Degradation Prediction Method of Proton Exchange Membrane Fuel Cell

Durability is one of the main obstacles to the large-scale application of proton exchange membrane fuel cell (PEMFC). Performance degradation prediction technology can effectively improve the durability of PEMFC. Through the study of PEMFC aging data, it is found that the actual PEMFC aging data is highly nonlinear, periodic and random, which makes it difficult for the prediction algorithm to extract the features effectively. In addition, in the problem of degradation prediction, the prediction algorithm needs to predict the degradation of PEMFC under different working conditions, which requires the prediction algorithm to have stronger generalization ability. To solve the above problems, a performance degradation prediction method of regularization stack long short-term memory combined with wavelet threshold denoising method (WTD-RS-LSTM) method is proposed. Firstly, the WTD method is used to process the original data, and the smooth data after eliminating noise and spikes is obtained by wavelet decomposition, threshold processing and data reconstruction. Then the RS-LSTM model is introduced to solve the problem of feature extraction caused by uncertainty and high nonlinearity of degraded data. The generalization ability of the model is improved by introducing parameter optimization algorithm. The model is stacked to enhance its learning ability. For increase the reliability of the model, Warmup strategy was used to dynamically adjust the learning rate of the network. Through the above operations, the overfitting phenomenon which may occur in the training of the model is effectively avoided, and the prediction accuracy and reliability of the prediction algorithm are improved. For verify the effectiveness of the proposed method, PEMFC aging data under two different working conditions are used for verification. The datasets under different working conditions are divided into five different lengths of training sets and test sets to train and test the proposed algorithm. The verification results show that under steady-state conditions, the maximum error of the proposed method is 0.016 3 V, and the error interval is within 0.5%. The prediction performance increases with the training length, and the best prediction performance is obtained at the training length of 1 000 h, when the RMSE and MAPE are 0.000 91 and 0.000 22, respectively. Under dynamic conditions, the maximum error is 0.006 4 V and the error interval is within 0.2%. The best performance was achieved when the training length was 550 h, when the RMSE and MAPE are 0.000 75 and 0.000 20, respectively. According to the above experimental results and the comparison with the existing traditional algorithms, the following conclusions are drawn: (1) the proposed method can make more accurate PEMFC degradation prediction under different working conditions and different training lengths, and has stronger generalization ability; (2) Comparing the prediction accuracy of the two conditions under different training lengths, it is found that the prediction of PEMFC degradation under dynamic conditions by the proposed method is better than that under steady-state conditions. Therefore, the proposed method has stronger prediction ability under dynamic conditions. (3) The proposed method has a simple structure, easy to deploy and is suitable for online application; (4) The aging of PEMFC under dynamic conditions will produce more randomness, which will have a great impact on the stability of the prediction algorithm. © 2024 China Machine Press. All rights reserved.