A Spatio-Temporal Attention Model for Short-Term Load Forecasting of Urban Electric-Vehicle Charging Stations and an Empirical Study of Spatial-Modeling Effectiveness

Abstract

Accurate short-term load forecasting for EV public charging stations is essential for grid and station operations. However, predictions are challenging because charging loads are non-stationary, spatially heterogeneous, and closely coupled with external factors such as weather and pricing. In this study, we forecast hourly regional charging energy using the open UrbanEV benchmark dataset, which includes hourly charging records from 1362 public charging stations across 275 traffic-analysis zones in Shenzhen from September 2022 to February 2023. We propose ST-Attention, a lightweight and modular forecasting model. It integrates temporal self-attention, spatial self-attention, an adjacency-matrix bias, and a residual prediction head. We compare ST-Attention with five baselines using a leakage-free rolling time-series evaluation protocol. For the 3 h horizon, ST-Attention achieves an MAE of 62.44 kWh, an RMSE of 291.8 kWh, and an MAPE of 5.91%, reducing the MAE by approximately 41% compared with the last-observation baseline. The model also maintains superior MAE performance at the 6 h and 9 h horizons. A modular ablation study shows that temporal attention and the residual head are the most stable sources of improvement, whereas dense spatial attention does not automatically provide benefits at hourly granularity with limited samples; removing it further reduces the 3 h MAE to 59.58 kWh. We present this as a cautionary finding: local temporal inertia dominates dense spatial coupling in hourly forecasting, and spatial model complexity must align with data granularity.