Research On Non-intrusive Appliance Load Monitoring And User's Behavior Based On Machine Learning

Energy conservation is an important social consciousness in today's world.Effective energy management can help reduce energy consumption.The popularization of smart grid and the development of smart home have effectively improved the efficiency of energy utilization,promoted the implementation of energy saving and emission reduction and sustainable development strategy,and also ensured the safety and reliability of power supply,and helped users live more intelligent and convenient.However,due to the huge cost of time and investment in traditional intrusive load monitoring systems such as smart household appliances and smart plugs,more economical and effective non-intrusive load monitoring research is gradually emerging.Non-intrusive appliance load monitoring technology,through monitoring and analysis of the total load,obtains the energy consumption of a single electrical appliance.The hardware facilities are simple,and easy to install,remove and maintain.It has a wide range of value for utilities,government departments,appliance manufacturers and other industries as well as user groups.Aiming at the problem that people are not paying enough attention to the application field of non-intrusive load monitoring,this paper proposes a new idea of recognizing the functions of branch circuits in household distribution box,and then implementing abnormal detection of user's behavior.Based on this idea,a new non-intrusive appliance load monitoring tool,Smart DB,is used to collect power data from an experimental family.And according to the characteristics of Smart DB which can simultaneously monitor the energy consumption of 10 branch circuits in the household distribution box,from the point of activities of daily living,the multiple branch circuits in household distribution box are divided into seven categories according to their functions.And then the functions of each branch circuit are identified.In order to improve the generalization ability of the recognition algorithm for different families,this paper integrates the experimental data with several common data sets published by other research groups.Four classification algorithms,support vector machine,decision tree,random forest and extreme gradient boosting,are used to train and model the data sets respectively.Finally,the performance advantages of the integrated classifier over the single classifier are verified.Then,the classification effect of the two integrated classification algorithms,random forest and extreme gradient boosting,for each category is analyzed in detail by comparison.After identifying the functional categories of each branch circuit,this paper selects the experimental data of two of them,combined with principal component analysis,K-means and isolation forest three algorithms,and conduct pattern learning and anomaly detection on the eating and bathing behaviors of experimental families respectively.It finds out the regularity of eating behavior and bathing behavior in the experimental family and detects the abnormal situations in the two behavior patterns.In order to verify the practicability and applicability of the anomaly detection algorithm,the final anomaly detection results are compared and analyzed with the activity log provided by experimental users.It is concluded that the abnormal points detected by the algorithm can remind users of safety,health or benefits to a certain extent,and help users improve their quality of life.