Explainable Multi-Fidelity Machine Learning Model for Ventless Hold Time Prediction in Vehicle Liquid Hydrogen Cylinder

Condition prediction for liquid hydrogen containers is crucial to their safety and efficient utilization. An explainable multi-fidelity machine learning model was established to predict the ventless hold time of 500 L vehicle liquid hydrogen cylinder, and the effects of machine learning algorithms and the proportion of high-fidelity data on model performance were analyzed. The results showed that ensemble learning algorithms exhibited superior accuracy and lower computational overhead compared to the multilayer perceptron. In terms of prediction accuracy, the random forest algorithm achieved the lowest root mean square and mean absolute error on the test set, with values of 4.04 and 2.10. The enhanced low-fidelity data strategy effectively reduced the discrepancies between data from different fidelity levels, while the light gradient boosting machine was the optimal choice to conduct multi-fidelity modeling. The accuracy of the enhanced multi-fidelity model greatly outperformed the multi-fidelity model without enhancement. As the proportion of high-fidelity data in the training set increased from 30% to 35%, the root mean square error of the enhanced multi-fidelity model decreased by 4.7%. In addition, an explainable analysis revealed that pressure increment and heat leakage were dominant promoters affecting ventless hold time. This multi-fidelity modeling study provided a method for predicting ventless hold time that combines accuracy with explainability, and offered insights into the state prediction of liquid hydrogen containers.