Local–Global Synergistic Forecasting of Electric Vehicle Charging Load Based on Frequency Separation and Feature Compression

The increasing penetration of electric vehicles introduces pronounced randomness, non-stationarity, and multi-scale fluctuations into charging load, making accurate short-term forecasting difficult. Existing methods often suffer from frequency mixing among heterogeneous driving variables, decomposition-induced feature redundancy, and insufficient coordination between local abrupt variation modeling and long-range dependency learning. To address these issues, this paper proposes a multi-scale feature decoupling and local–global synergistic forecasting mechanism. First, an ICEEMDAN-KPCA feature processing strategy is developed to separate stochastic disturbances, periodic fluctuations, and long-term trends in key charging-related variables, and then reorganize the decomposed components in a nonlinear kernel space to suppress redundant information. Second, a serial local–global forecasting architecture is constructed, in which a two-layer LSTM first captures charging-session continuity, short-term state transitions, and local ramping patterns, and the encoded sequence is subsequently modeled by an improved Informer. Different from the standard Informer, relative positional encoding is introduced to strengthen sensitivity to temporal shifts, while dilated causal convolutional distillation enlarges the historical receptive field and preserves abrupt peak-valley and session-boundary variations without future information leakage. Experiments on the Jiaxing EV charging dataset show that the proposed model achieves an RMSE of 56.120 kW, MAE of 42.280 kW, WAPE of 6.990%, and R2 of 0.9627. Compared with nine representative baselines, it achieves the best overall performance, reducing RMSE and MAE by 28.62% and 29.33%, respectively, compared to TimeXer. Ablation results further verify the effectiveness of the proposed model.