Spatio-Temporal Graph Neural Learning for Regional Traffic Flow Prediction and Optimization

Abstract

This study proposes a Spatio-Temporal Graph Transformer (STGT) framework for regional traffic-flow prediction and optimization. This framework integrates traffic, temporal, spatial, weather, and road-network information from metropolitan traffic monitoring data. A weighted graph represents road-segment connectivity, while graph-based spatial learning captures interactions among neighbouring segments. Transformer-based temporal attention models historical traffic dynamics and recurring variations. The resulting spatio-temporal representation is used to predict future traffic flow, estimate congestion conditions, and identify highly loaded road segments. These predictions support traffic-flow optimization by considering alternative connected segments, enabling improved traffic distribution and providing a foundation for intelligent transportation management and congestion mitigation.