Submeter based Training of Multi-class Support Vector Machines for Appliance Recognition in Home Electricity Consumption Data

In this paper we employ smart meter and support vector machines (SVM) for the problem of recognizing household appliances’ load patterns in measured load time series, which is an important step for various applications in energy consulting, process recognition or health care applications. We present an automated data collection and preprocessing approach that intrinsically avoids many privacy (and security) issues by keeping the whole process local to the household. In the experimental part we investigate multi-class SVMs in the problem domain of automatically recognizing appliances in load profiles of smart meters. For the learning phase, we use low intrusive submeters to automatically and locally generate household specific test data for the supervised training and validation of the SVMs. We analyze classifiers w.r.t. various training sets and feature spaces. Comparing data from household simulator and real household data, we find that excellent recognition rates can be achieved even with low resolution data and rather unsophisticated feature space.