Predictive Hybrid Model for Enhanced Remaining Useful Life Estimation of PEM Fuel Cells

Fuel cell systems are emerging as a strong alternative to traditional power sources due to their high energy efficiency, greater energy density, and zero emissions. However, their relatively short lifespan has limited their widespread commercial use. Predictive methods for estimating the state of health (SOH) and remaining useful life (RUL) have been effective in extending fuel cell life. Traditional diagnostic methods, such as electrochemical impedance spectroscopy, require system shutdowns, disrupting operations. To solve this problem, a nonintrusive hybrid machine learning model combining extreme gradient boosting (XGB) and random forest (RF) is proposed. This hybrid method addresses two main issues: improving prediction accuracy and making the model more reliable. The novelty lies in combining the strengths of both algorithms, achieving better prediction accuracy and reducing errors compared to single-model methods. It provides continuous and scalable SOH and RUL predictions, helping improve fuel cell durability without interrupting operations. The model is tested on a fuel cell aging dataset, which includes data from static and dynamic conditions, and is validated by yielding root mean squared errors (RMSEs) of 0.00026 and 0.00127 for SOH and RUL predictions, respectively. Compared to standalone XGB and RF models, the hybrid approach demonstrates superior prognostic performance.