An adaptive time-frequency fusion PatchTST model for electric vehicle charging load forecasting

Abstract

Accurate electric vehicle (EV) charging load forecasting is important for grid dispatch, peak-load management, and charging infrastructure planning. However, residential EV charging loads exhibit both multi-scale periodicity and stochastic fluctuations, making it difficult for fixed-window models to capture dynamic temporal patterns. To address this issue, this study proposes an Adaptive Time-Frequency Fusion PatchTST (ATF-PatchTST) model for EV charging load forecasting. The model combines an adaptive patch partitioning module with a time-frequency fusion module to jointly capture short-term variations and long-term periodic features. Experiments are conducted using real-world residential charging load data from Shanghai in August 2023 with a 15-min sampling interval. Compared with PatchTST, ATF-PatchTST reduces MAE, RMSE, RAE, and RSE by 6.6%, 10.4%, 6.9%, and 5.5% in the 8–30 day forecasting task, and by 7.6%, 10.2%, 11.9%, and 4.3% in the 7-day forecasting task, respectively. Time-series cross-validation further shows that ATF-PatchTST achieves the lowest average MAE of 0.372 kWh and average RMSE of 1.205 kWh, with an RMSE coefficient of variation of 3.2%. These results demonstrate that the proposed model provides a more accurate and stable forecasting framework for residential EV charging load management and distribution network operation.