the ANN model were the forecasted wind power time-series of each evaluated MS EPS member. The forecasted wind power time series generated from the first stage of the ANN model were used as inputs at the second stage of the ANN model. Finally, the second stage of the ANN model yielded the 67

85 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach forecasted wind power time-series in the 2 step ANN model. The structure of the ANN in both stages was the same. More details concerning the ANN models were discussed in the second chapter. The graphic below shows the structure of this model for the sample of three members (x, y and z). Figure 29: Sketch of the two step ANN combination model The results showed that wind power forecasting combination methods are better than a single NWP approach (please see the figure 30). The nrmse for the wind farm 1 (Meerhof) using single NWP model (from DWD) is 13 % of installed capacity. The effective usage of the EPS-MEAN (simple averaging method) means that the weather data was averaged prior to their use in an ANN module, where the wind power forecast was done. The nrmse of the same wind farm using EPS-MEAN is above 11.1 % of installed capacity. The nrmse of the simple averaging model is 11 %. Finally, the results of the two step ANN model are better than single member and single NWP results. The nrmse of the 2 step ANN model is about 10.5 % of installed capacity. 68

86 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach Figure 30: Comparison of the nrmse values of wind power forecasts with a single NWP model (from different meteorology service provider), the EPS Mean and the combination models for the wind farm 1 (Meerhof). 3.4 Conclusion In this study, different approaches regarding the multi wind power forecasting models were investigated to reduce the forecast error. As a first approach, three independent Numerical Weather Prediction (NWP) models were used in ISET s Artificial Neural Network ( ANN )-based forecasting system to perform wind power forecasts for the whole of Germany. Although the NWPs are completely independent, they show very similar wind power forecast errors, when they are used in an ANN model. A simple averaging based multi NWP model approach has already reduced the nrmse by about 20%. As an alternative to the use of independent NWP models, the use of the Multi- Scheme Ensemble Prediction System ( MS EPS ) was investigated. The MS EPS from WEPROG, which consists of 75 different members, was used in this study for the wind power forecast of two wind farms in the RWE (now Amprion) control zone. The results showed that, there is a considerable spread in the performance of the members. In the next step, different methods to combine 69

87 Improvement of Wind Power Forecasting Accuracy using the Multi-Model Approach ensemble members were developed and compared. Two methods were investigated: Averaging of the forecasted wind power of several ensemble members The two step model using ANNs for the optimized combination of the forecasted power of several members The use of ANNs to perform wind power forecasts for each member individually and subsequent optimizing of the forecasted power has yielded an improvement in the case of the wind farm 1 (Meerhof). The results obtained in this study with the ensemble weather prediction model MS-EPS are very promising. According to the investigations which were performed in this chapter, the utilization of the combination approaches such as the simple averaging and especially the double ANN approach has a great potential to reduce the wind power prediction error values. The usage of the two step model using ANNs (the double ANN approach) yielded approximately 8 % improvement in the prediction error value. In this first feasibility study only two onshore wind farms were investigated with all models and a limited amount of data was available. The results therefore clearly have to be verified in a study, where we use all wind farms and their corresponding upscaling factors that are used for the operational forecasting to show the full benefit of this approach. A longer data period will also be beneficial in order to verify the results of this study. The verification of this approach will be discussed in the coming chapter in the offshore environment. 70

88 Development of New Offshore-Specific Wind Power Forecasting Models 4 Development of New Offshore-Specific Wind Power Forecasting Models Abstract Advanced wind power prediction models based on Numerical Weather Prediction ( NWP ) are supposed to be a key technology for tackling network problems of wind energy. During this study, the development of new generation offshore-specific wind power forecasting models was realized by taking some additional input parameters such as the oceanographic variables, wave and wind measurements into account. The work was carried out with the measured data from the offshore wind farm Horns Rev in Denmark which was investigated under the project High Resolution Ensemble for the Horns Rev (HRensembleHR) funded by the Danish PSO Programme between 2006 and Introduction In the future, offshore wind energy will contribute significantly to the European electricity generation mix. According to the German government, the installed capacity of electricity generation from offshore wind power should increase by around 20 GW to 25 GW by the year The development of offshore wind power technology is an extension of wind power activities on land. The future development of offshore wind power will require new participants from the other sectors and scientific fields such as the offshore oil engineering industry and offshore meteorology. Currently, the offshore wind power has a small proportion (almost less than 1%) of the total installed wind power capacity in the world. However, the recent developments of the offshore wind power in some European countries increased general expectations from this type of energy source for the future. According to the 71

89 Development of New Offshore-Specific Wind Power Forecasting Models estimations made by European Wind Energy Associations ( EWEA ), 40 GW offshore wind power will be in operation by 2020 in Europe. [36]. Utilization of offshore wind power is becoming an increasingly important part of the electricity supply in many countries. Inspite of the delay of the offshore wind power activities in Germany, the German Government is aiming to achieve an offshore wind power capacity of up to 25 GW in The following figure illustrates the development of offshore wind power between 1998 and 2007 in Europe. Figure 31: Development of offshore wind power in the EU Currently, the six European countries (Germany, Denmark, Ireland, the Netherlands, Sweden and the UK) have been operating offshore wind farms. The allocation of the offshore wind power among the subject countries is shown in the following figure. 17 Source : ISET/ EWEA 72

90 Development of New Offshore-Specific Wind Power Forecasting Models Figure 32: Total offshore wind power installed by the end of 2007 according to EWEA`s records The potential of the offshore wind energy in Europe is very high. Besides the advantage of the high offshore potential, there are a number of challenges in relation to safe and reliable grid integration due to high concentration levels of large offshore wind farms. This is predominantly the case in the North and Baltic Seas, where only a small number of sea cables connect the wind farms to the electricity grid. Therefore, a relatively large amount of wind energy is produced in relatively small areas. Consequently, the smoothing effects are limited in comparison to onshore wind power. Thus, the balancing and control of fluctuating offshore wind power will be a critical issue for the usage of this energy. Such factors make the offshore wind power difficult to predict. On the top of the challenges mentioned above, the economic risks are also relatively higher than that of the onshore wind power, because of the higher capital costs. In order to overcome these technical and economic challenges a number of solutions will be proposed and developed. A sophisticated wind power management systems and integrated wind power forecasting system will help to integrate offshore wind power into the existing power grid and transmit it to the areas with higher power demand. 73

91 Development of New Offshore-Specific Wind Power Forecasting Models During this study, the development of a new offshore wind power forecasting approach including offshore specific variables (such as wave height and period) and online wind measurements was achieved. A wind power prediction system based on a neural network approach using input from a 75 member Multi-Scheme Ensemble Prediction System ( MSEPS ) was applied to the offshore wind farm Horns Rev. The multischeme ensemble prediction approach which is a technique that employs different parameterizations or schemes are used to vary the calculation of meteorological processes (e.g. vertical diffusion, condensation and dynamics). This technique gives the potential to provide a more realistic representation of the state of the atmosphere by taking the uncertainty of the forecasted weather into account. It is a common practice to use NWP data in wind power forecasting models. In this study, more than the commonly used meteorological parameters were fed into the neural network and examined. Initially, a sensitivity analysis was carried out in order to observe the effect of meteorological parameters on the wind power forecasting. As a next step, the influence of the forecasted and measured wave data on the offshore wind power forecasting was investigated. The most influential parameters were then integrated in the prediction system to improve the prediction accuracy. The results of the first trials include these new oceanographic and atmospheric parameters in the ANN to improve offshore wind power forecasts and will be presented in this chapter. 4.2 State of the Art in Offshore Wind Power Forecasting The State of the Art ( SoA ) in wind power forecasting was generally investigated generally in the previous chapters. In this light, there will only be a brief discussion on SoA in offshore wind power forecasting in this section. In general, it is possible to extend and adapt the models and approaches that are developed for the onshore wind farms to the offshore wind farms [37]. One of the main differences between onshore and offshore wind farms is the roughness of the environment. Therefore, it is necessary to model the 74

92 Development of New Offshore-Specific Wind Power Forecasting Models variability of the roughness of the sea surface 18. Beside the roughness factor of the offshore environment, the thermal stratification of the atmosphere is different from onshore conditions. The ensemble weather prediction based wind power forecasting method was investigated in [38] for the Horns Rev wind farm. This study states that The most complete information that can be provided today consists of probabilistic forecasts, the resolution of which may be maximized by using meteorological ensemble predictions as input. As discussed by Rugbjerg [39], the calibration of the wave forecast models is possible by using the wave measurements when these wave measurements are performed for the offshore wind farm area. The calibrations may improve the short-term wave forecast especially between 12 and 24 hours. Kariniotakis [40] studied the state of the art in short-term forecasting of wind power in the offshore environment and he mentioned that the major wind power developments are expected to take place on offshore and there exists a need to adapt the current wind power prediction systems to the offshore wind conditions. According to [41] the wind power prediction values between 5 and 40 hours varies between 10 to 15 % in terms of NMAE. The results presented in [61] also enforced the results represented in [41]. [42] and [6] have shown that the actual wind power forecasting models can be successfully modified for offshore wind farm conditions. In [43], an air-sea-interaction model called Inertially Coupled Wind Profiles ( ICWP ) was represented, where the Ekman layer profile of the atmosphere and the wave field were coupled. In this study, an aggregated wind power forecast of offshore wind farms with a total capacity of 25 GW in the German Bight was investigated, in which the prediction error was calculated to be between 9 and 17 % in terms of nrmse. In the same study the prediction errors for single wind farms varied between 15 to 20 % (nrmse). Pinson [44] discussed the possible benefits of using the forecasts of wave height or vertical temperature gradient information as additional input 18 [62] 75

93 Development of New Offshore-Specific Wind Power Forecasting Models parameters for the offshore specific wind power forecasting models without any detailed investigation. To date the author is not aware of any study concerning wind power forecasting using all of the available meteorological and oceanographic parameters such as forecasted wave, measured wind and wave data for offshore wind power predictions. Therefore, the investigations that were performed in this study will be pioneer evaluations in this field. 4.3 Wind Wave Interaction Wind - wave interaction can be explained by means of the transfer of momentum, heat and other currents at the sea-air interface. The behavior of wind is different on land than at offshore regions due to the influence of the sea surface on the air flow. The roughness of the sea is very low whereas the surface of ocean varies rapidly in space and time due to changing wave fields. [45]. Lange [45] [60] stated the importance of the stability effects due to the different thermal properties of water compared to land. The figure 33 shows the momentum transfer between wind and water influenced by the roughness of the sea surface. The modeling of the roughness variable of the sea surface is a very critical issue especially for the physical wind power forecasting models. Beside the challenge of the roughness variable, the influence of the wave effects on the offshore wind farm area and the effects of the coast have to be investigated. 76

94 Development of New Offshore-Specific Wind Power Forecasting Models Figure 33: Sketch of influences on the wind field over coastal waters 19 On the other hand, as discussed in [44], the statistical or artificial intelligencebased wind power forecasting models do not need very precise information of the offshore conditions in order to model an accurate wind power forecasting system. However, in this study it is emphasized that, additional information related to the phenomena concerning wind-air interaction such as measured and forecasted wave parameters influences the accuracy of the existing wind power forecasting models. When using this new method to predict offshore wind power generation, it is not necessary to model all the physical details for the AI or statistical based wind power forecasting models. However, it is very important to know which additional parameter is suitable as an input variable for the ANN based wind power prediction systems Wave Spectra Wave spectrum is a physical quantity derived from wave measurements that shows us how much energy is carried by components of different frequencies in real irregular sea waves. Irregular waves are often described as a spectrum that indicates the amount of wave energy at different wave frequencies. Mean wave period (T02) and spectral peak wave period (Tp) values are given on a 19 [45] 77

95 Development of New Offshore-Specific Wind Power Forecasting Models spectrum by plotting spectral density against frequency. A typical wave spectrum is shown below. Figure 34: Representation of wave spectra Spectral Characteristics In this section, some physical characteristics that are directly related to the wave spectral will be denoted. m n is a key characteristics in order to calculate the spectral moment [67], [70]. The n th moment describes the spectral moment and is calculated by using the following equation, where n is any positive integer value (n = 0, 1, 2 ) and ( f ) is the frequency. The frequency for which S( f ) reaches its maximum is called the peak wave frequency [70]. 20 Source: Some parts of the figure are taken from 78

96 Development of New Offshore-Specific Wind Power Forecasting Models n m = f S ( f ) df n f = 0 (4.1) The practical wave analysis may use the frequency ( f ) or the angular frequency ( ω = 2Πf ). Therefore it is possible to give the same equation above as follows: m n n ω S ) ζ 0 = ( ω dω (4.2) From these spectral moments it is possible to calculate many of the time series characteristics. In our case the zero th spectral moment, m 0, and second spectral moment m 2 are important to determine the average wave period (T 02 ort z ). The second (m 2 ) and fourth (m 4 ) spectral moments are used to determine the spectral peak wave period ( T p ). We can demonstrate the spectral peak wave period ( T p ) by using the equation below: T m m 2 p = 2Π (4.3) 4 The mean zero crossing periods (average wave period) (T z ) can be shown by using the following equation: T m m 0 z = 2Π (4.4) 2 The significant wave height (H mo ) can be calculated using the highest onethird of all the wave heights which is dependent on the bandwidth of the spectrum (ε ). The significant wave height (H mo ) may be also be shown ash / 3 1. H mo or H / 3 1 can be given as equation below: 79

97 Development of New Offshore-Specific Wind Power Forecasting Models H m ε 2 1 / 3 = (4.5) If we consider the spectrum to be wide band (ε = 1), H 1 / 3 can be given by: m H1 / 3 = 4 0 (4.6) If we consider the spectrum is wide band ε = 1): m H1 / 3 = (4.7) Finally, the last available measured wave parameter by the waverider is the maximum wave height (H max ). From observations, the highest wave height (H max ) is related to the significant wave height can be expressed by: H H max 1 / (4.8) 80

98 Development of New Offshore-Specific Wind Power Forecasting Models 4.4 Site Description of the Horns Rev Wind Farm In this chapter an investigation was performed at the offshore wind farm Horns Rev. Horns Rev is located in the western part of Jutland in the Danish North Sea. The following map shows the location of the Horns Rev Offshore wind farm. Figure 35: Location of offshore wind farm Horns Rev 21 The wind farm Horns Rev, which was constructed in 2002, is one the largest offshore wind farms in the world. The wind farm has 80 wind turbines with a total installed capacity of 160 MW and is capable of providing almost 2% of the minimum electricity consumption of Denmark. This corresponds to the total energy consumption of 150,000 private households. It was built with the offshore wind turbine type Vestas V80 2 MW which is equipped with the OptiSpeed technology. The cut-in wind speed is 4 m/s, and cut-off wind speed is 25 m/s. The full nominal power output of the wind turbines is reached at 13 m/s. That means if the wind speed exceeds 13 m/s, the turbines will automatically pitch the blades out of the wind in order to prevent overloads. 21 Source: ELSAM and 81

99 Development of New Offshore-Specific Wind Power Forecasting Models The wind farm is located at approximately km to the west of Blavandshuk Wind Measurement Masts There are three meteorological measurement masts installed close to the Horns Rev wind farm. The mast 2 (M2) was installed in 1999 in order to observe the wind and sea conditions. The mast 6 (M6) and the mast 7 (M7) are identical masts which were installed in order to observe the wind farm's wake effects especially when the wind comes from the west. The following figure shows the locations of the masts on the map. Figure 36: The location of the Horns Rev wind farm and the surrounding measurement platforms 23 The measurements of the wind and wave variables are used in this offshorespecific wind power forecasting model in order to investigate the effect of the additional information in increasing forecast accuracy. Since the additional collection of data is costly, the utilization of the mast information has to decrease the forecasting error significantly to justify the additional costs [47] 82

100 Development of New Offshore-Specific Wind Power Forecasting Models Wind Conditions in Horns Rev The annual wind conditions at Horns Rev were reported in [46] and [47]. The wind speed distribution of the M2 measurement mast at 62 m (April 1999 to November 2002) is shown in the following figure. Figure 37: Wind Speed Distribution of the M2 at 62m 24 The average wind speed is around 10 m/s at 62m of the M2 where the northwest and southwest winds are prevailing (western sectors). As it can be seen from the wind speed distribution, around 10 % of the wind speed observations is higher than 15 m/s. Dominating wind sectors can be observed more clearly in the wind rose graphics. [47] 24 [46], [47] 83

101 Development of New Offshore-Specific Wind Power Forecasting Models Figure 38: Wind distribution level 30 m at Horns Rev Wind Farm If we consider the location of the Horns Rev wind farm, the south-west and north-west areas can clearly be identified as the open sea sectors Wave Measurements in Horns Rev A commonly used instrument to measure sea waves is the so-called waverider. The following physical quantities are given by a waverider instrument at the Horns Rev offshore wind farm on the Danish west cost: Spectral estimate on mean wave period: T 02 or T z [s] Spectral peak wave period: T p [s] Significant wave height: H mo [m] Maximum wave height: H max [m] The wave parameters which correlate best with the wind speed in the surface layer were chosen as additional input variables for the wind power forecasting model. The criteria for the selection were set through a correlation analysis and linear interpolation. These wave parameters were selected among a large list of approximately 40 available wave forecasting parameters from the operational wave model ( WAM ) output at ECMWF, such as mean wave direction, coefficient of drag with waves, wave spectral kurtosis and peakedness, wave spectral directional width, mean wave period of the various 84

102 Development of New Offshore-Specific Wind Power Forecasting Models moments, mean wave period and directional width for swell, etc. The table below gives the marine statistics of the wind farm Horn Rev showing the H max, H mo, Tp and T02 parameters: Tab. 7: The marine statistics of the wind farm Horn Rev using Hmax, Hmo, Tp and T02 parameters 25 Unit Mean Max. Hmo [m] Waverider South Hmax [m] T02 [s] Tp [s] Hmo [m] Waverider North Hmax [m] T02 [s] Tp [s] Abs. Current [m/s] Current North Component [m/s] / 0.83 East Component [m/s] / 0.75 Water Level [m] / 2.8 The findings illustrate almost linear relationships between the wind speed and wave height values in the following figure. According to the [46], the wind speeds up to about 20 m/s are more strongly correlated with the measured wave height values. 25 [46] 85

103 Development of New Offshore-Specific Wind Power Forecasting Models Figure 39: Wind speed versus Hmax in Horns Rev 4.5 Evaluation of the Offshore-Specific Wind Power Forecasting Model using Ensemble Weather Predictions and Wave Parameters In this section, the evaluation of the offshore-specific wind power forecasting model using ensemble weather predictions and wave parameters will be investigated. Day ahead and 2 hours ahead very short-term wind power forecasting models will be developed by employing several meteorological and oceanographic parameters. In this day-ahead offshore wind power forecasting model, the NWP parameters from WEPROG's ( MSEPS Model ) and the forecasted wave parameters from the European Center for Medium-range Weather Forecasting ( ECMWF ) were used. The influence of wave parameters in the accuracy of day-ahead wind power forecasting was also analyzed. The tests were carried out for two year period ( to ) with hourly the NWP forecasts from the MSEPS Ensemble system and the 5 wave parameters from 86

104 Development of New Offshore-Specific Wind Power Forecasting Models the ECMWF wave model WAM. The wave forecast models including the water forecasts were all based on the WAM Cycle 4 model [48]. The original WAM information was at a 6 hours resolution. In the pre-processing phase, the data converted to a one hour resolution. The measured wind power information from the Horns Rev was available from February 2005 to July Hence, the common period for all variables was in this case from February 2005 to July The proposed very short-term offshore wind power forecasting model (2 hours ahead) consists of several meteorological and oceanographic input parameters such as the forecasted wind and other meteorological parameters from the NWP's, recently-measured wind power, measured wind parameters, measured wave parameters and forecasted wave parameters. All possible combinations of these parameters were investigated by employing sensitivity experiments in order to observe the influence of meteorological forecasts, wave forecasts, wind measurements and wave measurements on the accuracy of very short-term wind power forecasting models in the offshore environment. The following figure shows the possible combinations of the advanced offshore wind power forecasting model using all the parameters mentioned previously. 87

105 Development of New Offshore-Specific Wind Power Forecasting Models Figure 40: The sketch of all the combinations of the new offshore wind power forecasting models

106 Development of New Offshore-Specific Wind Power Forecasting Models 4.6 Day Ahead Wind Power Forecasting in the Offshore Environment The training and testing periods were conducted weekly. The odd-numbered weeks were used for training and the even-numbered weeks were used for testing of the model. There are also other several options for the selection of the training and testing of the datasets. To compare the forecasting errors, the other training and testing alternatives were performed. It was found that the nrmse values are almost the same. The weekly option was the most compatible option for the double ANN approach. The double ANN combination model will be discussed in the subsequent sections Influence of Several NWP Parameters on the Accuracy of the Wind Power Forecasting In this section, a sensitivity analysis is carried out using a number of NWP parameters in order to evaluate the influence of each NWP parameters on the accuracy of wind power prediction. The following sketch describes the structure of the offshore wind power forecasting model applying only standard NWP data. Before inserting the data into the ANN, a plausibility analysis has been carried out in order to eliminate inconsistent data. The standard artificial neural network based wind power forecasting model was then used to generate the day ahead wind power forecasting time series. Figure 41: The sketch of the day ahead offshore wind power forecasting model using only NWP variables 89

107 Development of New Offshore-Specific Wind Power Forecasting Models In the first experiment (DA_1), only the wind speed and direction parameters at 10m were used. In the second experiments (DA_2), it was seen that the employment of a simple regression method increased the accuracy of the prediction. According to the simple regression method, beside the current historical data (t=0) from the NWP dataset, the previous (t= -1) and the next values (t = +1) as additional inputs were also inserted into the system. In the third experiment (DA_3) we used a special function to optimize the wind direction values before inserting into the ANN module. The first part of the special function converts the wind direction values from grad (between 0 and 360 degree) into wind direction sectors (between 0 and 7). This conversion alone may cause some losses in the wind direction information. That is why, as a next step, it was necessary to form a special function to convert the wind direction sector values in to a special form (triangular) with two components. These two types of wind direction information were then inserted into the ANN module [27]. The following equations summarize the mathematical description of the special function. wd (1) o α - ( α ) - 1 = o (4.9) wd (2) (α) 90 o o -α o = o (4.10) Thereafter, the number of meteorological parameters (such as air pressure, temperature and humidity) the wind speed and direction values at different heights were increased progressively. The description of each experiment is summarized in the table 8. 90

108 Development of New Offshore-Specific Wind Power Forecasting Models Tab. 8: The description of the day ahead offshore wind power forecasting experiments (without wave forecasts) Parameters / Experiments Wind Speed at 10m x Wind Speed at 10m with regression x x x x x x x x Wind Direction at 10m x Wind Direction at 10m with regression x x x x x x x x Wind Direction at 10m with a special function x x x x x x x Wind Speed at 105 m (Level 31) x x x Wind Direction at 105 m (Level 31) x x x Wind Speed at 32 m (Level 30) x x Wind Direction at 32 m (Level 30) x x Wind Direction at 105 m with a special function x Wind Direction at 32 m with a special function x P (Air Pressure) x x x x x x T (Temperature) x x x x x H (Humidity) x Integration of Wave Forecasts into the Offshore Wind Power Forecasting Model In this section, the integration of the oceanographic variables (such as wave prediction parameters from a wave model) into the offshore-specific wind power forecasting model was performed Influence of the Forecasted Wave Parameters on the Day Ahead Wind Power Forecast Accuracy In this experiment, the influence of the forecasted wave parameters from ECMWF's wave model WAM on the accuracy of day ahead offshore wind power forecasting was investigated. The following oceanographic variables were used as input of this forecasting model: SWH: SHWW: MP1: MP2: Significant wave height Significant height of wind waves Mean wave period based on first moment Mean wave period based on second moment 91

109 Development of New Offshore-Specific Wind Power Forecasting Models MWP. Mean wave period Tab. 9: The description of the day ahead offshore wind power forecasting experiment DA_10 (with wave forecasts) Parameters / Experiments 9 10 Wind Speed at 10m Wind Speed at 10m with regression x x Wind Direction at 10m Wind Direction at 10m with regression x x Wind Direction at 10m with a special function x x Wind Speed at 105 m (Level 31) x x Wind Direction at 105 m (Level 31) x x Wind Speed at 32 m (Level 30) x x Wind Direction at 32 m (Level 30) x x Wind Direction at 105 m with a special function x x Wind Direction at 32 m with a special function x x P (Air Pressure) x x T (Temperature) x x Forecasted Wave Data from ECMWF x The following figure shows the structure of the wind power forecasting model including the additional wave forecasts. Figure 42: The structure of the day ahead offshore wind power forecasting model using NWP and forecasted wave variables 92

110 Development of New Offshore-Specific Wind Power Forecasting Models Results of the Day Ahead Offshore-Specific Wind Power Forecasting Model In the table 10, the results of the day-ahead offshore-specific wind power forecasting test are summarized. Please note that, these results demonstrate the sensitivity analysis of the first MS EPS member. After completing the sensitivity analysis for the first MS EPS members, we trained the other left 74 MS EPS members using the best model (DA_10). In the next sections more detailed results will be discussed by comparing the results with the combination methods. It can be seen in the table10 that the corresponding figures for the simplest model (DA_1) have the lowest forecast accuracy (22.86%). Beside the nrmse values RMSE, nmae, correlation and bias values are sorted in the following table. Finally, the wind power forecasting model including the forecasted wave parameter (DA_10) shows an increase in accuracy of the wind power forecasts (17.99%). Tab. 10: The results of the day ahead offshore wind power forecasting experiments for the Horns Rev Wind farm Exp. Code Correl RMSE nrmse nmae BIAS DA_ % 22.86% 16.58% DA_ % 22.83% 16.72% DA_ % 22.65% 16.46% DA_ % 22.52% 16.31% DA_ % 22.01% 15.90% DA_ % 21.95% 15.81% DA_ % 21.16% 15.10% DA_ % 21.03% 15.00% DA_ % 20.70% 14.67% DA_ % 17.99% 12.91%

111 Development of New Offshore-Specific Wind Power Forecasting Models Evaluation of the MS EPS Approach A detailed description of the Multi-Scheme Ensemble Prediction approach was given in the chapter 3. Thus, in this chapter, the focus will be on the performance of wind power forecasts with input from the MSEPS system at the Horns Rev wind farm. First, we applied the neural network based wind power model on all available 75 ensemble members of the MSEPS system. Next, the combination models, such as the simple averaging (arithmetical mean) and the double ANN approach were performed at the offshore wind farm. The DA_10 day-ahead wind power forecasting procedure was applied with input from the MSEPS system. During the investigation, the evaluation of both combination models was performed in a kind of post-processing evaluation. The following sketch shows the general structure of the system based on input from the MSEPS approach. 94

112 Development of New Offshore-Specific Wind Power Forecasting Models Figure 43: The general structure of the MS EPS approach

113 Development of New Offshore-Specific Wind Power Forecasting Models The experiment was carried out from February 2005 to July 2006 with hourly NWP forecasts from the MSEPS Ensemble system. The results of the wind power forecasts with input of all 75 ensemble members are demonstrated in the figure 44. The forecast error (nrmse) varies between 17.74% (35 th member) to 18.83% (19 th member) and the average correlation is approximately The prediction error (nrmse) distribution of the wind power forecasts based on the 75 ensemble members input is shown on the following graph. The X- axis represents the ensemble member number and the Y- axis represents the nrmse error. 96

114 Development of New Offshore-Specific Wind Power Forecasting Models Figure 44: Wind power forecasting error (nrmse) for forecasts based on input from 75 the MSEPS ensemble members for the Horns Rev wind farm

115 Development of New Offshore-Specific Wind Power Forecasting Models Combination Models to Improve the Accuracy of Wind Power Forecasting It has been found in chapter 3 that a combination of forecasts with different weather input always has a superior quality (in terms of performance and accuracy) than the individual one. The first combination method used in this study was (simple) averaging of the output time series of all forecasts from the 75 MS EPS members. In the first method, the averaging was done without weights, i.e. all members had equal weights. In the second combination methodology (so-called double ANN), the best 15 members of the MSEPS were used as meteorological input data for the first stage of the ANN model. The observations showed that after a saturation point, the nrmse values were increasing. In this case, after a combination of the best 9 wind power forecasting time series, the nrmse values started to increase again. Therefore, it made no sense to combine the rest of the wind power forecasting time series from the MS EPS models. The outputs of the first stage of the ANN model were the forecasted wind power time-series of each of the best 15 NWP inputs. The forecasted wind power time series generated from the first stage of ANN model were then used as input to the second stage of the ANN model. Finally, the second stage of the ANN model yielded the optimized forecasted wind power time-series. In addition to the generation of the arithmetic mean and the 2-step ANN method with the 15 the best members, a sensitivity study was carried out. The purpose of this sensitivity study was to find the optimal amount of ensemble members to be combined. The 15 best members were combined one by one till reaching the turning point at the nrmse values which is called as saturation point. In this way, a saturation point was found when combining the 9 best members. Both the simple averaging (Arithmetical) method with a nrmse of % and the double-ann method with a nrmse of 16.8% register the lowest level when combining the 9 best members at the Horns Rev offshore wind farm. A 98

116 Development of New Offshore-Specific Wind Power Forecasting Models 5.3 % improvement was achieved by using the 2-step ANN method in relation to results gained from the best MS EPS member. It is proved that it is possible to reach the onshore wind power accuracy levels, if the offshore wind power forecasting models are adjusted to the offshore environment by using additional oceanographic and meteorological parameters. In addition to this new adjustment approach, the combination models have a great potential to optimize the offshore wind power predictions and obtain the lowest level of the nrmse values. 99

117 Development of New Offshore-Specific Wind Power Forecasting Models Figure 45: Results of the combination models by employing the simple averaging and the double ANN (2ANN) methods

118 Development of New Offshore-Specific Wind Power Forecasting Models Comparison of the Day Ahead Wind Power Forecasting Results The results of the calculations clearly indicate that the new offshore wind power forecasting model brings improvements in forecasting accuracy. The improvement of the day ahead wind power forecasting model has been discussed in terms of nrmse. For the day ahead forecasting, we considered the simplest (DA_1) day ahead forecasting model as reference. The usage of the temperature values (DA_5) from MS EPS data decreased the nrmse value (22.01%). On the other hand, the air pressure (DA_4) and humidity (DA_6) parameters did not change the level of the forecast accuracy very much. The wind speed and direction information at 32 and 105 m (in addition to the wind speed and direction at 10m) lead to an improvement in the prediction accuracy % of improvement with respect to the DA_1 model is achieved by optimizing the NWP input variables (DA_9). A 21.3 % of improvement with respect to the DA_1 model is achieved by integrating the WAM input variables (DA_10 with % nrmse). The utilization of forecasted wave parameters results in a significant improvement with respect to forecast accuracy. The application of the multi-model approach (the simple averaging and 2 ANN models) improved the forecast accuracy. The improvements achieved by applying the simple averaging the simple averaging model and the double ANN approach are respectively % and %. It is seen that the level of performance of the multi-model approach is similar to the onshore wind farms in the offshore environment. 101

119 Development of New Offshore-Specific Wind Power Forecasting Models Figure 46: Comparison of the day ahead wind power forecasting results for the wind farm Horns Rev

120 Development of New Offshore-Specific Wind Power Forecasting Models 4.7 Evaluation of the Two Hours Ahead Very Short-term Offshore Wind Power Forecasting Model The main type of input variables in day-ahead wind power forecasting models generally comes from the NWP models. One of the problems with the NWP based input is that the data is already a few hours old when they become available for applications such as wind power forecasting. This implies a startup error of the forecast. In order to reduce this start-up error, the day-ahead forecast may be modified with the help of the observations. When conducting such ultra-short-term forecasts, it is also useful to reduce the forecast frequency to 15 minutes. In this study, two hours ahead wind power forecasts were generated. For this purpose, the online wind power production information was used in addition to the NWP input data in order to achieve a correction to the raw wind power forecasting value. The very short-term wind power forecasting hence allowed us to reduce the deviations between the measured and forecasted wind power values in the first few hours of the forecast. The following measured and forecasted variables were used to produce such a 2 hours ahead very shortterm wind power prediction time-series: Measured wind values (wind speed and direction) Measured wave parameters Forecasted wave parameters NWP dataset Measured wind power output 103

121 Development of New Offshore-Specific Wind Power Forecasting Models Descriptions of the Very Short-term Wind Power Forecasting Experiments In order to perform a sensitivity analysis, all combinations of the available meteorological and oceanographic parameters by eight experiments were investigated. The NWP and the actual wind power data were used together as the input dataset for all the experiments. The other available input variables were employed step by step systematically in to the new offshore wind power forecasting model. The test was carried out from to with hourly NWP forecasts from the MSEPS Ensemble system. In addition to the NWP data, five wave parameters from the ECMWF wave model WAM, the measured four wave variables and wind measurements from the site were utilized. The original WAM information was in 6 hours resolution. In the preprocessing phase, the data converted to one hour resolution. The measured wind power information from the wind farm Horns Rev was available from February 2005 to July Hence, the common period for all variables is in this case was from February 2005 to July The following table gives a short description of the short-term wind power forecasting experiments for the wind farm Horns Rev. Tab. 11: Overview of the short-term offshore wind power forecasting models Parameters / Exp. Code 1 2WF 3W 4 5WF 6W 7WWF 8WW Pers. NWP Data x x x x x x x x Measured Wind Power Data x x x x x x x x Forecasted Wave Data x x x x Measured Wave Data x x x x Measured Wind Data x x x x 2 Hours Persistence x ST2_1 Experiment The first very short-term offshore-specific wind power forecasting experiment is named ST2_1. This is the simplest model as it only uses NWP variables such as wind speed, wind direction, temperature, other available meteorological parameters and measured wind power values from the Horns Rev offshore wind farm as input parameters. After performing a plausibility analysis in the pre-processing part, the 104

122 Development of New Offshore-Specific Wind Power Forecasting Models input variables were inserted into the ANN module. The output of the model was the forecasted wind power time-series for the next 2 hours. ST2_2WF Experiment In this experiment, the forecasted wave parameters from the ECMWF WAM model were inserted into the offshore wind power forecasting model, in addition to the NWP input and measured wind power information. In the pre-processing part, a correlation analysis and linear interpolation were performed before inserting the forecasted wave parameters into the ANN model. The resolution of the WAM data was 6 hours while the forecasting model used data with 1 hour resolution. Thus, it was necessary to perform a linear interpolation in order to generate the 1 hour resolution forecasted wave data. ST2_3W Experiment In this experiment, measured wave parameters such as the Hmax, Hmo, To2 and Tp were added to the NWP input and measured wind power information in order to investigate the influence of measured wave parameters on the accuracy of the offshore-specific wind power forecasting model. In the preprocessing part, a plausibility analysis for the NWP input and the measured wind power values were performed. The measured wave values were logged in 30 minute resolution. Therefore the values were averaged to a 1 hour resolution in order to make them compatible with the wind power forecasting model. ST2_4 Experiment Measured wind parameters such as wind speed and wind direction are very important variables for a very short-term wind power forecasting model, if the wind measurements are plausible and the measurement station is located close to the wind farm. Usage of wind measurement parameters as additional input parameters for the wind power forecasting model has a potential to increase the accuracy of the predictions. In the case of the Horns Rev wind farm, the measurement masts are very close to the wind farm (about 2 km far away) and the environmental effects are very small in the offshore regions. 105

123 Development of New Offshore-Specific Wind Power Forecasting Models ST2_5WF Experiment The forecasted wave variables from the WAM model such as MP1, MP2, MWP, SWH, SHWW and the wind measurements from Mast 2 (M2) were used as inputs into the wind power forecasting model beside the NWP input and measured wind power data. ST2_6W Experiment In this experiment all the available types of the meteorological and oceanographic variables were used as inputs in addition to the NWP data and measured wind power values. ST2_7WWF Experiment In this experiment, the forecasted wave variables from the WAM and the measured wave parameters were used as inputs to the model in addition to the NWP variables and measured wind power information. ST2_8WW Experiment In this experiment the measured oceanographic (wave) and meteorological (wind) parameters were used as inputs into the system in addition to the NWP and measured wind power information Results of the Offshore Very Short-term Wind Power Forecasting Models According to the results which are illustrated in the table 12, the forecasting error value (nrmse) of the two hours ahead persistence value is 14.7 %. This error is higher than all the other very short-term wind power forecasting experiments, which indicates that it is beneficial to use more advanced methods in the very short-term forecasting than only the persistence. In addition to the nrmse values, correlation, RMSE, nmae and bias values are 106

124 Development of New Offshore-Specific Wind Power Forecasting Models also listed in the same table. The improvements in the correlation and nrmse values can be observed between the ST2_Pers and ST2_6W experiments. Tab. 12: Overview table of the very short-term offshore wind power forecasting results of the Horns Rev wind farm Exp. Code Correl RMSE nrmse nmae BIAS ST2_Pers % 14.70% 9.28% 0.00 ST2_ % 11.67% 7.73% ST2_2WF % 11.36% 7.66% ST2_3W % 11.35% 7.90% ST2_7WWF % 11.32% 7.68% ST2_ % 11.05% 7.50% ST2_5WF % 10.88% 7.30% ST2_8WW % 10.70% 7.50% 0.04 ST2_6W % 10.67% 7.30% In the following figure the results (in terms of nrmse and Improvements) of all the very short-term wind power forecasting experiments are shown. The results are sorted according to the nrmse values. 107

125 Development of New Offshore-Specific Wind Power Forecasting Models Figure 47: Prediction error values (in terms of nrmse) vs. improvements of 2 hours ahead wind power forecasting for the Horns Rev wind farm

126 Development of New Offshore-Specific Wind Power Forecasting Models If we take the persistence model as the reference model, we can reach a 20.61% improvement in the forecast accuracy value by using the ST2_1 model. The nrmse value for the ST2_1 experiment is %. As it can be observed from the results, the improvement of wind power forecasting accuracy using the forecasted wave variables (model ST2_2WF) is % while the nrmse is %. The influence of the measured wave parameters on the accuracy of the offshore (very short-term) forecasting model (ST2_3W) is a little bit more (22.79 % improvement) than using only forecasted wave information (ST2_2WF) from the WAM model. The improvement of wind power forecasting accuracy using the measured wind parameters is clearly more than the measured and forecasted wave parameters. The ST2_8WW experiment yielded the second best result in terms of nrmse (10.70 %). Finally, in the ST2_6W experiment, where all the available parameters (measured wave and wind information, forecasted meteorological and oceanographic variables) were used, yielded the lowest prediction error value with an nrmse of %. In this experiment, we can obtain the lowest nrmse value (10.67%) with a % improvement in the prediction accuracy. It is obvious that, utilization of the measured wave and wind information improved the wind power forecast accuracy. The economic and technical advantages of using each of the investigated parameters will be discussed in the next chapter. The graph below shows the probability density distribution of two very shortterm wind power forecasting models (ST2_1 and ST2_6W) versus persistence. As observed from the following figure, more values were collected from the model ST2_6W than the model ST2_1 in the center. That means that the prediction error values of the model ST2_6W are less than the ones of the model ST2_1. The influence of the wind and wave parameters on the new offshore wind power forecasting can be observed the following graphic: 109

127 Development of New Offshore-Specific Wind Power Forecasting Models 40% 35% 30% HR_ST2_1 HR_ST2_6W HR_ST2_PERS 25% 20% 15% 10% 5% 0% Figure 48: Comparison of probability density function values (ST2_1, ST2_6W vs. ST2_Pers.)

128 Development of New Offshore-Specific Wind Power Forecasting Models 4.8 Comparison of Day Ahead and Very Short-term Offshore Wind Power Forecasting Model The comparison of the day ahead single model, the day-ahead combination and the 2 hours very short-term offshore wind power forecasting results are compared in this section. The following figure summarizes the day ahead and short-term wind power forecasting results in terms of nrmse for the Horns Rev wind farm. Figure 49: Comparison of the prediction error values of the day ahead and 2 hours ahead wind power forecasting models The prediction error (nrmse) reduced from % to % by using additional wave parameters in addition to the classical NWP data. The results of this chapter clearly indicated that the new offshore day ahead wind power forecasting using additional oceanographic parameters model brings improvements with 21.3% with respect to the DA_1 model and this was achieved by integrating the WAM input variables (DA_10 with 17.99% nrmse). Besides integrating the WAM input variables, the application of the combination model approaches (such as simple averaging and the 2 ANN models) improves the forecast accuracy up to %. It can be concluded that the performance level of the multi-model approach is similar that of the onshore wind farms in the offshore environment. The two combination methods (the DA_Arit and DA_2ANN models) yielded also a reduction on the nrmse. Finally, the forecast error values were decreased down to % by employing the 2 hours ahead wind power forecasting models. 111

129 Development of New Offshore-Specific Wind Power Forecasting Models 4.9 Conclusion and Discussion In this chapter, a new approach of an offshore-specific wind power forecasting models using meteorological and oceanographic variables was presented. The potential benefits using new types of input variables such as measured and forecasted wave information in addition to the classical input information from NWPs were explored in order to improve the accuracy of the forecasting models in the offshore environment. In order to optimize the wind power prediction results, additional oceanographic, measured wind parameters and other advanced wind power forecasting techniques such as the ensemble prediction (MS EPS), the multi model (the double ANN model) and the simple averaging approaches were implemented. The results of this study in regards to offshore power forecasting indicate that, the error is relatively high in relation to the installed capacity, but relatively low in relation to the generated power. It was also found out that the forecast growth error is higher than for the error on land due to the uncertainty in the offshore weather forecast. These findings showed that the additional oceanographic and wind measurements in the vicinity of the offshore wind farms improve the grid integration capability of offshore wind power. In the next chapter, the energy economic benefits of the utilization of these new models will be covered. If we interpret the results and improvements achieved by using several meteorological and oceanographic input data, it can be obviously noted that it makes sense for the responsible parties such as TSOs and offshore wind farm operators to measure wind and wave conditions continuously in neighborhood of the planned the wind farm. The investment costs are relatively low in comparison to the benefits gained by using these useful parameters. A more 112

130 Development of New Offshore-Specific Wind Power Forecasting Models detailed discussion concerning the future of TSOs and the wind farm operators will be covered under the Outlook and Conclusion section. 113

131 Energy Economic and Technical Benefits of Wind Power Forecasting 5 Energy Economic and Technical Benefits of Wind Power Forecasting Abstract The goal of this chapter is to analyze the wind power forecasting systems from economic and technical point of views. In the final chapter, the investigations concerning the economic costs of wind power output fluctuations and the added value of wind power predictions will be evaluated. Beside the direct economic benefits of wind power forecasting, the possible reductions in the needed regulating or balancing power via utilization of the intra-day trading option will also be discussed. 5.1 Introduction The integration of intermittent energy sources into the electricity grid has become an important challenge for the TSOs, due to the fluctuating behavior of wind power generation. Wind power predictions improve the economic and technical integration capability of large amounts of wind energy into the existing electricity grid. Trading, balancing, grid operation, controllability and safety issues increase the importance of the forecasting power output from the wind power operators and TSOs point of view. The wind forecasts are relatively precise for the time period of only a few hours. An accurate forecast allows trading the wind energy and also allows grid operators to schedule economically efficient generation to meet the demand of their customers. Therefore, wind power forecast systems have to be integrated into the monitoring and control systems of the transmission system operator ( TSO ). According to the main outcomes and recommendations of the EU Policy Workshop [49], the utilization of the intra-day trading option for the offshore wind farms is listed among important recommendations from an energy policy point of view. 114

132 Energy Economic and Technical Benefits of Wind Power Forecasting It is noted that trading close to real time improves the integration of fluctuating renewable energy sources. Intra-day trading should therefore be developed throughout Europe. The influence of the wind energy on the power market is investigated in this chapter in order to spot a possible correlation between the generated or forecasted wind power and the electricity prices. Finally and most importantly, an improved integration of wind power by taking the intra-day markets into account is investigated in order to reduce the need of reserve power. The amount of the reserve power required depends on what amount of energy must be available as reserve in order not to exceed the given probability of imbalance. Day-ahead wind power forecasting models are used in the spot-markets with day-ahead trading option, such as in Germany, France and the Nordic countries in order to trade and schedule operations. In the countries with rolling markets, such as the UK, Australia, USA and Canada, where shorter forecast lengths can be used for trading and scheduling in addition to the day-ahead forecasts. The very short-term forecasts of wind power have significantly higher qualities than day-ahead forecasts. On the other hand, it does not mean that the very short-term wind power forecasts should be replaced with the day-ahead forecasts. It simply means that, they should complement the day-ahead forecasts. Without deploying the intra-day market, the deviations between actual and forecasted wind power values need to be balanced during the operational control by the control reserve. The price of controlling reserve power is very high in comparison to other sources of power. By using the very short-term wind power forecasting before the gate closure in the power market, it is possible to reduce the requirement of reserve power. These hourly trade options have been possible in the German intra-day power market since

133 Energy Economic and Technical Benefits of Wind Power Forecasting 5.2 The Influence of Wind Power Generation on the Electricity Prices Most European power markets have recently faced a significant increase in wind power generation. We can expect that this trend will continue in accordance with the given 2020 targets of the European Commission for the Renewable Energies. In this section, the influence of wind power generation on the electricity prices is explained using two approaches, one theoretical approach and practical approach Theoretical Approach The theoretical approach concerning the influence of the wind power on the power market can be explained by using the merit order effect. [50], [51] The main principle of the theoretical approach is that, if the amount of wind power generation increases, the electricity market prices will decrease due to a reduction of load power plants with a more expensive base. As presented in [52], when the wind generation is high, the supply curve shifts to the right hand-side. Therefore, the spot prices on the power market decrease, especially during the peak demand hours. The figure 50 illustrates a sample of the operational unit order of some power plants. 116

134 Energy Economic and Technical Benefits of Wind Power Forecasting Figure 50: A sample for the operational unit order of some power plants Practical Approach In this section, the day-ahead power market price and day ahead wind power forecasting values in Germany for the year 2006 are used in order to investigate the influence of the wind power generation on the electricity prices. The dependencies between the day ahead wind power forecast and the day ahead power market price values can be seen in the figure 51. If the amount of the wind power generation increases, the electricity price on the market decreases. In order to observe the same effect, a bin-averaging of the wind power prediction values using the daily average values of forecasted day ahead wind power (clustered in 1000MW groups) and the Phelix-based-price values of the day ahead power market values at the European Energy exchange ( EEX ) in Leibzig, Germany, were performed. The Phelix-based-price values are the 26 Source: RiSOE 117

135 Energy Economic and Technical Benefits of Wind Power Forecasting average price values of the auction prices for all hours of the day. The influence of the wind power generation on the power market prices can be observed more obviously in the following figure. Figure 51: Dependency between the day ahead wind power forecast results and the Phelix-based prices in 1000 MW classes (2006) 5.3 Overview of the European Electricity Markets Trading bilaterally on the Over-The-Counter ( OTC ) market is used to be the classical way of carrying out transactions in the single European electricity market. After liberalization of electricity markets in Europe, alternative exchange markets were established in many countries. They offer new trading options for the power market and have reached different levels of maturity. [53] As presented in [53], all contracts are characterized by three main components: A specified price, a time of delivery and a certain amount of electricity. The market can be divided according to distance to delivery, which shows a correlation with the market player s intentions: Trading long-term contracts as forwards, futures and options in the derivatives market minimize price risks and are thus used for hedging. If a derivative becomes mature, it can be counteracted either financially or integrated into the day ahead market for the physical delivery. In day-ahead 118

136 Energy Economic and Technical Benefits of Wind Power Forecasting markets, the seller is supposed to deliver the traded amount of electricity to the buyer in the following day. According to the rules of the spot (day-ahead) market, the offers are submitted using hourly or block contracts. Block bids can be flexible or standardized [53]. [54] stated that, typically, bids can be submitted for a time span between 12 to 36 hours before the physical delivery. The price equilibrium is achieved after collecting the generation and demand information for each delivery hour of the day in the regulation hour. Figure 52: Overview of the wholesale market The actual production may be different from the submitted amount on the dayahead market due to the time span between gate closure of the day-ahead market and physical delivery. Thus, participants need daily or even hourly physical contracts in order to decrease imbalances in their production portfolios. Intra-day markets are designed to provide trading options for participants that have to conduct shorter-term trading transactions. The intraday market covers contracts on the same day up to one hour before physical delivery. The balance of supply and demand has to be ensured in order to maintain a secure and stable operation of the electricity system. TSOs are responsible for the management of the balancing mechanism by adjusting production 119

137 Energy Economic and Technical Benefits of Wind Power Forecasting downwards and by maintaining additional capacity for upward regulation continuously during real-time operations. This is generally referred to as the balancing or the regulating market, which is again subdivided into several markets with different characteristics [53]. 5.4 State of the Art in Energy Economic Benefits of Wind Power Forecasting and Utilization of Intraday Trading As discussed in [55], the utilization of ensemble wind power forecasting models helps the integration of large-scale wind power production into the power system. Ensemble forecasting provides new opportunities for the market players to achieve more creative and efficient trading options. Therefore, ensemble wind power forecasting models are proposed in this study to optimize trading and the daily operations of the responsible parties in the power market. The economic benefits and importance of using short-term wind power forecasting were investigated [56] on the UK electricity market from the viewpoints of the wind farm operators and transmission/distribution system operators. According to [56], it is more beneficial to utilize the short-term wind power forecasting models for the purpose of power trading and it is possible to achieve a maximum benefit of forecasting of approximately 4.50/MWh (with regard to 2003 prices) calculated for a wind farm in the UK. Parkes [57] stated similar results, where the energy economic benefits are calculated at 5/MWh for a single 50 MW wind farm under UK electricity market conditions. Holttinen [58] is referred to as a successful study concerning the optimal electricity market for wind power from the point of view of wind farm operators. 8 % of gain is calculated by using better hourly wind power prediction models on the NordPool power market. The gains for trading the wind power are calculated by 7 % by using the ELBAS market as an after sales tool. The same study reported that, the estimated surplus or missing production of 2 hours before delivery would decrease the regulation costs by up to 70 %. 120

138 Energy Economic and Technical Benefits of Wind Power Forecasting 5.5 Wind Power Trading Evaluation and Utilization of Intraday Trading An improved integration of wind power by employing the intra-day markets is proposed in order to reduce the need of control reserve power and increase the value of wind power in the power market. Energy economic benefits of wind power forecasting were analyzed in the Nordic Electricity Market (NordPool) in the final part of the study. If only the day ahead wind power forecasting is available, the seller (trader or the wind farm operator) should only use the day ahead market option. Deviations between the day ahead wind power generation values and the real time generation values are compensated by using control reserve or regulation power market. On the other hand, if the seller has a very short-term wind power forecasting model beside the day head predictions, the intraday trading can be used as an additional trading option. The deviations between the very short-term and day ahead wind power forecasting values are compensated by using the Intraday market. Finally, the deviations between the very-short term wind power forecasting and the real time generation values are less than the trading option without the Intraday trading option. 121

139 Energy Economic and Technical Benefits of Wind Power Forecasting Figure 53: General structure of power market and utilization of intra-day market

140 Energy Economic and Technical Benefits of Wind Power Forecasting The wind power trading evaluation was performed for a Danish offshore wind farm (Horns Rev) and applied on the Nordic electricity market (NordPool). Although Germany and Nordic countries have different power markets with different regulations, both markets share some common basic trading principles which state that, the higher the wind power forecast accuracy, the higher the revenue from wind power. In this study the following assumptions are made: Trading calculations were based on persistence, perfect forecast case, day ahead and 2 hours ahead wind power forecasts. It was considered that the balance responsible entity (in most countries the TSO) is operating only the selected wind farms or the Direct Marketing Rules which are valid for the sample wind farm. Market prices were used in the NordPool ELSPOT and ELBAS area. Real-time market prices at Nordpool were used for up- and down regulations. The day ahead forecasted time-series were performed from optimal NWP batches on the previous day for 00:00-24:00 the next day. Therefore, day ahead wind power forecasts are ready for the Day Ahead electricity market (13-37h). Please note that, the fact is that indeed the 06UTC forecast in the Nordic market is used for the trading from the balance responsible parties of wind power which would be 18-42h. 2 hours ahead short-term wind power forecasts were applied on the Intra-day market where the bids can be traded up to 1 hour in advance before the closing time. Perfect forecasting was defined as the best case scenario where the forecasted and measured wind power is the same. To have a consistent study, it was important to use the data of the three markets (day-ahead, hour-ahead and imbalance market) for the same region and over a sufficiently long period to ensure accuracy of the results. After 123

141 Energy Economic and Technical Benefits of Wind Power Forecasting some investigations, it was decided to use the data from NordPool-Finland, where day-ahead, short-term and imbalance prices are freely available on the internet. The used data cover the period between and There were some missing hours or days in this analysis period due to the missing wind power forecasting time series. However, more than 7150 hours were simulated in this study Description of the Possible Trading Options Trading on the Day-Ahead (Spot) Electricity Market On the spot electricity market, the seller or the balance responsible party is supposed to deliver the traded amount of electricity to the grid on the next day to be available for the buyer. For instance, in the NordPool the bids for the market have to be given before 12:00 on the previous day in order to submit a generation schedule and a distribution schedule. The market participants predict their output (wind) power and trade that energy on the day-ahead market. The day ahead wind power predictions are used for this option. Figure 54: Wind power trading on the day ahead market 124

142 Energy Economic and Technical Benefits of Wind Power Forecasting As discussed by [58], in Nordic Countries, the so-called two price model is being operated. Only one regulated price (up or down) is defined depending on the direction of the system imbalance and the same price is used for both sales and purchases on the hour when settling imbalance of the individual market participant. That means that, if the imbalances due to wind power forecasting errors increase the system imbalance, the regulation price is applied. Otherwise, the imbalances are priced with the spot prices Trading on the Intraday Electricity Market Intra-day markets are designed to provide trading options for the balance responsible party or wind power operator that have to conduct the shorterterm trading transactions. Short-term trading is mostly organized by means of a power exchange. The intra-day market covers contracts on the same day up to one hour before physical delivery. The hour-ahead (intra-day) market or adjustment market is the best solution to trade energy in a flexible way until one hour before the time of production. If a short term wind power prediction model is available, the deviation between the day-ahead prediction and the latest (in our case 2 hours ahead short-term wind power forecasts) prediction is detected by the balance responsible party or wind power operator and he has then the opportunity to trade the difference on the hour-ahead spot market to correct its generation and distribution schedules, and minimize its imbalance. The trading mechanism is like the two price model of the dayahead markets. If the latest (short-term) wind power forecast value is above the day-ahead wind power prediction value (lacking or shortage of energy), the amount of energy which deviates from the day ahead forecasts is purchased in the hour-ahead (intra-day) electricity market. Otherwise, if the day- ahead wind power forecasting value is more than the short-term wind power value, the surplus energy is sold on the hour ahead electricity market. (See the figure 55) [58], [59] 125

143 Energy Economic and Technical Benefits of Wind Power Forecasting Figure 55: Wind power trading on the hour ahead market (adjustment by using short-term wind power forecasting model) Trade on the Regulating Market and Imbalance Settlement Mechanism The balance of supply and demand has to be ensured in order to maintain a secure and stable operation of the electricity system. TSOs are responsible for the management of the balancing mechanism by adjusting production downwards and by maintaining additional capacity for upward regulation continuously during real-time operations. This is generally referred to as the balancing or the regulating market, which is again subdivided into several markets with different characteristics. When the actual production is different from the last submitted (2 hours ahead wind power forecast) predictions, the price for surplus energy is usually respectively higher than in any other trading options. As a result, market players avoid trading on that market. Please note, as mentioned previously, a two-price model is used for the simulations in this study. 126

144 Energy Economic and Technical Benefits of Wind Power Forecasting 5.6 Cases studied Five different case scenarios are described in this section. The following case studies have been selected: Tab. 13: Overview of the different simulated scenarios Scenarios Case 1 Case 2 Case 3 Case 4 Case 5 Description Day Ahead Trading with prediction model DA1 (Ref. Scenario) Day Ahead with prediction model DA10 Day Ahead + 2 Hours Ahead Trading with DA10 and ST_2WF Day Ahead + 2 Hours Ahead Trading with DA10 and ST_6W Perfect Prediction Case 1 is the simplest scenario in which a prediction tool is used. The power schedule is filled up with the help of a day ahead wind power prediction model DA1. The DA1 is the simplest day ahead wind power forecasting model with only two input parameters (wind speed and wind direction at 10m). No corrections were performed and the prediction errors were dispatched by the imbalance mechanism. Case 2 is the second scenario in which a better day ahead wind power prediction tool is used. The power schedule was filled up with the help of the day ahead wind power prediction model DA10. The DA10 is the advanced day ahead wind power forecasting model with additional forecasted wave parameters from a wave model (the WAM model from ECMWF) as input variables. No corrections were made and the prediction errors were dispatched by the imbalance mechanism. Case 3 is the scenario in which advanced wind power prediction tools were implemented: a day ahead forecast model (DA10) was used to fill the power schedule and just before the time of operation. Corrections were made using the very short term wind power prediction model ST2_2WF. The ST2_2WF was a two hours ahead wind power forecasting model without wind speed measurements. 127

145 Energy Economic and Technical Benefits of Wind Power Forecasting Case 4 is the scenario in which the advanced wind power prediction tools were implemented: a day ahead forecast model (DA10) was used to fill the power schedule shortly before the time of operation and corrections were made using the short term wind power prediction model ST2_6W. The ST2_6W model was a two hours ahead wind power forecasting model with measured wind speed measurements, forecasted and measured wave parameters. Case 5 is the final scenario with a perfect prediction. That means that, the deviation between day ahead wind power forecast and the measured wind power value from the wind farm is the same. In that case, there is no need to consider the regulated power. The whole generated wind power is traded on the day ahead (spot) market. The following table summarizes the wind power forecasting error values in terms of nrmse (Normalized Root Mean Square Error) for the various models in use as a basis for the error computations resulting in imbalances. Figure 56 : Wind power forecasting results of the various experiments 128

146 Energy Economic and Technical Benefits of Wind Power Forecasting 5.7 Comparison of the Results Case 1 can be considered as a worst case scenario in this context, since it results in the largest amount of imbalance and smallest revenue. In this case, only a day-ahead wind power prediction model is used and there is no correction of the wind power forecasts in the intra-day electricity market. The final value of the produced wind energy in this case is /MWh. The amount of the up-regulation imbalance is GWh, which corresponds to 20.35% of the total contracted energy. The amount of the down regulation imbalance is GWh, which corresponds to % of the total contracted energy. Case 2 is the scenario that uses only the day ahead (spot) market option for the trading of wind energy. In this case, only a day ahead wind power prediction model is used and there is no correction of wind power forecasts in the intra-day electricity market. With the use of a day ahead prediction, (with additional wave parameters as input) a total produced energy price of /MWh was obtained. The amount of the up regulation imbalance is 72.3 GWh, which corresponds to % of the total contracted energy. The amount of the down regulation imbalance is 74.9 GWh, which corresponds to 15.08% of the total contracted energy. Case 3 is the scenario that utilizes an advanced wind power prediction model including a day-ahead prediction and short term prediction (simplest model) tool for the trading of wind energy. The final value of the produced energy in this case is /MWh. The amount of the up regulation imbalance is 47.4 GWh (corresponds to 9.55 % of the total energy contracted). The amount of the down regulation imbalance is 43.8 GWh (corresponds to 8.82 % of the total energy contracted). Case 4 is the scenario that utilizes an advanced wind power prediction model including a day-ahead prediction and a short term prediction (an innovative model with measured meteorological and oceanographic variables) tools for the wind energy trade. The final value of the produced energy in this case is 129

147 Energy Economic and Technical Benefits of Wind Power Forecasting /MWh. The amount of the up regulation imbalance is 35.9 GWh (corresponds to 7.24 % of the total energy contracted). The amount of the down regulation imbalance is 36.9 GWh (corresponds to 7.45 % of the total energy contracted). Case 5 refers to a perfect prediction scenario in which there is no deviation between the day ahead wind power predictions and measured wind power values. In this case, all the contracted wind power is traded on the day ahead electricity market without any correction needed in the intra-day and regulation markets. Thus the need of the regulation power is zero. The final value of the wind energy in this (ideal) case is /MWh. The energy economic investigations show that a monetary advantage can be achieved by employing short-term wind power forecasting on the intra-day electricity market per MWh additional revenues can be achieved by using an advanced prediction tool including the very short-term wind power forecasting module (Case 4) on the intra-day market compared to a prediction tool using a simple day ahead wind power forecasting module (Case 1). Figure 57: Average Energy revenue Euro per MWh (for the 7150 hours period) 130

148 Energy Economic and Technical Benefits of Wind Power Forecasting The reduction of the required balancing power can be achieved by using the intra-day trade option. Approx. 65 GWh (from 101 GWh to 35.9 GWh) of reduction in the requirement of the up regulating power and approx GWh (from 87.6 GWh to 36.9 GWh) of reduction in the requirement of the down regulating power can be achieved by utilizing intraday market. In other words, utilization of the Intraday trading option in combination with advanced (day ahead and very short-term) wind power forecasting models reduced the amount of needed up regulation power up to 64.5 % and down regulation power 57.8 %. Tab. 14: Needed Up and Down regulation values according to the different cases (for the 7150 hours) Scenarios Up Reg. [%] Down Reg. [%] Imbalance Up Imbalance Down Reg. [GWh) Reg. [GWh) Case % % Case % % Case % 8.82 % Case % 7.45 % Case % 0.00 % ,82 can be saved by using the Case 4 instead of the Case 1 during the evaluation time (7150 Hours) for the offshore wind farm Horns Rev with 160 MW installed capacity. The economic value of wind power forecasting system can be improved by using more advanced wind power forecasting models in combination with intra-day trading option. Figure 58: Total revenue values according to the various case scenarios 131

149 Energy Economic and Technical Benefits of Wind Power Forecasting A similar calculation can be performed for a capacity of 25 GW offshore wind power capacity, which is planned by the German government by If the following assumptions are taken into account; An average wind power capacity factor of offshore wind farms is 40 % There is a credit to the spatial smoothing effect that reduces the wind power forecasting error in the larger area, As presented in [43], the aggregated wind power prediction error is calculated between 9 to 17 % in terms of nrmse for a total capacity of the future offshore wind power capacity (25 GW), where the adjustment factor due to the spatial smoothing effects is given as A 2 per MWh additional revenue level is used for the calculation. Please note that, this investigation is performed in the NordPool market. Therefore the dynamic price structure may be different in the future (in 2020 or 2030) for the German power markets. According to the assumptions listed above the estimated gain attained by using an advanced wind power forecasting tool in the intra-day trading market is an annual revenue of up to million. If the adjustment factor due to the smoothing effects calculated by [43] is considered, we should downscale the possible revenue down to million. As known, the so-called direct marketing of electricity generation from wind farms was regulated by the revision of the renewable energy law (EEG) in Germany in 2009 ( ). According to this regulation, wind farm operators are allowed to trade their generation on the spot market over a monthly calendar period in full or as fixed portion of their installed capacity. According to the new regulations, offshore wind farms will receive an initial remuneration of 15ct/kWh for 12 years, if they are in operation by the end of 2015 and 13ct/kWh for 12 years after If we consider the increased amount of the remuneration for offshore wind energy, we can say that direct marketing option without any direct marketing bonus may not be the most attractive trading option for the wind farm owners at least for every month in a year. Therefore, the estimated value of additional revenue for offshore wind power of German wind farms in the future would rather be a reduction in costs of trading the wind power and may be considerably less than million. This amount has to be computed by 132

150 Energy Economic and Technical Benefits of Wind Power Forecasting taking different assumptions into consideration both on the market structure, the price development with 25GW installed offshore wind and the intra-hourly volatility of the prices. 5.8 Conclusion The goal of this economic analysis of the wind power forecasting system was to provide some insights to the economic costs caused by wind power output fluctuations and on the added value of wind power predictions. Beside the direct economic benefits of wind power forecasting, the reduction of needed regulating or balancing power via utilization of the intra-day trading option has been also investigated. This is a very important factor for the total system security of the power grid. The investigated cases were selected in order to obtain a better understanding of the possibilities that are available for wind power producers or balance responsible parties in the current market structure and how the efficiency and revenue of trading of wind power can be increased in future. As a conclusion, it is possible to reduce the demand for the regulation power by using very short-term wind power forecasting before the gate closure time in the power market. The utilization of the Intraday trading option in combination with advanced (day ahead and very short-term) wind power forecasting models reduced the amount of the needed up and down regulation power down to 64.5 and 57.8 % respectively. Savings become possible by the complementary employment of the shorter term hourly trading options in the Intraday market from the energy economic point of view. More than one million Euros calculated as additional revenue for the Hors Rev offshore wind farm (160 MW) for almost one year. This amount corresponds to approximately 2 per MWh. This hourly trade option has been possible in the German intra-day power market since In order to be able to use the intra-day market in the future securely, it will be necessary to increase the trade volume. 133

151 Conclusion and Outlook 6 Conclusion and Outlook 6.1 Conclusion As wind power is one of the fastest growing energy sources in Germany and most other countries, grid and market integration of wind power are becoming increasingly critical issues. In this study, the importance of the advanced wind power forecasting techniques and the usage of the wind power forecasting tools in the intraday trading were investigated. New approaches were also developed to increase reliability and accuracy of the wind power prediction models. The Multi-Model approach was developed to investigate the influence of merging different Numerical Weather Prediction ( NWP ) models on the accuracy of wind power forecast. The development of the multi NWP model approach using the data from several numerical weather prediction models clearly improved the forecast accuracy. It was observed that up to 20 % of improvement in terms of nrmse (in relation the to the best single NWP model) is possible by employing the Multi NWP model. Combination approaches such as the simple averaging and the double artificial neural network ( ANN ) models were also developed to optimize and reduce the forecast error. The use of ANN to perform a wind power forecast for each MS EPS (the ensemble weather prediction model) member individually and subsequent combination of the prediction output yielded an improvement. An about 8 % improvement in terms of nrmse (in relation to the best MS EPS member) was observed by employing an advanced combination model (the double ANN approach). Additionally, the scope of the study was extended to the offshore wind farm environment. A new offshore-specific wind power prediction system based on the NWP and oceanographic parameters was developed and investigated. The meteorological situation in the near-shore marine atmospheric boundary 134

152 Conclusion and Outlook layer differs from its over land counterpart. The atmospheric stability and the distance to the shore have a significant influence over the sea. On the other hand, wind power forecast of offshore wind farms has the potential to become more reliable than on land, if offshore-specific forecast models are adjusted for the offshore environment. The most influent parameters among the oceanographic parameters such as mean wave period and significant height of wind waves have been integrated into the prediction system to improve the prediction accuracy. These findings show that the usage of the additional oceanographic and wind measurements in the vicinity of the offshore wind farms improve the market and grid integration capability of offshore wind power in to the existing power system. According to the results of the experiments which were presented in the chapter four, the new offshore day ahead wind power forecasting using the additional oceanographic parameters (wave forecasts) brings 21.3% (in terms of nrmse) of improvements with respect to a simple traditional wind power forecasting model based on only the NWPs. The two combination methods (the simple averaging and the double step ANN approach) also yielded a reduction on the nrmse. Utilization of the advanced combination model (the double step ANN approach) improved the forecast accuracy of the offshore wind power forecasting model up to %. It can be concluded that, the performance level of the multi-model approach is similar to the onshore wind farms in the offshore environment. The influence of the additional wind and wave measurements on the very short-term (two hours ahead) wind power forecasting models were also investigated. The forecast error values were decreased down to % (in terms of nrsme) by employing the two hours ahead wind power forecasting models. Finally, an economic analysis concerning the wind power forecasting was carried out. Economic benefits of wind power forecasting and utilization of the intraday wind power trading option were investigated in the final part of the study. The improved forecasts for short time horizons will be needed for grid safety and intra-day trading transactions. By using very short-term wind power forecasting models before the gate closure time on the power market, it is possible to reduce the demand of regulation power. The evaluations show that, the utilization of the Intraday trading option in combination with advanced day ahead and very short-term wind power forecasting models reduce the amount of needed up and down regulation power respectively down to 64.5 % 135

153 Conclusion and Outlook and 57.8 %. The economical calculations prove that more than one million Euros can be gained additionally for the same (Horns Rev) offshore wind farm with 160 MW installed capacity within almost one year. This amount corresponds to approximately 2 per MWh. In order to be able to use the Intraday market in future securely, it is necessary to increase the trade volume. The study concludes that very short-term forecasting models should be a prerequisite which can be employed for safe operation of large amounts of onshore and offshore wind power. Effective utilization of the intraday trading option in combination with advanced very short-term forecasting tools increases the value of wind power. 6.2 Outlook The predictability of wind power is very important not only from the technical and economical but also environmental point of view. The findings of this PhD work can be applied in an environmental research space such as the investigations regarding the possible reduction in CO 2 emissions using advanced wind power forecasting models and the intraday trading option. The multi-model combined forecasting approaches have a high potential to minimize prediction errors. More advanced wind power prediction tools can be developed for the regional grid operations in order to overcome the possible challenges in relation to congestion management operations in the power grid. The decision between the fix priced feed-in tariff mechanism and the direct marketing option is a critical issue. Therefore the next generation forecasting tools could be developed to estimate month ahead long-term wind power yield and power market prices in order to decide the right trading options. The integration of the online wind and wave measurements from the measurement masts, remote monitoring techniques, new generation information and communication technologies (such as cloud computing and embedded systems) into the measurement masts and the turbines could improve prediction accuracy. These innovative concepts will definitely increase the interoperability of the wind power forecasting models in future. In future, the additional wave parameters, the inclusion of offshore specific wake effects and parameters related to the thermal stratification in the offshore environment 136

154 Conclusion and Outlook could be incorporated to the next generation offshore wind power forecasting models. The treatment of satellite images may increase the precision of wind power forecasting as well. During this study, only one wave foresting time series from the ECMWF were used for the investigations. The combination of different wave forecast models from different providers may increase the wind power prediction precision for the offshore wind farms. Integration of physical, statistical and artificial Intelligence - based wind power forecasting models has a potential to optimize the wind power predictions by using the advantage of these approaches. This study also aims to be considered as a useful reference for the researches who plan to integrate their physical wind power forecasting models into the offshore wind farms. The new offshore wind power forecasting system is supposed to support the transmission system operators and wind farm owners while optimizing their operational tasks from the technical point of view. Beside the technical benefits, this new approach helps the wind farm owners and the other related power market players to optimize their trading operations. According to the findings gained during this study, it can be concluded that it makes sense to measure the wind and wave conditions continuously close to the planned offshore wind farms for the balancing responsible parties such as the TSOs and wind farm operators. It is significant to gather all the measured information by means of the advantage of the information and communication technologies. The collected data can be monitored by the supervisory control and data acquisition ( SCADA ) systems. The regulatory authorities should set up some regulations and rules for the players who are supposed to perform such measurements to improve the accuracy of the offshore wind power predictions and the reliability of their daily operations. Due to the smoothing effects, the fluctuations between the high and low generation amounts can be vanished or minimized. Hence the power production becomes more constant. In practice this means that the aggregated wind farms (wind clusters) can start to behave like conventional power plants. The offshore wind power will guarantee a more constant output, which is generally considered as added value to this type of energy source. The generated wind power output in this way can be used for base load generation 137

155 Conclusion and Outlook and can replace (partially or fully) with especially nuclear power and other conventional power plants with higher CO 2 emissions by using additional techniques like advanced wind power forecasting techniques, energy storage and innovative energy management methods in future. Principally, this investigation shows that there are possibilities to enhance the economic and technical value of wind power from its current levels by further developing the existing forecasting methodologies. Such kind of enhancements may trigger the major players on the power market and policy makers from the environmental, political and economic perspectives. On the other hand, making plausible predictions into the future has become great challenge, if we consider the growing complexity and global interconnection of policies and practices. That means that, the integration of offshore wind power and in general large amounts of renewable based energy sources into the power system are crucial processes, where politics and economic aspects often are conflicted. Recently, there is an obvious trend in the world that wind power is moving increasingly towards a more controllable power source like the conventional counter-partners. Thus, this new capability of the wind power has created new potential application fields as well. Especially in many countries, where large percentages of wind power penetrates into the grid, there are some requirements for such new application areas of wind power in their liberalized power market mechanisms by employing incentives or additional bonus schemes to enable wind power to also be used as balancing or reserve power and ancillary services. The outputs of this PhD work support the new possible capability of wind power as well by demonstrating how the wind power forecasting systems can be utilized to convert wind power to a more reliable and economical feasible energy source in both directions (also the down regulation) of the power production. In conclusion, according to the investigations in the presented PhD thesis, the effective usage of advanced wind power forecasting techniques and utilization of intraday trading option have technical and economic potential to improve the market and grid integration capability of onshore and offshore wind power. 138

156 References References [1] European Wind Energy Technology Platform, TPWind; Strategic Research Agenda Market Deployment Strategy from 2008 to 2030 ; Annex A: Detailed Research Actions, [2] C.Tauber, Energie- und volkswirtschaftliche Aspekte der Windenergienutzung in Deutschland - Sichtweise von E.ON Kraftwerke, Regenerative Energien, Energiewirtschaftliche Tagesfragen 12 Page , Hannover, 2002 [3] G.Giebel, The state of the art in short-term prediction of wind power: A literature overview, Anemos Report v.1.1, 2003 [4] L. Langberg, Short-term prediction of local wind conditions, Technical Report RISOE, Denmark, 1994 [5] L. Langberg, Short-term prediction of the power production from wind farms, J. Wind Engineering Industry Aerodynamics, 1999 [6] Lange, M., On the Uncertainty of Wind Power Predictions Analysis of the forecast accuracy and statistical distribution of errors, Journal of Solar Energy Engineering 127: , 2005 [7] C. Ensslin, B. Ernst, K. Rohrig, and F. Schlögl Online monitoring and prediction of wind power in German transmission system operation centers" in European Wind Energy Conference. Madrid, Spain, 2003 [8] Kurt Rohrig, Online monitoring of 1700 MW wind capacity in a utility supply area, European Wind Energy Conference and Exhibition, Nice, France, March 1999 [9] Hans Georg Beyer, Detlev Heinemann, Harald Mellinghof, Kai Mönnich, Hans-Peter Waldl, Forecast of regional power output of wind turbines, 1999 European Wind Energy Conference and Exhibition, Nice, France, March 1999 [10] U.Focken, M.Lange and H.P. Waldl, Previento: regional wind power prediction system with innovative upscaling algorithm, in EWEC 2001, Copenhagen, 2001 [11] Ulrich Focken, Matthias Lange, Hans-Peter Waldl, Previento A wind power prediction system with an innovative upscaling algorithm,

157 References European Wind Energy Conference and Exhibition, Bella Center, Copenhagen, Denmark, 2-6 July 2001 [12] Landberg, L., Joensen, A., Giebel, G., Madsen, H., Nielsen, T.S., Short-term prediction towards the 21st century. in wind forecasting techniques, 33 Meeting of Experts, Technical Report from the International Energy Agency R&D Wind, Ed. S.-E. Thor, FFA,Sweden, 77-85, July 2000 [13] Lars Landberg, Prediktor an online prediction system, Wind Power for the 21st Century, Special Topic Conference and Exhibition, Convention Centre, Kassel, Germany, September 2000 [14] Nielsen, T.S., Madsen, H., Tofting, T., WPPT, A Tool for online wind power prediction. in wind forecasting techniques, 33 Meeting of Experts, Technical Report from the International Energy Agency, R&D Wind, Ed. S.-E. Thor, FFA, Sweden, , July, 2000 [15] H.Madsen, H.A. Nielsen, T.S.Nielsen, A Tool for Predicting the Wind Power Production of Offshore Wind Plants, Copenhagen Wind Conference, Copenhagen, 2005 [16] Gregor Giebel,Lars Landberg, Torben Skov Nielsen, Henrik Madsen, The Zephyr Project The next generation prediction system, EWEC, Paris, 2002 [17] T.S. Nielsen,H. Madsen, H. Aa. Nielsen,L. Landberg, G. Giebel, ZEPHYR THE PREDICTIONMODELS, Proceedings of the European Wind Energy Conference, Copenhagen, pp , July 2001 [18] I. Sanchez, Short-term prediction of wind energy production," International Journal of Forecasting, vol. 22, pp , 2006 [19] I. C. Martí, Daniel; Villanueva, Javier; Sanisdro, Maria Jesús; Loureiro, Yolanda ; Cantero, Elena; Sanz, Javier; Navarro, Jorge; Roldán, Antonio, LocalPred and RegioPred. advanced tools for wind energy prediction in complex terrain " in European Wind Energy Conference & Exhibition EWEC. Madrid, Spain, 2003 [20] M. A. Gaertner, C. Gallardo, C. Tejeda, N. Martínez, S. Calabria, N. Martínez, and B. Fernández, The Casandra project: results of wind 140

158 References power 72-h range daily operational forecasting in Spain," presented at European Wind Energy Conference, 2003 [21] B. Fay and L. Neunhauserer, Evaluation of very high-resolution simulations with the non-hydrostatic numerical weather prediction model Lokalmodell for urban air pollution episodes in Helsinki, Oslo and Valencia, 2005 [22] S.Haykin, Neural networks, a comprehensive foundation, New Jersey, 2nd edition, 1999 [23] A.P. Wieland, Evolving controls for unstable systems, Connectionist Models: 1990 Summer School, pp , California, 1990 [24] B. A. Pearlmutter, Dynamic recurrent neural networks, Technical Report CMU CU, pp , Carnegie Mellon University, 1990 [25] C. Elmas, Yapay Sinir Aglari Kuram, Mimari, Egitim, Uygulama (Artificial Neural Networks Theory, Architecture, Applications), Ankara, 2003 [26] K.Rohrig, Rechenmodelle und Informationssysteme zur Integration großer Windleistungen in die elektronische Energieversorgung, PhD Thesis, Kassel, 2003 [27] B.Ernst, Entwicklung eines Windleistungsprognosemodels zur Verbesserung der Kraftwerkeinsatzplanung, PhD Thesis, Kassel, 2003 [28] K.Rohrig and B.Ernst, Online-supervision and prediction of 2500 MW wind power, EWEA Special Topic Conference: Wind power for the 21th century, 2000, Kassel, Germany [29] B.Ernst, K.Rohrig, Online monitoring and prediction of wind power in German transmission system operation centers, first IEA Joint Action Symposium on Wind Forecasting Techniques, Norrköping, 2002 [30] Ue. Cali, B.Lange, J. Dobschinski, M.Kurt, C. Moehrlen, B.Ernst, Artificial neural network based wind power forecasting using a multimodel approach,7th International Workshop on Large Scale Integration of Wind Power and on Transmission Networks for Offshore Wind Farms, Madrid,

159 References [31] Ue.Cali, R.Jursa, M.Kurt, and B.Lange, Feasibility study for the use of WEPROG s ensemble weather prediction data for wind power forecasting at RWE, Kassel, Germany, February 2008 [32] C.Möhrlen, Description of MS EPS parameterization schemes,january 2003 [33] Holtslag, A.A.M., Boville, B.A., Local versus nonlocal Boundary-Layer Diffusion in a Global Climate Model. J. Climate 6, , 1993 [34] H.L. Kuo, Further studies of parameterization of the influence of Cumulus convection on large-scale flow, Journal of the Atmospheric Science 31. pp, , 1974 [35] M.Tiedke, Representation of clouds in large-scale models Monthly Weather Review, 121, , 1993 [36] EWEA, Delivering Offshore Wind Power in Europe, 2007 [37] H.Bergstrom, Estimating offshore wind potential - a more complex task than expected, EWEC 2001, Copenhagen, 2001 [38] P.Pinson, H. Madsen, Ensemble-based probabilistic forecasting at Horns Rev ; Wind Energy- Special issue on Offshore Wind Energy; August, 2008 [39] M. Rugbjerg, Ole René Sørensen and Vagner Jacobsen, Wave forecasting for offshore wind farms, 2001 [40] G. Kariniotakis, P. Pinson, G.Giebel, R. Barthelmie, The state of the art in short-term prediction of wind power from an offshore perspective, 2004 [41] Kariniotakis, G., I. Marti, D. Casas, P. Pinson, T. S. Nielsen, H. Madsen, G. Giebel, J. Usaola, I. Sanchez, A. M. Palomares, R. Brownsword, J. Tambke, U. Focken, A. Lange, P. Louka, G. Kallos, C. Lac, G. Sideratos and G. Descombes, What performance can be expected by short-term wind power prediction models depending on site characteristics?", European Wind Energy Conference, London, UK., 2004 [42] Tambke, J., M. Lange, U. Focken, J.-O. Wolff and J. Bye, Forecasting offshore wind speeds above the North Sea Wind Energy,

160 References [43] J. Tambke, L. von Bremen, R. Barthelmie, A. M. Palomares, T. Ranchin, J. Juban, G. Kariniotakis, R. A. [Brownsword, I. Waldl, Short-term forecasting of offshore wind farm production - Developments of the Anemos Project, European Wind Energy Conference, Athens, 2006 [44] P. Pinson, T. Ranchin, G. Kariniotakis, Short-term wind power prediction for offshore wind farms -Evaluation of fuzzy-neural network based models, 2004 [45] Lange, Offshore wind power meteorology, EUROMECH, 2005 [46] A. Sommer, Wind resources at Horns Rev, Eltra PSO Project, Programme for measurement wind, wave and current at Horns Rev ; Fredericia, Denmark, December 2002 [47] C. B. Hasager, A. Peña, T. Mikkelsen, M. Courtney, I. Antoniou, S.E. Gryning, P. Hansen (Risø) and P. B.Sørensen, 12MW Horns Rev experiment, Risø-R-1506(EN), Roskilde, Denmark, October 2007 [48] G. J., L. Komen, M. Cavaleri, K. Donelan, S. Hasselman, and P. E. A. M. Janssen, Dynamics and modeling of ocean waves. Cambridge University Press, 532 pp, 1994 [49] Berlin Declaration, Conclusions of the chair, European Policy Workshop on Offshore Wind Power Deployment, Berlin, February 2007 [50] Jürgen Neubarth, Oliver Woll, Christoph Weber und Michael Gerecht, Beeinflussung der Spotmarktpreise durch Windstromerzeugung, Energiehandel Heft 7, 2006 [51] Bode, On the impact of renewable energy support schemes on power prices, HWWI Research Paper, Paper 4-7, 2006 [52] Wind Energy The Facts: A guide to the technology, economics and future of wind power, European Wind Energy Association, Earthscan Publishing, London 2009 [53] Tony COCKER, Gunnar LUNDBERG, Integrating electricity markets through wholesale markets : EURELECTRIC Road Map to a Pan- European Market, Brussels,

161 References [54] F.H Boisseleau, The role of electricity trading and power exchanges for the constructions of a common European electricity market, France, the Netherlands, 2001 [55] Pahlow, M., Möhrlen, C, Jørgensen, J.U., Application of cost functions for large-scale integration of wind power using a multi-scheme ensemble prediction technique, Optimization Advances in Electric Power Systems, Ed. Edgardo D. Castronuovo, NOVA Publisher NY, 2008 [56] R.J. Barthelmiea, F. Murraya, S.C. Pryor, The economic benefit of short-term forecasting for wind energy in the UK electricity market, Energy Policy 36 (2008) , 2008 [57] Parkes, J., Wasey, J., Tindal, A., Munoz, L., Wind energy trading benefits through short-term forecasting, European Wind Energy Conference, Athens, 2006 [58] H.Holttinen, Optimal electricity market for wind power, Energy Policy Publication, 2005 [59] Saint-Drennan, Ümit Cali, W. Hicks, DISPOWER Task 5.4.b: Wind power forecast and adaptation to the BETTA market, 2005 [60] B.Lange, Modeling the marine boundary layer for offshore wind power utilization, PhD Thesis, 2002 [61] Marti, I., G. Kariniotakis, P. Pinson, I. Sanchez, T. S. Nielsen, H. Madsen, G. Giebel, J. Usaola, A. M. Palomares, R. Brownsword, J. Tambke, U. Focken, M. Lange, G. Sideratos and G. Descombes, Evaluation of advanced wind power forecasting models - Results of the Anemos Project, European Wind Energy Conference, Athens., 2006 [62] J. A. Halliday, G.M. Watson, Power- a methodology for predicting offshore wind energy resources, EWEC 2001, Copenhagen, 2001 [63] [64] [65] [66] 144

162 References [67] [68] [69] [70] Robert M. Sorensen, Basic Coastal Engineering 3 rd edition, Springer, Bethlehem, Pennsylvania, 2006 [71] Ümit Ciplak, Die Bedeutung von Windleistungsvorhersagen für die Elektrizitätswirtschaft, Master Thesis, University of Kassel, 2008 [72] Melih Kurt, Entwicklung eines offshore-spezifischen Windleistungsvorhersagesystems basierend auf der Ensemble Wetter Vorhersage, Master Thesis, University of Kassel, 2009 [73] Melih Kurt, Windleistungsvorhersage mit NWP und MS-EPS, Diploma Thesis, University of Kassel, 2008 [74] Zouhair Khadiri Yazami, Windleistungsprognose und Vergleich verschiedener Anwendungen und Programme zur Erstellung künstlicher neuronaler Netze, Diploma Thesis, University of Kassel,

163 Acknowledgement Acknowledgement Some of the results are investigated during HRENSEMBLEHR research project at ISET e.v.-kassel which is supported by the Danish PSO (Public Service Obligation) -F&U Please note that, Melih Kurt [72], [73], Ümit Ciplak [71], and Zouhair Khadiri Yazami [74], have co-operated with the author by developing some parts of this study during their master or bachelor thesis periods. The author was responsible to determine the scope of their thesis and supervise them as well. Therefore, thanks to Melih Kurt, Ümit Ciplak, and Zouhair Khadiri Yazami for your excellent assistance during numerous experiments. 146

164 ANNEX: Very short-term offshore wind power forecasting experiments ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_1 Figure 59: The structure of the very short-term offshore wind power forecasting model using the NWP and measured wind power data.

165 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_2WF Figure 60: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data and forecasted wave parameters.

166 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_3W Figure 61: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data and measured wave parameters

167 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_4 Figure 62: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data and measured wind parameters

168 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_5WF Figure 63: The structure of the very short-term offshore wind power forecasting model using the NWP, forecasted wave, measured wind power and measured wind data

169 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_6W Figure 64: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data, measured wave data, measured wind parameters and forecasted wave variables

170 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_7WWF Figure 65: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data, measured wave parameters and forecasted wave variables

171 ANNEX: Very short-term offshore wind power forecasting experiments Experiment ST2_8WW Figure 66: The structure of the very short-term offshore wind power forecasting model using the NWP, measured wind power data, measured wind parameters and measured wave variables