Modeling and prediction of solid oxide cell short-stack performance using an hybrid DOE–ANN modeling approach

This work presents a comprehensive modeling and validation study of a commercially available solid oxide cell short stack operating in both fuel cell and electrolysis modes, with the aim to predict the mean cell voltage, obtained by dividing the total stack voltage by the number of cells voltage output, and measured stack temperature through a combined statistical and machine-learning approach. A Design of Experiments framework guided 113 experimental tests under varied current densities, gas flow rates, and thermal conditions, enabling the development of regression-based predictive models using stepwise selection. To better capture nonlinear behaviors and improve accuracy, a three-node artificial neural network was trained on the same dataset. Both models were evaluated using adjusted R-squared, RMSE, MAE, and RSS metrics, along with residual and parity plots. For example, the ANN reduced the temperature RMSE from 32.6 °C (Design of Experiments) to 19.6 °C (training) and 31.1 °C (validation). Results confirm that integrating Design of Experiments and ANN modeling yields accurate and interpretable predictions, offering a robust methodology to optimize SOC stack performance and support the practical application of this reversible electrochemical technology for hydrogen production, energy storage, and renewable energy integration in decarbonization strategies.