Novosibirsk State Technical University (NSTU) is developing models for predicting the level of combustible gases in various operating modes of an industrial transformer. This will extend the service life of expensive equipment.
Existing monitoring systems for industrial transformers often lack the function of predictive analytics and predicting their condition. Using the accumulated data to solve the problem of developing models for predicting the most important parameters will make it possible to identify possible malfunctions in a timely manner, prevent emergencies, increase operational reliability, extend service life and reduce repair costs for this expensive and technologically important equipment, says Irina Yakovina, Associate Professor of the Department of Computing Technology at NSTU-NETI.
Various machine learning methods were used to build prediction models for one of the most important parameters — the level of combustible gases in transformer oil under various operating modes of an industrial transformer: logistic regression, decision trees, random forests, and neural networks of various architectures. The trained models made it possible to calculate the desired value of the predicted parameter with sufficient accuracy for different operating modes of the transformer. The transformer oil parameters were taken into account: water and hydrogen content, load data: the current and power of partial discharges, as well as data on the temperature and humidity of the outdoor air. A variant of calculating the desired values for different forecast horizons was considered: from several hours to several days from the observation point.
"When solving the forecasting problem, an ensemble of models is used, each of them has a specific application point. Depending on the transformer's operating mode (standard, emergency, emergency), a second-level model is selected. The combined use of a hierarchical chain of models provides a significant increase in the accuracy of the result," said Irina Yakovina.
"The results obtained allow us to draw conclusions about the degree of contribution of many factors to the predicted parameter and can be used in the future to create a collection of models for predicting the most important parameters of oil—filled transformers, studying and describing various modes of their operation," commented Alexander Dvortsevoy, Associate professor of the Department of Thermal Power Plants at NSTU-NETI.
The work is being carried out as part of the creation of a predictive analytics system for power transformers. The development became possible thanks to interfacult cooperation: to actively involve students and undergraduates of the Faculty of Automation and Computer Engineering in the work and to provide consulting support to specialists of the Faculty of Energy.
The scientists plan to expand the approach to the ensembling of machine learning models in order to maximize the use of the strengths of different algorithms, and to develop a technique that can be applied to a class of industrial transformers.