Real-time Prediction of Explosion Consequences during Gas Leakage in Tunnels Using a Two-stage Deep Learning Framework

Natural gas plays an essential role in the global energy landscape, and its safe transportation is critical to the reliability and efficiency of the energy supply chain. However, predicting the dispersion of leaked gas and resultant explosion hazards remains challenging due to uncertainties in environmental and leakage conditions. Such incidents not only cause significant energy loss but also pose severe threats to critical infrastructure. This study develops a novel two-stage deep learning framework to predict the consequences of explosions from natural gas leakage in tunnels using data from sparse sensors. This study leverages computational fluid dynamics (CFD) simulations to generate datasets that model gas leakage, dispersion, and explosion processes across a range of leakage rates and wind velocities. The proposed framework is structured around two sequential mappings. In the first stage, a hybrid Graph Attention Network–Convolutional Neural Network–Bidirectional Long Short-Term Memory (GAT–CNN–BiLSTM) architecture reconstructs three-dimensional gas concentration fields from time-series sensor measurements. In the second stage, a GAT–CNN network converts these reconstructed fields into distributions of maximum explosion hazards, including static pressure, dynamic pressure, and temperature. A systematic evaluation using the normalized root mean square error (NRMSE) and structural similarity index (SSIM) demonstrates prediction accuracy, with average values of 0.085 and 0.938, respectively. Furthermore, critical engineering parameters, including sensor spacing, input data length, sensor failure tolerance, and robustness to noise, are quantitatively analyzed to guide the design of reliable monitoring systems.