Classification of Direct Lightning Stroke according to its Polarity in Medium Voltage Networks based on Deep Learning Neural Networks

Abstract:
This paper presents the classification of direct lightning strokes according to its polarity in a medium voltage network composed of multiple branches based on the recording of transient overvoltages at the circuit header. The test network that allowed obtaining the recording of transient signals resulting from the direct lightning stroke was the IEEE 13 Node Test Feeder implemented in the EMTP-ATP software. The lightning current has an 8/20 mus waveform with a peak of 3 kA, 10 kA, 20 kA, 50 kA, 80 kA, and 100 kA. The training of two types of neural networks, GoogLeNet and AlexNet, was performed using the Matlab Deep Network Designer app. The contribution from this research is the identification of the reduction in the classification accuracy of direct lightning stroke according to its polarity based on the recording of transient overvoltages at the circuit header when the medium-voltage network is detuned.