Prognostic Modeling for PEM Fuel Cells: A Comparative Analysis Among Different Types of Approaches

This paper presents a comparative analysis among different types of approaches for prognostics of proton exchange membrane fuel cells (PEMFCs). The approaches are broadly categorized into three groups: 1) model-based; 2) data-driven; and 3) hybrid approaches. Each type of approach is discussed in detail. Model-based approaches, particularly semi-empirical models, utilize electrochemical principles to characterize PEMFC degradation through interpretable parameters associated with activation, ohmic, and concentration losses. These parameters are extracted from polarization data and modeled over time using curve-fitting techniques. While computationally efficient and physically grounded, model-based methods show limited adaptability under nonlinear or highly dynamic operating conditions. Data-driven methods such as long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), and support vector regression (SVR) offer high accuracy by learning degradation behavior directly from historical data. However, they typically require large training datasets, are sensitive to hyperparameter tuning, and lack physical interpretability. A hybrid prognostic approach is proposed by integrating a semi-empirical electrical equivalent circuit model with SVR. The hybrid model firstly predicts the voltage using the semi-empirical approach and then corrects residual errors using SVR. This structure combines the interpretability of physics-based models with the flexibility of machine learning. All approaches are evaluated on two publicly available aging datasets, covering both steady-state and quasi-dynamic conditions. Results demonstrate that the hybrid approach outperforms standalone models across varying training durations. The proposed hybrid approach offers a practical solution for real-time PEMFC health monitoring, providing a balanced trade-off between accuracy, computational efficiency, and interpretability.