Finding the Fingerprint of an PEMWE Electrolyzer Using Applied Machine Learning Techniques

Abstract

The energy transition is picking up pace and renewable energy sources are playing an increasing role in our energy supply. The use of green hydrogen is considered an important energy carrier, to store renewable energy and allow many industries to decarbonize. There are different ways to produce hydrogen, whereas low-temperature proton exchange membrane water electrolysis (PEMWE) is one of the more mature electrolysis technologies. The benefits of proton exchange membrane water electrolysis (PEMWE) include the ability to deliver pressurized hydrogen and/or oxygen, operate using intermittent and renewable power sources, and have a reduced footprint. To evaluate the characteristic performance of PEMWE, the voltage/current relation is an important metric. The voltage/current relation is commonly represented in jV-curves. jV-curves contain relevant information, and any issues or out-of-specification performance can be observed immediately. The shape of the curve, and indicated values allow to interpret the performance of the PEWME cell. Nevertheless, proper interpretation of jV-curves and understanding of the source leading to the out-of-specification performance requires expertise.This study aims to evaluate jV-curves and identify faults by applying machine learning (ML) techniques. A conventional approach has been used to create the model. In the model the Nernst voltage, activation voltage on the anode, and the voltage losses due to ohmic resistance are modeled. Implementation of optimization algorithms allow to fit measured jV-curve to the model. Literature has shown relevant techniques to apply these kinds of ML techniques to electrolysis data, for example Particle Swarm Optimization and a modified Honer Badger algorithm. The application of ML to electrolysis data enables the opportunity to develop online monitoring and automated interpretation of data. In this study particle Swarm Optimization (PSO) has been implemented to fit the model to the jV-curve and find the unknown parameters. Those parameters are unique, depending on the performance of the electrochemical cell, and can be interpreted at the fingerprint thereof.The data shown in the figure has been measured on a single-cell material screener using Nafion 117 CCMs. Although both CCMs have been produced in the same batch, a difference between the measurements can be observed. The jV-curve of sample 1 shows performance as expected. This expected performance is characterized by its linear behavior above 0.5 A/cm2. Observing sample 2 shows an increase in voltage at increasing current densities compared to sample 1. This increase in voltage results in increased energy consumption and therefore a less efficient electrolysis process. Both CCMs are measured under comparable conditions. The anode and cathode pressure were during the both tests close to atmospheric, and the average cell temperature was kept constant at 333.15 K.Despite high volume production processes for coating and manufacturing, the origin of the difference in performance remains unknown. The application of automated data analysis allows for the immediate evaluation of the parameters at fault. The preliminary results show a discrepancy in the modeling parameters relevant to the sub-model describing ohmic losses. Furthermore, ML allows for the creation of data sets that contain experiences from the past. This can help to develop faster performance evaluation protocols by reducing testing activities. For future work, alternative algorithms have to be considered as well. Utilizing ML can enable the opportunity for online monitoring, predictive maintenance, and optimizing control strategies. Figure 1