Gas dispersion modeling in stereoscopic space with obstacles using a novel spatiotemporal prediction network

Abstract

Gas leakage can lead to catastrophic consequences on both the environment and human health. To mitigate these losses, it is imperative to develop accurate and efficient spatiotemporal models for gas dispersion. The gas diffusion process occurs in a 3-dimensional (3D) space, but most research has been confined to flat-plane scenarios, neglecting the stereoscopic distribution of gas concentrations. To address this issue, we propose a novel method that combines 3D convolution with a long short-term memory neural network (3DConvLSTM) to forecast the 3D spatiotemporal concentration distribution of gas leakage in obstructed scenes. The 3D convolutional filters fully operate in the spatial domain, capturing spatial features horizontally and vertically. To provide data for the experiment, ethane leak scenarios with different sources, rates and wind directions are simulated by computational fluid dynamics (CFD). The results demonstrate that the 3DConvLSTM exhibits higher accuracy and requires fewer parameters, highlighting the effectiveness of the proposed method.