Research On Event Detection And Non-intrusive Load Identification Method

With the rapid advancement of science and technology,the consumption of electric energy continues to expand,and the situation of electricity consumption in various countries is becoming increasingly tense.In order to solve the problem of unbalanced power supply and demand,promote the construction of smart grids,and guide users to rationally plan electricity consumption,it is necessary to implement intelligent management on the power demand side.User-side energy consumption monitoring is an important part of demand-side management.The traditional user energy consumption monitoring method needs to install sensors and other devices at each power load in the monitoring area to collect user load power consumption data.This intrusive monitoring The method needs to consume a lot of manpower and material costs during installation and maintenance,and at the same time,the user acceptance is low,which is not easy to promote and use.Therefore,the non-intrusive identification method has received widespread attention.This method does not need to enter the user’s interior to install sensing devices,but only needs to collect bus data at the user’s entrance and analyze the load characteristic information to realize load identification.Compared with traditional intrusive technologies,non-intrusive identification methods have lower hardware costs,more convenient maintenance,and higher user acceptance,and will gradually replace intrusive monitoring technologies.In the non-intrusive load identification process,the performance of event detection and load identification algorithms is related to the accuracy of identification results,which is mainly studied in this thesis.In the event detection part,in the mainstream nonparametric CUSUM algorithm,the noise figure is a pre-set value.For loads with large irregular power oscillations during steady-state operation,false detection behaviors will occur.This thesis optimizes the noise figure The setting is changed to be dynamically updated with the sliding window.At the same time,in order to solve the problem that the interference signal may be detected as a load event during the detection process,a feedback verification link is introduced to effectively prevent the occurrence of false detection behaviors.At the same time,in view of the fact that the CUSUM algorithm only considers the time when the event occurs but ignores the problem of when the system enters the steady state after the event occurs,a method of adding a variance window is proposed to detect when the system enters the steady state after the event occurs,which is the steady state Waveform extraction is guaranteed.In the load identification part,the performance and applicable scenarios of statistical methods,machine learning methods and deep learning methods in the field of load identification were analyzed,the structural model of the deep learning network was improved,and the network hyperparameters were adjusted.After experimental verification,the improved Compared with the original model,the network optimizes the running time and improves the algorithm performance.This research work consists of the following parts:(1)Elaborate theories related to non-intrusive load identification,analyze and construct a load steady-state characteristic database.(2)This study investigates event detection algorithms in non-intrusive load identification and applies for the first time the Bernaola-Galvan heuristic segmentation algorithm from the field of climatology to this research domain.The accuracy of the algorithm for detecting event occurrence points in historical data is verified.Additionally,the CUSUM real-time online detection algorithm is used and traditional CUSUM algorithm is improved by dynamically updating the noise coefficient and introducing feedback verification to prevent false positives.After detecting an event,a variance window is introduced to determine when the system enters a stable state,facilitating subsequent identification work.(3)Research load identification algorithms based on statistics,machine learning,and neural networks,analyze the advantages and disadvantages of each algorithm and applicable scenarios,optimize the network model of deep learning,greatly shorten the running time,and improve the accuracy rate.Experiments are designed to verify the accuracy of the improved event detection algorithm and load identification algorithm in this thesis.