Enhanced performance prediction for proton exchange membrane fuel cells: A comprehensive study with different load profiles

Abstract

In this study, the Online System Identification (OSI) method is applied to Proton Exchange Membrane Fuel Cells (PEMFC) using a publicly available experimental dataset. The dataset contains two distinct load profiles: one representing a ramp-up current profile to assess general algorithm performance, and another representing urban driving behavior to reflect real conditions in a vehicle. Current and temperature values serve as inputs to various deep learning models to predict the corresponding voltage outputs. Initially, the Long Short-Term Memory (LSTM) model is employed, showing promising results. The Root Mean Squared Error (RMSE) for the LSTM model is 0.3252 for the first load profile and 0.2566 for the second profile. The correlation coefficient (R) is 0.996 and 0.979, respectively, indicating strong prediction accuracy. Additionally, the study expands to include the transformer network model and a Bayesian optimized Bidirectional-LSTM (Bi-LSTM) model. The Transformer model achieves an RMSE of 0.3395 on the training set and 0.3551 on the test set. The Bayesian Optimized Bi-LSTM Model shows the lowest RMSE of 0.2540, demonstrating superior accuracy compared to the other models. These results highlight the effectiveness of machine learning techniques, particularly LSTM and Bi-LSTM models, in predicting and optimizing PEMFC performance under real-world conditions. The study concludes that these models provide valuable insights for enhancing the efficiency and reliability of PEMFC systems.