Evaluation of supervised machine learning techniques for cavitation detection and diagnosis in a pump-as-turbine system

Abstract

The transition to sustainable and efficient energy systems has driven a rapid adoption of variable renewable energy sources such as wind and solar, increasing the demand for reliable and flexible power generation. Hydropower remains essential to grid stability; however, aging infrastructure and the need for more flexible operation present significant challenges. Digitalization has emerged as a key strategy to modernise hydropower systems with Machine Learning (ML) playing an increasingly important role in predictive maintenance. This study explores the application of ML techniques for cavitation detection in pump-as-turbine (PAT) systems using vibration data. Decision Trees (DT), Support Vector Machines and Artificial Neural Network (ANN) models were evaluated for applicability in a predictive maintenance system. A novel hybrid framework is proposed that integrates a DT model to provide transparent, interpretable rules for predictions with an ANN model that assures high accuracy cavitation state predictions. Experimental results show that the ANN model achieved the highest classification performance (accuracy: 99.86%, F1-score: 99.84%), while the DT offers valuable interpretability with competitive accuracy (99.66%). Agreement between the model’s predictions implies confidence while disagreements prompt further investigation. This hybrid framework supports informed decision-making, reduced downtime, and improved diagnostics in digital hydropower systems.

Topics & Concepts