Big Data and Neural Networks in Smart Grid - Part 1: The Impact of Measurement Differences on the Performance of Neural Network Identification Methodologies of Overhead Low-Voltage Broadband over Power Lines Networks

Abstract

Until now, the neural network identification methodology for the branch number identification (NNIM-BNI) and the neural network identification methodology for the distribution line and branch line length approximation (NNIM-LLA) have approximated the number of branches and the distribution line and branch line lengths given the theoretical channel attenuation behavior of the examined overhead low-voltage broadband over powerlines (OV LV BPL) topologies [1], [2]. The impact of measurement differences that follow continuous uniform distribution (CUDs) of different intensities on the performance of NNIM-BNI and NNIM-LLA is assessed in this paper. The countermeasure of the application of OV LV BPL topology databases of higher accuracy is here investigated in the case of NNIM-LLA. The strong inherent mitigation efficiency of NNIM-BNI and NNIM-LLA against CUD measurement differences and especially against those of low intensities is the key finding of this paper. The other two findings that are going to be discussed in this paper are: (i) The dependence of the approximation Root-Mean-Square Deviation (RMSD) stability of NNIM-BNI and NNIM-LLA on the applied default operation settings; and (ii) the proposal of more elaborate countermeasure techniques from the literature against CUD measurement differences aiming at improving NNIM-LLA approximations.