A hybrid model based on CapSA-VMD-ResNet-GRU-attention mechanism for ultra-short-term and short-term wind speed prediction

Accurate wind energy prediction is crucial for the safety and stability of power systems. Aiming to further improve the accuracy and robustness of wind speed prediction, this paper presents a hybrid model based on capuchin search algorithm (CapSA), variational mode decomposition (VMD), residual network (ResNet), gated recurrent unit (GRU) network and attention mechanism. In which VMD optimized by CapSA with a new joint fitness function is proposed to decompose original wind speed data, mitigating the non-stationary characteristics of the wind speed sequence, a ResNet based GRU framework is constructed, which alleviating the degradation problem resulting from the deep network layers, and an attention mechanism is introduced to emphasize the impact of key information. Based on the proposed model, a series of experiments on multiple datasets are conducted. The ablation experiments verify the effectiveness of the current model and the contribution of each component. The comparison experiments with three state-of-the-art hybrid models including both decomposition and prediction modules further demonstrate its superiority and stability. Experiments on the datasets collected from typical months of two regions suggest that the proposed model can capture the regional and monthly diversities. Detailed analysis for multiple scenarios provides insight into the correlations of input characteristic and learning capabilities.