Graph Neural Network and Reinforcement Learning–Based Framework for Real-Time Traffic Congestion Detection and Police Dispatch Using Multi-Source Heterogeneous Data

Abstract

The fusion of multi-source heterogeneous data in high-speed transportation networks is essential for real-time congestion detection and rapid police response. Existing methods remain limited in data consistency, spatio-temporal pattern extraction, and path planning stability. This study proposes a congestion detection and police response framework driven by multi-source heterogeneous data. A dataset integrating flow sensors, road cameras, and Internet of Vehicles signals is constructed, with unified node, edge, and temporal features modeled through graph mapping. A spatio-temporal graph convolutional network (STGCN) with attention is employed to capture dependencies and enhance key road section representations, while a multi-task framework enables deep congestion pattern extraction. For response, geometric constraints guide path decoding, and proximal policy optimization (PPO)-based reinforcement learning achieves dynamic police dispatch. Experiments on a real expressway network with 6,120 roads and 580,000 samples show 92.4% ± 0.5 Accuracy, 89.6% ± 0.6 Topology Score, and 91.7% ± 0.6 F1-Response Score, surpassing baselines. The novelty lies in STGCN-based cross-modal fusion, geometric constraints, and the integration of PPO-based reinforcement learning. Rather than being a first-time application, the contribution is reflected in the technical integration of GNN with RL and the incorporation of constraint modeling for traffic police response, which distinguishes this framework from prior studies in emergency dispatch.