Performance prediction of fuel cells using long short‐term memory recurrent neural network

Abstract

Performance prediction of proton-exchange membrane fuel cell (PEMFC) under dynamic conditions, especially for vehicle applications, is increasingly become the focus of attention. This article proposes a performance prediction method of PEMFC using long short-term memory (LSTM) recurrent neural network (RNN). In this article, polarization curve (current-voltage curve) and voltage degradation curve (current-time curve) are adopted as the main performance indexes of PEMFC. Both polarization curve prediction and performance degradation prediction of PEMFC can be effectively implemented based on the LSTM method. To investigate the voltage losses law of experimental and predicted results, the paper introduces an empirical equation of polarization curve. The perfect match between the experimental and predicted polarization losses of PEMFC can further validate the prediction performance of LSTM method. The proposed prediction method is validated by the PEMFC polarization curve data obtained from the designed aging experiment of a 4 kW stack operated under dynamic loading cycling situation during 600 hours. Then, LSTM network is compared with traditional RNN and back-propagation neural network (BPNN) to prove its superiority. The minimum values of rootmean-square error (RMSE) and the mean absolute percentage error (MAPE) of LSTM network with different training data are 0.0088 and 0.0101, respectively. All the coefficient of determination (R 2 ) of LSTM model with different training data is over 0.95, which is close to 1.0. The prediction accuracy of LSTM network is higher than that of two other networks. The result indicates that LSTM network outperforms two other networks in PEMFC performance prediction.Hence, the prediction method based on LSTM network is very suitable for PEMFC performance prediction.

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

“…LSTM can selectively store the intrinsic information of long-term data and capture the long time-scale correlation between time series data. In fact, as in [14][15][16][17][18][19][24][25][26][27][28][29][30][31][32][33], the LSTM-based prognostic framework is demonstrated to be capable of predicting the short-term PEMFC degradation behavior. Specifically, the basic network configuration of the LSTM contains three layers, namely the input layer, the single LSTM layer, and the output layer, as in Fig.…”

“…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.…”

A Critical Review on Artificial Intelligence for Fuel Cell Diagnosis

“…The types of hidden layers employed differentiate deep networks. Convolutional neural networks (CNNs) have been used to forecast stack voltages and polarization curves in PEM-FCs [86], while recurrent neural networks (RNNs) have been used to anticipate functionality decline [87]. Other uses include flaw categorization in FC and water management systems and surrogates for mesoscale simulations [88].…”