Estimate of three-phase distribution transformer losses through artificial neural networks

Abstract

This article presents a study of a neural network applied to estimate losses in core and winding of three-phase distribution network. The architecture of neural network used was the topology Multiple Layer Perceptron and training algorithm Levenberg-Marquard that use non-linear methods. From collate of data is create a database with all selected attribute for subsequently be used on simulation of artificial neural networks. All samples of learning are collected from transformers electric tests of automated software of routine testing used in diverse transformers industry and concessionaire of electric energy. The test stage represents 15% of samples, this part represents not supervision stage of training process where is possible observe ANN behavior after training stage (70% of samples) and validation (15% of samples). The evaluation of neural network was made by tools Mean Square Error, Linear Correlation Coefficient and graphic analyzer of cross-validation process. And in the training process obtain accuracy of 80 and 96% of data samples test of a transformer industry.