Data driven hybrid feature selection and GPR based modeling for identifying critical fault variables in PEMFCs

Abstract

This study identifies the most influential variables governing voltage prediction in Polymer Electrolyte Membrane Fuel Cells (PEMFCs) and quantifies their deviations under flooding related fault conditions. An experimental dataset of 33,899 records and 22 variables from an 80 W PEMFC system is analyzed. A hybrid feature selection framework that integrates correlation analysis, Least Absolute Shrinkage and Selection Operator regression, Random Forest and Gradient Boosting determines the dominant variables affecting cell voltage. 28 machine learning regression algorithms are compared, and the Exponential Gaussian Process Regression model achieves the best prediction performance, with a RMSE of 0.00362 and a coefficient of determination of 0.99949. Critical deviation analysis then determines the positive and negative deviation ranges of the key variables that lead to a +/- 5 % voltage change, defined as the fault threshold. In this study, a fault is interpreted as a performance deviation from the nominal operating voltage rather than a specific degradation mechanism alone. The results show that heater temperature, stack thermal profile and hydrogen inlet temperature are the most sensitive parameters.