A conditional and regularized approach for large-scale spatiotemporal wind power forecasting

Abstract

Large-scale wind power forecasting at intraday horizons of minutes-ahead up to hours-ahead is essential to secure important operations in transmission systems. It is clear that recent information collected about neighboring sites improve the predictive performance of autoregressive models. At the scale of a region or of a country, regularization or feature selection are needed to mitigate the high dimensionality of the autoregressive model. Unconditional approaches of regularization have shown limited added value compared to benchmark models in the context of wind power forecasting. This work proposes an intraday wind power forecasting method that predicts the production of any wind farm in the control area of a Transmission System Operator (TSO), taking into account the information collected from other wind farms. The method combines feature selection, regularization and local-learning via conditioning on recent production levels or on expected weather conditions. Improvements in Root Mean Squared Error (RMSE) with respect to other models, evaluated on a dataset with a large number of wind farms are comprised between 4% (10-min horizon) and 11% (3-h horizon). Interpretability of the forecasting model is demonstrated via an analysis of the model coefficients and a discussion of the performance in a challenging situation, namely a wind front.