A fast and multifactor evacuation method considering cumulative fatality rate based on deep reinforcement learning for urban toxic gas leakage

Abstract

Toxic gas leakage accidents negatively impact human health and the social economy, affecting the sustainability and resilience of cities. It is significant to provide safe evacuation paths timely, but most of the current evacuation methods do not consider the impact of multiple factors and have slow computation speed. In this paper, a fast and multifactor evacuation model based on deep reinforcement learning was proposed to quickly calculate evacuation paths with the lowest cumulative fatality rate. Specifically, the concentration distribution of carbon monoxide was acquired accurately by using a new solver based on buoyantBoussinesqPimpleFoam of OpenFOAM. The evacuation paths were calculated by a novel Double Dueling Deep Q Network, whose reward function was constructed by calculating high-risk areas based on an improved Wells-Riley model. To simplify the training of the model, the pedestrian was divided into leaders and followers, and Cellular Automata was coupled to simulate pedestrian collision and congestion. The results demonstrate the proposed method provides safe evacuation paths for urban toxic gas leakage faster. The study identifies the influence mechanism of multiple factors on evacuation, among which wind direction and pre-evacuation time have more significant impacts, providing valuable insights for urban planners to reduce risk and enhance urban sustainability.