Multi-objective optimization of a neural network based differential relay for power transformers

An application of artificial neural networks (ANNs) to the protection of three-phase power transformers is presented in this paper. The emphasis is placed on selecting the appropriate type of an ANN as well as pre- and post-processing algorithms. More than 30 essentially different architectures have been developed and tested in this paper: from natural relaying signals to highly pre-processed logic signals as far as the pre-processing is considered; from a single perceptron to "winner takes all" neural nets as far as the type of an ANN is considered; and from natural output from an ANN to median filters as far as the post-processing is considered. The comparative analysis of the selected designs is performed and the optimal ANN-based classifier is selected. The included operation examples illustrate high security and dependability of the optimized ANN-based protection for power transformers.