Development of a Neural Network for a Digital Twin of Three-Winding Transformer

This article discusses an approach to the design of a neural network for a three-winding transformer as part of a digital twin. The structure of the neural network is shown, important aspects for choosing the number of input neurons, hidden layers of the network, neurons located in the hidden layer, and neurons at the output of the network are described, the activation functions used are presented. A neural network is used to predict the values of transformer phase voltages. A transformer operating in idle mode is considered. The data for training the neural network is calculated using a digital model of a power transformer created in the Simulink application of the MatLab program. As the results of the experiment, a neural network training graph and a neural network testing error graph for five tests are presented. In the future, the neural network model is planned to be improved by training it for short-circuit mode and nominal mode. The predicted values obtained will be applied to subsequent classifications of emergency operation modes of transformers, namely “steel fire” of the transformer, single-phase ground faults, interturn, phase-to-phase and inter-winding short circuits, as well as to predict the remaining lifetime of equipment for scheduled preventive repairs.