Regional-scale day-ahead wind power forecasting using deep learning

Abstract

<p>Historically, electricity was provided by dispatchable sources that are able to adjust to demand variations. Adding an irregular source to the network is a challenge for the grid stability. Especially, wind turbine production varies depending on meteorological conditions. At a regional or country scale, Transmission System Operators (TSOs) are responsible to maintain supply demand equilibrium in their network. In this context, wind power production forecast is one of the tools needed to manage the network. Traditionally, physical models are used to predict power production based on turbine characteristics and numerical weather prediction models. Indeed, wind power production is strongly correlated to wind speed at turbine hub height and other meteorological parameters. One limit of those physical approaches is that they require precise knowledge on turbines characteristics and locations, in particular at a regional scale.</p><p>To overpass this limit, a statistical approach such as deep learning can be used but needs to be qualified in terms of performances. In this study, a supervised deep learning model is explored. This model does not require information on turbine location or characteristics but does require historical samples of weather parameters and associated production.</p><p>Our work focuses on day-ahead forecasts (horizons 24 to 48 hours) for a German TSO (region-scale). One physical model was selected as a reference and the goal was to combine deep learning and physical predictions to obtain the best possible forecast. Both the physical and the deep learning models use spatiotemporal meteorological inputs from the IFS (ECMWF) and GFS (NCEP) models. A convolutional neural network (CNN) was used to exploit the spatial information of maps of features. A LSTM was added to capture information from the time series evolution. Finally, several deep learning predictions were combined with the physical model prediction using a multi-layer perceptron. With this method, the MAE-based skill score of our final model, combining the physical one and deep learning ones, reaches more than 6% over a validation period of one year.</p>