A prediction interval framework-based spatial–temporal convolution block network for traffic demand prediction

Predicting urban travel demand intervals is crucial for improving traffic management and optimizing resource allocation in smart city transit systems. We introduce a Prediction Intervals framework-based Spatial-Temporal Convolutional Block Network (PI-STCBN) that forecasts demand intervals by integrating the attention mechanism and spatial–temporal convolutional blocks (ST-Conv block), with a prediction intervals (PIs) framework. The model effectively accounts for irregular region connectivity and captures dynamic spatiotemporal correlations through a spatiotemporal graph convolutional module. By incorporating attention mechanism, the global–local graph adjacent matrix containing correlation information serves as the input of model. Specifically, different hierarchical regions have been divided by designed regional division methods depending on administrative functions, transportation accessibility, and economic factors to enhance urban network spatial partitioning. Validation experiments on two real-world datasets from Chengdu, China, and New York, USA, demonstrate that PI-STCBN outperforms four advanced methods and three baselines, achieving state-of-the-art performance in reliability, loss function performance, and prediction accuracy of PIs. In addition, we utilize the PIs and proposed global–local attention adjacent matrices as effective tools for urban traffic management, and conduct simulation experiments to schedule the vehicle dispatching and improve the operational efficiency.