A case study to the prediction of erosion process in hydroelectric Kaplan turbines using machine learning models

The main sources of electricity generation in Brazil are hydroelectric plants. With high energy production capacity and more stability than other renewable sources, hydroelectric power plants are essential for the country's energy security. Thus, it is of strong importance to study actions that allow an increase in the operational availability of the generating units, in order to guarantee efficiency to the process and reduce operating costs that impact directly the value passed on to the final consumer. One of the problems with the greatest impact on the plants is the wear processes of hydraulic turbines, caused by abrasion and cavitation processes. Among the numerous possible solutions, the deposition of a coating by the thermal spraying process has been shown to be efficient. This paper demonstrates the application of machine learning algorithms for modeling the wear problem in Kaplan turbines, predicting the erosion rate of the coating based on the operational characteristics of the hydroelectric plant. The best results have accomplished symmetric mean absolute percentage error (SMAPE) below 10% through 5-fold cross-validation. The model allows for monitoring the wear of the coatings and calculating their remaining life, supporting better estimation for maintenance downtimes of the generating units.