Prediction and comparison of burning rate of n-heptane pool fire in open space based on BPNN and XGBoost

Abstract

Pool fires pose a threat to the safety and environmental protection of industrial production. The burning rate is one of the most important burning parameters that determines the behavior of pool fires. Back Propagation Neural Network (BPNN) and Extreme Gradient Boosting Tree (XGBoost) were used to predict the burning rate of n-heptane pool fire with the effect of fuel depth, ullage depth, area and perimeter of pool pans, atmospheric pressure and speed of cross air flow in open spaces. The results show that after optimizing the model parameters using RandomizedSearchCV, both models yielded higher precision evaluation results. Compared with BPNN, the proposed XGBoost model demonstrates better prediction performance, achieving a R2 of 0.9744, RMSE of 4.1068, and MAE of 2.5036 in test set. The impact of different hyperparameters and data normalization on model prediction accuracy was explored. The impact of fuel depth, ullage depth, atmospheric pressure, speed of cross air flow, area, and perimeter on burning rate was determined using SHapley Additive exPlanations (SHAP) sensitivity analysis. This analysis aids in conducting risk analysis for pool fires, and offers early warning and feasible reference for fire safety hazards.