Non-intrusive electrical load monitoring system applying neural networks with combined steady-state electrical variables

Abstract

This paper presents a full electrical load identification model that considers steady-state parameters obtained easily from low-cost residential smart meters. The model was developed using neural networks including combinations of real power, current, impedance and admittance variables to identify the best input parameters. The monitoring model was improved by training one neural network to identify changing events and another neural network to identify the load state. The proposed model was tested using two different groups of residential loads: residential appliances measured in the laboratory and a public database of electrical measurements. The results show that the impedance model and a feedforward neural network achieved the best performance to characterise the load. In addition, when combining the different input parameters, those that consider impedance as an input parameter produced better results. The output provides simultaneous information about the operation state of all the loads before and after an event occurs.