A Dual-input Multi-label Classification Approach for Non-Intrusive Load Monitoring via Deep Learning

Abstract

Non-intrusive load monitoring (NILM) is the process of obtaining appliance-level data from users' total electricity consumption data. These data can be of great benefit, especially in demand response applications. In this paper, a multi-label classification for NILM based on a two-input gated recurrent unit (GRU) is presented. Since the presented method is designed with a multi-label approach, great savings in training time are achieved. While a separate model is trained for each appliance in the literature, only one model is trained in the proposed model. Besides, the model was trained using two different inputs. The first is the total active power value consumed by the whole house. The second input is the Spikes obtained by analyzing this active power consumption. Simply put, spikes are obtained by analyzing the instant power changes in active power. Both inputs are evaluated with a convolutional layer and necessary features are extracted. Obtained features are fed into the GRU to be able to analyze time-dependent changes. The simulation results show that an additional input can slightly improve the analysis accuracy. Besides, it was found that the second input is useful especially in the analysis of short-term devices.