Contribuições para NILM baseadas em redes convolucionais profundas

Abstract

Non-intrusive monitoring of electrical loads (NILM) determines the profile of electrical energy consumption of each load of a consumer unit without the need to use independent sensors for each of these loads. NILM applications help the consumer save energy and assist in planning the distribution network. Since 2015, several feature extraction methods for NILM based on Deep Learning have been proposed, which have achieved state-of-the-art results for classification and disaggregation. Despite this, the results obtained by methods based on deep learning depend on the amount of training data, and the architectures of deep networks are empirically determined. With that in mind, this document proposes frameworks to extract and select features using the Scattering Transform to classify NILM signals, both individual loads and aggregated loads. We also introduce a new multi-label and multi-task convolutional network architecture for NILM, whose filters are not learned but analytically determined through wavelets. The classification results of the proposed method surpass the state-of-the-art techniques for several test scenarios, highlighting the scenarios with a reduced dataset. We also propose an expanded approach to NILM, including distributed generation (DG) photovoltaic (PV) data aggregated with electrical loads. We create a new public dataset containing DG-PV and NILM data, and perform an ablation study to evaluate both the influence of the presence of PV microinverters on the classification of aggregated loads and the influence of the existence of aggregated loads on the identification of PV microinverters in the distribution power grid.