Machine learning algorithms for predictive maintenance in hydroelectric plants

In recent years, power plants and electrical grids have been undergoing a progressive modernization that aims at making them "smarter". This modernization seeks to ensure high levels of quality, security, safety, efficiency and sustainability in power supply. Digitalisation plays a great role in this process, since the availability of vast amounts of data from various components of these systems, along with insights into their operational behaviour, enables the application of modern machine learning and statistical techniques to improve automation and assist human operators. One prominent example of digitalisation capabilities is predictive maintenance, that tries to estimate when maintenance should be performed based on real-time condition monitoring of system components. In this work, a machine learning model - specifically an autoencoder - was developed and used to detect anomalies in hydroelectric power plants, with promising results. The work was carried out as an internship in a company specialized in data analytics in the energy field. The model developed was intended to be applied to different mechanical and electrical components (bearings, generator, transformers, valves, etc. ) of many power plants present worldwide. The model will be later extensively tested on a larger dataset to validate its effectiveness as a tool for enhancing predictive maintenance.