DEVELOPMENT OF A PREDICTIVE MAINTENANCE SOLUTION FOR A PUMP-AS-TURBINE HYDROPOWER SYSTEM

Abstract

The increasing demand for reliable and cost-effective small-scale hydropower systems, alongside the evolving dynamics of renewable energy integrated electric grids, has led to a growing adoption of Pump-as-Turbine (PAT) systems. While PATs offer economic and operational advantages over conventional turbines, their practical deployment is often hindered by insufficient research into their operation and maintenance, particularly regarding hydrodynamic failures such as cavitation. This PhD research addresses this gap through the development of a predictive maintenance system tailored for PAT-based micro-hydropower systems, with a strong emphasis on the detection and diagnosis of cavitation. The study pursued four core objectives: numerical analysis of cavitation in PATs using Computational Fluid Dynamics (CFD), experimental investigations of PAT vibration responses under cavitating flows, evaluation of the suitability of Machine Learning (ML) models in cavitation diagnosis and the development of a software platform to integrate the findings of the study. CFD simulations revealed that cavitation onset varies between design and off-design conditions, and it evolves through four distinct stages, influencing PAT performance differently depending on the flow regime. Experimental investigations confirmed the narrow safe operating range of PATs and led to the development of a modified Deviation from Normal Distribution (DND) method for real-time cavitation detection, adapted from gearbox monitoring techniques. In the machine learning domain, Decision Trees (DT), Artificial Neural Networks (ANN) and Support Vector Machines (SVM) were assessed, with ANN achieving the highest prediction accuracy and DT being competitive while offering greater interpretability. A hybrid ANN-DT model was proposed to balance predictive performance with interpretability of predictions to improve transparency of decisions and trust in ML methods. These models alongside other findings from the study were embedded in a modular software platform built in MATLAB App Designer, structured according to ISO 13374:2003 and capable of time and frequency domain analysis for accurate cavitation classification. The developed system significantly enhances the operational reliability and sustainability of PAT-based micro-hydropower by enabling proactive maintenance through digital diagnostics. The outcomes of this research not only contribute new knowledge to the field of PAT operation and condition monitoring but also advance the digitalization of hydropower, supporting global efforts to build resilient, intelligent and sustainable energy infrastructure.