Improving Park-and-Ride Occupancy Prediction With Feature-Augmented Cross-Variate and Temporal Transformer

Accurate prediction of parking occupancy at Park-and-Ride (P+R) facilities is crucial for enabling efficient multimodal transportation systems and supporting dynamic urban mobility management. P+R parking behavior is influenced by multidimensional factors encompassing temporal patterns, meteorological conditions, socioeconomic indicators, and transit-related attributes. To address this challenging prediction task, this study proposes CvTFormer (Cross-variate and Temporal Transformer), a novel Transformer-based model specifically designed to jointly capture temporal dependencies and intricate cross-variate interactions within P+R occupancy data. CvTFormer employs a dual-attention architecture: self-attention mechanisms capture time-dependent patterns, while cross-attention mechanisms facilitate the deep integration of heterogeneous features. Experiments on a real-world dataset spanning 24 P+R facilities in Beijing demonstrate that CvTFormer consistently outperforms state-of-the-art baselines. It achieves average reductions in MAE, RMSE, and MAPE of 10.04%, 9.88%, and 13.45%, respectively, compared to the standard Transformer model. Further ablation studies highlight the importance of incorporating auxiliary features, yielding additional reductions of 6.05% in MAE, 5.34% in RMSE, and 3.15% in MAPE. These results underscore the effectiveness, robustness, and generalizability of CvTFormer for practical urban mobility applications.