A New Supervised Learning Approach for Predicting the Performance of Proton Exchange Membrane Fuel Cells

ABSTRACT The efficiency and performance of proton exchange membrane fuel cells (PEMFCs) are possible by operating in the maximum power point region. This is made possible by accurately measuring the output voltage. For this purpose, a hybrid machine learning (ML) model is proposed in this study to predict the output voltages of PEMFCs with minimum error. In the proposed model, optimal data augmentation was performed for the limited experimental data using polynomial fitting. Then, all models were trained in the ML‐based regression learner, and the models with the lowest error values were identified. This process was applied to different data sets where data augmentation was performed. As a result, the root mean square error (RMSE) value and training time were evaluated. The rational quadratic Gaussian process regression (GPR) model provided the best result with an RMSE of 0.01277 and a training time of 3.0288, followed by the Matern 5/2 GPR model with an RMSE of 0.01307 and a training time of 3.7758. Later, these two models were proposed as a hybrid model; the proposed model was tested with external test data of real current information and the results were compared. Then, a comparison of the proposed sequential stacked hybrid model with other regression learner methods and the literature is given. In this way, the total sum of the absolute voltage differences was calculated as 0.102, the RMSE value of the voltage as 0.0093, and the RMSE value of the power as 0.1064. These performance figures reflect the specific experimental conditions and the employed polynomial augmentation procedure; therefore, conclusions regarding universal generalization should be drawn cautiously until validated on larger, multisource datasets.