Improving Core Quality in Power Distribution Transformers Using Machine Learning Methods

Abstract

The estimation of individual core losses of wound core power distribution transformers areparticularly important since their core costs account for around 30% of their overall material cost and are one ofthe key determinants of their quality. In addition, accurate calculations of individual core actual losses areextremely difficult, since actual losses show a divergence of up to 20%, in relation to the theoretical individualcore losses. This paper demonstrates the use of Machine Learning (ML) techniques, namely Decision Trees(DTs) and the Learning Vector Quantization (LVQ) neural network to the enhancement of each core's quality inwound core power distribution transformers. The DTs method makes use of inductive inference to automaticallybuild decision rules and apply them to the power distribution transformers production procedure. In the LVQneural network, any set of input vectors can be classified by using supervised training of competitive layers.Real industrial measurements were used to create the learning and test set. Information includes measurementsof the production line's quality control as well as the electrical properties of grain-oriented steel. The resultingDTs present a success rate of 94%. Based on these DTs, rules comprising the most significant parameters andtheir threshold values can be derived. These are used to lower the actual losses of individual cores, hence raisingtheir quality. The LVQ neural network approach achieves a total classification success rate of 95%.