Advanced Machine Learning-Based System for Predictive Monitoring and Early Fault Detection in Dubrovnik Hydropower Plant

To detect faults early and reduce downtime and maintenance costs, an advanced monitoring system with early fault detection is critical. This paper presents an advanced monitoring and diagnostic system with a machine learning module added for early fault detection and behavioral deviation identification. The module leverages historical vibration data alongside operational parameters, including active and reactive power, stator temperatures, and water flow and other parameters to continuously monitor the normal behavior of the hydro generator set. Upon detecting abnormal behavior, the module quantifies and reports deviations of output parameters from their established reference values. In addition, the system presents a novel Key

Performance Indicator (KPI) that forecasts generator behavior over future intervals ranging from a few minutes to thirty days. At each forecasting horizon, the predicted values generated by the monitoring system for the previous horizon are evaluated, and based on the system’s behavior, its future horizon for a similar time interval is forecast. At each forecasting horizon, the monitoring system provides predicted values, assuming the generator conditions remain unchanged. To develop this system, domain experts first select an appropriate period representing normal generator operation. Then, correlations between input and output signals are analyzed, the most relevant signals are selected, the machine learning model is trained. Next, the forecasting horizons are determined, and the KPI for predicting the generator’s future behavior is configured. Finally, the model is validated using additional datasets to ensure accuracy and reliability. The suggested approach has been successfully applied and verified at the Dubrovnik Hydropower Plant (HPP Dubrovnik), proving its usefulness and practicality in early fault detection and behavioral deviation monitoring.