Research on rapid prediction of mine fire in full ventilation network based on BP neural network

Abstract Mine fires pose a serious threat to underground safety, and rapid prediction of fire-induced hazards is critical for timely rescue operations. Traditional numerical simulations using Computational Fluid Dynamics are accurate but time-consuming and limited to single or local tunnels. This study proposes a rapid prediction method for mine fires in a full ventilation network based on a backpropagation (BP) neural network. A 14-branch full-ventilation network mine model was established, and numerical simulations were conducted using Fire Dynamics Simulator (FDS) to generate a fire dataset. The BP neural network was constructed with combustion location, heat release rate, time, wind speed, ambient temperature, relative humidity, ambient pressure, and cross-sectional area as inputs, and the average CO concentration, temperature, visibility, and wind speed of each roadway as outputs. The trained model enables rapid prediction of mine fires under various conditions. The results show that the model achieves high prediction accuracy for temperature, visibility, and wind speed, with R² exceeding 0.99, while the CO concentration prediction yields a SMAPE of 110.40%, an RMSE of 0.43, and an MAE of 0.31. The inference time for a single prediction is approximately 3 s, much faster than traditional simulation methods. This study provides a methodological validation for rapid mine fire prediction in a full ventilation network and offers practical value for mine fire emergency decision-making and rescue operations.