Improving operational reliability in hydropower units using incremental learning-based monitoring

Reliable and efficient operation of hydropower plants is essential for ensuring a stable renewable energy supply. However, the growing demand for frequency regulation in modern power systems has led to more frequent start-stop cycles and varying load conditions, introducing operational stresses that can accelerate the degradation of critical components. To address these challenges, this study proposes a data-driven incremental learning (IL) framework for performance monitoring and predictive maintenance in hydropower generation systems. The framework incrementally updates a neural network model using sliding window data stream, while retaining prior knowledge through a freezing-based adaptation strategy. Key performance indicators (KPIs) are derived by comparing model predictions under Monte Carlo-simulated reference conditions, providing quantitative insights into the progression of equipment health. The proposed method is validated using over three years of full-scale operational data from a Swedish hydropower plant. Results demonstrate that the IL-based approach successfully tracks KPI increases from 0 to 0.1 over two years of operation and detects abrupt KPI drops following planned maintenance, as observed in the case study bearings. Compared to conventional retraining methods, the IL framework offers improved adaptability and stability. By providing a robust framework for quantifying both gradual degradation and abrupt health status shifts, this work presents a direct pathway toward more proactive, condition-based maintenance strategies, ultimately enhancing the operational reliability and economic viability of hydropower assets.