CN106055761A - Load model parameter verification method based on decision tree classification - Google Patents

Abstract

The present invention relates to a load model parameter verification method based on decision tree classification, comprising the following steps of 1) obtaining load module parameters of load nodes and voltage, current, active power and reactive power dynamic characteristic curves of the corresponding load node; 2) discretizing load model parameter sections; 3) obtaining final characteristic segments shapelet and corresponding split points d<bsp>; 4) selecting the final characteristic segment shapelet having the maximum information gain as the best characteristic segment shapelet; 5) dividing all samples into data subsets; 6) determining whether all the samples are correctly classified, if not, returning a step 7), and if so, returning a step 8); 7) respectively determining the samples in D1 and D2, and returning the step 3) for iterative search; 8) utilizing a C4.5 algorithm and successively obtaining decision tree models of pct, s0 and R2; and 9) testing classification accuracy of the obtained decision trees by using a 10-fold cross validation method, and further obtaining the integral accuracy of a load model parameter verification model.

Description

A kind of load model parameters method of calibration based on decision tree classification

Technical field

The present invention relates to a kind of load model parameters method of calibration, use in field of power especially with regard to one Load model parameters method of calibration based on decision tree classification.

Background technology

Power system load model structure and parameter has material impact to system emulation result.Load model structure, negative Lotus identification of Model Parameters and load model parameters verification are three main aspects of load modeling.And current loads modeling grind Study carefully the identification being concentrated mainly on load model parameters.The identification essence of load model parameters is curve matching, i.e. for a certain negative Lotus model structure obtains the relation between matching input and the output that one group of parameter can be best at load parameter space search, because of This, the identification acquired results of load model parameters still needs to verify further.The verification of tradition load model parameters is mainly by negative Lotus identification of Model Parameters curve obtained realizes, so, for gained load mould under a certain failure condition with simulation curve contrast Shape parameter is not necessarily suitable other failure situations, and therefore the verification of tradition load model parameters exists certain defect.

Along with the development of WAMS (Wide Area Measurement System, WAMS), PMU data is negative The identification of lotus model parameter provides another Data Source, utilizes these data to combine the method for data mining simultaneously and thinking can To realize the verification of load model parameters.Main thought in view of load model parameters identification is as the feelings inputted at voltage Under condition, obtaining corresponding load model parameters by the output of fitting power, use for reference this thinking, load model parameters is with dynamic Certain contact should be there is between state feature.Therefore, need badly by excavating between load model parameters and dynamic response curve Relation, realize in the case of breaking down, correctly carry out load model parameters verification.

Summary of the invention

For the problems referred to above, it is an object of the invention to provide a kind of load model parameters verification side based on decision tree classification Method, it is obtained between load model parameters and load bus dynamic response curve after a perturbation by the method for data mining Statistical relationship.

For achieving the above object, the present invention takes techniques below scheme: a kind of load model based on decision tree classification is joined Number method of calibration, it is characterised in that it comprises the following steps: 1) obtained the load model of load bus by system emulation result Parameter and the corresponding load bus behavioral characteristics of voltage curve, current curve, meritorious curve and idle curve when disturbance is bent Line, and then obtain comprising the data set D of N number of sample；2) load model parameters is carried out Interval Discrete, load model parameters master Sound load proportion pct in load model parameters to be included, rotor-side resistance R₂And initial slippage s₀Three parameters, according to respectively The total constant interval of parameter respectively obtains its concrete discrete segment and class number；3) according to gained class number, respectively From voltage curve, current curve, meritorious curve and idle curve, obtain final characteristic segments shapelet, and divide accordingly Point d_bsp；4) from step 3) final characteristic segments shapelet corresponding to voltage curve, final characteristic segments that current curve is corresponding Final characteristic segments shapelet that shapelet, final characteristic segments shapelet that meritorious curve is corresponding are corresponding with idle curve, Choose maximum final characteristic segments shapelet of information gain as best feature section shapelet；5) calculate all samples with Distance between good characteristic segments shapelet, and according to corresponding split point d_bsp, all samples are divided into D₁And D₂Two parts Data set；6) judge that all samples are the most correctly classified, then enter step 7 without all correct classification), otherwise enter step Rapid 8)；7) D is judged respectively₁And D₂In sample whether belong to same class, if D₁Middle sample is not belonging to same class and then makes D=D₁ Enter step 3) it is iterated search, otherwise stop iteration；If D₂Middle sample is not belonging to same class and then makes D=D₂Enter step 3) it is iterated search, otherwise stops iteration；8) N number of sample and iterative search gained all best feature section shapelet are calculated Distance, then using each distance as categorical attribute, use C4.5 algorithm to obtain pct, s successively₀, R₂Decision-tree model；9) it is The effectiveness of checking decision-tree model, uses the mode of 10 folding cross validations to survey the classification accuracy of gained decision tree Examination, respectively obtains pct, s₀And R₂Model accuracy, and then obtain the accuracy that load model parameters Knowledge Verification Model is overall.

Preferably, described step 3) in, as a example by voltage curve, obtain final characteristic segments shapelet, and accordingly Split point d_bspStep as follows: (3.1) choose all length sequence between [minlen, maxlen] from N number of sample Row, the number thus obtaining candidate feature section shapelet is:

$\underset{l=\min len}{\overset{l=\max len}{\&Sigma;}}\underset{i=1}{\overset{N}{\&Sigma;}}(m_i-l+1),$

Wherein, m_iFor i-th sample time-series length, minlen and maxlen is preset value；(3.2) pass through Formula below calculates distance d (s, T) between each candidate feature section shapelet s and all sample T, wherein, i-th sample And the distance between j-th candidates characteristic segments shapelet is designated as d (s_j,T_i)；

$d(s,T)=\underset{t\&Element;T}{min}{\left(\underset{i=1}{\overset{length\left(s\right)}{\&Sigma;}}{\left(s_i-t_i\right)}^2\right)}^{\frac{1}{2}},$

Wherein, length (s) represents the length of candidate feature section shapelet s；(3.3) for each candidate feature section Shapelet, according to d (s_j,T_i), select the meansigma methods of any two point of proximity distance as split point distance d successively_sp, calculate Obtain the information gain value of this candidate feature section shapelet in the case of different split point distances；(3.4) letter of maximum is chosen Breath gain is as the information gain of this candidate shapelet, and now split point is d_bsp；(3.5) information gain is chosen maximum Candidate feature section shapelet be final characteristic segments shapelet.

Preferably, in described step (3.3), the information gain value calculation procedure of candidate feature section shapelet is as follows: (a) It is p for comprising m classification and jth class sample proportion_jData set D, its entropy is defined as:

$E\left(D\right)=-\underset{j=1}{\overset{m}{\&Sigma;}}{p}_j{\log }_2\left(p_j\right);$

(b) in data set D, d (s_j,T_i) more than d_spSample be attributed to Sub Data Set D₁, d (s_j,T_i) less than d_sp's Sample is attributed to Sub Data Set D₂If, D₁And D₂Middle sample proportion is respectively f (D₁) and f (D₂), then after dividing, the entropy of data set D is fixed Justice is:C the information gain of () then this partition strategy is:

Preferably, described step 5) in, the distance between each sample with best feature section shapelet with corresponding d_bspComparing, its distance is more than corresponding d_bspSample be attributed to a class, its distance is less than corresponding d_bspSample be attributed to separately One class, i.e. obtains D₁And D₂Two data sets.

Preferably, described step 6) in, if each data subset only comprises the sample of a kind of class number, the most all Sample is the most correctly classified.

Due to the fact that and take above technical scheme, it has the advantage that 1, the present invention uses based on decision tree classification Load model parameters method of calibration, utilize all service datas to excavate corresponding dynamic characteristic values and load model parameters it Between dependency relation, therefore gained Knowledge Verification Model more statistical significance.2, the present invention uses load mould based on decision tree classification Shape parameter method of calibration, can be obtained the size of overall Knowledge Verification Model credibility, have actual electric network by the study of sample Certain directive significance.

Accompanying drawing explanation

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

Fig. 2 is the electrical network single line structural representation of embodiment in the present invention；

Fig. 3 is the sound load proportion decision tree mould that in the present invention, embodiment carries out that to training sample set classification learning obtains Type structural representation；

Fig. 4 is the initial slippage decision-tree model knot that in the present invention, embodiment carries out that to training sample set classification learning obtains Structure schematic diagram；

Fig. 5 is the rotor-side resistance decision-tree model that in the present invention, embodiment carries out that to training sample set classification learning obtains Structural representation.

Detailed description of the invention

With embodiment, the present invention is described in detail below in conjunction with the accompanying drawings.

As it is shown in figure 1, the present invention provides a kind of load model parameters method of calibration based on decision tree classification, it specifically walks Rapid as follows:

1) load model parameters of load bus and corresponding load bus are obtained when disturbance by system emulation result Voltage curve, current curve, meritorious curve and the dynamic characteristic values of idle curve, and then obtain comprising the data of N number of sample Collection D；

2) load model parameters being carried out Interval Discrete, load model parameters mainly includes sound in load model parameters Load proportion pct, rotor-side resistance R₂And initial slippage s₀Three parameters, respectively obtain according to the constant interval that each parameter is total Its concrete discrete segment and class number；

3) according to gained class number, obtain from voltage curve, current curve, meritorious curve and idle curve respectively Whole characteristic segments shapelet, and corresponding split point d_bsp；

As a example by wherein voltage curve, the step obtaining final characteristic segments shapelet is as follows:

(3.1) from N number of sample, choose all length subsequence between [minlen, maxlen], thus waited The number selecting characteristic segments shapelet is:

$\underset{l=\min len}{\overset{l=\max len}{\&Sigma;}}\underset{i=1}{\overset{N}{\&Sigma;}}(m_i-l+1)---\left(1\right)$

Wherein, m_iFor i-th sample time-series length, minlen and maxlen is preset value；

(3.2) by formula (2) calculate distance d between each candidate feature section shapelet s and all sample T (s, T), wherein, the distance between i-th sample and j-th candidates characteristic segments shapelet is designated as d (s_j,T_i)；

$d(s,T)=\underset{t\&Element;T}{min}{\left(\underset{i=1}{\overset{length\left(s\right)}{\&Sigma;}}{\left(s_i-t_i\right)}^2\right)}^{\frac{1}{2}}---\left(2\right)$

Wherein, length (s) represents the length of candidate feature section shapelet s；

(3.3) for each candidate feature section shapelet, according to d (s_j,T_i), successively select any two point of proximity away from From meansigma methods as split point distance d_sp, it is calculated this candidate feature section shapelet in the case of different split point distances Information gain value；

(3.4) choose the information gain information gain as this candidate shapelet of maximum, and now split point is d_bsp；

(3.5) candidate feature section shapelet choosing information gain maximum is final characteristic segments shapelet；

4) from step 3) final characteristic segments shapelet corresponding to voltage curve, final characteristic segments that current curve is corresponding Final characteristic segments shapelet that shapelet, final characteristic segments shapelet that meritorious curve is corresponding are corresponding with idle curve, Choose maximum final characteristic segments shapelet of information gain as best feature section shapelet；

5) distance between all samples and best feature section shapelet is calculated, and according to corresponding split point d_bsp, will All samples are divided into D₁And D₂Two parts data set；

6) judge that all samples are the most correctly classified, then enter step 7 without all correct classification), otherwise enter Step 8)；

7) D is judged respectively₁And D₂In sample whether belong to same class, if D₁In sample be not belonging to same class, then make D=D₁Enter step 3) it is iterated search, otherwise stop iteration；If D₂In sample be not belonging to same class, then make D=D₂ Enter step 3) it is iterated search, otherwise stop iteration；

8) distance of N number of sample and iterative search gained all best feature section shapelet is calculated, then by each distance As categorical attribute, C4.5 algorithm is used to obtain pct, s successively₀, R₂Decision-tree model；

9) for the effectiveness of checking decision-tree model, the mode using 10 folding cross validations is accurate to the classification of gained decision tree Exactness is tested, and respectively obtains pct, s₀And R₂Model accuracy, and then it is overall to obtain load model parameters Knowledge Verification Model Accuracy.

Above-mentioned steps 1) in, with a certain load bus institute on-load as simulation object, under different flow situations, when When power system occurs three phase short circuit fault, then with load model parameters and the relevant voltage behavioral characteristics of this load bus Curve, electric current dynamic characteristic values, meritorious dynamic characteristic values and idle dynamic characteristic curve are sample.

In above-mentioned steps (3.3), the information gain value calculation procedure of candidate feature section shapelet is as follows:

A () is p for comprising m classification and jth class sample proportion_jData set D, its entropy is defined as:

$E\left(D\right)=-\underset{j=1}{\overset{m}{\&Sigma;}}{p}_j{\log }_2\left(p_j\right)---\left(3\right)$

(b) in data set D, d (s_j,T_i) more than d_spSample be attributed to Sub Data Set D₁, d (s_j,T_i) less than d_sp's Sample is attributed to Sub Data Set D₂If, D₁And D₂Middle sample proportion is respectively f (D₁) and f (D₂), then it is divided the entropy of rear data set D It is defined as:

$\hat{E}\left(D\right)=f\left(D_1\right)E\left(D_1\right)+f\left(D_2\right)E\left(D_2\right)---\left(4\right)$

C information gain IG (sp) of () then this partition strategy is obtained by formula (5).

$IG\left(sp\right)=E\left(D\right)-\hat{E}\left(D\right)---\left(5\right)$

Above-mentioned steps 5) in, the distance between each sample with best feature section shapelet with corresponding d_bspCarry out Relatively, its distance is more than corresponding d_bspSample be attributed to a class, its distance is less than corresponding d_bspSample be attributed to another kind of, i.e. Obtain D₁And D₂Two data sets.

Above-mentioned steps 6) in, if only comprising the sample of a kind of class number in each data subset, the most all samples are all Correct classification.

Embodiment, as shown in Figure 2-5, utilizes load model parameters method of calibration based on decision tree classification to verify The system emulation result of Bus16 institute on-load, it specifically comprises the following steps that

1) with Bus16 institute on-load as simulation object, under different flow situations, Bus19-Bus21 occurs three-phase short During the fault of road, with Bus16 voltage dynamic characteristic values in the case of different load model parameter, electric current dynamic characteristic values, Meritorious dynamic characteristic values and idle dynamic characteristic curve are sample, obtain comprising the data set D of 135 samples；

2) load model parameters is carried out Interval Discrete, sound load proportion pct in main consideration load model parameters, Rotor-side resistance R₂And initial slippage s₀Three parameters, respectively obtain it according to the constant interval that each parameter is total concrete discrete Interval and class number, as shown in table 1:

Table 1 each parameter discrete interval and class number

3) according to gained class number, obtain from voltage curve, current curve, meritorious curve and idle curve respectively Whole characteristic segments shapelet, and corresponding split point d_bsp, it is as follows that it implements step:

(3.1) from 135 samples, choose all length subsequence between [10,200], thus obtain candidate The number of shapelet is:

$\underset{l=\min len}{\overset{l=\max len}{\&Sigma;}}\underset{i=1}{\overset{N}{\&Sigma;}}(m_i-l+1)---\left(6\right)$

Wherein, minlen=10, maxlen=200, N=135, m_i=290；

(3.2) distance d (s, T) between each candidate shapelet s and all sample T is calculated by formula (2), wherein, Distance between i-th sample and j-th candidates shapelet is designated as d (s_j,T_i)；

$d(s,T)=\underset{t\&Element;T}{min}{\left(\underset{i=1}{\overset{length\left(s\right)}{\&Sigma;}}{\left(s_i-t_i\right)}^2\right)}^{\frac{1}{2}}---\left(7\right)$

(3.3) for each candidate feature section shapelet, according to d (s_j,T_i), successively select any two point of proximity away from From meansigma methods as split point distance d_sp, it is calculated this candidate feature section shapelet in the case of different split point distances Information gain value；

(3.4) choose the information gain information gain as this candidate shapelet of maximum, and now split point is d_bsp；

(3.5) candidate feature section shapelet choosing information gain maximum is final characteristic segments shapelet；

4) from step 3) final characteristic segments shapelet corresponding to voltage curve, final characteristic segments that current curve is corresponding Final characteristic segments shapelet that shapelet, final characteristic segments shapelet that meritorious curve is corresponding are corresponding with idle curve, Choose maximum final characteristic segments shapelet of information gain as best feature section shapelet；

5) distance between all samples and best characteristic segments shapelet is calculated, and according to corresponding d_bsp, will be all Sample is divided into D₁And D₂Two parts；

6) judge that all samples are the most correctly classified, then enter step 7 without all correct classification), otherwise enter Step 8)；

7) D is judged respectively₁And D₂In sample whether belong to same class, if D₁In sample be not belonging to same class, then make D=D₁Enter step 3) it is iterated search, otherwise stop iteration；If D₂In sample be not belonging to same class, then make D=D₂ Enter step 3) it is iterated search, otherwise stop iteration；

8) it is calculated the distance of N number of sample and all shapelet of iterative search gained, then each distance feature is made For categorical attribute, obtain pct, s successively₀, R₂Decision-tree model, as seen in figures 3-5, wherein, decision tree leaf node represents The discrete segment of each parameter is identical with table 1, and the split point dqi in decision tree represents the ith feature section from idle q；Division Point dpi represents the ith feature section from meritorious p；Split point dvi represents the ith feature section from voltage v；Split point dii Represent the ith feature section from electric current i；

9) for the effectiveness of checking decision-tree model, the mode using 10 folding cross validations is accurate to the classification of gained decision tree Exactness is tested, and the precision respectively obtaining pct is 96.2963%, R₂Precision be 89.6296%, s₀Precision be 95.5556%, it can be seen that the classification accuracy of load model parameters entirety is of a relatively high, decision tree is used for power system and bears The verification of lotus model parameter has certain directive significance to load model parameters identification and load model parameters validation verification.

The various embodiments described above are merely to illustrate the present invention, the structure of each parts, size, arrange position and shape is all permissible Be varied from, on the basis of technical solution of the present invention, all improvement individual part carried out according to the principle of the invention and etc. With conversion, the most should not get rid of outside protection scope of the present invention.

Claims (5)

1. a load model parameters method of calibration based on decision tree classification, it is characterised in that: it comprises the following steps:

1) load model parameters of load bus and corresponding load bus voltage when disturbance are obtained by system emulation result Curve, current curve, meritorious curve and the dynamic characteristic values of idle curve, and then obtain comprising the data set D of N number of sample；

2) load model parameters being carried out Interval Discrete, load model parameters mainly includes sound load in load model parameters Ratio pct, rotor-side resistance R₂And initial slippage s₀Three parameters, respectively obtain its tool according to the constant interval that each parameter is total The discrete segment of body and class number；

3) according to gained class number, obtain final respectively from voltage curve, current curve, meritorious curve and idle curve Characteristic segments shapelet, and corresponding split point d_bsp；

4) from step 3) final characteristic segments shapelet corresponding to voltage curve, final characteristic segments that current curve is corresponding Final characteristic segments shapelet that shapelet, final characteristic segments shapelet that meritorious curve is corresponding are corresponding with idle curve, Choose maximum final characteristic segments shapelet of information gain as best feature section shapelet；

5) distance between all samples and best feature section shapelet is calculated, and according to corresponding split point d_bsp, will be all Sample is divided into D₁And D₂Two parts data set；

6) judge that all samples are the most correctly classified, then enter step 7 without all correct classification), otherwise enter step 8)；

7) D is judged respectively₁And D₂In sample whether belong to same class, if D₁Middle sample is not belonging to same class and then makes D=D₁Enter Enter step 3) it is iterated search, otherwise stop iteration；If D₂Middle sample is not belonging to same class and then makes D=D₂Enter step 3) It is iterated search, otherwise stops iteration；

8) calculate the distance of N number of sample and iterative search gained all best feature section shapelet, then using each distance as Categorical attribute, uses C4.5 algorithm to obtain pct, s successively₀, R₂Decision-tree model；

9) for the effectiveness of checking decision-tree model, the mode classification accuracy to gained decision tree of 10 folding cross validations is used Test, respectively obtain pct, s₀And R₂Model accuracy, and then obtain the standard that load model parameters Knowledge Verification Model is overall Exactness.

A kind of load model parameters method of calibration based on decision tree classification, it is characterised in that: institute State step 3) in, as a example by voltage curve, obtain final characteristic segments shapelet, and corresponding split point d_bspStep such as Under:

(3.1) from N number of sample, choose all length subsequence between [minlen, maxlen], thus obtain candidate special The number of the section of levying shapelet is:

$\underset{l=\min len}{\overset{l=\max len}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left(m_i-l+1\right),$

Wherein, m_iFor i-th sample time-series length, minlen and maxlen is preset value；

(3.2) distance d (s, T) between each candidate feature section shapelet s and all sample T is calculated by formula below, its In, the distance between i-th sample and j-th candidates characteristic segments shapelet is designated as d (s_j,T_i)；

$d(s,T)=\underset{t\&Element;T}{min}{\left(\underset{i=1}{\overset{length\left(s\right)}{\&Sigma;}}{\left(s_i-t_i\right)}^2\right)}^{\frac{1}{2}},$

Wherein, length (s) represents the length of candidate feature section shapelet s；

(3.3) for each candidate feature section shapelet, according to d (s_j,T_i), select any two point of proximity distance successively Meansigma methods is as split point distance d_sp, it is calculated the letter of this candidate feature section shapelet in the case of different split point distances Breath yield value；

(3.4) choose the information gain information gain as this candidate shapelet of maximum, and now split point is d_bsp；

(3.5) candidate feature section shapelet choosing information gain maximum is final characteristic segments shapelet.

A kind of load model parameters method of calibration based on decision tree classification, it is characterised in that: institute Stating in step (3.3), the information gain value calculation procedure of candidate feature section shapelet is as follows:

A () is p for comprising m classification and jth class sample proportion_jData set D, its entropy is defined as:

$E\left(D\right)=-\underset{j=1}{\overset{m}{\&Sigma;}}{p}_j{\log }_2\left(p_j\right);$

(b) in data set D, d (s_j,T_i) more than d_spSample be attributed to Sub Data Set D₁, d (s_j,T_i) less than d_spSample It is attributed to Sub Data Set D₂If, D₁And D₂Middle sample proportion is respectively f (D₁) and f (D₂), then the entropy definition of data set D after dividing For:

$\hat{E}\left(D\right)=f\left(D_1\right)E\left(D_1\right)+f\left(D_2\right)E\left(D_2\right);$

C the information gain of () then this partition strategy is:

$IG\left(sp\right)=E\left(D\right)-\hat{E}\left(D\right).$

A kind of load model parameters method of calibration based on decision tree classification, it is characterised in that: institute State step 5) in, the distance between each sample with best feature section shapelet with corresponding d_bspCompare, its distance More than corresponding d_bspSample be attributed to a class, its distance is less than corresponding d_bspSample be attributed to another kind of, i.e. obtain D₁And D₂Two Individual data set.

A kind of load model parameters method of calibration based on decision tree classification, it is characterised in that: institute State step 6) in, if only comprising the sample of a kind of class number in each data subset, the most all samples are the most correctly classified.