A Dynamic-Inner Manifold Broad Learning System With Coupled Time-Frequency Domain for Wind Speed Prediction

Wind speed prediction plays an important role in operating renewable energy power systems that incorporate wind power. However, wind speed prediction remains a significant challenge due to the non linearity, chaos, and nonstationarity of wind speed data. In this article, we propose a novel wind speed prediction model based on the dynamic-inner manifold broad learning system with coupled time-frequency domain (DiM-BLS-CTFD). We first propose the autovariational-mode decomposition (AVMD) algorithm to automatically transform nonstationary wind speed series data into a relatively more stationary series. Then, the dynamic-inner manifold learning (DiML) algorithm has been developed to extract the dynamic manifold structure with regression properties from the decomposed results by AVMD. Finally, building upon AVMD and DiML, DiM-BLS-CTFD is designed for wind speed prediction. DiM-BLS-CTFD is a unified framework that combines signal processing, feature extraction, and prediction. It effectively couples the multimodal time-frequency features of wind speed data, enhancing the accuracy of wind speed prediction. Compared with the original BLS model, DiM-BLS-CTFD showed an average reduction of 50.63% in root-mean-square error, 50.49% in mean absolute error, and 51.90% in symmetric mean absolute percentage error.