阅读说明：本技术 一种基于图信号处理的非侵入式设备负载监测方法 (A kind of non-invasive apparatus load monitoring method based on figure signal processing ) 是由 赵生捷 张冰 张荣庆 于 2019-07-05 设计创作，主要内容包括：本发明涉及一种基于图信号处理的非侵入式设备负载监测方法,包括数据预处理、起始点获取和目标函数整体优化三个步骤,首先对数据进行预处理,为各已知设备选取合适的时间段作为训练集；然后确定目标函数,对目标函数的正则化项进行最小值求解,获得解析解；最后将已获得的解析解作为起始点,执行梯度投影优化算法对目标函数进行整体优化求解。与现有技术相比,本发明具有提高了算法性能,获得更好的电力分离结果,实现更高精度的非侵入式设备的负载监测等优点。(The non-invasive apparatus load monitoring method based on figure signal processing that the present invention relates to a kind of, including data prediction, starting point obtains and three steps of objective function global optimization, data are pre-processed first, choose the suitable period as training set for each known device；Then it determines objective function, minimum value solution is carried out to the regularization term of objective function, obtains analytic solutions；Finally using acquired analytic solutions as starting point, executes gradient projection optimization algorithm and global optimization solution is carried out to objective function.Compared with prior art, the present invention has the advantages that improving algorithm performance, obtaining better electric power separating resulting, realize the load monitoring of the non-invasive apparatus of higher precision.)

1. a kind of non-invasive apparatus load monitoring method based on figure signal processing, which is characterized in that including data prediction, Starting point obtains and three steps of objective function global optimization, pre-processes first to data, chooses and closes for each known device The suitable period is as training set；Then it determines objective function, minimum value solution is carried out to the regularization term of objective function, is obtained Analytic solutions；Finally using acquired analytic solutions as starting point, executes gradient projection optimization algorithm and entirety is carried out to objective function Optimization Solution.

2. the non-invasive apparatus load monitoring method according to claim 1 based on figure signal processing, which is characterized in that Specifically includes the following steps:

S1, total load active power p of the user equipment in 1~N of set period of time is obtained_i, as monitoring data；It obtains each Independent active power of the equipment in 1~n of set period of timeAs training data, wherein n and N is positive integer and n Less than N, i is a certain moment, and m is the number of equipment, all devices constitution equipment collection

S2, the operating status average power consumption that each equipment is calculated based on training dataAnd according to average power consumptionFrom big to small Descending arrangement；

S3, the total load active power changing value Δ p that each moment is obtained according to monitoring data_i, for each equipment m, believed by figure Number processing mode structure figures For the vertex set comprising multiple vertex, one wattful power of each vertex correspondence Rate changing value Δ p_i, A is weighted adjacent matrix, weight A_i,jFor corresponding vertex v_iAnd v_jBetween connection side right weight, i and j table At the time of showing different；

S4, the independent active power changing value that each equipment of each moment is obtained according to training dataBased on the figure constructed Calculate all figure signals in 1~n-hourIts calculation expression are as follows:

Wherein, Thr_mBe for judging whether equipment m changes the threshold value of operating status,It is equipment m in set period of time Independent active power changing value in 1~n；

S5, according to all figure signals in 1~n momentCalculate the figure signal for obtaining n+1 moment to n-hourPass through WithBetween mapping, obtain the actual motion change of power consumption value of each equipment；

S6, step S3 to step S5 is repeated, every time after one equipment of separation, the operating status of the equipment is removed in monitoring data Average power consumption, until separation all devices；

Starting point P is obtained after S7, separation all devices_n+1,...,P_N, wherein

S8, starting point P is arrived according to what is obtained_n+1,...,P_NSolving optimization model, obtain and export optimization after at the n+1 moment to N Moment

3. the non-invasive apparatus load monitoring method according to claim 2 based on figure signal processing, which is characterized in that In the step S3, weighted adjacent matrix A_i,jCalculation expression are as follows:

Wherein, Δ p_iIndicate the total load active power changing value at i moment, Δ p_jIndicate the total load active power variation at j moment Value, σ indicate scale factor.

4. the non-invasive apparatus load monitoring method according to claim 2 based on figure signal processing, which is characterized in that In the step S5, figure signalCalculation expression are as follows:

Wherein, L is figure Laplacian Matrix, is calculated by the adjoining weight matrix A of figure.

5. the non-invasive apparatus load monitoring method according to claim 2 based on figure signal processing, which is characterized in that In the step S7, mapping process is as follows:

IfThen judge that equipment m changes operating status in moment i, as Δ p_iWhen < 0, judge that equipment is closed at the i moment It closes, thenAs Δ p_iWhen > 0, judge that equipment is opened at the i moment, then

IfThen judge that equipment m does not change operating status and corresponding

6. the non-invasive apparatus load monitoring method according to claim 2 based on figure signal processing, which is characterized in that In the step S8, the expression formula of Optimized model are as follows:

Wherein, Δ p_iIndicate the total load active power changing value at i moment,Indicate the active power variation of i moment equipment m Value, ω indicate regularization term weight parameter.

Technical field

The present invention relates to non-invasive apparatus load monitoring fields, are invaded more particularly, to a kind of based on the non-of figure signal processing Enter formula apparatus of load monitoring method.

Background technique

Non-invasive apparatus load monitoring (Non-intrusive Appliance Load Monitoring, NILM), i.e., The total electricity power consumption at the given family's a certain moment obtained by way circuit ammeter measurement is obtained using some analyses, the mode calculated The practical electric power power consumption of each electrical equipment is corresponded to the moment.In this process, without measuring spy in dividing ammeter The power consumption of locking equipment.Therefore, NILM have it is easy to use, inexpensive, without equipment interference, do not influence the advantages such as family life.From Energy saving angle, NILM play irreplaceable key player.For ordinary family user, NILM can be helped They understand the power consumption situation of each electric appliance in family, and user can close high power consumption electric appliance in time on this basis, this is invisible In excite the economy consciousness of people.For policy maker, NILM can help them to generate household electricity situation One more macroscopical, deep understanding facilitates it and formulates more reasonable energy feedback mechanism and demand response strategy.

In existing research, when by figure signal processing applications into the field NILM, about figure signal (Graph Signal definition) be all a value be -1,0,1 discrete function, this make final objective function be one it is discrete not Guidable function limits the use of optimization algorithm.Simultaneously as objective function is a discrete not guidable function, therefore Most existing methods are all that objective function is regarded as to two parts, are then split, step-by-step processing.I.e. first to as target letter The regularization term of number second part carries out minimum value solution, is then solved as starting point, heuristic using simulated annealing etc. Algorithm optimizes objective function first part, therefore will lead to calculating complexity, and electric power separating resulting precision is not high.

Summary of the invention

It is an object of the present invention to overcome the above-mentioned drawbacks of the prior art and provide one kind based at figure signal The non-invasive apparatus load monitoring method of reason.

The purpose of the present invention can be achieved through the following technical solutions:

A kind of non-invasive apparatus load monitoring method based on figure signal processing, including data prediction, starting point obtain Take with three steps of objective function global optimization, data are pre-processed first, choose suitable time for each known device Duan Zuowei training set；Then it determines objective function, minimum value solution is carried out to the regularization term of objective function, obtains analytic solutions； Finally using acquired analytic solutions as starting point, executes gradient projection optimization algorithm and objective function progress global optimization is asked Solution.

Further, specifically includes the following steps:

S1, total load active power p of the user equipment in 1~N of set period of time is obtained_i, as monitoring data；It obtains Independent active power of each equipment in 1~n of set period of timeAs training data, wherein n and N is positive integer And n is less than N, and i is a certain moment, and m is the number of equipment, all devices constitution equipment collection

S2, the operating status average power consumption that each equipment is calculated based on training dataAnd according to average power consumptionFrom Small descending arrangement is arrived greatly；

S3, the total load active power changing value Δ p that each moment is obtained according to monitoring data_i, for each equipment m, lead to Cross figure signal processing mode structure figures For the vertex set comprising multiple vertex, each vertex correspondence one Active power changing value Δ p_i, A is weighted adjacent matrix, weight A_i,jFor corresponding vertex v_iAnd v_jBetween connection side right weight, i At the time of different with j expression；

S4, the independent active power changing value that each equipment of each moment is obtained according to training dataBased on what is constructed FigureCalculate all figure signals in 1~n-hourIts calculation expression are as follows:

Wherein, Thr_mBe for judging whether equipment m changes the threshold value of operating status,It is equipment m in setting Between independent active power changing value in 1~n of section；

S5, according to all figure signals in 1~n momentCalculate the figure signal for obtaining n+1 moment to n-hourIt is logical It crossesWithBetween mapping, obtain the actual motion change of power consumption value of each equipment；

S6, step S3 to step S5 is repeated, every time after one equipment of separation, the operation of the equipment is removed in monitoring data State average power consumption, until separation all devices；

Starting point P is obtained after S7, separation all devices_n+1,...,P_N, wherein

S8, starting point P is arrived according to what is obtained_n+1,...,P_NSolving optimization model, obtain and export optimization after in n+1 It is carved into n-hour

Further, in the step S3, weighted adjacent matrix A_i,jCalculation expression are as follows:

Wherein, Δ p_iIndicate the total load active power changing value at i moment, Δ p_jIndicate the total load active power at j moment Changing value, σ indicate scale factor.

Further, in the step S5, figure signalCalculation expression are as follows:

Wherein, L is figure Laplacian Matrix, is calculated by the adjoining weight matrix A of figure.

Further, in the step S7, mapping process is as follows:

IfThen judge that equipment m changes operating status in moment i, as Δ p_iWhen < 0, judge equipment in i Quarter is closed, thenAs Δ p_iWhen > 0, judge that equipment is opened at the i moment, then

IfThen judge that equipment m does not change operating status and corresponding

Further, in the step S8, the expression formula of Optimized model are as follows:

Wherein, Δ p_iIndicate the total load active power changing value at i moment,Indicate the active power of i moment equipment m Changing value, ω indicate regularization term weight parameter.

Compared with prior art, the invention has the following advantages that

The present invention uses figure signal processing technology, establishes model for NILM problem, and solve target letter using optimization algorithm Number realizes the separation to electric power total power consumption to obtain optimal solution.The present invention improves the definition of figure signal, is determined Justice is value in the continuous function in [- 1,1] section, this makes final objective function be a continuous guidable function.Therefore, Objective function can execute GRADIENT PROJECTION METHODS as starting point, to target after obtaining the solution for minimizing regularization term Function integrally optimizes, and improves algorithm performance, obtains better electric power separating resulting, realizes that the non-of higher precision is invaded Enter the load monitoring of formula equipment.

Detailed description of the invention

Fig. 1 is overall flow schematic diagram of the invention.

Specific embodiment

The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention Premised on implemented, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to Following embodiments.

As shown in Figure 1, present embodiments providing a kind of non-invasive apparatus load monitoring method based on figure signal processing. It is broadly divided into three steps, i.e. data prediction, starting point acquisition and objective function global optimization.Data are carried out first pre- Processing chooses the suitable period as training set for each known device；Then objective function is determined, to the canonical of objective function Change item and carry out minimum value solution, obtains analytic solutions；Finally using acquired analytic solutions as starting point, gradient projection optimization is executed Algorithm carries out global optimization solution to objective function.

Application drawing signal processing technology solves NILM problem, is that NILM is regarded as to classification problem, i.e., for a certain The variation of moment total power consumption, judgement are thus to can determine whether out the operating status of equipment as caused by which known device.This reality It applies example and chooses suitable data as training set.

In order to preferably capture the operation conditions of each equipment, raising Decomposition Accuracy, the selection principle of equipment training set is followed: Select the period of the only equipment isolated operation as its training set, to reduce the interference of unrelated running equipment.In addition, for Each has selected the equipment of training set, and training set individually stores, and only need to be new equipment structure when installing new electrical equipment Training set is built, the training set of remaining known device without updating again.

The present embodiment uses REDD (can refer to energy decomposition data set) database, is total in one family it includes data The active power of body active power and each equipment.I.e. for a certain moment i, overall active power is p_i, a certain equipment m corresponds to this The active power at moment isWhole known device constitution equipment collection

The present embodiment specifically includes the following steps:

Step S1, total load active power p of the user equipment in 1~N of set period of time is obtained_i, as monitoring data； Obtain independent active power of each equipment in 1~n of set period of timeAs training data, wherein n and N are positive Integer and n are less than N, and m is the number of equipment.

Step S2, the operating status average power consumption of each equipment is calculated based on training dataAnd by equipment according to average Power consumptionDescending is arranged from big to small.

Step S3, the total load active power changing value Δ p at each moment is obtained according to monitoring data_i, obtain each moment Power change values Δ p_i=p_i+1-p_i。

Step S4, using figure signal processing technology structural mapWherein,For the vertex comprising multiple vertex Collection, one active power changing value Δ p of each vertex correspondence_i, A is weighted adjacent matrix, weight A_i,jFor corresponding vertex v_i And v_jBetween connection side right weight, at the time of i and j indicate different.Weight A_i,jValue reflect the similar of two adjacent vertexes Degree, i.e., weight is bigger, and the similitude between adjacent vertex is higher.Weight computing formula is usually using Gaussian kernel weighting function, it may be assumed that

Wherein, Δ p_iWith Δ p_jThe total load active power changing value at corresponding i moment and j moment is respectively indicated, σ indicates ratio The example factor.

Step S5, based on the figure constructedCalculate all figure signals in 1~n-hour

Figure signal s is key concept used in the present invention, it is defined by a mapping of vertex set to set of digits, Each vertex v_iOne figure signal s of label_i.Based on the figure constructedThe figure signal s of equipment m^mFor a length be N to Amount, value are defined as follows

Wherein, Thr_mIt is for judging whether equipment m changes the threshold value of operating status, value is usually that equipment m respectively connects The half of difference between the mean power of continuous state.As tag along sort, in training set (i≤n), list of the equipment m in moment i Exclusive function power change valuesKnown, if therefore the total power consumption at i moment variation be as equipment m state change caused by, Then haveLevel off to+1, otherwiseLevel off to -1；In test set (n < i≤N),It is unknown, thereforeTemporarily it is set as 0.

Step S6, based on total smoothness minimum is schemed, initial analytic solutions are obtained.Figure is defined using figure Laplacian Matrix Total smoothness of signal, are as follows:

Wherein, L is figure Laplacian Matrix, is defined as L=D-A, is the real symmetric matrix of a N × N；D is one right Diagonal matrices, i.e. element value on diagonal line are defined as D_i,i=∑_J=1:NA_i,j, off diagonal element 0.Based on above-mentioned definition, Similarity number strong point v_iAnd v_jBetween connection weight weight values A_i,jGreatly, and its corresponding figure signal value s_iAnd s_jIt is close.Therefore, for number For classification, figure signal constitutes a smooth signal, and when figure signal is sectionally smooth, schemes the value of total smoothness very It is small.

For the invention solves NILM problem, corresponding mathematical model indicates are as follows:

N therein_iIt mainly include the power consumption of circuit base load power consumption and other unknown devices for the noise at i moment. By the angle of change of power consumption, formula (4) may be modified such that following forms

For training set (i≤n), Δ p_iWithIt is known；For test set (n < i≤N), Δ p_iIt is known andNot Know.

Based on figure signal processing technology, use total smoothness of figure signal as regularization term, then final NILM target Function is defined as:

Wherein,s^mFor the figure signal of corresponding equipment m.Therefore NILM task is asked as an optimization Topic, looks for optimal solution, i.e. change of power consumption value of each equipment m in each moment i (n < i≤N)

In order to obtain initial point P_n+1,...,P_N, an optimization problem will be solved to each equipment m, obtained most smooth Solution.The corresponding regularization term of equipment m is minimized, is indicated are as follows:

Due to matrix L be it is diagonal symmetrical,It is known constant vector, utilizes matrix in block form, above formula solution procedure Are as follows:

Solution obtainsAfterwards, pass throughWithBetween mapping, required starting point P can be obtained_n+1,...,P_N.It reflects It is as follows to penetrate process: ifThen deducibility equipment m changes operating status in moment i, and otherwise equipment m does not change fortune Row state and correspondingIf equipment m changes operating status, as Δ p_iWhen < 0, infer that equipment is closed at the i moment It closes, hasOtherwise as Δ p_iWhen > 0, infers that equipment is opened at the i moment, have

The above process constantly repeats, and separates an equipment every time, and equipment is separated according to its running average powerSuccessively decrease Sequence carries out.When an equipment is separated and is obtained correspondingAfterwards, by the operating status average power consumption of the equipment It is subtracted from total power consumption, reduces interference as far as possible for the separation of subsequent equipment.

After obtaining the solution for minimizing regularization term as starting point, GRADIENT PROJECTION METHODS is executed to overall goals function (6) it optimizes, Solve problems at this time can be regarded as a nonlinear constrained programming, be expressed as

Wherein Ω constrains the range of disaggregation, to ensure that gained solution, i.e., the change of power consumption of each equipment are not got higher than practical fair Perhaps range.Finally obtain separating resulting

Performance evaluation

The common Performance Evaluating Indexes in the field NILM be recall rate (recall, RE), accurate rate (precision, PR) and F1-score.Its is respective to be defined as follows

RE=TP/ (TP+FN)

PR=TP/ (TP+FP)

F1=2 × (PR × RE)/(PR+RE),

That is the state change that TP (true positive) represents equipment is correctly validated, FP (false positive) generation Table equipment is misidentified as state change, and the state change that FN (false negative) represents equipment is not correctly validated.

Therefore recall rate (RE) has measured the event proportion being correctly validated out in all equipment state change events, Accurate rate (PR) has measured the recognition correct rate that equipment state changes event, and F1-score is then above-mentioned tradeoff between the two Value, the separating property of the higher illustration method of value are better.

By being tested on the House 1 and House 2 of REDD data set, the performance of the present embodiment proposed method It is verified.Tables 1 and 2 is then the present invention and other 3 existing NILM methods --- it is based on the side of figure signal processing (GSP) The performance comparison of method, the method based on Hidden Markov Model (HMM) and the method based on decision tree (DT).It can by Tables 1 and 2 Know, the electric power data separating resulting accuracy of the method for the present invention is more preferable.

The comparison of 1 REDD House of table, 1 electric power data separating resulting F1 performance number

The comparison of 2 REDD House of table, 2 electric power data separating resulting F1 performance number

The preferred embodiment of the present invention has been described in detail above.It should be appreciated that those skilled in the art without It needs creative work according to the present invention can conceive and makes many modifications and variations.Therefore, all technologies in the art Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea Technical solution, all should be within the scope of protection determined by the claims.