Abstract
Accurate prediction of urban electric vehicle (EV) charging demand is critical for infrastructure planning and dynamic pricing strategies. Although various methods have been developed, most existing studies focus primarily on spatiotemporal dependencies, paying limited attention to interactions among multivariate features. Furthermore, conventional serial spatiotemporal architectures typically extract features dimension-by-dimension, which may impede cross-dimensional information flow and lead to imbalanced representations. To address these challenges, we propose the Multi-Dimensional Feature Aggregation Network (MDFANet). MDFANet is designed to enhance multivariate representations while embedding spatiotemporal attention to strengthen relational modeling. Specifically, we introduce a Multi-Dimensional Feature Aggregation Module (MDFAM) that conducts fine-grained aggregation along both temporal and variable dimensions. By fusing these aggregated features with raw inputs, the model preserves distributional and semantic heterogeneity. Extensive experiments on real-world datasets demonstrate that MDFANet outperforms competitive baselines in prediction accuracy while reducing computational costs by approximately 50%. For reproducibility, the source code is available at https://github.com/kion-86/MDFANet .