Performance prediction of proton-exchange membrane fuel cell based on convolutional neural network and random forest feature selection. (1st September 2021)

• Record Type: Journal Article
• Authors: Huo, Weiwei; Li, Weier; Zhang, Zehui; Sun, Chao; Zhou, Feikun; Gong, Guoqing
• Is Part Of: Energy conversion and management. Volume 243(2021)
• Publication Date: 2021-09-01
• Subjects: Performance prediction -- Fuel cell -- Deep learning -- Random forest; Direct energy conversion -- Periodicals; Energy storage -- Periodicals; Energy transfer -- Periodicals; Énergie -- Conversion directe -- Périodiques; Direct energy conversion; Periodicals; 621.3105
• DOI: 10.1016/j.enconman.2021.114367
• Languages: English
• ISSNs: 0196-8904

Abstract

Highlights: Random forest algorithm is used to select important factors. Performance prediction method of the FC employing convolutional neural networks (CNN) Dropout layer and batch normalization are utilized to avoid model overfitting and improve model generalization. Abstract: For optimizing the performance of the proton exchange membrane fuel cells (PEMFCs), the I–V polarization curve is generally used as an important evaluation metric, which can represent many important properties of PEMFCs such as current density, specific power, etc. However, a vast number of experiments for achieving I-V polarization curves are conducted, which consumes a lot of resources, since the membrane electrode assembly (MEA) in PEMFCs involves complex electrochemical, thermodynamic, and hydrodynamic processes. To solve the issues, this paper utilizes deep learning (DL) to design a performance prediction method based on the random forest algorithm (RF) and convolutional neural networks (CNN), which can reduce unnecessary experiments for MEA development. In the proposed method, to improve the high quality of the training dataset, the RF algorithm is adopted to select the important factors as the input feature of the model, and the selected factors are further verified by the previous studies. CNN is used to construct the performance prediction model which outputs the I-V polarization curve. In particular, batch normalization and dropout methods are applied to enhance model generalization. The effectiveness of the CNN-based prediction model is evaluated on the real I-V polarization curve dataset. Experiment results indicate that the prediction curves of the proposed model have good agreement with the real curves. Our study demonstrates the deep learning technologies are powerful complements for optimizing the PEMFCs.