A wind power prediction method fusing deep learning with rolling adaptive successive variational mode decomposition

Data modal decomposition combined with deep learning networks has demonstrated high efficacy in short-term wind power prediction. However, traditional data modal decomposition methods are plagued by issues such as data leakage and struggle to balance multiple objectives. This study proposes a novel wind power prediction model integrating rolling adaptive successive variational mode decomposition with deep learning networks. In the proposed model, a rolling window mechanism is employed to safeguard future data during the decomposition process. Building upon successive variational mode decomposition, a multi-objective adaptive stopping criterion is incorporated to realize multi-objective optimization throughout the decomposition process. Subsequently, deep learning networks are utilized to fulfill the prediction tasks, with a specific analysis of the coupling relationship between data modal decomposition algorithms and deep learning networks. Based on open-source datasets, an extensive set of comparative experiments were carried out: The optimal rolling window size was determined via window irrelevance verification. Compared with coupled prediction models utilizing various alternative data modal decomposition algorithms, the proposed model demonstrates superior performance. Among models ensuring data reconstruction accuracy, it outperforms the second-ranked model by reducing RMSE, MAPE, and MSE by 26.81 %, 11.48 %, and 18.3 % respectively, while shortening the data modal decomposition time by 60 %. Among models exhibiting excellent predictive performance, it reduces the reconstruction error by approximately 80 % compared to the second-best model. Experiments investigating the coupling of modal decomposition algorithms with deep learning networks confirm that the proposed model and a simple deep learning network constitutes an optimal coupled prediction model.