Wind power generation prediction model based on secondary reconstruction decomposition and CNN-Transformer-AcfPSO-BiLSTM in extreme weather conditions

Abstract Extreme weather can easily cause drastic fluctuations in renewable energy output, which seriously endangers the operational stability and power supply reliability of power systems. Existing ultra-short-term prediction methods struggle to accurately characterize the nonlinear dynamic features of renewable energy under extreme weather conditions. Aiming at the sharp output variations and the decline in prediction accuracy caused by extreme weather, this paper proposes an ultra-short-term prediction method for renewable energy considering extreme weather factors. Firstly, the maximum information coefficient (MIC) is adopted to screen key climatic features affecting renewable energy output. Secondly, quadratic reconstruction decomposition and denoising techniques are employed to extract multi-band features, reduce data dimensionality and optimize input sequences. Meanwhile, the particle swarm optimization (PSO) algorithm is modified to avoid falling into local optima. The improved PSO algorithm is utilized to optimize the hyperparameters of the Bidirectional long short-term memory (BiLSTM) network. Furthermore, the BiLSTM model is combined with Convolutional Neural Network (CNN) and Transformer to forecast renewable energy output. Verified on the collected extreme weather data from a certain region in Xinjiang, the results show that the proposed method achieves higher accuracy in ultra-short-term renewable energy prediction under extreme weather, and possesses promising prospects for engineering application.