Data-driven prognostic methods aim at estimating the remaining useful life (RUL) of the proton exchange membrane fuel cell (PEMFC) system and contribute to its lifespan extension and reliability improvement. Lack of dynamic health indicator (HI) and low prediction accuracy is the weaknesses for current RUL prediction methods, especially when the PEMFC system works at the variable load profiles. To overcome the drawbacks of static HIs (e.g., voltage and power), a novel dynamic HI of relative power-loss rate (RPLR) is put forward to represent the degradation state under the variable load profiles. Besides, the discrete wavelet transform and ensemble echo state network (DWT-EESN) approach is proposed to deal with the multi-timescale features of RPLR and enhance the prediction performance. First, the decomposition method of discrete wavelet transform (DWT) is applied to decompose the RPLR into several sub-waveforms. Second, several independent echo state networks (ESNs) are utilized to predict the sub-waveforms separately. Finally, the prediction results in different timescales are integrated to achieve the overall RUL prediction. Three micro-combined heat and power ( $\mu $ -CHP) experiments with different dynamic profiles are carried out, and results show that the DWT-EESN has a higher long-term prediction accuracy than the traditional ESN structure.
18 October 2021 in IEEE Transactions on Transportation Electrification
Limited durability is one of the major issues that hinder the large-scale commercialization of the proton exchange membrane fuel cells system. Based on the prognostic technique, predicting the remaining useful life (RUL) efficiently and accurately can help prolong its residual life, especially on the long-term horizon and under different mission profiles. Thus, a data-driven approach of discrete wavelet transform-echo state network-genetic algorithm (DWT-ESN-GA) is proposed to improve the RUL prediction performance. First, the historical datasets are compressed by the DWT. Second, the approximation components of the original data are predicted in the compressed space by ESN. Rather than predicting the degradation data themselves, their shortened coefficients are evaluated to decrease the prediction data points, i.e., from 2016 data points to 253 data points. Besides, a GA is used to optimize the key parameters of ESN, and it can further increase the prediction accuracy. Finally, the inverse DWT is utilized to reconstruct the coming data based on the estimated approximation components. The performance of the proposed approach is evaluated by three different experimental tests under steady-state, quasi-dynamic, and full dynamic operating conditions separately.
The dataset proposed two long term tests to evaluate the impact of the current ripples of the fuel cell durability. To be able to compare both degradations, the first ageing test is performed under steady-state conditions at constant current. This time series could be used as the durability reference. While the second long-term test is dedicated to degradation under rippled current with high frequency triangular perturbations, similar to the load imposed by a power converter. Thus, the second time series considers the actual degradation that could appear on an embedded system. Both tests are running on two fresh proton exchange membrane fuel cell (PEMFC) stacks composed of 5-cells during around 1000 h, to make degradation comparable. An in-lab test bench is used to monitor all the fluidic, thermal and electric parameters thanks to a National Instruments acquisition device and an in-lab designed Labview software. Some polarization curves and electrochemical impedance spectra (EIS) are performed periodically during both durability tests. Both datasets provide to study the performance degradation along time for steady-state and rippled solicitations. These datasets can be directly used to validate some prognostics approaches and estimate the remaining useful lifetime (RUL).