Forecaster as a simulator: Simulating multi-directional pedestrian flow with knowledge-guided Graph Neural Networks

Crowd dynamics, particularly within multi-directional pedestrian flows, present a complex system that has been a focal point in simulation studies due to its intricate complexity. In addition to the continuous improvements being made to traditional crowd dynamics simulators, emerging technologies such as Machine Learning (ML) have injected new vitality into this domain. By leveraging the ML forecaster's capability to discern complex mechanisms, this study introduced the concept of ‘Forecaster as a Simulator’ (FaaS) and developed a knowledge-guided Graph Neural Network-based crowd simulator (KGCS). The GNN-based predictor's understanding of the scene is enhanced by a dedicated hierarchical graph converter, local hierarchical subgraph encoder, and global attention mechanism, while its predictive accuracy is improved with the assistance of a mixture-of-experts projector. Based on this, knowledge guidance helped suppress error accumulation that may occur in rolling simulations (i.e., the process of implementing FaaS), thus further ensuring the accuracy of trajectories over extended periods. Comparative experiments and analyses against various baselines, including knowledge-based and data-driven simulators, demonstrated that KGCS significantly elevates trajectory prediction accuracy in multi-directional pedestrian flow simulations. Additionally, the simulator offers more precise simulations at both mesoscopic and macroscopic scales, effectively pinpointing regions with potential crowd hazards and reduced traffic capacity more accurately than traditional models. This enhanced simulator provides a robust scientific tool for layout streamlining design and crowd safety management in large-scale facilities, marking a significant step forward in the field of crowd dynamic simulation studies. © 2024 Elsevier Ltd.