Interpretable Spatiotemporal Kolmogorov–Arnold Network for Urban-Scale Electric Vehicle Charging Forecasting

Accurate and interpretable electric vehicle (EV) charging forecasting is increasingly important for understanding and operating urban charging networks. Existing deep and graph-based models can capture complex spatiotemporal dependencies yet typically rely on latent representations that obscure how observable factors shape charging forecasts. Meanwhile, existing Kolmogorov--Arnold network (KAN)-based approaches generally apply KANs to hidden or intermediate features, limiting direct and faithful feature-level interpretation. We develop an intrinsically interpretable spatiotemporal framework that maps observable features directly to future EV charging trajectories. It applies univariate KAN functions to temporal, contextual, spatial, and interaction features and combines their effects through horizon-conditioned semantic trajectory bases. The charging forecast is constructed as an exact additive decomposition, enabling each factor's contribution to be directly quantified and inspected at both feature and trajectory levels. Experiments demonstrate competitive predictive performance against state-of-the-art deep-learning baselines while revealing horizon-dependent attribution patterns and distinct roles of contextual factors in shaping charging trajectories.