A Non-intrusive Residential Load Identification Method Considering Transient Features

Fine-grained appliance consumption information plays an important role in guiding residential building electricity consumption and energy efficiency.Non-intrusive load monitoring(NILM)can reduce the cost of obtaining load information on a large scale and avoid interfering with the daily electricity consumption behavior of residential customers.At present,NILM is difficult to effectively detect the start and end of load events,which affects the accurate extraction of load features,and has low accuracy in load identification for appliances with similar steady-state features.In practical application,the performance of load identification is affected by the unknown appliances.In view of the above problems,it is a practical research topic to accurately extract load transient process information,improve the accuracy of load identification by making comprehensive use of load transient and steady-state features,and design a non-intrusive load monitoring system that can adapt to monitoring scenarios with unknown appliances.In this paper,we focus on the load event detection,load recognition and design of load monitoring system in NILM,and the main work is as follows:(1)A non-intrusive load event detection method based on an improved Moving Average Change and a two-way moving t-test is proposed.Firstly,the traditional Moving Average Change(MAC)algorithm is analyzed and the problem of false detection and missed detection caused by power fluctuation,fixed detection threshold,and duration of load events is solved.Then,an improved MAC-based change-point detection algorithm is proposed to improve the event detection accuracy.To solve the problem that existing load event detection methods cannot effectively detect the start and end points of long transient load events,which affects the accuracy of subsequent feature extraction,this chapter uses a two-way moving t-test to determine the start and end points of load events based on the improved MAC variable point detection method.The detection effectiveness of the proposed method is compared and validated on a publicly available dataset.(2)A non-intrusive load recognition algorithm considering transient features is proposed.To solve the problem of low accuracy of existing load identification methods for identifying loads with similar steady-state features,the transient and steady-state voltage and current data of the appliances are obtained by the load separation algorithm and the typical transient and steady-state features are extracted.Considering the intra-class variety of the appliances and the unbalanced datasets decrease the load recognition accuracy,the mean-shift clustering algorithm and adaptive synthetic oversampling(ADASYN)were used respectively.To solve the problem that the traditional recursive feature elimination(RFE)algorithm is ineffective in selecting feature sets with high redundancy,an improved RFE-RF feature selection algorithm is proposed,which uses the maximal information coefficient(MIC)to measure the redundancy of features and selects the optimal feature subset.Finally,to solve the problem that the voting mechanism of the random forest classifier does consider the classification performance of decision trees,which decreases the accuracy of load recognition,the classification margin is used to evaluate the classification performance of decision trees.Finally,the proposed method is compared and validated on the PLAID dataset.(3)An unknown appliance detection algorithm based on isolated forest and an unknown appliance annotation algorithm combined with manual naming are proposed.To solve the problems of low detection accuracy of unknown appliance detection,an isolated forest-based unknown appliance detection algorithm is proposed,which utilizes the advantages of low computational complexity and strong global detection performance of isolated forests and uses typical steady-state and transient features.The proposed algorithm is validated on the PLAID dataset.In terms of unknown appliance annotation,a two-layer mean-shift clustering algorithm is proposed,which can effectively distinguish different unknown appliances by transient power feature.Then take advantage of the high accuracy of manual naming methods to label unknown appliances.Considering the relationship between known appliance monitoring and unknown appliance learning,during the actual operation of NILM,an adaptive non-intrusive residential electricity load monitoring framework is proposed to coordinating the monitoring process of known appliances and the learning process of unknown appliances.A non-intrusive load monitoring system is designed on this basis.The monitoring system can learn the common household appliances of users and provide accurate feedback on the electricity consumption information after a short period of operation in a test user,which has certain practical significance.