An Empirical Study on Energy Disaggregation via Deep Learning

Abstract

Abstract-Energy disaggregation is the task of estimating power consumption of each individual appliance from the whole-house electric signals. In this paper, we study this task based on deep learning methods which have achieved a lot of success in various domains recently. We introduce the feature extraction method that uses multiple parallel convolutional layers of varying filter sizes and present an LSTM (Long Short Term Memory) based recurrent network model as well as an auto-encoder model for energy disaggregation. Then we evaluate the proposed methods using the largest dataset available. And experimental results show the superiority of our feature extraction method and the LSTM based model.

Convolutional Neural Network Applied to the Identification of Residential Equipment in Nonintrusive Load Monitoring Systems

“…In the context of the deep neural networks cited in [7][8][9][10], this study performed better than the evaluation metrics used. This is due to the difficulty of other methods in classifying multi-state appliances, such as the dishwasher and the washing machine.…”

“…In [8] the author sought to make an analysis of the various methods of deep learning to improve the performance of a NILM system. In [9], the authors used convolutional neural networks for the task of load disaggregation, promoting the individual identification of equipment loads based on the time series of the aggregate load. In [10], it is shown that CNN networks can also be used in the NILM context for equipment classification based on the VI path of an equipment.…”

“…Confusion Matrix: Allows an effective measure of the classification model, presenting the number of correct classifications versus classifications predicted for each class, on a set of examples [9]. The main diagonal presents for each class the correct classification number and the percentage that this number represents within the complete number of data of the class.…”

Home Appliance Identification for Nilm Systems Based on Deep Neural Networks

“…In [10] the author sought to make an analysis of the various methods of deep learning to improve the performance of a NILM system. In [11], the authors used convolutional neural networks for the task of load disaggregation, promoting the individual identification of equipment loads based on the time series of the aggregate load. In [12], it is shown that CNN networks can also be used in the NILM context for equipment classification based on the VI path of an equipment.…”

“…Due to this advancement in the area, some researchers have sought to apply as Deep Neural Networks to equipment identification problems in NILM systems. Some works were used in Long Short Term Units (LSTM), Auto-encoder Neural Network and Convolutional Neural Network (CNN) [9][10][11][12], with satisfactory results.…”