Gradient boosting algorithm for current-voltage prediction of fuel cells

Abstract

The developed catalyst performs well in a half-cell test, but since the catalyst performance is not completely shown in a single-cell test, single-cell optimization through repeated experiments is essential. This is because single-cell is a complex process system affected by various factors, not just a catalyst evaluation system. To reduce the number of laborious tests, we used a machine learning algorithm to prioritize an alkaline fuel cell's operational factor from types of catalyst to fuel concentration for developing an overall performance prediction model. We selected seventeen input features from more than 80 I-V curves and 8000 data points and established prediction models based on two error-scoring modes (mean absolute error and root mean square error), which focused on operational conditions rather than catalytic characteristics. Both models from the two modes charted the factor importance and predicted the overall fuel cell performance with high R2 values (> 0.95) before experiments.