A Long-Term Prediction of Fuel Cell Performance Degradation Based on Deep Reinforcement Learning

In this paper, a hybrid model based on deep reinforcement learning (DRL) is proposed for predicting the degradation process of the fuel cell stack. The model integrates the interpretability of mechanism models with the strengths of data-driven approaches in capturing nonlinear dynamics. Voltage is selected as an indicator for predicting the performance degradation of the stack. By utilizing DRL, a dynamic weighting process is achieved, enhancing both the accuracy and robustness of the model. The model is validated by the IEEE 2014 dataset. The results show that the hybrid model achieves high accuracy with the R² value of 0.875 (30% of the data used as a training set). Moreover, when the training set is 7:3 compared to the test set, the accuracy of the hybrid model is 14.18% higher than that of the long short-term memory network (LSTM) model. The DRL model has the highest accuracy for different percentages of the training set in the total data set, which further verifies the universality of the hybrid model. In addition, feature selection using the SHapley Additive exPlanations (SHAP) method reduces the number of input features, reducing the number of input data types from 19 to 7. The most influential factor is voltage, followed by time, which is consistent with the laws of the mechanism model. The dependence of the model on large data sets is minimized without compromising accuracy. The DRL model demonstrates strong potential as a reliable tool for fuel cell degradation prediction, particularly in long-term forecasting applications that demand high accuracy and reduced cumulative error.