Day-ahead Wind Power Prediction Based on Corrected Transformer Network

In the electric power system, wind energy is widely used due to its clean, inexhaustible and free characteristics. The accurate prediction of wind power in the day-ahead plays a vital role in both dispatching and energy market, as it can potentially increase wind energy permeability and the profit of power plants. Currently, numerical weather prediction (NWP) is used in day-ahead wind power prediction. However, the NWP data are not currently under fully used. For NWP commences at UTC 12:00, and there is a certain period before the prediction result needs to be uploaded, which provides both the input data, NWP and output wind power, for correction. Furthermore, with the development of GPU, there is ample time for correction without including the model training time. To improve the day-ahead prediction accuracy, this paper presents a hybrid deep learning model with correction by using historical data. The proposed model is built with a transformer network that accesses the encoder with historical data to enhance itself with correction information. The results of experiments conducted in a real wind farm, using four commonly used benchmark models, demonstrate the accuracy and robustness of the proposed model.