Machine Learning Prediction of Smoke Motion Parameters in a Half-U-shaped Underground Space with Different Structural Parameters
A comprehensive analysis of smoke movement during a fire in a half-U-shaped underground space is con-ducted by using a combination of Fire Dynamics Simulator(FDS)and machine learning techniques.The findings indicate that the Backpropagation Neural Network(BPNN)outperforms Support Vector Regression(SVR)in forecasting the length of smoke backlayering and the maximum temperature increase,with a determination coefficient exceeding 95%and relative errors predominantly within the 20%range,marking a significant improvement over the SVR method.By elucidating the machine learning model through SHAP values and integrating the outcomes of FDS numerical simulations,it is determined that the slope height is the pivotal factor influencing the length of smoke backlayering.An increase in slope height,a reduc-tion in width,or an escalation in heat release rate are all found to curtail the smoke reflux length.Concurrently,the heat re-lease rate is identified as the primary factor affecting the maximum temperature rise of the smoke,with the slope height ex-erting a substantial influence.Although a decrease in width can marginally diminish the maximum temperature rise of the smoke,its impact is not pronounced.This research not only broadens the scope of predictive methods for fire smoke motion parameters in underground spaces but also presents an innovative approach to forecasting the dynamics of fires in under-ground environments and to the optimization of ventilation and smoke exhaust systems.
国家科技期刊平台
NSTL
万方数据
