Degradation Prediction of PEMFCs Based on Discrete Wavelet Transform and Decoupled Echo State Network

Predicting the degradation process of proton exchange membrane fuel cells (PEMFCs) under diverse operational conditions is crucial for their maintenance planning and health monitoring, but it is also quite complex. The variability in dynamic conditions and the shortcomings of short-term forecasting methods make accurate predictions difficult in practice. To strengthen the precision of deterioration predictive methods, this study introduces a degradation prediction of PEMFCs incorporating discrete wavelet transform (DWT) and a decoupled echo state network (DESN). The high-frequency noise is shielded by wavelet decomposition. Within data-driven approaches, an echo state network (ESN) can estimate the decline in PEMFC performance. To address the issue of low forecasting precision, this paper introduces a novel DESN with a lateral inhibition based on the decreasing inhibition (DESN-Z) mechanism. This enhancement aims to refine the ESN structure by mitigating the impact of other neurons and sub-reservoirs on the currently active ones, achieving initial decoupling. The lateral inhibition mechanism expedites the network's acquisition of pertinent information and refines predictions by intensifying the rivalry among active neurons while suppressing others, thereby diminishing neuron interconnectivity and curbing redundant internal state data. Overall, combining DWT with DESN-Z (DDESN-Z) bolsters feature representation, promotes sparsity, mitigates overfitting risks, and enhances the network's generalization capabilities. It has been demonstrated that DDESN-Z significantly elevates the precision of long-term PEMFC degradation predictions across static, quasi-dynamic, and fully dynamic scenarios.