Predictive maintenance in hydropower plants : a case study of valves and servomotors

Digitalization has opened the opportunity for a fourth industrial revolution and the hydropower industry is taking charge of enabling digitalization in their operation. There are a lot of studies on predictive maintenance, however, there are, to our knowledge no studies on system-specific predictive maintenance for hydropower. To bridge this gap, the idea of system-specific, Machine Learning driven Predictive Maintenance is explored. Two systems are chosen as a use-case for this thesis: valves and servomotors. With the increasing amount of intermittent renewable energy resources entering the power system, the need for flexibility in the power grid is unequivocal. Valves and servomotors are key components of hydropower control and thus will play a pivotal role in securing flexibility to the grid. The first system assessed is the main valve. In order to make this analysis easily applicable, the data that is already being collected at Nore 1 hydropower plant is analyzed in order to assess the possibility of maintenance prediction from limited data. Unfortunately, this did not achieve the desired results for the data collected from the valve sensors. This is due to the fact that only one variable was measured, in this case, the opening and closing time-lag of the valve. However, this thesis presents a framework for data collection that allows the use of Machine Learning for predictive maintenance. Various sensors are suggested based on several published works on predictive maintenance. The second system assessed is the servomotor that controls the guiding vanes in a Francis turbine. Servomotors are key components of hydropower control. Due to the data not being collected by Statkraft at the time of the study, this data was provided by one of Statkrafts suppliers. By making use of the historical data of pressure as a function of the piston position, a boundary for where new values should be expected is computed by making use of One Class Support Vector Machine. Another embodiment of this case is presented where force is given as a function of piston position, which yielded better results. When new values are being measured, the data is presented as a bullet chart that visualizes the distance of new values compared to the boundary computed by the One Class Support Vector Machine. This tool could easily be applied to other servomotors which perform other tasks such as controlling water injection to a Pelton turbine or opening and closing of the valve, whether they are butterfly or ball valves. Suggestions for further data collection are presented in order to make use of more data for the use of Machine Learning in Predictive Maintenance