Estimating thermal radiation of vertical jet fires of hydrogen pipeline based on linear integral and machine learning

Accurate and efficient prediction of thermal radiant of hydrogen jet fire is important to schedule safety design and emergency rescue program for hydrogen pipelines. In response, this paper proposes a novel Optuna-improved back propagation neural network (Optuna-BPNN) to estimate hydrogen jet flame radiation. A linear integral approach incorporating leakage rate and jet flame length is theoretically derived to establish dataset for machine learning. Then the Optuna tool is employed to optimize the initial weights and thresholds of the BP neural network. Input matrix of the Optuna-BPNN model includes pipeline diameter, leakage aperture size and hydrogen pressure. 8 sets of experimental data are employed to verify its correctness. When the abnormal data is excluded, the predicted thermal radiation of hydrogen jet fire agrees quite well with experimental results, with average and maximum deviations being 12.4% and 24.4% respectively. Using the linear integral approach, 32,670 thermal radiation data points are generated to train and test the Optuna-BPNN model. The maximum deviation between predicted and theoretical radiant heat flux for training and testing sets are only 4.5% and 6.2%, respectively. Parallel comparison trials using 6 different machine learning algorithms show that the Optuna-BPNN model gives the best mean absolute error, root mean square error and determination coefficient, which proves the effectiveness and feasibility of the developed Optuna-BPNN model in predicting thermal radiation of hydrogen pipeline jet fires.