High accuracy and adaptability of PEMFC degradation interval prediction with Informer-GPR under dynamic conditions

Abstract

Proton exchange membrane fuel cells (PEMFCs) are pivotal components within green energy systems; However, commercialization and large-scale application of these techniques are constrained by the performance degradation prediction problem. Existing prediction methods mainly focus on the performance degradation under static and quasi-dynamic conditions, yet point estimation uncertainty quantification and interval estimation under dynamic conditions would contribute to safe and efficient operation. In this research, a fusion method named Informer-GPR is proposed, which accurately quantifies the uncertainty of performance degradation prediction and demonstrates good adaptability under various dynamic conditions and prediction step sizes. This method utilizes Informer as a point estimation method to circumvent the deficiency of global information extraction in recurrent neural networks and employs sparse self-attention mechanisms to allocate weights, enhancing the extraction of crucial information in the aging process. Furthermore, gaussian process regression (GPR) is employed to quantify the uncertainty of point estimation process, providing safer confidence interval estimation. Experimental findings demonstrate that the Informer-GPR method reduces the RMSE of point estimation by 20.8 %–64.4 % on dynamic cycle condition datasets, offering precise confidence interval estimations. Moreover, its robust performance across diverse dynamic conditions and multiple prediction steps underscores its versatility, thereby enhancing prediction efficacy in dynamic scenarios.