Context-aware temporal fusion transformer for short-term electric vehicle charging stations demand prediction

The rapid growth of electric vehicles (EVs) has significant implications for power grid stability and charging station operations. Accurate short-term prediction of charging demand is critical for efficient resource allocation, load management, and service planning. Existing studies usually aggregate data across locations, overlooking the spatial variability in charging behavior that arises from differences in site types, user patterns, and temporal trends. This study develops a Context-Aware Temporal Fusion Transformer (CATFT) for short-term EV charging demand forecasting, incorporating both location-specific and weather-based contextual features. The dataset consists of two years of charging transactions collected from multiple stations across Jiaxing City, Zhejiang Province of China. The 1-hour and 1-day ahead predictions are performed using 24-hour and 7-days rolling windows. The performance comparison is performed with baselines including deep learning-based individual and hybrid models. The CATFT model performs better than LSTM, Transformer, and CATFormer for both 1-hour ahead and daily demand predictions. Compared to LSTM, CATFT achieves up to 65.0% MSE and 42.7% RMSE reductions in hourly predictions and 87.5% MSE and 69.2% RMSE reductions in daily predictions. When compared to Transformer, CATFT reduces MSE by 9.7% in CS1, 60.4% in CS2, and 43.9% in CS3, with RMSE reductions of 32.7% and 20.6% in CS2 and CS3, respectively. Compared to CATFormer, CATFT provides better performance in CS3 with 34.8% MSE and 21.1% RMSE improvements, but underperforms in CS2. Compared to the Transformer, CATFT is a more balanced and robust model, which works well at stations with stable and variable demand patterns.