Forecast of Aging of PEMFCs Based on CEEMD-VMD and Triple Echo State Network

Accurately forecasting the degradation trajectory of proton exchange membrane fuel cells (PEMFCs) across a spectrum of operational scenarios is indispensable for effective maintenance scheduling and robust health surveillance. However, this task is highly intricate due to the fluctuating nature of dynamic operating conditions and the limitations inherent in short-term forecasting techniques, which collectively pose significant challenges to achieving reliable predictions. To enhance the accuracy of PEMFC degradation forecasting, this research proposes an integrated approach that combines the complete ensemble empirical mode decomposition with the variational mode decomposition (CEEMD-VMD) and triple echo state network (TriESN) to predict the deterioration process precisely. Decomposition can filter out high-frequency noise and retain low-frequency degradation information effectively. Among data-driven methods, the echo state network (ESN) is capable of estimating the degradation performance of PEMFCs. To tackle the problem of low prediction accuracy, this study proposes a novel TriESN that builds upon the classical ESN. The proposed enhancement method seeks to refine the ESN architecture by reducing the impact of surrounding neurons and sub-reservoirs on active neurons, thus realizing partial decoupling of the ESN. On this basis of decoupling, the method takes into account the multi-timescale aging characteristics of PEMFCs to achieve precise prediction of remaining useful life. Overall, combining CEEMD-VMD with the TriESN strengthens feature depiction, fosters sparsity, diminishes the likelihood of overfitting, and augments the network's capacity for generalization. It has been shown that the TriESN markedly improved the accuracy of long-term PEMFC degradation predictions in three different dynamic contexts.