Feature Mapping based Deep Neural Networks for Non-intrusive Load Monitoring of Similar Appliances in Buildings

Energy management plays an important role in the smart sustainable cities development programme to utilise energy resources in a responsible manner for conserving the environment and improving well-being of the society. Building sector is one of the major sectors that consumes more energy in commercial and residential buildings. Recently, non-intrusive load monitoring technique (NILM) has become popular among the researchers for its capability in disaggregation of energy at appliance/load level from the measured aggregated energy. Appliance signatures are learned using machine learning and deep learning approaches for effective detection of appliance events and energy consumption. However, appliance detection becomes challenging when appliances in the electrical network are similar or same type. Therefore, effective feature learning methodologies need to be developed for distinguishing the events of similar loads more accurately. In this paper, we used the open source dataset1 that consists of fundamental electrical features extracted from the four fluorescent lamps having same technical specifications. From the preliminary experiments, it is observed that the baseline system performance with the support vector machine (SVM) and deep neural networks (DNN) is not much encouraging due to the overlapping and nonlinear characteristics of similar loads. To overcome this problem, we expresses the original feature vectors in terms of appliance independent basis vectors in a higher dimensional space using a feature mapping technique, locality constrained linear coding (LLC) and then used machine learning classifiers for similar load identification. From the experiments and results, it is observed that feature mapping based deep neural networks (LLC-DNN) outperforms the baseline, LLC-SVM and other reported approaches from the literature significantly for similar appliances detection in NILM system.