Posters

Abstract

General summary The development of wind energy is crucial to the building and development of a low-carbon and sustainable future worldwide. The intermittency of wind power output, related to the nature of atmospheric wind flow, remains a non-negligible aspect of wind farms, and the increasing share of wind energy in electricity generation and supply is adding more weight to the question of reliability. Power forecasting based on machine learning methods is a technique to predict power output of a wind farm related to different horizons, as in short term (1 day ahead) or super short term (several hours ahead). The forecast is useful for wind farm operators and power grid management, and with the help of power storage technologies, results in better anticipation and regulation of variations in wind power output. Method Two meso-scale weather forecast data sources have been used, the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Global Forecast System (GFS) run by the U.S. National Weather Service. Both datasets provide data of wind speed, wind direction as well as pressure , humidity and temperature at different heights above ground. The wind farm production data used in this study derive from ENGIE's open-access SCADA dataset [1] from the La Haute Borne wind farm in France. The wind farm has an installed capacity of 8200 kW total. For this study, SCADA data between January 2014 and January 2018 have been selected for building and evaluating machine learning models. Multi-Layer Perceptron neural networks based on TensorFlow-Keras framework is built to conduct a regression task that use meso-scale variables as features to predict wind farm power output at each 10-minute time step. The RMSE (Root Mean Square Error) is calculated daily (144 data points) and then averaged. The accuracy score is obtained from one minus the ratio between the average RMSE and the installed capacity. Results To measure the model performance, a split of the whole dataset has been made to obtain a train dataset and test dataset, which represents a little longer than 1 year of data for the test dataset. In the search of hyperparameters, it is found that the accuracy score obtained based on cross-validation is the best indicator of model accuracy on unforeseen test data. Based on power output data between January 2017 and January 2018 during which both GFS and ECMWF data are available (about 50,000 samples), a comparison has been made on the accuracy score of models trained exclusively with one of the mesoscale data sources. Preliminary results show that the ECMWF-based model has higher accuracy scores than the GFS-based one, whereas the hybrid model is slightly more accurate than ECMWF. Conclusions Using weather forecast data from ECMWF and GFS databases, this study focuses on the day-ahead prediction of wind farm power output. For wind farm and grid operators, forecast results from an appropriate mesoscale weather forecast data source can help anticipate wind farm power output and thus reduce the uncertainty in decision making related to integration of wind energy into power systems.