AI-Driven Predictive Maintenance and Power Optimization Strategy for Distributed Hydropower Systems

Abstract

Distributed hydrokinetic systems produce sequential operational telemetry that can support predictive condition monitoring and power-control optimization. This study combines short-horizon prediction of exact-zero-power onset with field-informed evaluation of a maximum-power-point controller multiplier. The classifier uses power, torque, and torque-to-power ratio from the two preceding per-rotor observations. A leakage-resistant chronological evaluation was performed using 36,636 non-overlapping windows: 279 positive episode onsets and 36,357 normal-operation windows. The untouched final period contained 35 positive onsets and 7365 normal windows (7400 total). On this period, the Random Forest showed strong discrimination, with ROC-AUC of 98.3%, PR-AUC of 81.4%, and precision of 81.2% at the conventional 0.50 classification threshold. Corrected analysis using Shapley additive explanations (SHAP) identified the most recent torque-to-power ratio as the leading model attribution. In the archived field excerpts, the selected 0.82 setting was associated with 13.6% higher mean instantaneous power than the prior 1.0 configuration. The observed onset rate of the power-loss events used in the study was approximately 76-fold lower at 0.82 than at 1.0 (0.23 versus 17.29 events per rotor-hour). This observational estimate remains uncertain because only one event occurred during the 4.40 rotor-hours recorded at 0.82. The results provide a practical framework linking interpretable condition monitoring with field-informed power optimization while retaining a clear distinction between predictive and observational evidence.