Application of neural networks in the thermal ageing prediction of transformer oil

Abstract

Studies on transformer oil thermal ageing were carried out at the ENP Laboratory. The oil, named BORAK22, is used by the Algerian national electric and gas company (SONELGAZ). Experiments were performed at different temperatures with a maximum ageing duration time of 2000 hours. The objective is to build a neural network that gives a good prediction of the nonlinear property variations of the material versus the ageing time, and whose learning time is clearly less than the laboratory test time. The chosen network is a radial basis function Gaussian network (RBFG) trained by the ROM (random optimisation method) and uses the FFN pattern and the batch learning techniques. The designed network gave a good prediction with a relative error of 5% and 3% for the two learning techniques respectively.