Multi-Model Ensemble for day ahead PV power forecasting improvement

Similar documents
SOLAR RADIATION FORECAST USING NEURAL NETWORKS FOR THE PREDICTION OF GRID CONNECTED PV PLANTS ENERGY PRODUCTION (DSP PROJECT)

Importance of Numerical Weather Prediction in Variable Renewable Energy Forecast

The document was not produced by the CAISO and therefore does not necessarily reflect its views or opinion.

A Community Gridded Atmospheric Forecast System for Calibrated Solar Irradiance

Current best practice of uncertainty forecast for wind energy

Solar Nowcasting with Cluster-based Detrending

CARLOS F. M. COIMBRA (PI) HUGO T. C. PEDRO (CO-PI)

A SOLAR AND WIND INTEGRATED FORECAST TOOL (SWIFT) DESIGNED FOR THE MANAGEMENT OF RENEWABLE ENERGY VARIABILITY ON HAWAIIAN GRID SYSTEMS

SOLAR POWER FORECASTING BASED ON NUMERICAL WEATHER PREDICTION, SATELLITE DATA, AND POWER MEASUREMENTS

Solar spectral irradiance measurements relevant to photovoltaic applications

Probabilistic Energy Forecasting

The Center for Renewable Resource Integration at UC San Diego

GL Garrad Hassan Short term power forecasts for large offshore wind turbine arrays

FORECASTING: A REVIEW OF STATUS AND CHALLENGES. Eric Grimit and Kristin Larson 3TIER, Inc. Pacific Northwest Weather Workshop March 5-6, 2010

AN INTERNATIONAL SOLAR IRRADIANCE DATA INGEST SYSTEM FOR FORECASTING SOLAR POWER AND AGRICULTURAL CROP YIELDS

A methodology for DNI forecasting using NWP models and aerosol load forecasts

Short and medium term solar irradiance and power forecasting given high penetration and a tropical environment

Speedwell High Resolution WRF Forecasts. Application

ANN and Statistical Theory Based Forecasting and Analysis of Power System Variables

Application and verification of ECMWF products 2009

David John Gagne II, NCAR

Systems Operations. PRAMOD JAIN, Ph.D. Consultant, USAID Power the Future. Astana, September, /6/2018

Short- term solar forecas/ng with sta/s/cal models and combina/on of models

2014 HIGHLIGHTS. SHC Task 46 is a five-year collaborative project with the IEA SolarPACES Programme and the IEA Photovoltaic Power Systems Programme.

Application and verification of ECMWF products 2008

COMPARISON OF CLEAR-SKY MODELS FOR EVALUATING SOLAR FORECASTING SKILL

An Adaptive Multi-Modeling Approach to Solar Nowcasting

Wind power and management of the electric system. EWEA Wind Power Forecasting 2015 Leuven, BELGIUM - 02/10/2015

An Integrated Approach to the Prediction of Weather, Renewable Energy Generation and Energy Demand in Vermont

Multi-Plant Photovoltaic Energy Forecasting Challenge: Second place solution

Short-Term Irradiance Forecasting Using an Irradiance Sensor Network, Satellite Imagery, and Data Assimilation

Improving the accuracy of solar irradiance forecasts based on Numerical Weather Prediction

Multi-Plant Photovoltaic Energy Forecasting Challenge with Regression Tree Ensembles and Hourly Average Forecasts

Application and verification of ECMWF products 2009

Application and verification of ECMWF products 2016

COMPARISON OF CLEAR-SKY MODELS FOR EVALUATING SOLAR FORECASTING SKILL

AMPS Update June 2016

CS 229: Final Paper Wind Prediction: Physical model improvement through support vector regression Daniel Bejarano

CLOUD VELOCITY ESTIMATION FROM AN ARRAY OF SOLAR RADIATION MEASUREMENTS

Modelling Wind Farm Data and the Short Term Prediction of Wind Speeds

CHAPTER 6 CONCLUSION AND FUTURE SCOPE

About Nnergix +2, More than 2,5 GW forecasted. Forecasting in 5 countries. 4 predictive technologies. More than power facilities

Forecast solutions for the energy sector

Temporal Wind Variability and Uncertainty

Energy produc-on forecas-ng based on renewable sources of energy

Wind Power Production Estimation through Short-Term Forecasting

Predic've Analy'cs for Energy Systems State Es'ma'on

Heat Load Forecasting of District Heating System Based on Numerical Weather Prediction Model

Short term forecasting of solar radiation based on satellite data

An Operational Solar Forecast Model For PV Fleet Simulation. Richard Perez & Skip Dise Jim Schlemmer Sergey Kivalov Karl Hemker, Jr.

Sun to Market Solutions

DRAFT: CLOUD VELOCITY ESTIMATION FROM AN ARRAY OF SOLAR RADIATION MEASUREMENTS

Probabilistic forecasting of the solar irradiance with recursive ARMA and GARCH models

S e a s o n a l F o r e c a s t i n g f o r t h e E u r o p e a n e n e r g y s e c t o r

Probabilistic forecasting of solar radiation

Short-Term Power Production Forecasting in Smart Grid Based on Solar Power Plants

Power System Seminar Presentation Wind Forecasting and Dispatch 7 th July, Wind Power Forecasting tools and methodologies

Day-Ahead Solar Forecasting Based on Multi-level Solar Measurements

Bankable Solar Resource Data for Energy Projects. Riaan Meyer, GeoSUN Africa, South Africa Marcel Suri, GeoModel Solar, Slovakia

Feature-specific verification of ensemble forecasts

Review of solar irradiance forecasting methods and a proposition for small-scale insular grids

Energy Forecasting Customers: Analysing end users requirements Dec 3rd, 2013 Carlos Alberto Castaño, PhD Head of R&D

FORECAST OF ENSEMBLE POWER PRODUCTION BY GRID-CONNECTED PV SYSTEMS

Post-processing of solar irradiance forecasts from WRF Model at Reunion Island

Country scale solar irradiance forecasting for PV power trading

Forecasting demand in the National Electricity Market. October 2017

DNICast Direct Normal Irradiance Nowcasting methods for optimized operation of concentrating solar technologies

Ensemble Trajectories and Moisture Quantification for the Hurricane Joaquin (2015) Event

Shadow camera system for the validation of nowcasted plant-size irradiance maps

Solar Irradiance Prediction using Neural Model

Data Analytics for Solar Energy Management

Solar Power Forecasting Using Support Vector Regression

Bayesian Based Neural Network Model for Solar Photovoltaic Power Forecasting

Uncertainty of satellite-based solar resource data

FIRST CORRELATIONS FOR SOLAR RADIATION ON CLOUDY DAYS IN ITALY

Modelling wind power in unit commitment models

Numerical Weather Prediction. Meteorology 311 Fall 2010

TOWARDS A GENERIC ONTOLOGY FOR SOLAR IRRADIANCE FORECASTING

Stochastic methods for representing atmospheric model uncertainties in ECMWF's IFS model

CONTROL AND OPTIMIZATION IN SMART-GRIDS

JOINT WMO TECHNICAL PROGRESS REPORT ON THE GLOBAL DATA PROCESSING AND FORECASTING SYSTEM AND NUMERICAL WEATHER PREDICTION RESEARCH ACTIVITIES FOR 2016

32nd European Photovoltaic Solar Energy Conference and Exhibition

ANN based techniques for prediction of wind speed of 67 sites of India

SOLAR ENERGY FORECASTING A PATHWAY FOR SUCCESSFUL RENEWABLE ENERGY INTEGRATION. (An ANN Based Model using NARX Model for forecasting of GHI)

Temporal global solar radiation forecasting using artificial neural network in Tunisian climate

Radar data assimilation using a modular programming approach with the Ensemble Kalman Filter: preliminary results

Satellite-based solar irradiance assessment and forecasting in tropical insular areas

Short-term wind forecasting using artificial neural networks (ANNs)

Lake parameters climatology for cold start runs (lake initialization) in the ECMWF forecast system

High Wind and Energy Specific Models for Global. Production Forecast

Modelling residual wind farm variability using HMMs

ECMWF global reanalyses: Resources for the wind energy community

0-6 hour Weather Forecast Guidance at The Weather Company. Steven Honey, Joseph Koval, Cathryn Meyer, Peter Neilley The Weather Company

MODELING ATMOSPHERIC SIGNALS IN SPACEBORNE INTERFEROMETRIC SAR DATA

Analysis and validation of 24 hours ahead neural network forecasting of photovoltaic output power

Estimation of Forecat uncertainty with graphical products. Karyne Viard, Christian Viel, François Vinit, Jacques Richon, Nicole Girardot

Integrated Electricity Demand and Price Forecasting

Energy Resource Group, DIT

Impact of light soaking and thermal annealing on amorphous silicon thin film performance

ESTIMATION OF SOLAR RADIATION USING LEAST SQUARE SUPPORT VECTOR MACHINE

Transcription:

Multi-Model Ensemble for day ahead PV power forecasting improvement Cristina Cornaro a,b, Marco Pierro a,e, Francesco Bucci a, Matteo De Felice d, Enrico Maggioni c, David Moser e,alessandro Perotto c, Francesco Spada c, a Department of Enterprise Engineering, University of Rome Tor Vergata, Via del Politecnico 1, 00133 Rome, Italy, e-mail: cornaro@uniroma2.it, marco.pierro@gmail.com, frabucci@gmail.com b CHOSE, University of Rome Tor Vergata, Via del Politecnico 1, 00133 Rome, Italy 7 c Ideam Srl, via Frova 34 Cinisello Balsamo, Italy, e-mail: alessandro.perotto, enrico.maggioni, francesco.spada@ideamweb.com d Casaccia R.C., ENEA Climate Modelling Laboratory, Rome, Italy e-mail: matteo.defelice@enea.it e EURAC Research, Viale Druso, 1, 39100 Bolzano, Italy e-mail: david.moser@eurac.edu

Why day ahead PV power forecast Large share of PV power introduces into the electric demand a stochastic variability dependent on meteorological conditions: residual load = load-pv generation Reserve PV Power ramp example of regional load and PV generation trend with 8.3% PV penetration

Why day ahead PV power forecast Day ahead PV power forecast could mitigate these effects PV POWER FORECAST to improve the capability of residual load tracking and transmission scheduling to obtain a better match between the day-ahead market commitment and the real PV production, reducing the energy imbalance costs.

Why day ahead PV power forecast PREDICTION INTERVALS to reduce uncertainty in the electric demand so that lower energy reserves are needed for energy trading issues

Aim of the work to develop and test several data-driven models for day ahead site PV power forecast (with hour granularity) using different NWP forcing to build up an outperforming Multi-Model Ensemble with its prediction intervals.

Data driven approach for day ahead PV production forecast In the last years a data-driven approach has been extensively tested for PV power generation forecast from 24 to 72 hours horizon. This approach involves a wide range of machine learning techniques that can be built making use of Numerical Weather Prediction (possibly corrected by Model Output Statistic) and weather and PV generation historical data. These algorithms try to reconstruct relationships between input and output through a training and validation procedure on historical data Hybrid models could be obtained using different models in series. While combining together different forecast models a Multi-Model Ensemble can be built.

Data used for training and test Historical weather and PV power production data Four years of monitored irradiance, temperature and production data (2011-2014) from a 662 kwp Cadmium Telluride PV plant, located in Bolzano (Italy), were employed to train and test the models. Data were acquired every 15 minutes and then averaged each hours Daily reference and final yield Monthly average of daily power yield

Data used for training and test Two Numerical Weather Prediction data were used as models input 1) NWP generated by the Weather Research and Forecasting (WRF ARW 3.6.1) mesoscale model developed by National Center of Atmospheric Research (NCAR) Forecast horizon: 24 hour Temporal output resolution: 20 minute and then averaged each hours Spatial resolution 3 km centered on the region of interest Initial and contour data for model initialization: GSF model Radiation scheme: Rapid Radiative Transfer Model (RRTM) Global Horizontal Irradiance (GHI) provided by WRF was post processed with an original Model Output Statistic called MOSRH. 2) NWP generated by the Integrated Forecasting System (IFS) the global weather forecasting model from the European Centre for Medium-Range Weather Forecasts(ECMWF). Forecast horizon: 24 hour Temporal output resolution: 1 hour Spatial resolution 16 km Radiation scheme: RRTM

Data driven techniques Two data-driven techniques were adopted to built the forecast models 1. Qualified ensemble of 300 MLPNNs with one hidden layer 500 MLPNN with the optimal hidden neuron (S) were generated using a Sub-Sample Random Validation Procedure on the training data A qualified ensemble was selected (around 300 ANNs), choosing all the ANNs with the MSE lower than the average MSE of the 500 networks Forecast was obtained by averaging the ensemble outputs. 2. Support Vector Regression method called ε-svr, Gaussian Kernel was adopted an extensive grid search on more than 400 combinations was performed to set the model parameters: regularization parameter (C), insensitive zone (ε), std (γ)

Data driven forecasting models 1 Based on Ensemble of MLPNNs using NWP inputs from WRF PV power forecast 2 Hybrid model based on MOSRH + ANNs Ensemble using NWP inputs from WRF

Data driven forecasting models 3 Based on Ensemble of MLPNNs using GHI inputs from ECMWF 4 Based on Support Vector Machine using GHI inputs from ECMWF

Results: forecast models accuracy

Results: Multi Model Ensemble construction and evaluation Since all the models show similar errors in different predicted typologies of days (identifies by daily clear sky predicted by WRF), the Multi-Model ensemble was built just averaging the different prediction trajectories

Results: Multi Model Ensemble construction and evaluation The MME outperforms the best model of the ensemble GTNN(ECMWF) MME reaches a skill score with respect to the RMSE of PM of 46% while the best forecast model GTNN(ECMWF) obtains a skill score of 42%. It was proved that the best performance of multi-model approach could be achieved averaging the higher variety of different algorithms and different NWP models with the only condition that all the ensemble members should have similar RMSE (RMSE difference less than 1% measured on one year data)

Results: MME prediction intervals construction and evaluation The prediction Intervals could be calculated forecasting the standard deviation of the residuals (σ for ) under the hypothesis that the residuals are normally distributed with zero expected value Ensemble of MLPNNs using MME power forecast

Results: MME prediction intervals construction and evaluation The frequency of observations falling inside the prediction interval is greater or equal than the confidence level associated to that interval for all years considered. Observation (dots), MME forecast (white line) and prediction intervals (grey lines) for five days of 2011

Conclusions Models based on different non linear machine learning algorithms (stochastic or statistic) making use of the same NWP data provide forecast with similar accuracy. The best performance of multi-model approach could be achieved averaging the higher variety of different algorithms and different NWP models with the only condition that all the ensemble members should have similar RMSE (RMSE difference less than 1% measured on one year data). The MME reaches a skill score with respect to the RMSE of PM of 46% while the best forecast model obtains a skill score of 42%.

References C. Cornaro, F. Bucci, M. Pierro, F. Del Frate, S. Peronaci, A. Taravat, 2015. 24-H solar irradiance forecast based on neural networks and numerical weather prediction. J. Sol. Energy Eng. 2015; 137(3). C. Cornaro, M. Pierro, F. Bucci, 2015. Master optimization process based on neural network ensemble for 24h solar radiation forecast. Solar Energy, 111, 297-312, 2015. M. Pierro, F. Bucci, C. Cornaro, E. Maggioni, A. Perotto, M. Pravettoni, F. Spada, 2015. Model Output Statistics cascade to improve day ahead solar irradiance forecast. Solar Energy, Volume 117, July 2015, Pages 99-113. M. Pierro, F. Bucci, M. De Felice, E. Maggioni, D. Moser, A. Perotto, F.Spada, C.Cornaro, 2016. Multi-Model Ensemble for day ahead prediction of photovoltaic power generation. Solar Energy, Volume 134, September 2016, Pages 132 146. M. Pierro, F. Bucci, M. De Felice, E. Maggioni, D. Moser, A. Perotto, F.Spada, C.Cornaro, 2016. Deterministic and stochastic approaches for day-ahead solar power forecasting.published online J. Sol. Energy Eng., doi: 10.1115/1.4034823.

Thank you for your attention Visit ESTER lab at www.ester.uniroma2.it