Urban intersection traffic flow prediction: A physics-guided stepwise framework utilizing spatio-temporal graph neural network algorithms

Abstract

Accurate traffic flow forecasting at urban intersections is critical for optimizing transportation infrastructure and reducing congestion. This manuscript introduces a novel framework, the Physics-Guided Spatio-Temporal Graph Neural Network (PG-STGNN), specifically designed for traffic flow prediction. By integrating the principles of traffic flow physics with advanced spatio-temporal graph neural network algorithms, the framework captures complex spatio-temporal dependencies in traffic networks. PG-STGNN adopts a stepwise approach, addressing key performance metrics like queue formation and signal timing complexities at intersections. To validate its effectiveness, the model was applied to real-world traffic data from the Yizhuang District of Beijing. Compared to traditional models such as ARIMA, KNN, and Random Forest, PG-STGNN significantly improves prediction accuracy, achieving MAPE reductions of 19.9 %, 18.6 %, 6.1 %, 20.7 %, 5.0 %, 1.8 %, and 1.1 % against KNN, ARIMA, RF, BP, T-GCN, STGCN, and ST-ED-RMGC, respectively. With the lowest MAPE (9.452 %), MAE (2.485), and RMSE (4.364), PG-STGNN demonstrates superior prediction performance. These results underscore its potential to provide reliable short-term traffic forecasts, offering essential insights for the strategic planning and management of urban intelligent transportation systems.