Recalibration of recurrent neural networks for short-term wind power forecasting

Abstract

This paper is focused on the day-ahead prediction of the onshore wind generation. This information is indeed published each day, ahead of the market clearing, by European Transmission System Operators (TSOs) to help market actors in their scheduling strategy. In that regard, our first objective is to improve the forecast performance by efficiently capturing the complex temporal dynamics of the wind power using recurrent neural networks. Practically, advanced architectures of Long Short Term Memory (LSTM) networks are implemented and compared. Secondly, in order to continuously refine the prediction tool, different techniques for recalibrating the model during its practical utilization are analyzed. This procedure consists in adjusting the parameters of the neural networks by taking advantage of the new information revealed over time, without the (time-consuming) need to retrain the model from scratch using the whole available dataset. Finally, the financial savings from the improvement of the forecast accuracy are estimated. Outcomes from the Belgian case study show that an optimal model recalibration can significantly improve forecast reliability, thereby decreasing the balancing costs of the system.