Hierarchical graph neural network-based generalized graph partitioning for accelerated large-scale microscopic traffic parallel simulation

This study aims to develop an efficient traffic parallel simulation framework by using hierarchical graph neural network for graph partitioning, improving the scalability and accuracy of large-scale microscopic traffic simulations. To fully utilize road network information for partitioning, graph embedding learning is first introduced to enrich graph node features representation. Based on embedded network graph, aiming for balanced partitioning and minimized communication, graph partitioning model uses a hierarchical graph neural network architecture to infer partitioning result. Additionally, in the design of traffic information transmission mechanism, overlapping states and crossing vehicle groups are employed to ensure the synchronization of macro and micro traffic information between partitions. Experiments conducted on large-scale highway network in Henan Province, China, demonstrate significant improvements in simulation efficiency and accuracy over traditional methods. The framework can achieve efficient acceleration under different traffic demands, with a maximum speed up of 13.23 times in 16 partitions, while ensuring higher load balancing and lower communication cost. Meanwhile, parallel simulation can also achieve nearly 100% accuracy compared to the original simulation. Moreover, in real engineering scenarios, proposed framework can also highly reproduce traffic state changes to replace the original simulation while ensuring a speed up of over 5 times. This study demonstrates the great potential of our parallel simulation framework for large-scale traffic environments, providing robust support for addressing future traffic demand growth and complex road network management.