Artificial intelligence for time-efficient prediction and optimization of solid oxide fuel cell performances. (15th February 2021)

• Record Type: Journal Article
• Authors: Subotić, Vanja; Eibl, Michael; Hochenauer, Christoph
• Is Part Of: Energy conversion and management. Volume 230(2021)
• Publication Date: 2021-02-15
• Subjects: Artificial neural networks (ANN) -- Solid oxide fuel cells (SOFC) -- Performance optimization -- Performance prediction -- Model validation; 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.2020.113764
• Languages: English
• ISSNs: 0196-8904

Abstract

Highlights: Time-efficient prediction of SOFC performance employing artificial neural networks (ANN). Model training and validation using data from both very detailed multi-physics model and experimental data. Optimal parameters for time-efficient SOFC performance optimization and prediction employing ANN identified. Simulation of electrochemical impedance data is possible using ANN. Very good agreement between the computed and measured data. Abstract: Reliability and durability are main issues that must be addressed in order to accelerate the commercialization of solid oxide fuel cells (SOFCs). Time-efficient and exact prediction of system performance as a function of an operating environment could reduce the time required to find the operating optimum within a wide range of parameters. For this purpose, a prognostic framework based on artificial neural network (ANN) is designed within this study to predict SOFC performance presented by polarization curves and electrochemical impedance spectra. In order to train and validate the ANN developed two approaches are followed to generate the data sets required: very detailed multi-physic model and experimental data. Very good agreement between the ANN model and the measured data is observed, with an exception for very low current densities lower than 20 mA cm −2 . The polarization model with 1–3 hidden layers and 3–5 neurons as well as a patience parameter 5–20 resulted in a very good accuracy. Increasing the system complexity, e.g. required prediction of the overall cell impedance as a function of the operating temperature, the system complexity increased thus increasing the number of neurons per hidden layer up to 10–30 and a patience of up to 500–1000 epochs.