Development And Evaluation Of Deep Neural Network Models For Non-Intrusive Load Monitoring

Energy conservation in residential and commercial buildings through smart electrification has been a hot topic for researchers in recent years.Because of the large deployments of smart meters,non-intrusive load monitoring(NILM)has become a very valuable tool to achieve this objective.NILM or simply energy disaggregation system estimates power consumption of individual household appliances or other electrical apparatus from an aggregated power signal,which is acquired through single-point sensing from a smart meter using some supervised/unsupervised technique.A practical NILM system can provide real-time actionable feedback to consumers that gives them an idea about an individual appliance operation state,its power consumption,and cumulative energy usage.Studies have shown that appliance specific feedback urges the consumer to use energy wisely,which as a result can save up to 12%energy in residential and commercial buildings.For utility companies,this information can be useful in foreseeing the power demands,and efficiently operate electric power facilities.The most efficient and cost-effective approach to achieve energy conservation goals through NILM is to use intelligent yet practically feasible algorithms.Recent works have indicated that deep neural network(DNN)models can be made generalized and highly accurate if they are trained on either large quantity of data or trained on sufficient data comprising of most effective features.However,the practical feasibility of DNN models,from generalization and accuracy point-of-view,is still an open problem.A part from practical feasibility of DNN based NILM algorithm;another biggest challenge is their evaluation using a more practical and robust evaluation metric(s).There is an urgent need for NILM application specific metrics that fully demonstrates algorithm's capability to identify appliance states(events),and estimation of consumed energy altogether.This forms the basis of our research to develop highly accurate and generalized NILM algorithm using DNNs,and to propose a practical evaluation method to test that algorithm so that it can be used in real-world.One of the biggest obstacles in using DNNs is the unavailability of sufficient training data,which eventually aids in getting high accuracy and lowest training error.In real-world,getting hands to huge amount of training data at once is very difficult,so it is highly desirable to develop a NILM solution,which can get acceptable accuracy when trained with limited amount of data.This was our first contribution where we tried to develop an accurate NILM algorithm by exploiting features from limited training data and trained LSTM-based deep neural network models accordingly.In our second contribution,we proposed a solution to deal with generalization problem of DNN based NILM algorithms.We proposed a data augmentation algorithm to synthetically generate training data by combing data from cross-domain datasets and then trained CNN-based deep neural network models.Our third contribution was related to algorithm's evaluation domain,where we proposed domain-specific evaluation metrics to evaluate NILM algorithms.In our first contribution,we proposed a long short-term memory(LSTM)network based NILM solution using multi-feature input space and post-processing.In the proposed approach,the mutual information method was used to select electrical parameters that had the most influence on the power consumption of each target appliance.Selected steady-state parameters based multi-feature input space(MFS)was used to train the 4-layered bidirectional LSTM model for each target appliance.Finally,a post-processing technique was used at the disaggregation stage to eliminate irrelevant predicted sequences,enhancing the classification and estimation accuracy of the algorithm.The proposed approach was compared with state-of-the-art NILM algorithms using unseen data taken from publicly available datasets.The results proved that if DNN models are trained using most relevant features without having sufficient data,then highest estimation accuracy can still be achieved,which is prerequisite for practical application of NILM solution.In our second contribution,we proposed a deep convolutional neural network(CNN)based NILM algorithm that used data augmentation to improve the generalizability of NILM system.Proposed algorithm worked in two steps.First,a unified and comprehensive synthetic data was prepared by combining on and offdurations of a target appliance from various datasets.The synthetic aggregate and submeter data were comprised of real activations of target appliances and had same distribution as original data.In second step,the synthetic data was used to train 8layer deep CNN model for each target appliance.The trained models were tested on unseen data from new datasets(not used during training)and compared with modern NILM algorithms to prove its effectiveness.The models trained on one dataset,and tested on another dataset achieved better estimation accuracy then other state-of-theart algorithms.Results proved that DNN models trained on synthetic data can improve the generalization capability of the algorithm.Testing of machine learning/deep learning algorithms is integral part of any ML/DL based project;therefore,choosing most relevant and right evaluation metrics is very critical.In our third contribution,we proposed NILM-specific evaluation metrics,which used overlapping regions of ground-truth energy and predicted energy of an appliance to give insights about individual appliance model's performance.Proposed metrics are capable of providing detailed insights about algorithm's performance using four indicators:total predicted energy,total overlapping energy,missing energy and extra energy predicted by the algorithm.These metrics revealed very important information about NILM algorithm's performance in different test scenarios.Since,these metrics are specific to NILM problem;therefore,they can be used to evaluate any NILM algorithm.