Performance analysis of optimized machine learning models for hydrogen leakage and dispersion prediction via genetic algorithms

Hydrogen, a leading alternative to petroleum fuels, poses significant risks, necessitating accurate leakage and dispersion predictions. Numerous Machine Learning (ML) models have been developed to address this challenge; however, these models often exhibit two major limitations: they are generally tailored to specific leakage scenarios and their hyperparameters are selected empirically. This research aims to evaluate various ML models optimized with Genetic Algorithms (GA) for predicting hydrogen dispersion in typical leakage scenarios. Using a dataset of 6561 scenarios generated by PHAST, considering both source and dispersion properties, the study finetuned each model's hyperparameters with GA. k-fold cross validation was used to verify the optimized ML models' reproducibility, while statistical metrics such as R2 assessed the models' performance. Ultimately, the GA-DNN model is determined to be the best appropriate for hydrogen dispersion prediction. This methodology offers a comprehensive framework for developing hydrogen dispersion prediction models, encompassing data selection, model design, and execution.