A graph neural network surrogate for traffic simulation in probabilistic wildfire evacuation planning

Abstract

Agent-based traffic simulation is one of the main computational bottlenecks in probabilistic wildfire evacuation frameworks that rely on Monte Carlo analysis to populate Bayesian network models. This paper presents a graph neural network (GNN) surrogate that replaces it within the WiSE (Wildfire Safe Egress) framework. The surrogate operates on a graph of the evacuation danger zone, built automatically from openly available road network and population data. A two-branch architecture processes static distance features through graph convolutional layers and dynamic vehicle occupancy through a graph convolutional LSTM, predicting next-step node occupancy. It is trained on one agent-based scenario from the 2018 Camp Fire in the Paradise-Magalia area of California and validated on a second scenario with a different destination configuration. The surrogate reproduces occupancy dynamics, generates plausible routes, and produces departure-travel time distributions for Bayesian network calibration. It conserves 97.4% of the network traffic volume over 24 h, with a mean absolute error of 0.40 vehicles per node, and cuts the cost of one evacuation run by a factor of 30 to 45. Integrated with WiSE, it yields a safe evacuation estimate of 15.9%, consistent with the 16% of the agent-based reference, confirming that the decision-relevant outputs are preserved. • A GNN surrogate replaces agent-based traffic simulation in wildfire evacuation. • Graph construction is automated from open road network and population data. • Two branches combine graph convolutions with a graph convolutional LSTM. • The surrogate preserves 97.4% of network traffic volume against the reference. • Inference runs 30 to 45 times faster, making Monte Carlo risk analysis tractable.