A crucial piece of the puzzle for the long-term success of the cost-effective supply of green hydrogen through electrolysis is the description of the time-dependent degradation behaviour in order to be able to consider this, for example, in predictive maintenance concepts. Without this information, it cannot be guaranteed that customers will be able to fulfil their supply contracts for the agreed amounts of produced hydrogen. For this purpose, knowing the current state-of-health (SoH) and also describing its future development over time play essential roles. However, using conventional methods it is only possible to obtain the SoH if characterization measurements are carried out at regular intervals, such as the recording of polarization curves or electrochemical impedance spectroscopy. In real applications, this has the major disadvantage that the desired current operation mode has to be interrupted, which leads to a reduction in system utilization and undesired operation conditions. Consequently, intensive efforts are being made to find model-based alternatives for describing the SoH at a certain point in time or even predicting the future degradation trend purely on the use of begin-of-life (BoL) measurements and anyway available operating data without additional characterization intervals. Since the physics-based description of degradation mechanisms is not yet sufficiently sophisticated, data-driven approaches are currently receiving increased attention. These have the great advantage that they can use information in the available data to learn underlying unknown relationships independently, without the need for prior physico-chemical knowledge [1].For this reason, we tackle the aforementioned problem of predicting the degradation trend using a machine learning (ML)-based approach and investigate it in the context of a proton exchange membrane water electrolysis (PEMWE) cell.The central relationship of PEMWE related to ageing is the change in polarization behaviour, which essentially describes the connection between the operating conditions and the resulting cell voltage. The current density corresponds to the electrical load, while the voltage is a measure of the specific energy consumption (SEC) and the efficiency. The degradation-related drift of the entire polarization curve means that the cell voltage depends not only on the operating conditions but also on time and thus has two core dimensions. Although the temporal development of the cell voltage only provides integral information about ageing, this is the key variable in the context of economic operation and maintenance due to the direct correlation to the above-mentioned SEC and thus with the supply agreements.The goal of this study is therefore to predict future polarization behaviour based on historical data. In principle, there are various ways of specifying this desired model output. The possible spectrum of describing the polarization behaviour ranges from the cell voltage at a selected reference current density to a cell voltage analysis including a voltage breakdown over the entire current density range. Depending on the data used for model training, this results in a data-driven degradation matrix (see Figure 1a), which was experimentally supported by performing a specially tailored accelerated stress test (AST) [2] and investigated using selected practice-relevant showcases. We were able to demonstrate that the applied ML workflow can be used to build neural network-based models, which are capable of covering different cases of this degradation matrix. This applies both to cases that are characterized by the identical information content of input and output and lie on the diagonal (cases 1 and 2) and to a case in which the model is provided with less information via the training data than the desired model output should contain (case 3).Another unique feature of these investigations is the different experimental background of the data used for model training in the individual cases. Cases 1 and 3 are based on data from discrete load levels, which are extracted exclusively from the operating profile of the AST, while case 2 is based on data of polarization curves obtained from separate measurements during characterization phases. Accordingly, case 3 clearly demonstrates that future polarization curves, i.e. information that is typically only available via specific characterization measurements interrupting the operation, can be predicted solely based on past operating data (Figure 1b).The authors gratefully acknowledge funding by BMBF in the framework of projects Segiwa/Deriel, FKZ. 03HY121G/03HY122G.
Read full abstract
Jan 24, 2025
Idriss El-Thalji 
Open Access
Just Published
