A novel method for accurate and enhanced prediction of proton exchange membrane fuel cell performance based on graph neural networks

Abstract

This study introduces a novel prediction method that utilizes graph neural networks (GNN) to enhance the prediction accuracy and thereby improve the durability of PEM fuel cell. This method leverages the ability of GNN to extract features from graph-structured data, effectively capturing the implicit variables and time dependencies in the operating parameters. The primary findings indicate that when the number of graph convolution layers, the number of LSTM layers, the batch size, and the learning rate are 3, 1, 32, and 0.000272 respectively, the mean square error after 18 iterations of the model is 4.8957 × 10⁻⁶. At this point, the performance of the GNN model is the best. This method is compared with five other classic benchmark models. The root mean square error on the training set is 2.213 × 10⁻³, and the error on the test set is 5.798 × 10⁻³. Compared with traditional methods, the error is reduced by 74.12% and 62.63% respectively. Its accuracy and fit are also superior to other models. By comparing the actual values of the output voltage of PEM fuel cell under dynamic load conditions with the output voltage predictions based on the GNN model, the GNN model is able to accurately fit the output voltage of the fuel cell under dynamic load conditions over a time span of 1008 h. By analyzing the average weight distribution of the graph attention layer, the GNN not only can capture the instantaneous correlation through short-term time information, but also extract long-term time information to predict the turning point of irreversible damage in PEM fuel cell, ultimately achieving highly accurate prediction results.