Physics-constrained generative digital twin for early leakage detection in nitrogen accumulators in governor control systems at hydroelectric power plants

Abstract Nitrogen pre-charge loss in hydropower governor accumulators is a slow but critical degradation process that weakens hydraulic energy storage, reduces pressure-retention time, increases pump cycling, and accelerates thermal–mechanical stress. Conventional pressure-threshold SCADA alarms can miss the early stage because oil pressure may remain inside the operating band while waveform dynamics deteriorate. This paper proposes a SCADA-calibrated physics-constrained generative digital twin for early nitrogen leakage diagnosis. Limited anonymized field-SCADA observations are used only as calibration anchors for pressure set-points, pump-ON duration, pump-OFF retention, temperature range, and the fast-rise/slow-decay waveform; they are not presented as a complete labeled leakage dataset. The evaluation uses 80,000 physics-constrained synthetic training records and 20,000 independently generated synthetic test records per run, balanced across eight degradation months. Train and test records are produced with disjoint seeds and no shared cycles or residual draws. Across five independent generation-and-training runs, Random Forest gives the best pressure-prediction performance, with R 2 = 0.999731 ± 0.000004, MAE = 0.0856 ± 0.0007 bar, RMSE = 0.1074 ± 0.0007 bar, and 96.78 ± 0.18% of samples within 1% of the 23-bar operating span (0.23 bar). The corrected health index is bounded by 1.0, and cycle-level indicators identify the Month-4 warning transition before a conventional pressure-threshold response. The framework therefore supports earlier inspection and nitrogen-recharge planning, helping operators reduce avoidable pump cycling, maintenance escalation, and the risk of unplanned governor-system outages, while explicitly distinguishing SCADA calibration from synthetic diagnostic evaluation.