A novel multivariable prognostic approach for PEMFC degradation and remaining useful life prediction using random forest and temporal convolutional network

Data-driven methods are effective in predicting future degradation trends (FDT) and remaining useful life (RUL) of proton exchange membrane fuel cells (PEMFCs). However, the complex and dynamic degradation behaviour of PEMFCs, influenced by diverse operational variables, poses significant challenges to existing prognostic approaches. This paper proposes a novel multivariable prognostic approach, termed RF-TCN, which combines random forest (RF) with temporal convolutional networks (TCN) to address these challenges. The approach incorporates three key innovations: (1) A hybrid RF and recursive feature elimination (RFE) method is employed to automatically select features most relevant to fuel cell degradation, reducing manual intervention and enhancing input robustness. (2) An improved TCN-based model is developed to effectively capture temporal degradation patterns, enabling accurate FDT and RUL predictions. (3) Particle swarm optimization (PSO) is utilized for automatic hyperparameter configuration, further boosting predictive performance. Empirical validation on ageing durability datasets demonstrates that the RF-TCN approach achieves superior prediction accuracy with selected optimal features and outperforms baseline TCN, CNN, RNN, and existing methods in the literature. This work advances prognostic methodologies, contributing to extending fuel cell lifespan and optimizing control strategies. • A novel multivariable data-driven approach is proposed for PEMFC prognostics. • The RF-RFE method is employed to automatically select optimal features. • The baseline TCN is enhanced to develop an improved predictive model. • PSO is utilized to automate hyperparameter configurations. • By leveraging the optimal features selected, the developed predictive model achieves greater accuracy and demonstrates superior performance.