An attention-based dynamic graph model for on-street parking availability prediction

As cities grow denser, the need for sustainable urban transport solutions intensifies. Effective management of on-street parking is critical in addressing traffic congestion and promoting environmental sustainability. This study presents a machine learning model that leverages complex spatiotemporal dependencies and incorporates essential exogenous factors to accurately predict on-street parking availability. Our approach employs a combination of graph representations-static, dynamic time-warping, and hidden state-generated graphs-to capture distinct aspects of urban parking dynamics. An attention-based fusion mechanism integrates these graphs into a cohesive dynamic graph, providing a refined understanding of parking behavior. The inclusion of external temporal features through advanced gated recurrent units enhances the model’s predictive accuracy. Rigorous testing on real datasets demonstrates the model’s superior performance, achieving a mean absolute error of 0.0379 and a mean square error of 0.0067, thereby surpassing existing benchmarks. Our results highlight the model’s efficacy as a decision-support tool for urban planners and policymakers, facilitating the development of more efficient and sustainable transport systems. Additionally, the model’s interpretability and adaptability make it a valuable tool for better understanding the intricate dynamics of urban parking. We further explore the effects of prediction accuracy and the availability of predictive information on the efficiency of the parking search process, emphasizing the critical role of accurate and timely parking data in minimizing cruising time and enhancing urban mobility.