A Spatiotemporal Multi-Graph Convolutional Approach to Ride-Pooling Demand Prediction with Carpooling Features

Predicting ride-pooling demand is essential for optimizing ride-hailing services and shared transportation systems, as it balances supply and demand while enhancing vehicle utilization. Current studies often overlook the relationships among shared ride orders. To address this, we propose a spatiotemporal multi-graph convolutional neural network model that incorporates ride-pooling features for OD (origin-destination) demand forecasting. We analyze the spatiotemporal characteristics of ride-pooling demand based on origin-destination points, examining temporal patterns and spatial movement across different cycles. An OD matrix is constructed using regional divisions, with adjacency matrices illustrating geographic and semantic relationships. We introduce semantic graphs that capture ride-pooling features, leveraging multi-graph convolutional neural networks to extract complex spatiotemporal and nonlinear characteristics. A time series model is then applied for further optimization. Finally, the proposed model improves prediction accuracy by 7.26% compared to state-of-the-art models.