Non-intrusive Load Monitoring Using Imaging Time Series and Convolutional Neural Networks

In recent years, more than 50 million advanced (smart) metering infrastructure units have been installed by the U.S electric utilities. Although, smart metering can provide hourly or sub-hourly customer load, it has failed to directly benefit and provide actionable information to consumers and engage them in energy savings. Using nonintrusive load monitoring techniques, the smart metering data can be disaggregated to individual components for each appliance which consequently can be used to monitor the performance of the appliances, inform consumer about future failures and help them to reduce energy consumption. In this study, the time series data was encoded to images using Gramian Angular Summation Fields (GASF). This enabled us to train Convolutional Neural Networks (CNNs) on the encoded images. The developed model was used for energy disaggregation. To determine the performance of the developed model, mean absolute error and relative error in total energy was calculated for the obtained results from the deep neural net architecture. The convolutional neural network generalized well on unused test data without significant effort in data cleaning and feature engineering.