Abstract: Aiming at the problem that traditional prediction models struggle to simultaneously capture global and local features of multi-feature load data, a hybrid prediction model based on sparrow search algorithm (SSA)-optimized variational mode decomposition (VMD) and improved squeeze-and-excitation (ISE) module, and temporal convolutional network (TCN) and Pyraformer was proposed. Firstly, the SSA was employed to optimize VMD parameters, enabling VMD to highly volatile load sequences into a set of relatively stable modal components to reduce the non-stationarity of the original data. Then, the obtained intrinsic mode functions was input into the TCN model to capture the local features of the data, while the ISE module adaptively assigned appropriate weights to the extracted features, thereby reducing the impact of redundant information on the prediction results. Finally, the weighted data was fed into the Pyraformer model to capture the global features and generated the final prediction results.To validate the mode’s performance, real-world power load datasets from two regions were used for simulation experiments. The results show that in both cases, the proposed model achieves the coefficients of determination is 0.9949 and 0.9842, respectively, outperforming other comparative models. This verifies the proposed model’s superiority in simultaneously capture global and local features of multi-feature load data, demonstrating higher prediction accuracy and stability.