Using Machine Learning and Artificial Neural Networks to Recognize Inter-Turn Faults in Power Transformers

Abstract

The power transformer relay protection is among those featuring the poorest selectivity. It is only in 70–80 % of fault cases that the relay protection responds in a correct and timely manner, preventing further development of the emergency. One of the main problems is the recognition of internal faults (inter-turn faults in the transformer winding). Neural network algorithms are suggested used to be used to recognize inter-turn faults in transformers. Simulation-type Matlab models have been developed to study the operation of a transformer as part of an electrical network in normal mode, as well as in case of inter-turn faults. Based on the simulation results, a large amount of statistical data has been obtained that serves to train an artificial neural network (ANN). Two options of transformer protection aimed at recognizing inter-turn faults are proposed: (i) the use of ANN directly for decision making and (ii) the use of ANN to calculate the coefficients of multiparametric relay protection. The proposed protection options were tested using neural network algorithms. It has been determined that the second protection option is more preferable for implementation, because the ANN is used only for calculating the weighting factors. In comparison with the classical differential protection, the proposed ANN-based protection is distinguished by better recognition of inter-turn winding faults with a small number of short-circuited turns at the early fault development stages.