Ricky Jay Gomez, Dahlia Apodaca, Michelle Almendrala

Abstract: This work focused on developing a Physics-informed ML model using Gaussian Process Regression (GPR) in predicting the open-circuit voltage (OCV) against temporal hydrogen crossover (HCO) current by employing the Shapley value analysis to explain the model predictions. First, the GPR-based model developed was seen to exceptionally perform well from model training to deployment based on the fit results (RMSE = 6.78E-05, R2 = 1.0000), correlation analysis (R = 1.0000), and statistical validation (p-value = 0.3216). The uncertainty range, as part of the results of a GPR-based model, suggests that the probability that the model predictions represent the actual OCV of the unseen data is high. Second, the global model interpretation suggested that both the HCO and time have strong influence on the OCV values although a positive impact was observed based on the direction of influence given by the Shapley summary which subjects the data used to test the Shapley algorithm to ambiguity. Anyhow, this finding was eventually contrasted as the Shapley dependence implied that majority of the Shapley values were observed under the zero-value Shapley, indicating that both predictors negatively impacted the OCV. Lastly, the local Shapley inspection suggested that predictors have weak influence over the OCV at around 40,000 to 60,000 hours where great decline in the OCV values were recorded. HCO greatly dominated the OCV decline at the near end of the AST program.

Keywords: Gaussian process regression, Shapley, PEMFC

Date Published: December 4, 2024
DOI: 10.11159/jffhmt.2024.041

View Article