Neural networks for the prediction of magnetic transformer core characteristics

Because the performance of power transformers is by various distinct parameters of the magnetic core, the prediction of relevant characteristics such as no-load losses P by analytical methods is impractical. This paper reports first attempts to predict the dependence of P on several parameters of core design by means of artificial neural networks (ANN's). Investigations of several ANN versions showed good results for simple backpropagation networks equipped with several output neurons for an adaptive version of Gaussian coarse coding. A main problem arises from the fact that an increase of input parameters is linked with a large increase in training batches established by time-consuming model core experiments. As a compromise, first ANN's were trained for the prediction of the losses P/sub J/ of "linearized" joint regions as a function of the most relevant parameters, including the number of overlap steps and the mean air-gap length of joints. This yields rough estimations of the joint's contribution to the building factor for small cores. For larger cores, an ANN cascade structure was tested. It includes a second ANN that considers indirect effect of joint designs on the global distribution of losses. The major problem with an ANN-based prediction system is establishing representative training data. Modified versions of the ANN method can be applied to various tasks, including the prediction of losses and noises of full-sized cores.