Performance degradation decomposition-ensemble prediction of PEMFC using CEEMDAN and dual data-driven model

Proton exchange membrane fuel cells (PEMFCs) are essential modern sustainable energy generation devices. Since such an electrochemical system has a limited lifetime, accurately estimating its performance degradation is critical for practical applications. When a large amount of measurement data is available, many nonlinear forecasting methods can be used to predict the performance degradation of a PEMFC system, and the prediction accuracy can be improved by optimizing the structure and parameters of the algorithm. However, the voltage recovery phenomenon would pose a challenge to the classical data-driven methods. In this work, we propose a novel hybrid data-driven PEMFC performance prediction framework by exploring the extensive degradation information buried in the voltage decay data. With complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), the raw voltage data are first decomposed into sequences of multiple time scales. Then, the linear and nonlinear components in the decomposed sequences are predicted by autoregressive integrated moving average (ARIMA) and the attention-based gated recurrent unit (GRU), respectively. Comparative studies show that the proposed method can improve the prediction performance by 42.6%-84.2% on FC1 and 35.0%-90.6% on FC2, compared to state-of-the-art algorithms on the basis of an open-source dataset of PEMFCs.

Automated summary

In this work, a novel hybrid data-driven PEMFC performance prediction framework is proposed, which decomposes the raw voltage data into sequences of multiple time scales using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and predicts the linear and nonlinear components in the decomposed sequences by autoregressive integrated moving average (ARIMA) and attention-based gated recurrent unit (GRU), respectively. Comparative studies on an open-source dataset of PEMFCs show that the proposed method can improve the prediction performance by 42.6%-84.2% on FC1 and 35.0%-90.6% on FC2, compared to state-of-the-art algorithms.