Machine learning-based optimizing of heavy-duty fuel cell performance: Experimental data-driven prediction models for voltage and stack temperature

This paper tackles the significant issue of enhancing heavy-duty fuel cell performance, specifically concentrating on the prediction of voltage and stack temperature, two essential determinants of fuel cell efficiency. The aim is to create resilient prediction models that enhance fuel cell longevity and operational efficacy, crucial for minimizing carbon emissions in the transportation industry. Two artificial neural network models, constructed on multilayer perceptron architecture, were created using experimental data. The input parameters were temperature, relative humidity, and current, whilst voltage and stack temperature were forecasted as outputs. The dataset was partitioned into training, validation, and testing stages, allocating 70%, 15%, and 15% of the data, respectively. The findings indicate that the artificial neural network models provide exceptional predictive accuracy, with coefficient of determination values surpassing 0.99 and mean squared error values as low as 4.87 × 10−2 for voltage and 8.53× 10−3 for stack temperature. Index of agreement values reached 0.999, indicating robust concordance between predicted and experimental data. This integrated methodology, which concurrently forecasts voltage and stack temperature, offers a full model for fuel cell optimization. The study's innovation lies in its comprehensive, dual prediction methodology, which, unlike prior research, often focused on isolated parameters or components, concurrently forecasts both voltage and stack temperature. This integrated approach offers a more holistic model for heavy-duty fuel cell optimization, holding considerable potential for future design and performance enhancements.