Load Identification from Aggregated Data using Generative Modeling

Abstract: In the view of an exponential increase in demand for energy, there is a need to come up with a sustainable energy consumption system in residential buildings. Several pieces of research show that this can be achieved by providing real-time energy consumption feedback of each appliance to its residents. This can be achieved through Non-Intrusive Load Monitoring System (NILM) that disaggregates the electricity consumption of individual appliances from the total energy consumption of a household. The state-of-art NILM have several challenges that preventing its large-scale implementation due to its limited applicability and scalability on different households. Most of the NILM research only trains the inference model for a specific house with a limited set of appliances and does not create models that can generalize appliances that are not present in the dataset. In this Master thesis, a novel approach is proposed to tackle the above-mentioned issue in the NILM. The thesis propose to use a Gaussian Mixture Model (GMM) procedure to create a generalizable electrical signature model for each appliance type by training over labelled data from different appliances of the same type and create various combinations of appliances by merging the generated models. Maximum likelihood estimation method is used to label the unlabeled aggregated data and disaggregate it into individual appliances. As a proof of concept, the proposed algorithm is evaluated on two datasets, Toy dataset and ACSF2 dataset, and is compared with a modified version of state-of-the-art RNN network on ACS-F2 dataset. For evaluation, Precision, Recall and F-score metrics are used on all the implementations. From the evaluation, it can be stated that the GMM procedure can create a generalizable appliance signature model, can disaggregate the aggregated data and label previously unseen appliances. The thesis work also shows that given a small set of training data, the proposed algorithm performs better than RNN implementation. On the other hand, the proposed algorithm highly depends on the quality of the data. The algorithm also fails to create an accurate model for appliances due to the poor initialization of parameters for the GMM. In addition, the proposed algorithm suffers from the same inaccuracies as the state of art.