Back to the Future: A LSTM Based Approach for Fuel Cell Degradation Prediction and Anomaly Detection

Besides costs, the long-term durability is currently the biggest challenge for the commercialization of Proton Exchange Membrane Fuel Cells (PEMFC). Precise monitoring mechanisms that can identify anomalies at an early stage are essential to support the long-term operation. Cell voltage is a well suited monitoring parameter, but its prediction is challenging due to its non-linear dependence on numerous different operating conditions as well as on the history of the fuel cell. An existing machine learning proposes the division of the prediction task into two components: an existing stationary model and a conceptual degradation model. This work addresses this gap by developing and implementing a degradation model that complements the existing stationary model. A methodology is devised to prepare the training data and to calculate the targets for the degradation model based on the predictions of the stationary model. Three different neural network models are developed, trained and subsequently compared for their prediction accuracy. Transfer learning is then employed to evaluate the generalization of the best model model to three different validation data sets. The results demonstrate that the newly developed degradation model successfully identifies anomalies thus leading to significant improvements in long-term fuel cell monitoring.