Prediction of PEMFC stack efficiency using recurrent neural networks

Hydrogen-related technologies are a promising solution for energy storage systems. The energy surplus from renewables or the off-peak power in a centralized electric system can be used to feed an electrolyzer to produce hydrogen. Then, the stored hydrogen can be converted to electric energy using proton exchange membrane fuel cells (PEMFCs), which have higher efficiency than other conventional energy conversion technologies. However, a PEMFC stack might be subject to variations in its energy conversion efficiency due to undesirable changes in the internal operating conditions, which may lead to a relatively high loss of generation capacity and an increase in the hydrogen consumption. A system identification method using recurrent neural networks (RNNs) able to build prediction models for the efficiency is proposed in this paper aiming at a further use in a control application to maximize the stack's efficiency in real-time. A 3 kW commercial PEMFC stack operating under a typical residential load profile is employed in this work to obtain experimental data. Three models are developed with the aforementioned method using data sets with different sizes for training the RNNs. The results show good agreement between the predicted and the measured efficiency, with a coefficient of determination (R
²
coefficient) higher than 0.999 and a validation mean squared error (MSE) lower than 0.02 for most of the models. The results also reveal that, for a fixed prediction horizon of 6.4 h, all the three models present similar performance, with the size of the training data sets making little difference in the performance indicators.