Spatiotemporal Predictions of Toxic Urban Plumes Using Deep Learning

TL;DR

This work tackles the challenge of rapidly predicting toxic urban plume dispersion to support emergency response, where high-fidelity CFD are computationally intensive. It introduces ST-GasNet, a spatiotemporal deep learning model with a novel ST-LSTM++ unit that explicitly captures first- and second-order information flows inspired by the advection-diffusion process. The method is trained on large-eddy-simulation–generated plume data and demonstrates superior or comparable performance to PredRNN, even when wind information is unavailable, achieving about 90% accuracy over a 9-minute horizon and enabling near real-time inference. The approach offers a practical pathway to real-time plume forecasting in urban environments, with potential applicability to other spatiotemporal prediction tasks beyond plume dispersion.

Abstract

Industrial accidents, chemical spills, and structural fires can release large amounts of harmful materials that disperse into urban atmospheres and impact populated areas. Computer models are typically used to predict the transport of toxic plumes by solving fluid dynamical equations. However, these models can be computationally expensive due to the need for many grid cells to simulate turbulent flow and resolve individual buildings and streets. In emergency response situations, alternative methods are needed that can run quickly and adequately capture important spatiotemporal features. Here, we present a novel deep learning model called ST-GasNet that was inspired by the mathematical equations that govern the behavior of plumes as they disperse through the atmosphere. ST-GasNet learns the spatiotemporal dependencies from a limited set of temporal sequences of ground-level toxic urban plumes generated by a high-resolution large eddy simulation model. On independent sequences, ST-GasNet accurately predicts the late-time spatiotemporal evolution, given the early-time behavior as an input, even for cases when a building splits a large plume into smaller plumes. By incorporating large-scale wind boundary condition information, ST-GasNet achieves a prediction accuracy of at least 90% on test data for the entire prediction period.

Figures (10)

• Figure 1: Schematic of the proposed workflow. The spatiotemporal ML model is trained and tested to replace part of the expensive high-fidelity physics-based model, allowing for real-time predictions of plume evolution from a few initial observations
• Figure 2: Example of the difficulties when estimating the dispersion process within two adjacent steps. The dispersion process happens simultaneously in multiple directions if spatial discontinuities (such as buildings) exist. Thus, considering enough early time steps is crucial when training the ML model
• Figure 3: Aerial image of the Aeolus urban grid used for this problem
• Figure 4: Structure of PredRNN. Arrowed lines in yellow and gray demonstrate information flow within PredRNN. Figure is adjusted from wang2021predrnn
• Figure 5: Two formulations of the problem. Left: Problem formulation without any wind information; Right: Problem formulation with wind information