Research On The Non-Intrusive Load Monitoring Based On Machine Learning

With the strategic goal of developing ubiquitous power Internet of things,the perception and acquisition of the user's appliance Information is the basis for providing new smart services,while Non-intrusive load monitoring is a suitable intelligent sensing technology.Non-intrusive load monitoring(NILM)refers to the technology of decomposing the aggregated power con-sumption data obtained at a single measurement point into the operational p ower consumption of each electrical equipment.This thesis studies two different types of NILM algorithms:event-based and state-based.It proposes different algorithms for the four steps of event-based method(event detection,feature extractiong load identification,and energy decomposition)based on Machine Learning,and verify the effectiveness of the algorithm through scene experiments and public data sets BLUED and ECO.Besides,it uses the Hidden Markov Model as an example to study the state-based method to explore the advantages and disadvantages of two different types of methods.The whole thesis consisted of the following three parts.Firstly,it proposes a hybrid event detection method for the event detection step of event-based NILM.It uses different event detectors to resolve the interference of noise on event de-tection.After case study and experimental analysis,it has better detection effect in complex scenes than the single event detection method.Secondly,for the step of load identification and consumption decomposition in the event-based NILM,it proposes an unsupervised load identification method based on DBSCAN and a method of consumption decomposition based on active window.The unsupervised load recog-nition method based on DBSCAN introduces consideration of noise in the clustering process,and divides the appliances into periodic appliances and user-driven appliances.The second method is used to address the adverse effects of the events that are missed or incorrectly identi-fied in the consumption decomposition.The effectiveness has been verified by the public data set.Finally,a hidden Markov model is used as an example to study the state-based NILM.Through experiments,it compares the monitoring effect with the event-based method under different acquisition frequencies and analyses their applicable scenarios.