Abstract
Hydropower turbine–generator units operate under strongly coupled hydraulic, mechanical, electrical, and thermal dynamics, making reliable fault diagnosis and predictive maintenance challenging under uncertain and time-varying conditions. This study proposes an uncertainty-aware physics-informed digital twin that integrates a reduced-order multi-physics model, learned residual correction, online state synchronization, ensemble-based uncertainty quantification, out-of-distribution detection, fault classification, severity estimation, and remaining-useful-life prediction within a unified real-time framework. The method was evaluated under healthy operation, load transitions, progressive degradation, multiple fault classes, measurement noise, and unfamiliar operating conditions. Over 10 independent runs, the proposed digital twin achieved a state-reconstruction RMSE of 0.0025±0.0004 and an R2 of 0.997±0.002, while reducing the normalized physics residual to 0.0094±0.0011. Diagnostic performance reached a macro-F1 score of 0.950, an MCC of 0.940, and an AUPRC of 0.970. The calibration strategy reduced the expected calibration error from 0.0947 to 0.0115, while out-of-distribution detection achieved an AUROC of approximately 0.954. The prognostic module produced an RUL RMSE of approximately 2.24 h. End-to-end processing required 21.4 ms per synchronized measurement window, supporting operation within a 100 ms monitoring interval. These results demonstrate that the proposed framework combines high-fidelity reconstruction, physically consistent learning, reliable uncertainty estimation, robust diagnosis, actionable prognostics, and real-time computational feasibility for intelligent hydropower asset management. The framework therefore offers a practical foundation for risk-informed condition monitoring and maintenance decision support in modern hydropower plants.