Spatio-temporal deep learning for day-ahead wind speed forecasting relying on WRF predictions

We focus on deep learning algorithms, improving upon the weather research and forecasting (WRF) model, and we show that the combination of these methods produces day-ahead wind speed predictions of high accuracy, with no need for previous-day measurements. We also show that previous-day data offer a significant enhancement in a short-term neural network for hour-ahead predictions, assuming that they are available on a daily basis. Our main contribution is the design and testing of original neural networks that capture both spatial and temporal characteristics of the wind, by combining convolutional (CNN) as well as recurrent (RNN) neural networks. The input predictions are obtained by a WRF model that we appropriately parameterize; we also specify a grid adapted to each park so as to capture its topography. Training uses historical data from five wind farms in Greece, and the 5-month testing period includes winter months, which exhibit the highest wind speed values. Our models improve WRF accuracy on average by 19.4%, and the improvement occurs in every month; expectedly, the improvement is lowest for the park where WRF performs best. Our neural network is competitive to state-of-the-art models, achieving an average MAE of 1.75 m/s. Accuracy improves for speed values up to 20 m/s, which are important in wind energy prediction. We also develop an RNN model and show that MAE reduces to less than 1 m/s for short-term predictions if actual data is employed.