A hybrid deep learning architecture for wind power prediction based on bi-attention mechanism and crisscross optimization

Abstract Accurate wind power forecasting is of great significance for power system operation. In this study, a triple-stage multi-step wind power forecasting approach is proposed by applying attention-based deep residual gated recurrent unit (GRU) network combined with ensemble empirical mode decomposition (EEMD) and crisscross optimization algorithm (CSO). In the data processing stage, the EEMD is used to decompose the wind power/speed time series and a bi-attention mechanism (BA) is applied to enhance the sensitivity of model to the important information from both time and feature dimension. In the prediction stage, a series-connected deep learning model called RGRU consisting of the residual network and GRU is proposed as the forecasting model, aiming to make full use of extracting the static and dynamic coupling relationship among the input features. In the retraining-stage, a high-performance CSO algorithm is adopted to further optimize the fully-connected layer of RGRU model so as to improve the generalization ability of the model. The proposed method is validated on a wind farm located in Spain and the experimental results demonstrate that the proposed hybrid model has significant advantage over other state-of-the-art models involved in this study in terms of prediction accuracy and stability.