Semantics-Aware Spatial-Temporal Dynamic Graph Transformer Network for On-Street Parking Occupancy Prediction

Accurately predicting on-street parking occupancy is essential for urban traffic management and smart city applications. Previous studies have largely overlooked the importance of semantic correlations among parking locations, while also struggled to effectively capture long-term temporal dependencies and large-scale spatial interactions. To address these limitations, we propose a Semantics-aware Spatial-Temporal Dynamic Graph Transformer Network (SaSFormer) for on-street parking occupancy prediction. Specifically, a graph-based convolutional approach is employed to capture semantic correlations among parking locations; a ProbSparse self-attention mechanism is utilized to model temporal dependencies; and a dynamic self-attention mechanism is designed to capture spatial interactions. Experimental results on real-world datasets demonstrate that SaSFormer significantly outperforms baseline models in prediction accuracy and robustness. Notably, incorporating semantic correlations results in an additional 5.38% reduction in MAE and 6.01% reduction in RMSE, underscoring the critical role of semantic correlations in enhancing on-street parking occupancy prediction.