Non-intrusive load monitoring system for similar loads identification using feature mapping and deep learning techniques

Abstract

In the recent years, non-intrusive load monitoring (NILM) technique has received much attention among researchers because of its effective monitoring of events and extraction of energy consumption of individual loads from aggregated energy consumption. The performance of the existing NILM algorithms may not be adequate while multiple similar type/same kind of loads are connected in an electrical network which cause overlapping of signatures of the individual loads of similar type/same kind. In the present study, multiple air conditioners of same model from the three different makes have been considered for NILM in a commercial/residential environment. Preliminary investigations by considering the general features such as apparent power and power factor for the purpose of load classification indicated that there is scope for improvement in their performance of the load classification. Hence, a new feature set consisting of power and intrinsic features which are derived from the electrical signals during transient and steady state operation of the loads has been proposed. The geometric mean between all the considered electrical features has been computed to capture the relationship between the features for more effective load classification. These studies with the new feature set indicated enhanced performance in load classification of same model and same make. Further, a feature mapping technique, viz. locality constrained linear coding (LLC) has been explored to represent the new set of features as a linear combination of basis vectors to make the features load independent for further improving the performance of load classification of loads of same model and same make. These additional studies indicated that the LLC based approach achieved further enhanced performance as compared to the baseline systems for load classification of loads of same model and same make in NILM systems.