TabM-LightGBM Model for Electrolyzer Stack Degradation Diagnosis Based on Polarization Curves and Pressure Data

Source abstract

Electrolyzer stack degradation diagnosis is difficult because voltage rise, polarization-curve distortion, gas pressure fluctuation and water-gas transport disturbance can occur simultaneously under variable load and regulation conditions. This paper introduces a TabM-LightGBM model for stack degradation diagnosis based on polarization curves and pressure data. The method first extracts curve-slope, voltage-residual, pressure-variation, operating-state and pressure- reliability features from segmented operating windows. TabM is used to learn multiple tabular representations of feature interactions, while LightGBM performs interpretable degradation classification and maintenance-priority ranking. The Pressure Reliability Index is introduced to reduce interference from transient pressure pulsations during electrochemical ageing and balance-of-plant regulation. A simulated stack monitoring dataset of 6,055 operating windows is constructed for normal, activation-loss, ohmic-growth, mass-transfer-risk and mixed-degradation conditions. The proposed model achieves 93.4% overall accuracy, a macro-F1 score of 0.921 and an expected calibration error of 0.048. Compared with a standard LightGBM baseline, mixed-degradation recall increases by 14.7%, and the false warning rate under pressure disturbance decreases by 11.2%. These results indicate that combining polarization-curve evidence with pressure-feature analysis improves degradation-state recognition, probability calibration and engineering-review reliability. The output reports degradation class, confidence level, dominant evidence and maintenance priority, supporting the distinction among membrane ageing, contact-resistance increase, mass-transfer limitation and uncertain pressure disturbance.