Case C6 - Condition monitoring of Kaplan turbine hydraulic system

The aim of this case was to develop algorithms for monitoring the condition of the hydraulic regulation system for a Kaplan turbine using SCADA data. The work was carried out by professor Miguel Sanz-Bobi from Comillas Pontifical University in Madrid, Spain, in close collaboration with the hydropower plant operator and MonitorX industry partners Glitre and Vattenfall. Part of the motivation for this work is that the Kaplan propeller and hub is not accessible for inspection during production. A method for online condition monitoring without the need for unwanted production stops is therefore beneficial. The hydraulic system is of special interest as it is vital for the control of the turbine, and because e.g. oil leakage is a known issue.

A Kaplan turbine is regulated by adjusting the position of the wicket gates and the turbine runner blades. This is done by a high-pressure hydraulic system, typically consisting of an oil tank, oil pumps, valves, filters, coolers, and accumulator banks for the wicket gates and runner blades. To enable dynamical condition monitoring of this system, a normal behaviour model was developed for the level in the oil tank, using artificial neural networks (ANN). This model predicts the normal state of a variable, in this case the oil level, from other explanatory variables. Based on a physical understanding of the system, the explanatory variables were chosen to be the power, the oil tank temperature, and the oil level in the accumulators. Before the model can be used for anomaly detection, the model first learns the normal behaviour from carefully selected historical data. Once trained, the model can be used to detect anomalies, i.e. deviations from normal behaviour; see references [1] and [2] for further details.

Selected results are shown in the figure. The left diagram illustrates the training of the model, and the right diagram testing of the model for anomaly detection. It can be seen that the model accurately predicts the systems normal behaviour for the training set (left), and that an apparent anomaly is detected in the test set (right). The increasing deviation between the model (estimated value) and real data (real value) in the test set indicates a possible oil leakage. This was confirmed by the plant operator to be a leakage in one of the accumulators.

Estimated value from the ANN model and real measured value for the oil tank level for the training data set (left) and the test data set (right). Courtesy of Prof. M. A. Sanz Bobi.

The model was also tested on data from the hydropower plant operator and MonitorX industry partner Vattenfall. The test confirmed the ability of ANNs to accurately predict the normal behaviour of the hydraulic system. The ANN model must however be rebuilt and trained for the hydraulic system at hand, showing that significant work is required to deploy such models for multiple turbines.

[1] M. A. Sanz-Bobi, T. M. Welte, and L. Eilertsen, "Anomaly indicators for Kaplan turbine components based on patterns of normal behavior," in Proceedings of ESREL, Trondheim, 2018.
[2] M. A. Sanz-Bobi, "Normal behavior modelling oriented to diagnosis and prognosis," Comillas Pontifical University, Santa Cruz de Marcenado, Madrid, Technical Report 4.0, Feb. 2019.