A fast fuel cell parametric identification approach based on machine learning inverse models. (15th January 2022)

• Record Type: Journal Article
• Authors: Guarino, Antonio; Trinchero, Riccardo; Canavero, Flavio; Spagnuolo, Giovanni
• Is Part Of: Energy. Volume 239:Part C(2022)
• Publication Date: 2022-01-15
• Subjects: Parametric identification of equivalent circuit model -- Inverse model -- Machine learning regression; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2021.122140
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: In this paper a computationally efficient optimization approach to the parametric identification of a fuel cell equivalent circuit model is presented. It is based on the inverse model and on machine learning regressions. During the training phase, the inverse model is built numerically by means of advanced regression approaches, i.e., the support vector machine regression, the least squares support vector machine regression and the Gaussian process regression. The training set is synthetically generated to the aim of exploring the parameter space and to characterize different stack operating conditions, including normal and faulty ones. The accuracy of the considered approaches is investigated by employing a test set including many experimental data, consisting of impedance spectra measured through the electrochemical impedance spectroscopy and referring to very different stack operating conditions. The results show that all the considered machine learning methods allow to identify the parameters of the fuel cell model with a low computational burden, so that they fit with the hardware resources of low cost embedded processors. This feature allows to envisage that the proposed approaches are good candidates for a model-based on-line diagnosis of fuel cell stacks. Graphical abstract: Image 1 Highlights: Parameter identification of equivalent circuit model performed through Inverse Model. Parameter search space is first constrained in feasibility regions. Inverse Models built through GPR, SVM and LS-SVM machine learning regression models. The identification results are shown for both synthetic and experimental data. Performance and Resource Usage of the embedded implementation are explored.