Research On Load Disaggregation Algorithms For User Environment Of Energy Internet

In recent years,with the development of Internet of Things(Io T)and machine learning in smart grid,Energy Internet(EI)has become a new form of energy development with the deep integration of internet and power energy production,transmission,storage,and consumption.As the user environment is at the end of energy consumption,intelligent load monitoring has become an important application scenario of Energy Internet on the user side and the main goal of Energy Internet construction.Non-intrusive intelligent load monitoring is an important breakthrough to realize Energy Internet and has become a research hotspot in the field of energy monitoring in recent years.As an important part of intelligent load monitoring,non-intrusive load disaggregation algorithms have become the research content that many researchers are vying to carry out and have made great development.However,there are still huge challenges in the research of load disaggregation algorithms in user environment that need to be solved urgently.For example,the accuracy of load disaggregation algorithms based on optimization theory algorithm still have room to be improved;the accuracy of load disaggregation algorithms based on factor hidden Markov model have been improved,but the models still have the problem of large computational time complexity;the high-precision load disaggregation algorithms based on machine learning still have the problem of excessive dependence on the labeled dataset in practical applications of the user environment.Therefore,how to improve the accuracy of the load disaggregation algorithms,improve the computational efficiency of the models,and reduce the amount of labeled data that the models depend on has become a key issue in optimizing the load disaggregation algorithms,which has greater theoretical value and practical significance.This work adopts the energy internet users' environment as the application background,by means of machine learning technology,to study the accuracy of the load disaggregation algorithms based on optimization theory,the computation complexity based on the factor hidden Markov model,and the labeled sample data problem in neural network model of the research on the key techniques.The main work and contributions are as follows:(1)Aiming at solving the problem of low accuracy of load disaggregation algorithms based on optimization theory,an improved load disaggregation algorithm based on sparse dictionary learning is proposed.The problems of load disaggregation algorithms are as follows,firstly,the load disaggregation algorithm based on state-based combinatorial optimization models the electrical load signal as a state quantity,which makes the model lose part of the electrical information.At the same time,traditional statistical signal processing algorithms only consider the orthogonal basis functions,making load disaggregation algorithms with low-frequency load data more challenging.Thus,by studying the sparse dictionary learning to optimize each step of the modeling process,the problem of slow convergence speed was found.By optimizing the optimal matrix segmentation period,the speed and accuracy of the load disaggregation algorithm were improved,and the load disaggregation algorithm was used in safety application of electrical scenes.Experimental results show that the algorithm can improve the accuracy of load disaggregation on the REDD dataset.The mean absolute error of the load disaggregation algorithm is 8.26,which is 70%lower than the benchmark algorithms;the root mean square error of the load disaggregation algorithm is 97.75,which is obvious better than the benchmark algorithms.(2)Aiming at solving the problem of high time complexity in the load disaggregation algorithm for real-time system,a load disaggregation algorithm based on fast event monitoring is proposed.To meet the requirement of energy-saving and safety in the user environment,the state-based load disaggregation model has the advantage of low computational time complexity under the premise of ensuring the accuracy of load disaggregation.However,most of these studies mainly focus on the optimization and simplification of factorial hidden Markov models,the computational time complexity of such models has not been deeply applied in practice.In this work,the electrical appliances in the user environment are modeled as a unified super-state hidden Markov model and based on the sparse Viterbi algorithm for load disaggregation algorithm,the attributes of aggregated data under multiple time slots are considered,and fast event detection is adopted,the method realizes the fast solution of sparse Viterbi algorithm in solving such problems.Experimental results show that the algorithm can greatly improve the calculation efficiency of load disaggregation.For example,when the number of appliances on the reference load dataset is19,the maximum calculation time is 7.2 seconds,compared with the calculation of the traditional sparse Viterbi load disaggregation algorithm,which is 300 seconds,it is reduced to2.4% of the comparison algorithm.The average calculation time is 2.83 seconds,which is2.6% of the traditional Viterbi algorithm.(3)Aiming at solving the problem of high cost and difficult to obtain labeled data in the promotion and deployment of load disaggregation algorithm,a load disaggregation algorithm based on unsupervised learning algorithm with operating curve reconstruction is proposed.Firstly,through the idea of event detection and event pair matching,the typical operating curve of electrical appliances is directly generated on the aggregated data.Then,the typical operating curve and typical operating cycle of each electrical appliance are reversely synthesized into the operating cycle dataset of the electrical appliance.Then,the dataset is superimposed on a single appliance according to a certain rule.Finally,the new synthetic dataset is used to train the load disaggregation field with higher accuracy and better application of supervised learning algorithms,to obtain the neural network model parameters of each electrical appliance,to realize the unlabeled load disaggregation algorithm.Experimental shows that the algorithm can realize the load disaggregation algorithm of the unsupervised learning algorithm under the premise of the data of the typical electrical appliance operating curve and achieve the result of approximate accuracy with the real dataset.(4)Aiming at the situation that the edge computing ability of the load disaggregation algorithm is enhanced in actual deployment,an experimental method of the load disaggregation algorithm enhanced by the edge appliance identification algorithm is proposed.With the enhancement of computing power of edge devices and inspired by experiments with actual high-frequency current waveform data,current load disaggregation algorithms mostly rely on cloud training and computational inference.However,the demand for energy saving and safety has greatly increased the demand for local inference on edge devices.Therefore,this work studies the experimental method to support the data collection,data collection,model training,model reasoning and model generation,model download and model deployment of the electrical appliance identification and load decomposition algorithm,to meet the requirements of the experimental simulation system.Based on an open and extensible microservice simulation software architecture,a simulation interface that supports machine learning data access is opened.Through training configuration,reasoning demonstration provides support for the experimental simulation system of electrical appliance identification and load disaggregation.Based on this system,the experiment simulation of some electric appliance identification and load disaggregation algorithms was carried out,and the visual verification was carried out.The simulation results show that the system supports the use of small neural network model algorithms to train electrical appliance identification algorithms,and the accuracy of electrical appliance identification on the edge device side can reach 84%,and the load disaggregation algorithm accuracy can be increased by 10%,which promotes appliance identification and application of load disaggregation algorithm.