Low-Cost Ensembling for Deep Neural Network based Non-Intrusive Load Monitoring

Non-Intrusive Load Monitoring (NILM) is the process of monitoring the power consumption of individual appliances by disaggregating the aggregate power consumption data from a single sensor, which is usually the main meter. The increase in adoption of smart meters facilitates large scale NILM. Appliance-level load monitoring could provide utilities and users with useful information which could lead to significant energy savings as well as better demand-side management. In this paper, we propose a low-cost method for ensembling deep neural network models trained for the task of load disaggregation, which does not require the training of multiple different models. Additionally, we analyze the output characteristics of the resultant ensembled model in relation to the outputs of its component models. The UK-DALE dataset is used for training the models and evaluating the effectiveness of our ensembling technique. The results show that the proposed technique provides a considerable improvement in load disaggregation performance.