Urban traffic networks exhibit highly dynamic and nonlinear spatiotemporal interactions that require predictive modeling and adaptive control mechanisms capable of operating under low-latency constraints. This study proposes a GNN–RL-based Edge–Cloud Digital Twin framework for real-time traffic forecasting and adaptive signal control. At the network edge, multi-source traffic data collected from roadside sensors are processed on distributed edge devices to perform multi-step prediction of traffic flow, vehicle density, and congestion states. The forecasting module integrates Graph Convolutional Networks (GCNs) to capture spatial dependencies across the road topology with Long Short-Term Memory (LSTM) units and a Transformer-based predictor to model short- and long-range temporal dynamics. These predicted traffic states are transmitted to a cloud-level Digital Twin Engine, which performs data fusion, state estimation, calibration, and scenario-based simulation to maintain a continuously updated virtual representation of the physical traffic network. Using the forecasted states as inputs, a deep reinforcement learning optimization module performs adaptive signal phase control to minimize average vehicle delay and maximize intersection throughput. The overall framework operates as a closed feedback loop integrating edge-level spatiotemporal forecasting, cloud-level synchronization and simulation, and reinforcement learning–based control policy optimization. Experimental evaluation demonstrates a 17% reduction in average vehicle waiting time and significant improvements in forecasting performance measured using MAE and RMSE, with strong robustness to missing and noisy data conditions. The proposed architecture provides a scalable and low-latency solution for data-driven traffic prediction and signal control within an edge–cloud digital twin environment.
Alanazi et al. (Mon,) studied this question.