CN110501568B - Non-invasive equipment load monitoring method based on graph signal processing - Google Patents

Non-invasive equipment load monitoring method based on graph signal processing Download PDF

Info

Publication number
CN110501568B
CN110501568B CN201910606143.XA CN201910606143A CN110501568B CN 110501568 B CN110501568 B CN 110501568B CN 201910606143 A CN201910606143 A CN 201910606143A CN 110501568 B CN110501568 B CN 110501568B
Authority
CN
China
Prior art keywords
time
graph
active power
equipment
moment
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910606143.XA
Other languages
Chinese (zh)
Other versions
CN110501568A (en
Inventor
赵生捷
张冰
张荣庆
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tongji University
Original Assignee
Tongji University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tongji University filed Critical Tongji University
Priority to CN201910606143.XA priority Critical patent/CN110501568B/en
Publication of CN110501568A publication Critical patent/CN110501568A/en
Application granted granted Critical
Publication of CN110501568B publication Critical patent/CN110501568B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01RMEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R21/00Arrangements for measuring electric power or power factor
    • G01R21/133Arrangements for measuring electric power or power factor by using digital technique
    • G01R21/1331Measuring real or reactive component, measuring apparent energy

Abstract

The invention relates to a non-invasive equipment load monitoring method based on graph signal processing, which comprises three steps of data preprocessing, initial point acquisition and target function integral optimization, wherein the data is preprocessed firstly, and a proper time period is selected for each known equipment to be used as a training set; then determining a target function, and solving the minimum value of the regularization item of the target function to obtain an analytic solution; and finally, taking the obtained analytic solution as a starting point, and executing a gradient projection optimization algorithm to carry out overall optimization solution on the objective function. Compared with the prior art, the method has the advantages of improving the algorithm performance, obtaining a better power separation result, realizing the load monitoring of the non-invasive equipment with higher precision and the like.

Description

Non-invasive equipment load monitoring method based on graph signal processing
Technical Field
The invention relates to the field of non-intrusive equipment load monitoring, in particular to a non-intrusive equipment load monitoring method based on graph signal processing.
Background
Non-intrusive application Load Monitoring (NILM), that is, the total power consumption of a household at a certain moment measured by a total circuit ammeter is given, and the actual power consumption of each electrical equipment corresponding to the moment is obtained by applying some analysis and calculation modes. In this process, the power consumption of a particular device need not be measured by means of a power meter. Therefore, the NILM has the advantages of easy use, low cost, no equipment interference, no influence on family life, and the like. From the viewpoint of energy saving, NILM plays an irreplaceable important role. For common household users, the NILM can help the users to know the power consumption conditions of various household appliances, and the users can timely turn off the high-power-consumption appliances on the basis, so that the saving consciousness of people is invisibly aroused. For policy makers, NILM can help them to develop a more macroscopic and in-depth understanding of home electricity usage, helping them to make more reasonable energy feedback mechanisms and demand response strategies.
In the existing research, when the Graph Signal processing is applied to the NILM field, the definition of the Graph Signal (Graph Signal) is a discrete function with values of-1, 0 and 1, so that the final objective function is a discrete and non-derivable function, and the use of an optimization algorithm is limited. Meanwhile, the objective function is a discrete and non-derivable function, so most of the existing methods regard the objective function as two parts, and then carry out segmentation and step processing. Firstly, solving the minimum value of the regularization term which is taken as the second part of the target function, then, taking the solution as a starting point, and optimizing the first part of the target function by adopting heuristic algorithms such as simulated annealing and the like, so that the calculation is complex, and the accuracy of the electric power separation result is not high.
Disclosure of Invention
The present invention is directed to overcoming the above-mentioned drawbacks of the prior art and providing a non-intrusive device load monitoring method based on graph signal processing.
The purpose of the invention can be realized by the following technical scheme:
a non-invasive equipment load monitoring method based on graph signal processing comprises three steps of data preprocessing, initial point obtaining and target function integral optimization, wherein data are preprocessed firstly, and proper time periods are selected for known equipment to serve as training sets; then determining a target function, and solving the minimum value of the regularization item of the target function to obtain an analytic solution; and finally, taking the obtained analytic solution as a starting point, and executing a gradient projection optimization algorithm to carry out overall optimization solution on the objective function.
Further, the method specifically comprises the following steps:
s1, acquiring the total load active power p of the user equipment in the set time period from 1 to NiAs monitoring data; obtaining the single active power of each device in a set time period 1-nAs training data, where N and N are positive integers and N is less than N, i is a time, m is the number of the device, and all devices constitute a device set
S2, calculating the average power consumption of each device based on the training dataAnd according to average power consumptionArranging in descending order from big to small;
s3, according to the monitoringData obtaining total load active power change value delta p at each momentiConstructing a graph for each device m by graph signal processing A vertex set comprising a plurality of vertexes, each vertex corresponds to an active power variation value delta piA is a weighted adjacency matrix with a weight Ai,jFor corresponding vertex viAnd vjThe weights of the connecting edges between, i and j represent different time instants;
s4, obtaining the individual active power change value of each device at each moment according to the training dataBased on constructed graphsCalculating all graph signals within 1-N timeThe calculation expression is as follows:
wherein, ThrmIs a threshold value for determining whether the device m has changed the operation state,is the single active power change value of the device m in a set time period 1-n;
s5, according to all graph signals within 1-n timeCalculating and acquiring graph signals from N +1 time to N timeBy passingAndmapping between the devices to obtain the actual operation power consumption change value of each device;
s6, repeating the steps S3 to S5, and removing the average power consumption of the running state of the equipment in the monitoring data after separating one equipment each time until all the equipment is separated;
s7, separating all the devices to obtain a starting point Pn+1,...,PNWherein
S8, obtaining the starting point Pn+1,...,PNSolving the optimization model to obtain and output the optimized time from N +1 to N
Further, in step S3, the adjacency matrix a is weightedi,jThe calculation expression of (a) is:
wherein, Δ piRepresenting the value of the change in the active power of the total load at time i, Δ pjAnd expressing the change value of the active power of the total load at the moment j, and expressing a scaling factor.
Further, in step S5, the graph signalThe calculation expression of (a) is:
where L is the graph laplacian matrix, which is calculated from the graph adjacency weight matrix a.
Further, in step S7, the mapping process is as follows:
if it is notIt is judged that the device m has changed the operation state at the time i when Δ piIf < 0, the judgment device is turned off at the moment i, thenWhen Δ piWhen the current time is more than 0, the judgment device is turned on at the moment i, and then
If it is notIt is judged that the device m does not change the operation state and accordingly
Further, in step S8, the expression of the optimization model is:
wherein, Δ piRepresenting the value of the total load active power change at the moment i,and the active power change value of the equipment m at the moment i is represented, and omega represents a regularization term weight parameter.
Compared with the prior art, the invention has the following advantages:
the invention uses the graph signal processing technology to establish a model for the NILM problem, and adopts an optimization algorithm to solve an objective function so as to obtain an optimal solution and realize the separation of the total power consumption of the electric power. The invention improves the definition of the graph signal, and defines the graph signal as a continuous function with the value in the range of [ -1,1], so that the final objective function is a continuously derivable function. Therefore, after the solution of the minimized regularization term is obtained by the objective function, the objective function can be used as a starting point to execute a gradient projection method, the objective function is optimized and solved integrally, the algorithm performance is improved, a better power separation result is obtained, and the load monitoring of the non-invasive equipment with higher precision is realized.
Drawings
FIG. 1 is a schematic overall flow diagram of the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
As shown in fig. 1, the present embodiment provides a non-intrusive device load monitoring method based on graph signal processing. The method mainly comprises three steps of data preprocessing, initial point acquisition and target function overall optimization. Firstly, preprocessing data, and selecting a proper time period for each known device as a training set; then determining a target function, and solving the minimum value of the regularization item of the target function to obtain an analytic solution; and finally, taking the obtained analytic solution as a starting point, and executing a gradient projection optimization algorithm to carry out overall optimization solution on the objective function.
The solving of the NILM problem by applying the graph signal processing technology is to regard the NILM as a classification problem, that is, to the change of the total power consumption at a certain time, it is determined which known device caused the change, so that the operation state of the device can be determined. The embodiment selects suitable data as the training set.
In order to better capture the running condition of each device and improve the decomposition precision, the selection principle of the device training set follows: the time period in which only the device operates alone is selected as its training set to reduce interference from unrelated operating devices. In addition, for each device with the selected training set, the training set is stored separately, when a new electrical device is installed, the training set is only required to be built for the new device, and the training sets of the rest known devices are not required to be updated again.
The present embodiment employs a REDD (reference energy decomposition data set) database containing data of total active power in a home and active power of each device. I.e. for a certain time i the total active power is piThe active power of a certain device m corresponding to the moment isAll known devices constitute a device set
The embodiment specifically includes the following steps:
step S1, obtaining the total load active power p of the user equipment in the set time period 1-NiAs monitoring data; obtaining the single active power of each device in a set time period 1-nAs training data, where N and N are both positive integers and N is less than N, and m is the number of the device.
Step S2, calculating the average power consumption of each device based on the training dataAnd to make the device power consumption on averageArranged in descending order from large to small.
Step S3, obtaining the total load active power change value delta p at each moment according to the monitoring dataiObtaining the power change value delta p at each momenti=pi+1-pi
Step S4Construction diagram of signal processing technology using diagramWherein the content of the first and second substances,a vertex set comprising a plurality of vertexes, each vertex corresponds to an active power variation value delta piA is a weighted adjacency matrix with a weight Ai,jFor corresponding vertex viAnd vjThe connecting edge weights between, i and j represent different time instants. Weight Ai,jThe value of (b) reflects the similarity of two adjacent vertices, i.e., the greater the weight, the higher the similarity between adjacent vertices. The weight calculation formula typically uses a gaussian kernel weight function, i.e.:
wherein, Δ piAnd Δ pjAnd respectively representing the change values of the active power of the total load corresponding to the time i and the time j, wherein sigma represents a scaling factor.
Step S5, based on constructed graphCalculating all graph signals within 1-N time
Graph signal s is an important concept used in the present invention and is defined by a mapping of a set of vertices to a set of numbers, each vertex viMarking a picture signal si. Based on constructed graphsPicture signal s of device mmFor a vector of length N, the values are defined as follows
Wherein, ThrmIs a threshold value used to determine whether device m has changed operating conditions, typically at a value of half the difference between the average powers of successive states of device m.As class labels, in the training set (i ≦ n), the individual active power change value of device m at time iIs known, so if the change in total power consumption at time i is caused by a change in state of device m, then there isApproaching +1, otherwiseApproaching to-1; in the test set (N < i ≦ N),is unknown, thereforeTemporarily set to 0.
Step S6, based on the graph total smoothness minimization, obtains an initial analytic solution. The overall smoothness of the graph signal is defined by the graph laplacian matrix as:
wherein, L is a graph laplacian matrix, defined as L ═ D-a, and is an N × N real symmetric matrix; d is a diagonal matrix, i.e. the value of the element on the diagonal is defined as Di,i=∑j=1:NAi,jAnd the off-diagonal element is 0. Based on the above definition, similar data points viAnd vjA connection weight value ofi,jIs large and itCorresponding map signal value siAnd sjAnd (4) approaching. Thus, for data classification, the graph signal constitutes a smooth signal, and when the graph signal is piecewise smooth, the overall graph smoothness value is small.
For the NILM problem to be solved by the present invention, the corresponding mathematical model is expressed as:
wherein n isiThe noise at the moment i mainly comprises the power consumption of a circuit-based load and the power consumption of other unknown devices. From the viewpoint of variation in power consumption, the formula (4) can be modified into the following form
For the training set (i ≦ n), Δ piAndare all known; for the test set (N < i ≦ N), Δ piIs known asIs unknown.
Based on the graph signal processing technique, using the overall smoothness of the graph signal as the regularization term, the final NILM objective function is then defined as:
wherein the content of the first and second substances,smis the map signal of the corresponding device m. Therefore, the NILM task is used as an optimization problem to find the optimal solution, namely the power consumption change value of each device m at each moment i (N is more than i and less than or equal to N)
To obtain an initial point Pn+1,...,PNFor each device m, an optimization problem is required to be solved to obtain the smoothest solution. I.e., the regularization term corresponding to the minimization device m, is expressed as:
since the matrix L is diagonally symmetrical,is a known constant vector, and utilizes a block matrix, and the solving process of the formula is as follows:
solved to obtainThen throughAndcan obtain the required starting point Pn+1,...,PN. The mapping process is as follows: if it is notIt can be concluded that device m has changed operating state at time i, otherwise device m has not changed operating state and is accordinglyIf the device m changes its operating state, when Δ pi< 0, the inference device is turned off at time i, haveWhen Δ p is otherwiseiWhen > 0, the inference device is turned on at time i, have
The above process is repeated each time a device is separated according to its average operating powerIn decreasing order. When a device is detached and gets the correspondingThe average power consumption of the device in the operating state is then subtracted from the total power consumption to minimize interference for subsequent device separation.
After the solution of the minimized regularization term is obtained as a starting point, a gradient projection method is executed to carry out optimization solution on the whole objective function (6), and the solution problem at the moment can be regarded as a constraint nonlinear programming problem and is expressed as a constraint nonlinear programming problem
Where Ω constrains the range of the solution set to ensure that the resulting solution, i.e., the power consumption of each device, does not vary beyond what is actually allowed. Finally obtaining the separation result
Evaluation of Performance
Commonly used performance evaluation criteria in the NILM field are Recall (RE), Precision (PR), and F1-score. Each of which is defined as follows
RE=TP/(TP+FN)
PR=TP/(TP+FP)
F1=2×(PR×RE)/(PR+RE),
Namely, tp (true positive) represents that the state change of the device is correctly recognized, fp (false positive) represents that the device is erroneously recognized as the state change, and fn (false negative) represents that the state change of the device is not correctly recognized.
Therefore, the recall Rate (RE) measures the proportion of correctly identified events in all device state change events, the accuracy rate (PR) measures the correct rate of identification of device state change events, and F1-score is a trade-off between the two, with higher values indicating better separation performance of the method.
The performance of the method proposed in this example was verified by performing experiments on House 1 and House 2 of the REDD data set. Tables 1 and 2 are a comparison of the performance of the present invention with 3 other prior NILM methods, Graph Signal Processing (GSP) -based, Hidden Markov Model (HMM) -based, and Decision Tree (DT) -based. As can be seen from tables 1 and 2, the accuracy of the power data separation result of the method of the present invention is better.
Table 1 comparison of REDD House 1 power data separation results F1 performance values
Table 2 comparison of REDD House 2 power data separation results F1 performance values
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.

Claims (3)

1. A non-invasive equipment load monitoring method based on graph signal processing is characterized by comprising three steps of data preprocessing, initial point obtaining and target function integral optimization, wherein data are preprocessed firstly, and proper time periods are selected for known equipment to serve as training sets; then determining a target function, and solving the minimum value of the regularization item of the target function to obtain an analytic solution; finally, taking the obtained analytic solution as a starting point, and executing a gradient projection optimization algorithm to carry out overall optimization solution on the objective function; the method specifically comprises the following steps:
s1, acquiring the total load active power p of the user equipment in the set time period from 1 to NiAs monitoring data; obtaining the single active power of each device in a set time period 1-nAs training data, where N and N are positive integers and N is less than N, i is a time, m is the number of the device, and all devices constitute a device set
S2, calculating the average power consumption of each device based on the training dataAnd according to average power consumptionArranging in descending order from big to small;
s3, obtaining the total load active power change value delta p at each moment according to the monitoring dataiFor each device mConstructing a graph by graph signal processing A vertex set comprising a plurality of vertexes, each vertex corresponds to an active power variation value delta piA is a weighted adjacency matrix with a weight Ai,jFor corresponding vertex viAnd vjThe weights of the connecting edges between, i and j represent different time instants;
s4, obtaining the individual active power change value of each device at each moment according to the training dataBased on constructed graphsCalculating all graph signals within 1-N timeThe calculation expression is as follows:
wherein, ThrmIs a threshold value for determining whether the device m has changed the operation state,is the single active power change value of the device m in a set time period 1-n;
s5, according to all graph signals within 1-n timeCalculating and acquiring graph signals from N +1 time to N timeBy passingAndmapping between the devices to obtain the actual operation power consumption change value and graph signal of each deviceThe calculation expression of (a) is:
wherein, L is a graph Laplace matrix which is obtained by calculating an adjacent weight matrix A of the graph;
s6, repeating the steps S3 to S5, and removing the average power consumption of the running state of the equipment in the monitoring data after separating one equipment each time until all the equipment is separated;
s7, separating all the devices to obtain a starting point Pn+1,...,PNWherein
S8, obtaining the starting point Pn+1,...,PNSolving the optimization model to obtain and output the optimized time from N +1 to NThe expression of the optimization model is as follows:
wherein, Δ piRepresenting the value of the total load active power change at the moment i,and the active power change value of the equipment m at the moment i is represented, and omega represents a regularization term weight parameter.
2. The non-invasive device load monitoring method based on graph signal processing according to claim 1, wherein in step S3, the adjacency matrix a is weightedi,jThe calculation expression of (a) is:
wherein, Δ piRepresenting the value of the change in the active power of the total load at time i, Δ pjAnd expressing the change value of the active power of the total load at the moment j, and expressing a scaling factor.
3. The method for non-intrusive device load monitoring based on graph signal processing as defined in claim 1, wherein in step S5, the mapping procedure is as follows:
if it is notIt is judged that the device m has changed the operation state at the time i when Δ piIf < 0, the judgment device is turned off at the moment i, thenWhen Δ piWhen the current time is more than 0, the judgment device is turned on at the moment i, and then
If it is notIt is judged that the device m does not change the operation state and accordingly
CN201910606143.XA 2019-07-05 2019-07-05 Non-invasive equipment load monitoring method based on graph signal processing Active CN110501568B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910606143.XA CN110501568B (en) 2019-07-05 2019-07-05 Non-invasive equipment load monitoring method based on graph signal processing

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910606143.XA CN110501568B (en) 2019-07-05 2019-07-05 Non-invasive equipment load monitoring method based on graph signal processing

Publications (2)

Publication Number Publication Date
CN110501568A CN110501568A (en) 2019-11-26
CN110501568B true CN110501568B (en) 2020-11-27

Family

ID=68585401

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910606143.XA Active CN110501568B (en) 2019-07-05 2019-07-05 Non-invasive equipment load monitoring method based on graph signal processing

Country Status (1)

Country Link
CN (1) CN110501568B (en)

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104408480B (en) * 2014-11-28 2018-05-04 安徽师范大学 A kind of feature selection approach based on Laplacian operators
CN105913141A (en) * 2016-04-06 2016-08-31 上海电器科学研究院 Load prediction method supporting big data
CN106096726B (en) * 2016-05-31 2018-12-18 华北电力大学 A kind of non-intrusion type load monitoring method and device
CN106443233A (en) * 2016-08-26 2017-02-22 北京电力经济技术研究院 Non-invasive steady-state load monitoring method
CN106447085A (en) * 2016-09-05 2017-02-22 华北电力大学 Short-term power load forecasting method and device and computation equipment
CN106600074B (en) * 2016-12-28 2020-08-18 天津求实智源科技有限公司 DFHSMM-based non-invasive power load monitoring method and system
CN107546855B (en) * 2017-09-28 2020-05-08 广东工业大学 Non-invasive decomposition method for residential electricity load
CN108899892A (en) * 2018-06-08 2018-11-27 上海电力学院 One kind being based on CNN non-intrusion type power load decomposition method
CN109462231B (en) * 2018-11-15 2020-09-01 合肥工业大学 Load optimization scheduling method, system and storage medium for residential micro-grid
CN109508908A (en) * 2018-12-25 2019-03-22 深圳市城市公共安全技术研究院有限公司 Non-intrusion type load identification model training method, load monitoring method and device

Also Published As

Publication number Publication date
CN110501568A (en) 2019-11-26

Similar Documents

Publication Publication Date Title
Guo et al. Home appliance load modeling from aggregated smart meter data
Jung et al. Prediction of building energy consumption using an improved real coded genetic algorithm based least squares support vector machine approach
Zhao et al. Improving event-based non-intrusive load monitoring using graph signal processing
Stankovic et al. A graph-based signal processing approach for low-rate energy disaggregation
Pindoriya et al. Composite reliability evaluation using Monte Carlo simulation and least squares support vector classifier
Yunpeng et al. Multi-step ahead time series forecasting for different data patterns based on LSTM recurrent neural network
CN110610280B (en) Short-term prediction method, model, device and system for power load
CN102855634A (en) Image detection method and image detection device
WO2012160062A1 (en) Transition detection method for automatic-setup non-intrusive appliance load monitoring
CN108021736B (en) Load switching action monitoring method based on sliding window residual error model
CN109359665B (en) Household appliance load identification method and device based on support vector machine
CN109813978B (en) Non-intrusive load type identification method integrating transient characteristic and change characteristic between stable states
De Baets et al. VI-based appliance classification using aggregated power consumption data
Goodwin et al. A pattern recognition approach for peak prediction of electrical consumption
Quek et al. Load disaggregation using one-directional convolutional stacked long short-term memory recurrent neural network
Chen et al. NeuCast: Seasonal Neural Forecast of Power Grid Time Series.
Wang et al. Big data analytics for price forecasting in smart grids
CN110501568B (en) Non-invasive equipment load monitoring method based on graph signal processing
Dinesh et al. Non-intrusive load monitoring based on low frequency active power measurements
Kumar et al. Data-driven electrical load disaggregation using graph signal processing
Tongta et al. Long short-term memory (LSTM) neural networks applied to energy disaggregation
CN106897553B (en) Home intelligent power method for establishing model based on single-sensor
CN110837915A (en) Low-voltage load point prediction and probability prediction method for power system based on hybrid integrated deep learning
Leonardi et al. Detecting smart plug configuration changes in smart homes
Gómez-Boix et al. Consumer segmentation through multi-instance clustering time-series energy data from smart meters

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant