Prediction of the breakdown voltage in a point-barrier-plane air gap using neural networks

Abstract

We have developed a neural network as tool for prediction of the breakdown voltage in a point-barrier-plane air gap for longest air gaps. We have used an array of figures from the experiments (tests) done in the HV laboratories of the Polytechnique University of Warsaw by A.Boubakeur (1979) changing many parameters of the air gap system. This shows promise for use in industry. The application of the radial basis function Gaussian network (RBGF) method trained by random optimisation method (ROM) is found to be very effective in its predictions. The choice of the RBFG method is argued by the fact that it's local characteristic avoids divergence problems. In practice, it would be very economical to use artificial neural networks in the investigations on high voltage insulation breakdown predictions. In fact, we may reduce the laboratory tests and let the network predict the remained breakdown voltages at longest distances. We may propose the application of this method to the prediction of other parameters (barrier width, barrier conductivity, air gap length, barrier hole diameter...); in general, in the case where we need to extrapolate non-linear functions giving their variation versus a given parameter.