GRNN approach to estimate a smooth mean curve on high voltage impulse waveforms

Abstract

One use of evolutionary computational approaches, generalized regression neural networks (GRNNs) in mean curve approximation of lightning impulse waveforms is presented. This paper explores in-depth details on artificial neural nets (ANNs) especially in biased term and probabilistic transfer function adapted with special linear layer. Evaluation in impulse parameters of standard/non-standard lightning impulse superimposed by oscillation, overshoot or noise generated from IEC-TDG software is discussed. As a result, the fitness is computed by spreading variance of Gaussian function in radial basis layer and it does not assume any model for estimating the mean curve. This can be easily implemented by adding initial input (samples) and desired target (impulse waveforms) then the training mean curve can be obtained through a black box. However, the problem of this approach is still existed when waveforms have an overshoot, which requires further study. In addition, advantages and disadvantages of this algorithm are as well investigated