Recurrence Plots and Convolutional Neural Networks Applied to Nonintrusive Load Monitoring

The need to increasingly promote the efficient use of electricity for residential consumers motivated some researches to obtain solutions that allow the identification of loads for use in demand response programs and in the energy consumption disaggregation. Recently, the use of Convolutional Neural Networks for load identification has been common. However, it is necessary to transform the temporal signal into an image. In this sense, the present paper proposes a novel feature engineering stage, based on Recurrence Plots, to obtain the images for Nonintrusive Load Monitoring. In order to prove its effectiveness, the Recurrence Plots were compared to the Gramian Angular Difference Fields, which had already been applied in this context. For this purpose, it was used the REDD dataset, in which only the low frequency data measured in the main distribution panel were considered. Individual load measurements were only used for the labelling of the 90-seconds windows. The Convolutional Neural Network was trained and validated, considering two different strategies (multi-label and binary models), obtaining an F1-score of 66.05% and 72.65%, respectively.