Development of the Quantitative Property–Consequence Relationship Model for Prediction of Hydrogen Leakage and Dispersion Using Response Surface Method and Artificial Neural Network Approaches

Abstract

Hydrogen, an environmentally friendly and highly regarded future energy source, can form flammable vapor clouds upon leakage, which may transition into explosion. Predicting the dispersion behavior of hydrogen is crucial for preventing such incidents. This study aims to develop a quantitative property–consequence relationship (QPCR) model using the response surface method (RSM) and artificial neural network (ANN) to swiftly and accurately predict dispersion behavior. Initially, 8 variables were defined from source and dispersion models, constructing a data set through 6,561 PHAST simulations. Subsequently, the RSM-BBD (Box-Behnken design) and ANN-BPNN (Backpropagation neural network) models were developed, alongside a hybrid model incorporating BPNN after excluding four low-influence variables based on analysis of variance (ANOVA). All models achieved an R² value exceeding 0.99. The hybrid model notably reduced computational costs by 97% compared to ANN-BPNN and exhibited lower mean square error (MSE). These results introduce a cost-effective approach for high-accuracy QPCR modeling and highlight the viability of diverse statistical methods.