Development of an optimized proton exchange membrane fuel cell model based on the artificial neural network

Numerical studies have been considered as a vital method to optimize the system design and the control strategy of proton exchange membrane (PEM) fuel cells practically. Given that the engineering application of multi-dimensional physics-based simulations is very challenging in terms of efficiency, this presents a unique opportunity for modeling approaches based on the artificial neural network (ANN). As a supplement to traditional statistical methods, the ANN technique demonstrates advantages in dealing with arbitrary nonlinear relations between the independent and dependent variables. In the present study, an optimized model using a feed-forward back-propagation (BP) network has been developed. By integrating with the genetic algorithm, the risk of overfitting could be reduced. The automatic process of searching for the most suitable network structure algorithm has also been adopted. Moreover, to figure out appropriate input variables, a feature dimension reduction methodology has been implemented in the proposed input variable determination (IVD) sub-model during the pre-processing procedure. The data points required for training, validating, and testing are obtained from comprehensive sensitivity tests. The active area of the membrane electrode assembly (MEA) in the present experiment is around 220 cm² which is the same order of magnitude as commercial products. The optimized model has been thoroughly validated against experimental measurements, results show that simulations could accurately reproduce the effect of multiple operating parameters on the fuel cell performance. This new model is applicable to both interpolation and extrapolation. Furthermore, by activating the IVD sub-model, the maximum and average relative errors of extrapolation simulation results could be reduced up to 63 % and 37 %, respectively. In addition, by reasonably selecting the input variables in the order of priority, the mean relative error remains under 1 % with fewer input variables. The number of required training data points could be reduced up to 53 %.