Research On Non-intrusive Load Decomposition Technology Based On Data Modeling

With the rapid development of China’s construction of a resource-saving society,the implementation of national big data strategy,and smart grid construction,the non-intrusive power load monitoring system has become the development direction of power demand side management,energy saving,and safe power use in the future,and is also used by refined users.The basis of energy conservation and emission reduction is realized by electric management and intelligent electricity interaction.The existing non-intrusive power load monitoring systems mostly identify the operation of a single power load.However,in the actual power load operating environment,the types and quantities of loads are relatively large,and the existing algorithms have poor recognition effects on the scene.The accuracy is low and the load characteristics are seriously lost.Aiming at this problem,this paper designs an intrusive power load decomposition method with full voltagecurrent waveform as the power load characteristic and implements hardware based on STM32 microcontroller.Firstly,this paper analyzes the advantages and disadvantages of typical non-intrusive load monitoring framework and various load feature selection methods.Under the data-driven framework,the load voltage-current waveform characteristics are used as load characteristics,and each load model is established.The way to replace the extraction of load characteristics in a typical framework.Based on this,an optimized non-intrusive load monitoring system framework is presented.Secondly,BP neural network,recurrent neural network and RBF neural network based on support vector machine are used to model the voltage-current characteristics of power load and no-load,respectively.Using four indicators quantitative evaluation,we find RBF nerve based on support vector machine.The network load modeling method has the highest accuracy and has a good effect on power load modeling.The current model of each load is obtained by RBF neural network training to form a load model library.And then,the genetic algorithm is used to find the combination of the current and actual current of the load model combination.The optimal combination is the type and quantity of the actual load.In order to accurately evaluate the similarity of the two current waveforms,this paper designs a current similarity evaluation function,which uses the weighted values of root mean square error,correlation coefficient and correlation entropy as the basis for evaluation.Information such as the degree of deviation,shape,and trend of the waveform provides a quantitative basis for the load monitoring system.In the experimental environment of this paper,the average accuracy of this method is 99.0%,which has a high practical price.Finally,based on STM32 single-chip microcomputer,a non-intrusive load monitoring system hardware platform is designed.The hardware circuit design,main software design and PCB design of the monitoring system are introduced in detail.The non-intrusive load decomposition of this paper is further verified by STM32 single-chip hardware platform.The effectiveness of the algorithm verifies the practicability of the monitoring system.