Author: Vincent Köhler (Fraunhofer IWES) - The Hydrogen Lab Leuna (HLL) provides a large-scale testing environment for electrolysis systems, supporting stack tests from 3 kW up to 2 MW (planned) and system-level tests up to 5 MW. Although sensor data is continuously recorded, conventional threshold-based monitoring often fails to detect early deviations. Gradual degradation, such as slow voltage increases over thousands of hours, or short-term anomalies—for example, faulty thermal regulation like unexpected temperature behavior under constant current—typically remain unnoticed until critical thresholds are reached. This work presents a machine learning framework using Long Short-Term Memory (LSTM) networks for real-time anomaly detection. The framework is validated on a 1-year dataset from a high-temperature 800 kW electrolysis system with a multi-stack array. It includes modular components for data preprocessing, feature selection (PCA, permutation), and model training. LSTM architectures are varied systematically (32–128 units), with prediction horizons from 5 to 600 s and window sizes between 60 and 300 s. Feature reduction shows that 10–20 inputs are sufficient to achieve R² > 0.992 for short-term predictions. Trained models were applied to unseen data in a simulated real-time scenario and successfully detected known anomalies via elevated prediction errors. Future work includes long-term drift detection.
