Big Data and Neural Networks in Smart Grid - Part 2: The Impact of Piecewise Monotonic Data Approximation Methods on the Performance of Neural Network Identification Methodology for the Distribution Line and Branch Line Length Approximation of Overhead Low-Voltage Broadband over Powerlines Networks

Abstract

Τhe impact of measurement differences that follow continuous uniform distributions (CUDs) of different intensities on the performance of the Neural Network Identification Methodology for the distribution line and branch Line Length Approximation (NNIM-LLA) of the overhead low-voltage broadband over powerlines (OV LV BPL) topologies has been assessed in [1]. When the αCUD values of the applied CUD measurement differences remain low and below 5dB, NNIM-LLA may internally and satisfactorily cope with the CUD measurement differences. However, when the αCUD values of CUD measurement differences exceed approximately 5dB, external countermeasure techniques against the measurement differences are required to be applied to the contaminated data prior to their handling by NNIM-LLA. In this companion paper, the impact of piecewise monotonic data approximation methods, such as L1PMA and L2WPMA of the literature, on the performance of NNIM-LLA of OV LV BPL topologies is assessed when CUD measurement differences of various αCUD values are applied. The key findings that are going to be discussed in this companion paper are: (i) The crucial role of the applied numbers of monotonic sections of the L1PMA and L2WPMA for the overall performance improvement of NNIM-LLA approximations as well as the dependence of the applied numbers of monotonic sections on the complexity of the examined OV LV BPL topology classes; and (ii) the performance comparison of the piecewise monotonic data approximation methods of this paper against the one of more elaborated versions of the default operation settings in order to reveal the most suitable countermeasure technique against the CUD measurement differences in OV LV BPL topologies.