A hybrid methodology for assessing hydropower plants under flexible operations: Leveraging experimental data and machine learning techniques

Abstract

In recent years, there's been a demand for more flexible hydropower to offset power fluctuations from intermittent renewables, leading to increased wear in turbines due to extended off-design operations and transients. The present paper proposes novel methods to analyze and assess experimental data of a hydropower plant for a better monitoring over time. Measurements are performed in a run-off-river plant with Francis turbines in Switzerland. During two experimental campaigns, high-frequency acquisitions up to 51.2 kHz are realized with more than 30 sensors while the SCADA data are collected at 10 Hz. In the physics-based analysis, statistical metrics and time-frequency decompositions are deployed to assess the operating conditions. To aggregate all information, the Vibrational Content Index is introduced, which unifies all sensors data by normalizing their frequency spectra with the Best Efficiency Point (BEP). This provides a single-value indicator of overall vibration, which is more sensitive than traditional metrics and easier to track over time. In the data-driven approach, the operating points are segmented into three clusters with distinct physical interpretations by applying dimensionality reduction algorithms with only two coded dimensions. This method correctly detects an abrupt change in the machine behavior for a slight power shift from $1.23P_{\mathrm{BEP}}$ to $1.30P_{\mathrm{BEP}}$ – due to full-load vortex self-excitation. Finally, the concept of virtual sensors is developed by corroborating the coded high-frequency experimental data with low-frequency SCADA using machine learning. The trained model uses SCADA data as input to estimate the sensors high-frequency response in real-time. This promising approach enables improved continuous monitoring without the need for permanent installations.