Advanced Demand Prediction for Demand‐Responsive Transit ( DRT ) Systems: Integrating Graph Convolutional Networks and Deep Learning Techniques

Abstract

ABSTRACT Demand‐Responsive Transit (DRT) systems require reliable short‐term demand forecasts to support fleet allocation, service planning, and operational decision‐making under spatially heterogeneous and temporally variable travel demand. This study formulates short‐term DRT demand forecasting as a graph‐based spatial–temporal learning problem using real‐world Metro Connect operations in Kalamazoo County, Michigan. Historical trip requests are aggregated into 30‐min operational intervals and represented across 251 predefined one‐square‐mile zones, with each zone modeled as a graph node and inter‐zone relationships encoded through distance‐based adjacency. A Graph Convolutional Network (GCN), implemented using PyTorch Geometric, is combined with a multi‐view temporal representation inspired by DMVST‐Net to incorporate short‐term, recurrent service‐period, and longer‐term historical demand patterns. The framework is evaluated under a strictly chronological train–validation–test protocol against statistical, machine‐learning, and deep‐learning baselines. Results show that Random Forest achieves the strongest network‐level forecasting performance (MAE = 1.51, RMSE = 2.87, and R 2 = 0.915), demonstrating the predictive strength of lagged demand and calendar‐based features for aggregate DRT demand. The graph‐based models, while not outperforming Random Forest at the network level, provide an explicit representation of spatial relationships among service zones that is absent from conventional aggregate forecasting approaches. The findings demonstrate that model performance depends strongly on the forecasting target and highlight the distinction between aggregate predictive accuracy and spatially structured demand representation. The proposed framework establishes a reproducible foundation for spatially explicit DRT demand forecasting and future integration with zone‐level prediction, adaptive service zoning, fleet allocation, routing, and transit decision‐support systems.

The authors' abstract, as published at the source. Engineering Reports, 2026 · DOI ↗