Combustible gas leakage and diffusion prediction based on graph neural network

Gas leakage and explosion accidents pose a serious threat to public safety, and a prerequisite for accurately predicting the explosive effects of combustible gas leakage is to determine the concentration distribution after leakage. To develop a real-time, full-field spatiotemporal prediction model for combustible gas leakage and diffusion and to achieve efficient prediction of the equivalent gas cloud volume, a novel graph neural network model based on a dual-neural-network architecture and a multi-stage training strategy, named Multi-Stage Dual Graph Neural Network (MSDGNN), has been proposed in this study. The MSDGNN model consists of two synergistic sub-networks: (1) the Concentration Network (<italic>N</italic>_con), which establishes the mapping relationship between the concentration fields of two consecutive timesteps, and (2) the Volume Network (<italic>N</italic>_vol), which generates the equivalent gas cloud volume at each timestep to provide a quantitative metric for explosion risk assessment. To further enhance model performance, a multi-stage progressive training strategy has been developed to jointly optimize the dual networks. To evaluate the performance of MSDGNN, a dataset considering different leakage rates, leakage points, and leakage durations was constructed. Experimental results demonstrate that compared with traditional single-network architectures (e.g., MGN), the dual-network architecture effectively decouples the tasks of concentration field prediction and risk assessment. This approach significantly mitigates the interference of weight factors in single-objective loss functions during the training process. The multi-stage training strategy, through stepwise parameter optimization, successfully addresses the issue of insufficient data fitting encountered in traditional methods. Compared with MGN, the Mean Absolute Percentage Error (MAPE) for concentration fields and equivalent gas cloud volumes is reduced from 49.47% and 108.93% to 7.55% and 9.07%, respectively. Furthermore, the generalization error (MAPE) of MSDGNN for concentration fields and equivalent gas cloud volumes is reduced from 41.18% and 38.81% to 8.01% and 14.92%, respectively. In addition, MSDGNN exhibits robust prediction performance even when key parameters such as leakage rate, leakage location, and leakage duration exceed the range of training data. Compared with CFD numerical simulation methods, the proposed model achieves a three-order-of-magnitude improvement in computational efficiency while maintaining prediction accuracy, providing an effective real-time analytical tool for industrial safety monitoring.