Symbolic deep learning based prognostics for dynamic operating proton exchange membrane fuel cells

Abstract

Fuel cell (FC) is a promising alternative energy source in a wide range of applications. Due to the unsatisfactory durability performance, FC has not yet been widely used. Prognostics and health management (PHM) has been demonstrated to be an effective solution to enhance the FC durability performance by predicting FC degradation characteristics and adopting health condition based control and maintenance. As the primary task of PHM, prognostics seeks to estimate the remaining useful life (RUL) of FC as early and accurately as possible. However, when FC faces dynamic operating conditions, its degradation characteristics are often hidden in the complex system dynamic behaviors, which makes prognostics challenging. To address this issue, a hybrid prognostics approach is proposed in this paper. Specifically, the health indicator of FC is extracted using a degradation behavior model and sliding-window model identification method. Subsequently, a symbolic-based long short-term memory networks (LSTM) is used to predict the health indicator degradation trend and estimate the RUL. The experimental and simulation results show that the proposed model is able to describe the dynamic behavior of the FC stack voltage and the extracted health indicator show a significant degradation trend. Moreover, health indicator prediction and RUL estimation performance can be improved by deploying the proposed symbolic-based LSTM prognostics model. The proposed approach provides a prognostic horizon approaching 50% of the FC life-cycle, and the average relative accuracy of estimated RUL is close to 90%.

A fusion prognostics strategy for fuel cells operating under dynamic conditions

“…In particular, methods in long short-term memory networks (LSTM) framework have demonstrated their strong performance in short-term SoH prediction [14,15,16,22,37]. However, LSTM performance becomes unsatisfactory with increased prediction horizon length which is observed and indicated in our previous work [15,23,36]. Interestingly, the cause of this issue may stem from the powerful "memory" ability of the LSTM, which incorrectly records irrelevant features in the training set.…”

“…To describe this process more clearly, a series of notation marks involved in the ABBA-LSTM prognostics model are shown in Table 1. The detailed operations in each step listed in Figure 3 could be found in our previous work [23].…”

A novel long short-term memory networks-based data-driven prognostic strategy for proton exchange membrane fuel cells

“…In recent years, three types of PEMFC prognostic strategies [6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][24][25][26][27][28][29][30][31][32][33] have been proposed: model-based, data-based, and hybrid methods based on model and data fusion. Among them, the model-based strategy describes the degradation processes by constructing a physical model [6][7][8], which is beneficial to the prediction accuracy in the condition that the physical degradation model is sufficiently accurate.…”

“…Among different DNN structures, long short-term memory (LSTM), a recurrent neural network (RNN) paradigm, has been considered as a potentially effective tool to handle time series prediction problems [14][15][16][17][18][19][20][24][25][26][27][28][29][30][31][32][33]. Several recent studies have been proposed to explore the LSTM application in fuel cell prognostics.…”

“…Meanwhile, the degradation trend can be effectively predicted only for a short horizon of a few hours [14][15][16][17][18][30][31][32][33]. To achieve RUL estimation, multi-step prediction with long prediction horizon is required, particularly for the prognostics of aging degradation [19,20,[24][25][26][27][28][29]. However, several issues arise as the prediction horizon is extended.…”