A multi-scale component feature learning framework based on CNN-BiGRU and online sequential regularized extreme learning machine for wind speed prediction

Wind energy has been widely used as a clean energy source, and short-term wind speed predictions have also played a crucial role in wind power generation.However, non-stationarity and randomness make it difficult for traditional methods to achieve accurate predictions. To obtain accurate wind speed prediction results, a new multi-scale component feature learning framework is proposed. Specifically, symplectic geometric mode decomposition (SGMD) is used to decompose the data into a series of sub-sequences of different complexity, then the components are divided into high frequencies and low frequencies by permutation entropy (PE). Then according to the characteristics of different models, the high-complexity sequences are predicted by the convolutional neural network-bidirectional gated recurrent unit (CNN-BiGRU) combined model, the hyperparameters of CNN-BiGRU are optimized by the improved dandelion optimizer (DO) algorithm, while the low-complexity sequences are predicted by the online sequential regularized extreme learning machine (OSRELM) model. Finally, all component predictions are summed. In this experiment, the wind speed data of four representative months in Inner Mongolia, China in 2020 were selected to verify the proposed method. Results show that the proposed model is better than that of the benchmark models.