Lifetime Prediction of Direct Methanol Fuel Cells Based on Hybrid Predictive Modeling

Direct methanol fuel cells (DMFCs) face durability challenges that limit their widespread adoption. Traditional data-driven degradation models excel in short-term predictions but lack capability for long-term forecasting. To mitigate disturbances caused by the voltage rebound phenomenon in fuel cells, this paper proposes a hybrid prediction model that combines mechanistic modeling and machine learning. The core idea is to use mechanistic modeling to extract degradation-relevant internal impedance features from electrochemical impedance spectroscopy (EIS), and then leverage machine learning to predict the evolution of these features, thereby avoiding direct fitting of noisy voltage signals. Specifically, the model first fits EIS using a full-circuit equivalent circuit model (ECM) to extract key internal degradation parameters. It then combines these parameters with a convolutional neural network (CNN) to predict impedance parameter trends. Finally, by establishing a linear relationship between the predicted impedance and the output voltage, the remaining useful life (RUL) of the methanol fuel cell is reliably estimated. Comparison results between the hybrid prediction model, Gated Recurrent Unit (GRU) network, and iTransformer network show predicted failure threshold time errors of 1 hour, 11 hours, and 5 hours, respectively. These results demonstrate that the proposed hybrid model achieves the smallest time prediction error (1 hour), indicating superior performance for long-term prediction tasks. The hybrid model improves data tracking ability and prediction accuracy, and reduces dependence on large amounts of degradation data. This model provides a feasible solution for DMFC lifetime prediction.