(143b) CFD Simulation and Machine Learning Surrogate Modeling for Consequence and Risk Assessment of Leakage, Dispersion, Fire, and Explosion of FCEV Hydrogen Storage Tanks

The use of hydrogen fuel is increasing as a sustainable and eco-friendly energy source in modern society. Accordingly, the importance of the safety of hydrogen fuel in various usage environments is also increasing. In order to assess the risk of hydrogen fuel, this study proposes FLACS CFD simulation and machine learning surrogate model-based prediction to predict quantitative consequence from gas leaks, diffusion, fire, and explosion scenarios that may occur in hydrogen storage tanks of FCEVs. The pros and cons of the two proposed methods are compared and analyzed.

CFD simulations simulate the behavior of gas leaks and dispersion, especially in closed environments where overpressure effects can be amplified, such as underground parking lots and tunnels, and quantitatively analyze the risk of various leak locations and directions. CFD simulation results for 27 scenarios are obtained by varying the environmental conditions that affect the vapor cloud formation and explosion overpressure of leaked hydrogen: vehicle location, leak direction, hydrogen storage tank capacity and pressure, and number of obstacles.

However, as various accident environments and hydrogen storage tank leak structural conditions change, a surrogate modeling is proposed as a more convenient and simple method to replace the detailed, large-capacity CFD simulation that is separately required. Using the environmental conditions of CFD simulation, such as vehicle location, leakage direction, hydrogen storage tank capacity and pressure, and number of obstacles, as input, the developed prediction model predicts the maximum explosion overpressure for each coordinate as output. This surrogate model is based on XGBoost, a decision tree-based machine learning method, and implemented using the ‘xgboost’ module (v. 2.0.0) in Python. Once a predictive machine learning model is designed and trained, and the model development is completed, the impact of various conditions and variables on leakage, fire and explosion results can be analyzed in about a few seconds (on NVIDIA GeForce RTX 3090 D6X 24GB).

Existing CFD simulations can accurately calculate the explosion overpressure for the conditions, but take approximately 2 hours. In contrast, the surrogate prediction model learns and predicts simulation results, and has an RMSE of about 0.33, which shows high accuracy, but has the disadvantage of being slightly less accurate than CFD simulation, and has the advantage of being able to analyze within a very short time. In addition, the surrogate model can quickly predict explosion overpressure in various environmental conditions and situations, which is useful for optimizing design structures with optimal risk while changing various conditions. In addition, it is confirmed that the dispersion path and explosion possibility of hydrogen gas are accurately predicted by cross-validating the model's prediction performance using the Leave-One-Out Cross-Validation verification method. This analysis is expected to provide practical guidelines necessary for accident prevention and safety management of hydrogen vehicles and related facilities, and to contribute to improving the safety of hydrogen fuel use.