A neural network framework for predicting transformer core losses

In this paper a neural network based framework is developed for predicting core losses of wound core distribution transformers at the early stages of transformer construction. The proposed framework is also used to improve the grouping process of the individual cores so as to reduce the variation in core loss of assembled transformer. Several neural network structures and the respective training sets have been stored in a database, corresponding to the various magnetic materials. Selection of the most appropriate network from the database is relied on the satisfaction of customers' requirements and several technical and economical criteria. In case that the network performance is not satisfactory, a small adaptation of the retrieved network weights is performed. A decision tree methodology has been adopted to select the most appropriate attributes used as input vectors to the neural networks. Significant improvement of core loss prediction is observed in comparison to the current practice.