Uncovering Operational Patterns in H2 Electrolyzer Sensor Data

Efficient operation of PEM electrolyzers is essential for maximizing hydrogen production while optimizing power consumption. Monitoring current and voltage, crucial sensors in electrolyzer systems, aids in the creation of polarization curves, enabling performance evaluation. To plot the polarization curve, both current and voltage must exhibit step-wise steady-state behavior, facilitating reliable current over voltage plotting. Achieving such conditions necessitates conducting staircase profile tests. Extracting the occurrence time of this specific pattern and the average voltage and current is vital for analysis. Implementing automatic pattern recognition methods to detect these patterns and classify stair-case and normal operation not only reduces manual effort but also enhances the accuracy and efficiency of the analysis.

This thesis focuses on leveraging machine learning techniques to address the classification task, alongside utilizing the similarity matching method to identify similar patterns based on the matrix profile. Two supervised machine learning algorithms, one based on statistical features and the other utilizing a CNN neural network, are employed as training methods to distinguish the pattern. The results from the supervised machine learning methods reveal limitations in detecting low-occurrence patterns. Conversely, the similarity matching method exhibits insensitivity to unbalanced data and, as it does not require training, delivers accurate results with low computational cost.