A learning-based MPC framework for gridlock-aware over-saturated network traffic control

The surge in traffic demand poses challenges for network traffic signal control to sustain operational efficiency under over-saturated traffic. Existing over-saturated network signal control approaches struggle to balance throughput maximization and gridlock prevention. The prevailing perimeter control provides an alternative to avoid over-saturation in the core area of the network. However, the shift problem of the macroscopic fundamental diagram makes it difficult to guarantee optimal performance under heterogeneous traffic conditions and varying demand patterns. To address the research gaps, this study proposes a learning-based model predictive control (MPC) framework for signal coordination in over-saturated networks with the aim of gridlock prevention and throughput maximization. The basic MPC framework is formulated, and its ability to prevent gridlock is theoretically proven. A graph neural network (GNN)-based predictor is designed to approximate the long-term throughput component in the objective function of MPC for computational efficiency, which also enables the model to perform well in heterogeneous traffic conditions. A Taylor expansion-based linearization is applied to the GNN-based throughput component in the objective for solutions, and the linearization reveals that the gradients of GNN provide guidance on regulating network traffic state. Simulation experiments demonstrate that the proposed model significantly outperforms the benchmark models in terms of gridlock prevention and throughput improvement, especially in heterogeneous networks and under unbalanced demand patterns. Further analysis highlights the model’s potential to identify bottlenecks in networks, which offers valuable insights for traffic management.