Short term wind speed prediction based on optimization decomposition and dynamic feature selection

Wind speed prediction in wind farms is often hindered by insufficient feature extraction and low prediction accuracy. To address these challenges and enhance both the accuracy and reliability of wind speed forecasts, a novel integrated prediction model is proposed, combining optimized variational mode decomposition (VMD), convolutional neural networks (CNN), long short-term memory networks (LSTM), and Kolmogorov Arnold networks (KAN). Initially, the maximum mutual information coefficient (MMIC) is selected as the fitness function, and the grey wolf optimization algorithm (GWO) is employed to globally optimize the modal decomposition layers and penalty factors of the VMD. The optimized VMD is then used to effectively decompose the raw wind speed data, followed by the reconstruction of the decomposed components. Finally, the dynamic feature selection mechanism of the KAN network is integrated with CNN and LSTM to construct a CNN-LSTM-KAN prediction model for wind speed forecasting. Experimental results demonstrate that the KAN network significantly improves the extraction of wind speed features, thereby enhancing the predictive capability of the model. The effectiveness and generalization of the proposed approach are further validated through simulations using measured data from various wind farms.