Uncertainty-quantified performance degradation prediction for Proton Exchange Membrane Fuel Cells under real-world driving scenarios

Predicting performance degradation trends provides essential information for developing energy management and maintenance strategies for automotive fuel cell systems. The systems’ performance is influenced by unstable external environments and changing operating conditions, making it challenging for conventional point estimation algorithms to achieve reliable prediction results. To address this issue, we proposed a probabilistic prediction algorithm to quantify the uncertainty of the automotive fuel cell’s performance degradation by adopting the diffusion model and Transformer. Firstly, we augmented the vanilla conditional diffusion model with a self-guiding sampling method to enhance the generation control of the fuel cell’s health indicator. Secondly, we integrated the Transformer architecture into the diffusion model’s denoising network to remove noise in the diffusion model. Additionally, we conducted a comparative analysis of the sampling effects of three guiding methods and compared them with advanced predictive model. The experiment results demonstrate that our guidance method and prediction model exhibit superior performance, with an average improvement of 35.10% in the Continuous Ranked Probability Score (CRPS) across two datasets compared to the advanced predictive model. Our predictive model, provides reliable probabilistic predictions on the automotive fuel cell performance degradation, offering a novel approach to applying the diffusion model in probabilistic prediction.