A Short-Term EV Charging Load Forecast Method Based on Kolmogorov–Arnold Spatiotemporal Attention Recurrent Network

Short-term electric vehicle charging load forecasting is essential for distribution grid operation, charging service planning, and near-term resource coordination. This paper proposes a Kolmogorov–Arnold Spatiotemporal Attention Recurrent Network (KASTAR) for multi-target forecasting in urban charging systems. The model combines a cross-correlation-guided graph attention encoder to capture meaningful spatial dependencies among charging zones with a gated recurrent decoder enhanced by Kolmogorov–Arnold temporal-pattern attention to model nonlinear lag effects. A lightweight Kolmogorov–Arnold output head is further used to generate multi-horizon forecasts while preserving interpretable shape functions. Experiments on multiple city-scale real-world datasets show that the proposed method achieves state-of-the-art accuracy for 1–4 h-ahead forecasting. Compared with a structurally similar spatiotemporal baseline, KASTAR improves forecasting accuracy by 10.5% while using more than 50% fewer parameters. The model also maintains strong robustness under additive noise and missing data. Ablation and basis-replacement studies show that the temporal-pattern attention design and the localized B-spline basis are the main sources of the performance gains. These results demonstrate that KASTAR is an accurate, robust, and interpretable solution for short-term electric vehicle charging load forecasting.