5 th International Symposium Topical Problems in the Field of Electrical and Power Engineering, Doctoral School of Energy and Geotechnology Kuressaare, Estonia, January 14 19, 2008 Wind Power Production Estimation through Short-Term Forecasting Hannes Agabus Tallinn University of Technology, Department of Electrical Power Engineering hannes.agabus@4energia.ee, Abstract The paper aims to introduce different wind energy generation forecasting modes and forecasting under uncertainty considerations. Another aim of the paper is to present the actual wind park shortterm forecast results and analyse theirs correspondence to different performance statistics. Any utility getting over a few percent wind power penetration is buying a forecasting system or have a service providers on the market. Therefore, accurate wind park production forecasting is highly important and can be very sophisticated task as it is influenced by the initial choice of the prediction approach or prediction system, the level of penetration, the intended use of the forecasts, the acceptance system operators may have for wind energy and by many other factors Keywords Wind power forecast, forecasting methods and models, forecast inaccuracy, uncertainty of wind, day-ahead forecasting. Introduction While there are good wind resources throughout Europe, the intermittency of the wind represents a major problem for the deployment of wind energy into the electricity networks [1]. Nevertheless, wind power has been undergoing a rapid development in recent years. Several countries have reached already a high level of installed wind power capacity, such as Germany, Spain, Denmark, while others follow with high rates of development. The development of wind energy use has led to a noticeable contribution to the energy supply in Estonia. Currently in Estonia, there is installed capacity ca 56 MW. Correspondingly, net available capacity of Estonian power plants is ca 2300 MW. The annual electricity consumption is ca 6 TWh (excluding losses). Domestic peak load in wintertime reaches 1600 MW and low load in summer decreases to 400 MW. In the near future, current amount of installed wind generators should be doubled. Already, Estonian Transmission System Operator (TSO) has wind park grid connection proposals close to 2000 MW. It is well know fact that wind is a variable resource that is difficult to predict wherefore the intermittence of wind energy presents a special challenge for TSO-s and wind park operators. While conventional power plants produce a near constant output, then the output of a wind power plant fluctuates. To the extent the fluctuations cannot be predicted, they create costs (balance power purchase) for the electricity system and consumers as well as potential risks to the reliability of electricity supply. For a time scale from some hours to two days additional conventional reserves have to be kept ready to replace the wind energy share in case of decreasing wind speeds. For example in Denmark, wind power generation can decrease from 90% to 10% within a half an hour (from gross installed capacity). In Estonia, there is possibility that 80% of installed wind capacity can change rapidly during very short time. To ensure power grid security a TSO-s needs today for each kilowatt of wind energy either an equal amount of spinning reserve (gas turbine etc.) or a forecasting system that can predict the amount of energy that will be produced from wind over a period usually of 1 to 72 hours (3days). Accurate next week and month forecast are already bonuses. 1 Wind energy generation forecasting 1.1 General The wind energy generation forecasting problem is closely linked to the problem of forecasting the variation of specific atmospheric variables (wind speed and direction, air density) over short time intervals and small spatial scales for a small volume of the atmosphere (the wind plant) for a variety of time horizons. Due to the wide range of spatial and temporal scales that determine the variations in the wind energy power generation, it is necessary to use a diverse mix of data sources and types to achieve the best possible forecast performance. For wind energy forecasting, the most fundamental type of data is the time series of meteorological parameters and power generation from the wind plant itself [2]. Wind predictions are not used in an automated way as for simple load forecasts. Given the uncertainty 119
they involve, users need to develop expertise on the optimal decisions to make as a function of the current or expected power system state or market conditions. Therefore, there is a need emerging to fully integrate predictions and information on their uncertainty in management functions (unit commitment, economic dispatch, reserves estimation etc.). 1.2 Forecasting tool There are two fundamental types of tools used in the forecasting process: 1) data gathering and 2) data processing. The data gathering tools include the wide range of measuring devices that provide data to the forecast process. These include measurements made at the wind power plant itself as well as those in the localarea, regional and even global environment of the power plant. The data processing tools serve to transform the measurement data into a forecast for the desired period of time. The data processing tools include physical and statistical atmospheric models as well as models of the relationship between meteorological conditions within the wind plant volume and the plant output (usually referred to as a plant output model) [2]. 1.3 Forecasting models There are four basic types of wind power prediction models used in the wind energy forecasting process: 1) physical models, 2) statistical models, 3) wind plant output models, and 4) forecast ensemble models. There are many types of models within each of these major categories. Physical models are based upon the fundamental physical principles of conservation of mass, momentum and energy and the equation of state for air. These models are actually models that have been specially adapted to simulate the atmosphere. Statistical models use statistical techniques to arrive at the connection between predicted weather parameters and power output. A large number of such techniques exist, some very complex. All statistical models require information about the true wind power output for calibrating the model parameters to a particular site or collection of turbines [3]. The statistical models tend to perform slightly better than the physical models, as the statistical adjustments act to keep the systematic errors at a low level. Their main drawback is their dependency on measured data, either historically or on line, as sometimes such data are difficult to get hold of. Wind plant output models are the relationships between the meteorological variables at the wind plant site and the plant s energy output. They can be formulated as physical models, statistical models or a hybrid of both types of models. Forecast ensemble models are statistical models that produce an optimal forecast by compositing forecasts from a number of different techniques. 120 1.4 Best practice The most effective way in the use of short-term forecasting of wind power can be summarized as: 1) get a basic model; 2) for improvement, operate with another model (numerical weather prediction (NWP) and/or short-term forecasting model); 3) balance all errors together (not just wind); 4) take account of the uncertainty; 5) for better income, use intraday trading; 6) use longer forecasts for maintenance planning; 7) arrange meteorological training for the operators and 8) meteorological hotline for special cases (for stormy cyclone detection etc.). 2 Estimation of forecasting models accuracy 2.1 Prediction preciseness Good wind power predictions increase the value of the wind power and makes wind power more competitive. On the side of the actual short-term prediction model, typical error sources are the power curve modeling and the taking into account of the stability of the atmosphere. In a typical short-term prediction model, the largest source of error is the NWP input. Within the weather forecast, the largest error possibilities are due to the (limited) horizontal and vertical resolution of the model, the number of weather observations used (especially upstream) and the quality of the data assimilation, plus the actual model physics as well [4]. Predictions of the power production for a wind power plant or collection of wind power plants can be improved by using so-called self-adaptive models. These models learn continuously from past experience and are thus able to adjust to changing circumstances, such as seasonal changes, small changes in weather model set-up, or even changes to the numbers or types of wind turbines [3]. The main drawback of the neural network model and other self-adaptive models is that they need measured values of the power output in real time or near real time in order to perform the ongoing adjustments. 2.2 Forecast evaluation issues There are many options for performance statistics available: mean error (bias); mean absolute error (MAE); mean absolute percentage error (MAPE); root mean square error (RMSE); median error (MDE); skill score (% improvement over a reference forecast); correlation coefficient; full error distributions etc. There are always possibilities to attempt to tune system for a specific statistic and we certainly must examine what statistics are relevant to the forecast user. In addition to the issue of different performance statistics providing a different picture of performance, there is also the issue that a forecast system can be tuned to produce better performance for a specific statistic while possibly degrading the performance for other statistic.
3 Practice of wind generation forecast 3.1 Day-ahead based forecasting Predicting the wind power production on time horizons from hours to weeks ahead is simple in principle. What you needed is a general-purpose weather model, which predicts the wind and other weather parameters at the site or sites in question, and a wind power model, which converts the predicted weather parameters into power. Short-term (day-ahead) forecast methods use essentially the same tools as very short-term forecast techniques. However, there are two important differences: 1) the importance of real-time data from the wind plant and its immediate environment is significantly reduced; 2) regional and sub-regional simulations with a physics-based atmospheric model play a much more significant role in the forecast process. 3.2 Basis In this paper, Pakri wind park four months forecast and actual production data will be investigated. Pakri wind farm itself is located in Paldiski, at the tip of the Pakri peninsula. The park consist eight 2,3 MW Nordex N-90 turbines. First wind turbine was connected with the power grid on the 15th of December in 2004.. Full capacity was achieved in the March of 2005. Pakri wind park is operated by the Nelja Energia OÜ (4energy) who is the largest wind emery developer in Estonia. The day-ahead forecasting of wind park output power came relevant when in May 2007, new Energy Market Act came to force witch enacts that beginning of the year 2009, every wind park must obtain its balance and therefore become a balance provider itself or buy the service in. Meanwhile, wind park owner can choose between two support schemes. In first case, simple obligation to buy (ca 73,5 eur/mwh) is applied by the TSO. Alternatively, you can choose subsidiary (ca 53,7 eur/mwh) and additionally sell entire production on open market. As 4energy chose second scheme and therefore possibly accurate forecasting is crucial for wind park operator income (balance expenses etc.). Physically, the day-ahead forecasting for 4energy began in June 2007. So practically it has been done only four months for now and it s been quite challenging task ever since. Almost all short-term forecast procedures begin with the grid point output from a regional scale physicsbased atmospheric model/numerical weather data. For Pakri wind park, this is executed by the Estonian Meteorological and Hydrological Institute (EMHI). This model ingest data from a wide variety of sources over a large area and produce forecasts of regional scale weather system for a three day period. However, this model does not resolve the physical processes occurring in the local or mesoscale areas around individual wind plant. Therefore wind power model with EMD (well known Danish Software and Consultancy Company) help, is provided and all the best practice of the world leading software for wind energy planning (Wind Pro and its components) are used. In wind power model, based on wind farm layout, terrain, assumed wind distribution and power curves, we established the transfer function between local wind data and power production. The output data is transmitted to balance manager (currently Eesti Energia AS) who will plan all its balance region production and refer it to TSO (OÜ Põhivõrk). 3.3 Accuracy of Pakri wind park forecast It is obvious that if there is big variation between forecasted wind data and actual wind, then remaining errors from forecasting model and from human factor are trivial. For example, EMHI provided July wind prediction and actual wind data is showed in Figure 1, where you can easily sight notable variations during many hours. 18 16 July forcast vs. actual forecasted wind actual wind 14 wind speed m/s 12 10 8 6 4 2 0 1 31 61 91 121 151 181 211 241 271 301 331 361 391 421 451 481 511 541 571 601 631 661 691 721 hours Fig. 1. Forecasted and actual wind values during July 2007 121
In July 2007, wind mean error in Pakri was 1,9 m/s. As wind generator output power is basically cube root reliance from the wind speed then even a slight difference in wind prognosis could cause wide production difference. For example, if there was hourly forecast 5 m/s but actual wind speed was 2,5 m/s then production bias was already 1 MWh, which was covered by balance power. A summary of the four months forecast performance results is presented in the following table. Table 1. Pakri wind park four months forecast performance results Parameter June July August September Mean error (MWh) 2,569 2,475 2,300 2,922 MAE (MWh) 61,653 59,397 55,203 70.135 MAPE (%) 39,4 4,5 9,4 24,8 The performance statistics in this table are for all forecast hours (24) and for the entire month evaluation period. The mean absolute error (MAE) as a percentage of installed capacity for entire four month period is ca 40%, what is much higher than typical range for 1 to 2 day forecast performance. Also, two months mean absolute percentage error (MAPE) values are higher than allowed. Commonly, it is acceptable, if MAPE is 20% at the maximum. The two month inaccuracy is caused manly by two factors. Firstly, in the beginning (June) whole dayahead wind and production forecasting was totally new challenges for both sides (EMHI/4energy). As it took some time to trim the weather forecast and calibrate the power model then intense forecast roughness was unavoidable. During September, there was period when EMHI stopped wind forecast service (one week) because there were misunderstandings within the valid service contract and 4energy had to used then simple public online weather service. After new contract was signed, it took some time to re-establish the forecast service. All these shortfalls leaded finally to very rough production forecast. Illustrative wind park production forecast and actual production is showed in Figure 2. August 2007 18 16 14 12 MW 10 8 6 4 2 0 01.08 04.08 07.08 10.08 13.08 16.08 19.08 22.08 25.08 28.08 31.08 Date Prognose, MWh Actual prod. MWh Fig. 2. Wind park production prognosis and actual production during August 2007 In brief, the main apparent forecasting errors are caused by: 1) inaccurate wind prognosis; 2) not so accurate power forecast model usage; 3) human factor. Has the prognosis is currently made manually then there is no day-ahead forecasting during weekends. Sunday and Monday prognosis is put together also on Friday witch increase the forecast inaccuracy; 4) insufficient wind park maintenance planning. The future main challenge is to minimize negative impact of these above described factors and improve whole forecast system accuracy. Some vital actions for this achievement are already in progress. For example, 4energy has already developing new power forecast model in conjunction with Vejr2 AS. This new fully automated neural model should more effectively manage the daily forecast challenges and decrease the cost of imbalance and makes available very short-term (next hour) forecasting. As the MAE of very short-term forecasts is typically in the range of 5% to 15% therefore such prognosis implementation in near future has large impact for accurate forecasting [3]. According to Vejr2, they are able to offer even half hour granularity for the day-ahead forecast witch would be substantial improvement. 122
4. Conclusions This current brief overview of forecast tools and models indicates that there is a large and diverse pool of tools that can be used to generate wind energy forecasts. The future challenge is to use the optimal set of tools and configurations for a specific forecast application. The majority of operational prediction models were initially designed to provide deterministic forecasts, in the form of a unique value for each hour of the prediction horizon. As the wind penetration increases, end-users require complementary information on the uncertainty of such forecasts. Therefore accurate forecasting model without uncertainty estimations is nowadays unthinkable feature. In areas with large amounts of wind energy, the wind power needs to be balanced by power produced from other energy sources at times of low wind. The more accurately the wind power can be predicted, the smaller the costs of the balancing production will be, environmentally as well as economically. A summary of the Pakri wind park four months forecast performance results were presented. Based on these results, lots of improvement for higher accuracy must be accomplished. Generally, however, for time horizons beyond a few hours it is the quality of the input data from the weather models that limits the quality of the wind power predictions. Therefore it is substantial that any error made by metrology is critical for wind power prediction and therefore when input wind data is not accurate then there will be big variations witch leads a inaccurate generation prediction. Same principles stands for wind power model accuracy and the adaptation of self-adaptive model should increase it. Very sophisticated challenge for 4energy will be implementation of short-time forecasting witch is an important key factor for future accurate wind energy production forecast. References 1. Möhrlen, C. Uncertainty in wind energy forecasting. Thesis (PhD) Department of Civil and Environmental Engineering, University College Cork, Ireland, DP2004MOHR, May 2004. 2. Zack, J. Overview of wind energy generation forecasting. Draft report for NY State Energy Research and Development Authority and for NY ISO, True Wind Solutions LLC, December 17 2003, Albany, NY, USA. 3. Improvements in wind power prediction. Vejr2 exclusive report for Nelja Energia OÜ, September 2007, Roskilde, Denmark. 4. Giebel, G., Kariniotakis, G. Best practice in short-term forecasting. Proceedings of the Wind Energy Conference 2007, 7-10 May 2007, Milan, Italy. 5. Juban, J., Fugon. L., and Kariniotakis, G. Probabilistic short-term wind power forecasting based on kernel density estimators. Proceedings of the Wind Energy Conference 2007, 7-10 May 2007, Milan, Italy. 6. Agabus, H., Landsberg, M., Liik, O. Optimal investment strategies for energy sector under uncertainty. Publications of the 3 rd International Symposium Topical problems of education in the field of electrical and power engineering. Doctoral school of energy and geotechnology, Tallinn University of Technology, Kuressaare, 2006 (ISBN 9985-69-036-2), pp 120-123. 123