LAUSR

Abstract This paper represents a systematic approach to develop a virtual hydrogen sensor for predicting both the incidence and the extent of hydrogen starvation in high-temperature solid oxide fuel cells (SOFCs). Fuel starvation would occur in a fuel cell when the fuel is consumed at a faster rate than is fed to the cell. A previously developed and validated pseudo-2D numerical model was used for finding the hydrogen distribution along the channels of a single cell of a high-temperature SOFC. Nine different input parameters of the model were changed and a dataset of nearly half a million operating points was generated using the fast and accurate pseudo-2D model. The dataset was randomly divided into the training set (70%) and the test set (30%). Four different binary classifiers including K-Nearest Neighbors (KNN), Artificial Neural Network (ANN), Naive Bayes, and Logistic Regression were employed to determine if the cell operates in normal (0) or starved (1) at a given set of input parameters. It was found that KNN and ANN outperform the other methods with an F1-Score above 0.97 if the parameters are appropriately tuned. Another ANN was then trained using the starved data to estimate the percentage of the cell being starved with an accuracy of 97.5% based on the mean absolute error (MAE). Therefore, the proposed set of classifier-regressor can be successfully employed as a virtual hydrogen sensor for online tracking of hydrogen concentration along the cell.