Learning Dynamic Spatial Interactions for Network-Wide On-Street Parking Demand Forecasting

Accurate forecasting of on-street parking demand is critical for urban traffic management and curbside resource allocation. Existing spatiotemporal models often have limited ability to capture dynamic spatial interactions across parking networks, particularly under longer prediction horizons. To address this issue, this study proposes a spatiotemporal adaptive graph neural network for network-wide on-street parking demand forecasting. The proposed model dynamically learns spatial dependencies among parking locations through an adaptive graph convolution mechanism, reducing reliance on predefined adjacency structures. Temporal demand patterns are modeled using a recurrent learning architecture that captures both short-term fluctuations and longer-term trends. The framework is evaluated using real-world on-street parking data under multiple prediction horizons and compared with several benchmark methods. Experimental results show that the proposed approach consistently improves prediction accuracy and stability, especially for multi-step forecasting tasks. These results indicate that the model is well suited to support operational on-street parking management and data-driven urban mobility applications.