CN103778470B - There is the distributed power generation island detection method of automatic measure on line ability - Google Patents

Abstract

The present invention relates to a kind of distributed power generation island detection method with automatic measure on line ability, comprise the following steps: 1) utilize microgrid SCADA system to obtain original sample online；2) utilize the on-line talking method containing micro-bunch, adaptively original sample is carried out re-sampling；3) calculate the weight of each sample set according to multi-categorizer model and select the superior and eliminate the inferior, eliminating classification accuracy less than the sample set setting threshold value, it is thus achieved that preferably sample set；4) according to preferred sample set, on-line training generates sorter model；5) sorter model utilizing above-mentioned automatic measure on line to obtain, updates the sorter model used by the detection of real-time isolated island in an asynchronous manner.Compared with prior art, the present invention has the advantages such as accuracy height, good stability, robustness is high, adaptability is good.

Description

There is the distributed power generation island detection method of automatic measure on line ability

Technical field

The present invention relates to a kind of distributed generation technology, especially relate to a kind of there is the distributed of automatic measure on line ability Generating island detection method.

Background technology

Distributed power generation (distributed generation, DG) is the important component part of intelligent grid.For people The consideration of the aspects such as member's equipment safety, system stable operation and the quality of power supply, common demands distributed power generation possesses isolated island inspection Brake.The threshold value that can effectively solve in isolated island detection due to data mining technology is adjusted a difficult problem, in recent years based on data mining Island detection method paid attention to.In existing research, the sorter model for isolated island detection is all to be instructed by off-line Practice acquisition.Due to distributed power source and the undulatory property of local load, and power distribution network topological structure containing distributed power source and There is change in the method for operation, the statistical property of sample can be made to elapse over time or environment changes and unpredictable change occurs Changing, its result can cause the classification accuracy of the sorter model that off-line obtains to be gradually reduced.This phenomenon is in data mining It is referred to as concept drift.Due to the power distribution network containing distributed power source belong to time-varying and astatic environment, concept drift is difficult to Avoid, so, solving concept drift is the key making data digging method be used practically.

Summary of the invention

Defect that the purpose of the present invention is contemplated to overcome above-mentioned prior art to exist and provide a kind of accuracy high, stable The distributed power generation island detection method with automatic measure on line ability that property is good, robustness is high, adaptability is good.

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

A kind of distributed power generation island detection method with automatic measure on line ability, it is characterised in that include following step Rapid:

1) microgrid SCADA system is utilized to obtain original sample online；

2) utilize the on-line talking method containing micro-bunch, adaptively original sample is carried out re-sampling；

3) calculate the weight of each sample set according to multi-categorizer model and select the superior and eliminate the inferior, eliminating classification accuracy Less than the sample set setting threshold value, it is thus achieved that preferably sample set；

4) according to preferred sample set, on-line training generates sorter model；

5) sorter model utilizing above-mentioned automatic measure on line to obtain, updates used by the detection of real-time isolated island in an asynchronous manner Sorter model.

Described utilize microgrid SCADA system to obtain original sample online particularly as follows:

If sample set is:

E={(x_i, C (x_i), i=1 ... N (1)

N is total sample number, x_iIt is characterized vector, C (x_i) { 0,1} is the class label of this sample to ∈, and 0 is non-isolated island, and 1 is Isolated island；

Using frequency, voltage magnitude, current amplitude, voltage-to-current phase angle difference, active power and reactive power as feature to Amount, is designated as:

In formula (1), characteristic vector x_i6 eigenvalues reported by each distributed generation system DG, and class label C (x_i) Obtain from PCC on off state, in SCADA system, when being realized the network pair of low cost by simple network agreement SNTP, transmit x_iWith C (x_i) report all contain time tag, with C (x_i) timestamp on the basis of, to x_iCarry out linear interpolation, i.e. obtain same The sample instance of one time cross-section.

Described utilize the on-line talking method containing micro-bunch, adaptively original sample is carried out re-sampling particularly as follows:

Step 1, under off-line state, carry out cluster analysis, if Q is bunch number set, produce C₁...C_Q；For each Bunch C_q, form micro-bunch of corresponding M_q；

When step 2, on-line operation, when new point?After moment arrives, calculateWith each bunch of C_qThe distance of barycenter

(1) ifThen create new bunch of C_new, turn to step 3；

(2) otherwise, by pointIt is merged into a bunch C_pIn, update bunch C_pMicro-bunch of corresponding M_p, return step 2；

Step 3, generalIncrease in training sample set, complete once to sample；Meanwhile, find out there is minimum CF1^tValue Bunch, abandoned, thus the sum of holding bunch is Q, returned step 2；

Wherein above-mentioned parameter is defined as follows:

Bunch point that C is tieed up by a series of dThe set constituted, the timestamp that each point is corresponding isDefine its micro-bunch First ancestral that M ties up for (2d+3):

$(\stackrel{\→}{{\mathrm{CF}2}^x},\stackrel{\→}{{\mathrm{CF}1}^x},{\mathrm{CF}2}^t\mathrm{CF}{1}^t,n)---\left(3\right)$

Wherein,For d dimensional vector, the value of its pth item is For d dimensional vector, the value of its pth item is ${\mathrm{\Σ}}_{j=1}^n{x}_{i_j}^p;{\mathrm{CF}2}^t={\mathrm{\Σ}}_{j=1}^n{\left(T_{i_j}\right)}^2;{\mathrm{CF}1}^t={\mathrm{\Σ}}_{j=1}^n{T}_{i_j};$ N is the number at bunch midpoint；

The barycenter of bunch CIt is defined as:

$\stackrel{\→}{C_c}=-(\stackrel{\→}{{\mathrm{CF}1}^x}/n)--\left(4\right)$

The maximum boundary of bunch C is defined as τ σ, wherein, τ > 0 it is the distance factor；σ be in bunch C each point to barycenter C_cDistance Root-mean-square error, its computing formula is:

$\mathrm{\σ}=\sqrt{\underset{p=1}{\overset{d}{\mathrm{\Σ}}}(\frac{{\mathrm{CF}2}^p}{n}-{\left(\frac{{\mathrm{CF}1}^p}{n}\right)}^2)}---\left(5\right)$

The described sorter model updated in an asynchronous manner used by the detection of real-time isolated island is specific as follows:

501) setting data window and comprise k sub-sample set the most simultaneously, each subsample collection generates a grader M_i；

502) after cycle T, the PCC on off state report Tong Bu with on-line sampling data block that SCADA system is sent is received Accuse；

503) M is used₁, M₂..., M_kK grader to new subsample collection c_k+1Carry out classification prediction respectively, according to switch State reports error rate of can classifying, with Error (M_i) represent M_iClassification error rate, wherein i=1,2 ..., k, M_k+1Carry out It is Error (M that cross validation obtains its error rate_k+1), Error (M_i) more than 0.5 grader weight be assigned to 0, calculate its remain-power Weight:

${\widehat{\mathrm{\ω}}}_i={\log }_{10}\left(\frac{1-\mathrm{Error}\left(M_i\right)}{\mathrm{Error}\left(M_i\right)}\right)---\left(6\right)$

504) M is rejected₁, M₂..., M_k+1The grader that middle weight is minimum, if having two or more weight minimum and phase With, then reject time grader the earliest, the subsample of other k grader is integrated into preferred sample set；

505) grader using the training of preferred sample set to generate updates dividing for real-time grading prediction in an asynchronous manner Class device model；

506) circulation step 502) to 505).

Compared with prior art, the invention have the advantages that

There is concept drift in microgrid, the sorter model that existing off-line training is formed is difficult to meet the inspection of real-time isolated island in running The needs surveyed, automatic measure on line method of the present invention can effectively solve the problem that concept drift problem, and the present invention combines microgrid SCADA system The online acquisition original sample containing tag along sort, effectively coordinates on-line study.

Utilizing on-line talking method, sample original sample adaptively, concept and the application of micro-bunch can bases The speed of concept drift adjusts sample rate, and has superior time and space performance.

Propose the on-line study method of preferred sample set strategy, utilize multi-categorizer evaluate the quality of sample and carry out winning Bad eliminating, constitute optimum sample set, it is achieved on-line study, simulation result shows, the method compared to Case-based Reasoning and based on Integrated on-line study, has the advantage of Stability and veracity.

It addition, on-line study and real-time grading run in an asynchronous manner, communication disruption in short-term had the highest robust Property.Method in this paper, applies also for the on-line tuning of protection, has relatively broad adaptability.

Accompanying drawing explanation

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

Fig. 2 is the cluster preprocessing containing micro-bunch and re-sampling schematic diagram；

Fig. 3 is preferred sample automatic measure on line system schematic；

Fig. 4 is the distribution network systems schematic diagram containing multiple DG；

Fig. 5 is the event distribution schematic diagram of subsample collection；

Fig. 6 is strategy of on-line and off-line strategy effectiveness comparison diagram；

Fig. 7 is the flow chart of the present invention.

Detailed description of the invention

The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.

As it is shown in fig. 7, a kind of distributed power generation island detection method with automatic measure on line ability, including following step Rapid:

S100, microgrid SCADA system is utilized to obtain original sample online；

S200, utilization, containing the on-line talking method of micro-bunch, carry out re-sampling to original sample adaptively；

S300, foundation multi-categorizer model calculate the weight of each sample set and select the superior and eliminate the inferior, and eliminate classification standard Really rate is less than the sample set setting threshold value, it is thus achieved that preferably sample set；

S400, according to preferred sample set, on-line training generates sorter model；

S500, utilize the sorter model that above-mentioned automatic measure on line process obtains, update the inspection of real-time isolated island in an asynchronous manner Sorter model used by survey.

Its general principles is as shown in Figure 1.Being divided into two parts in figure, wherein, top is on-line study system, its Business is that detection concept is drifted about and constantly updates sorter model；Bottom is real-time detecting system, and its task is from sampled data Extract characteristic quantity, and utilize sorter model to carry out real-time isolated island detection.

2, the online acquisition of training sample

2.1, original sample acquisition methods based on SCADA

In microgrid, each DG is configured with local protection and control unit, and these protections, control unit have data acquisition concurrently And communication function.Micro-capacitance sensor is the most all configured with energy management system of micro-grid [9], can with the electric parameters of each DG of online acquisition and PCC on off state.According to the requirement of general SCADA system, above-mentioned electric parameters and on off state with in report (Report) mode Send, above send reason mainly to have two kinds:

1), send in change.When there is switch changed position or the event such as analog quantity is out-of-limit, automatically on deliver newspaper literary composition.

2), send on the cycle.Sending the cycle on Yi Ban is 2s.

Utilize SCADA system can obtain the training sample of isolated island detection online.If sample set is:

E={(x_i, C (x_i), i=1 ... N (1)

Wherein, N is total sample number, x_iIt is characterized vector, C (x_i) { 0,1} is the class label of this sample to ∈, and 0 is non-orphan Island, 1 is isolated island.Make with frequency, voltage magnitude, current amplitude, voltage-to-current phase angle difference, active power and reactive power herein It is characterized vector, is designated as:

In formula (1), characteristic vector x_i6 eigenvalues reported by each DG, and class label C (x_i) obtain from PCC on off state Take.Obviously, so, x_iWith C (x_i) same time cross-section must be taken from.In SCADA system, by simple network agreement SNTP When realizing the network pair of low cost, precision can reach 2ms.Transmit x_iWith C (x_i) report all contain time tag.With C (x_i) On the basis of timestamp, to x_iCarry out linear interpolation, the sample instance of same time cross-section can be obtained.Owing to sending the cycle in report For 2s, and during stable state, the change of each eigenvalue is more slow, and said method can meet the demand of data mining.

Although it should be noted that said process needs communication network, but this examines with isolated island based on Real-Time Switch state Survey method is diverse.There is essential distinction to the use of communication network in both, the former is served only for obtaining sample, and the latter Then for detection in real time.On off state method needs to set up end to end between PCC switch and each DG, highly reliable in real time logical Letter, communicates and once lost efficacy, and isolated island detection function the most just lost efficacy.And in the present invention, owing to report containing time tag, so The lowest to network delay and bandwidth requirement, wireless mode can be used to realize.SCADA function belongs to the basic training of micro-grid system Can, it is not necessary to additionally invest.Owing to concept drift is a relatively slow process, even if so communication disruption, time delay occurring in short-term The abnormal conditions such as excessive, also will not have a strong impact in generation properly functioning to system.

2.2, specimen sample method based on cluster

Under offline mode, distribution of all categories in training sample is through well-designed, to reach unbiasedness.But Under online mode, original sample is to obtain in real time, has the biggest randomness.For example, it is assumed that long period system is transported continuously Line mode varies less, according to the isolated island detection standard under the isolated island detection model reply surging condition that the sampled data of system builds Really rate will reduce；If long period system is constantly in and net state continuously, model will fall for reliability of islanding detection Low.In these cases, in sample, categorical distribution is single, will be unfavorable for classification learning.Set forth herein that following method is to original sample Originally sampling processing is carried out.

Definition 1: set the point that bunch C is tieed up by a series of dThe set constituted, the timestamp that each point is corresponding isFixed Its micro-bunch of M of justice is first ancestral that (2d+3) ties up:

$(\stackrel{\→}{{\mathrm{CF}2}^x},\stackrel{\→}{{\mathrm{CF}1}^x},{\mathrm{CF}2}^t\mathrm{CF}{1}^t,n)---\left(3\right)$

Wherein,For d dimensional vector, the value of its pth item is For d dimensional vector, the value of its pth item is ${\mathrm{\Σ}}_{j=1}^n{x}_{i_j}^p;{\mathrm{CF}2}^t={\mathrm{\Σ}}_{j=1}^n{\left(T_{i_j}\right)}^2;{\mathrm{CF}1}^t={\mathrm{\Σ}}_{j=1}^n{T}_{i_j};$ N is the number at bunch midpoint；

Definition 2: the barycenter of bunch CIt is defined as:

$\stackrel{\→}{C_c}=-(\stackrel{\→}{{\mathrm{CF}1}^x}/n)--\left(4\right)$

Definition 3: the maximum boundary of bunch C is defined as τ σ, wherein, τ > 0 be apart from the factor；σ be in bunch C each point to barycenter C_c The root-mean-square error of distance, its computing formula is (derivation is shown in appendix C):

$\mathrm{\σ}=\sqrt{\underset{p=1}{\overset{d}{\mathrm{\Σ}}}(\frac{{\mathrm{CF}2}^p}{n}-{\left(\frac{{\mathrm{CF}1}^p}{n}\right)}^2)}---\left(5\right)$

Specimen sample step based on cluster is as follows:

Step 1, carrying out cluster analysis under off-line state, if InitNums is initially counting for cluster, Q is for setting Bunch number, produce C₁...C_Q；For each bunch of C_q, form micro-bunch of corresponding M_q；

When step 2, on-line operation, when new point?After moment arrives, calculateWith each bunch of C_qThe distance of barycenterIf distance C_qRecently,

(1) ifThen create new bunch of C_new.Turn to step 3；

(2) otherwise, by pointIt is merged into a bunch C_pIn.Update bunch C_pMicro-bunch of corresponding M_p, return step 2.

Step 3, generalIncrease in training sample set, complete once to sample；Meanwhile, find out there is minimum CFl^tValue Bunch, abandoned (this bunch is the longest away from modern average time, is assessed as the oldest bunch), so, the sum of holding bunch is Q.Return Step 2.

If inputting highly stable, then new pointAlways fall in existing bunch without producing new bunch.And when data stream During unexpected the or progressive concept drift of middle generation, the appearance of new bunch can be caused；For periodic concept drift, can be old After bunch being dropped, reenter in the way of new bunch again.

Can dynamically adjust the value of distance factor τ, thus change the speed producing new samples.τ is the least, produces new samples Speed the fastest.Take τ=2 herein；Q-value also needs to choose as the case may be, and Q-value takes 10 herein.

In step 2, new bunch of first sample extraction is brought in on-line training sample, on-line talking sample set sample Number maximum is set to M, and new samples collection then included in by the sample more than M.Therefore, the method can be adjusted automatically along with the speed of concept drift The speed of whole sampling.By the definition of micro-bunch, and the formula of barycenter, maximum boundary is visible, and above-mentioned calculating is all increment recursive side Formula, therefore amount of calculation is minimum；Without storing data stream, internal memory only need to preserve Q micro-bunch.To sum up, algorithm have splendid time, Space efficiency, is suitable for carrying out online.

3, on-line study

In order to tackle concept drift, generally there are two kinds of on-line study methods: Case-based Reasoning and based on integrated method.Based on The method of example uses single classifier.The method supposes the oldest sample, and the concept wherein contained more can not be adapted to up-to-date feelings Condition, then uses slip data window to carry out naturally the sample that " forgeing " is old.But it is true that old sample is likely to contain important Rule.For example, it is assumed that micro-capacitance sensor runs the most steady within a very long time so that in the recent period sample only contains limited Event, the old sample that more equalizes of distribution is then constantly squeezed away by slip data window mechanism.So, outstanding knowledge is by gradually Forgeing, the distribution of overall sample is by gradually disequilibrium.Different, based on integrated method by integrated by history number Learn according to the multiple graders formed, construct new integrated framework by multi-categorizer weighted sum, to improve the general of grader Change ability.The analysis of existing document shows, has higher accuracy based on integrated method than the method for Case-based Reasoning.But Simulation analysis herein but shows, can not effectively promote the accuracy of isolated island detection based on integrated method, and reason is, micro- In the actual motion of net, various isolated islands, pseudo-isolated island event are unbalanced, need sample set based on integrated method Close cutting be multiple sample set to constitute multiple graders, This further reduces the sample size of each grader, aggravate The disequilibrium of sample distribution in each subset, thus reduce the accuracy of sub-classifier.

According to the feature of isolated island test problems, set forth herein a kind of preferably sample set strategy, be based on each grader point Class effect is screened, and collects the subsample collection preferred sample set of composition after screening.

Subsample collection produces with the form of data stream.Slip data window forward slip in time, accommodates k son the most simultaneously Sample set.On-line study system and real-time detecting system are in fact divided into two asynchronous procedures: one is based on multi-categorizer model Classifying up-to-date online data stream sample, under SCADA system coordinates, the cycle updates threshold value of adjusting.Two is for up-to-date Real-time stream sample, judge according to above-mentioned threshold value of adjusting.

The present invention is through considering, and learning algorithm uses support vector machine (SVM, Support Vector Machine).Generate grader Mi according to subsample collection training, utilize weighting algorithm that grader is screened, it is thus achieved that preferably sample This collection, idiographic flow is as follows:

1) setting data window and comprise k sub-sample set the most simultaneously, each subsample collection generates a grader M_i

2) after cycle T, PCC on off state Tong Bu with the on-line sampling data block report that SCADA system is sent is received.

3) M is used₁, M₂..., M_kK grader is to new subsample collection c_k+1Carry out classification prediction respectively, according on off state Report can be classified error rate, with Error (M_i) represent M_i(i=1,2 ..., classification error rate k).M_k+1Carry out cross validation Obtaining its error rate is Error (M_k+1)。Error(M_i), (i=1,2 ..., k+1) grader weight more than 0.5 is assigned to 0, meter Calculate remaining weight:

${\widehat{\mathrm{\ω}}}_i={\log }_{10}\left(\frac{1-\mathrm{Error}\left(M_i\right)}{\mathrm{Error}\left(M_i\right)}\right)---\left(6\right)$

4) M is rejected₁, M₂..., M_k+1The grader that middle weight is minimum.If having two or more weight minimum and phase With, then reject time grader the earliest, the subsample collection of other k grader is merged into preferred sample set.

5) in an ensuing on-line sampling cycle T, the grader using the training of preferred sample set to generate carries out reality Time classification prediction；

6) circulation step 2 to 5.

Constantly select according to above method more excellent sample to be retained in data window, eliminate bad；Utilize preferred sample This collection generates sorter model online, and updates real-time grading device model, thus completes the monitoring of real-time isolated island.

The important information that preferably sample set strategy contains in can either retaining old sample, can overcome again integrated classifier increment The problem that the sample size of this collection is very few, strengthens the event detection accuracy Han concept drift.

Embodiment 1

Instance system comprises 3 distributed power source DG1～3.Utilize the model of PSCAD analogous diagram 4.Main electrical network uses nothing Poor big power supply, distributed power source all uses Synchronous Machine Models.Real-time sampling frequency is 4000Hz, and the isolated island detection time limit is set to 250ms, the on-line sampling cycle is 2 minutes.

As a example by G1, emulation considers isolated island event and local load and other DG switching events herein, refers to table 1, this kind of event can cause the concept drift of data stream.It addition, unbalanced power degree (PI, power imbalance) is not only The classification accuracy of isolated island detection is had a significant impact, and slow concept drift degree can be reflected.Herein PI is defined as:

$\mathrm{PI}=\frac{P_{\mathrm{SYS}}}{P_{\mathrm{SYS}}+P_{\mathrm{DG}}}\×100\%---\left(7\right)$

Wherein, P_SYSIt is that main electrical network flows into the active power of local power distribution network, P through point of common coupling (at cbl)_DGFor this locality The active power that distributed power source is sent, it is clear that PI ∈ [0,1].Adjust due to isolated island detection and there is check frequency, in theory PI is the lowest, is more unfavorable for passive detection, so only considering the situation of PI >=0.1 in emulation herein.

Table 1

Sample generates way: carry out event sets [A1, A2, A3, A4, B1, B2, B3, B4, B1, B2, B3, B4] at random Sampling.Owing to non-isolated island event has been replicated portion in above-mentioned set, therefore non-isolated island event occurrence rate can be higher than isolated island thing Part one times.In simulation process, it is considered to the fluctuation (L3 active reactive random fluctuation) of local load, to local unbalanced power degree Impact is for-5%～5%.Carry out clustering re-sampling, until finally obtaining 20 groups of training sample subsamples for on-line study Collection, often group comprises 70 groups of samples.Event distribution is as shown in Figure 5.Both the plays such as random switch motion, DG and load switching had been simulated Strong concept drift event, simulates again that unbalanced power degree is slowly varying and the slow concept drift event that causes.

c₁-c₂₀Represent the subsample collection of different event random distribution, referring specifically to Fig. 5.The difference body of unbalanced power degree Show the degree of slow concept drift.In order to verify the effect of the concept drift of strategy reply herein, data window size selects k=8.

Preferably sample set policy validation:

Table 2

New subsample collection	Eliminate grader	Preferably sample set
			c9	M4	M1, M2, M3, M5, M6, M7, M8, M9
c10	M2	M1, M3, M5, M6, M7, M8, M9, M10
			C11	M3	M4, M5, M6, M7, M8, M9, M10, M11
c12	M10	M1, M5, M6, M7, M8, M9, M11, M12
			c13	M5	M1, M6, M7, M8, M9, M11, M12, M13
c14	M12	M1, M6, M7, M8, M9, M11, M13, M14
			c15	M1	M6, M7, M8, M9, M11, M13,M14, M15
c16	M14	M6, M7, M8, M9, M11, M13, M15, M16
			c17	M9	M6, M7, M8, M11, M13, M15, M16, M17
c18	M11	M6, M7, M8, M13, M15, M16, M17, M18

c19	M15	M6, M7, M8, M13, M16, M17, M18, M19
			c20	M13	M6, M7, M8, M16, M17, M18, M19, M20

Subsample collection c₁-c₂₀Sequentially enter data window, c₁-c₈After all entering, data window is the fullest, needs afterwards constantly to include in newly Subsample collection, eliminates worst subsample collection, to update preferred sample set.

Preferably each grader of sample set carries out classification prediction to new subsample collection, receives the report that SCADA system is sent Error (the M of each grader is understood after announcement_i).Thus obtain weightAbandon the grader that weight is minimum, retain remaining 8 points The subsample collection of class device constitutes optimum sample set.Then start to accept online data and flow down a sub-sample set, more than repetition Flow process.Table 2 lists data stream c₉-c₂₀During data window eliminate grader and preferred sample set more news.

Table 3

Contrasting preferred sample set strategy and off-line training classification prediction, the online classification of simple slip data window (is simply lost Forget), the method for integrated multi-categorizer.The algorithm of above strategy all uses SVM, and its effect sees table 3 and Fig. 6.

c₉-c₂₀Instantaneous and concept drift slowly occurring, elapses over time, off-line classification accuracy nearly all exists Less than 80%, c₁₇-c₂₀Go up.Because the training subset that off-line training is chosen is c₁, send out in wink, slowly concept drift make Threshold value of adjusting gradually lost efficacy, and c₁₇-c₂₀Event distribution and unbalanced power degree move closer to again c₁, therefore accuracy rate is carried Rise.

On the other hand, simply forget classification policy accuracy rate and Ensemble classifier accuracy rate is higher, but deficient in stability, ripple Dynamic bigger.Although Ensemble classifier strategy can correct threshold value of adjusting in time, but owing to sub-classifier lacks enough training samples, leads Cause prediction instability；But simple forgetting strategy is owing to can not retain older important information, it is easily caused accuracy rate and continues It is difficult to improve, such as c₁₅-c₂₀Situation.

And preferred sample set strategy in this paper can the most simply forget sliding window and the advantage of integrated multi-categorizer, Most classification accuracies are all more than 90%, and it is little to fluctuate, good stability.Although accuracy rate exists compared to other two at the beginning Line strategy, without clear superiority, elapses over time, and advantage is the most substantially changed.Because it can constantly be washed in a pan according to last samples sample Eliminating the most bad grader, update preferred sample set, the isolated island adapting to new situation is adjusted, and ensures typical case's training sample of enough numbers simultaneously This.Therefore the ability of its reply concept drift is strong, is independent of accidentalia.

Preferred sample set strategy in this paper has a clear superiority on the isolated island test problems tackling containing concept drift, energy Enough real-time optimizations are adjusted threshold value.

The time of off-line training is concentrated mainly on each subset and generates in the grader training to new subsample collection.With c₉Enter As a example by data window, c₁-c₈Off-line training is time-consumingly 0.03s, 0.02s, 0.02s, 0.02s, 0.02s, 0.01s, 0.01s respectively, 0.03s, amounts to 0.16s.For off-line training, 0.1 second rank belongs to can accept scope.

The time of on-line checking concentrates in the grader classification to new subsample collection that preferred sample set generates, and the time is Nanosecond rank, fully meet the time requirement of automatic measure on line.It is therefore preferable that sample set strategy has good real-time.

There is concept drift in microgrid, the sorter model that existing off-line training is formed is difficult to meet the inspection of real-time isolated island in running The needs surveyed.Automatic measure on line method can effectively solve the problem that concept drift problem.Obtain online herein in conjunction with microgrid SCADA system Containing the original sample of tag along sort, effectively coordinate on-line study.

Utilize on-line talking method, adaptively original sample is sampled.Concept and the application of micro-bunch can bases The speed of concept drift adjusts sample rate, and has outstanding time and space performance.

Propose the on-line study method of preferred sample set strategy, utilize multi-categorizer evaluate the quality of sample and carry out winning Bad eliminate, constitute optimum sample set, it is achieved on-line study.Simulation result shows, the method compared to Case-based Reasoning and based on Integrated on-line study, has the advantage of Stability and veracity.

It addition, on-line study and real-time grading run in an asynchronous manner, communication disruption in short-term had the highest robust Property.Method in this paper, applies also for the on-line tuning of protection, has relatively broad adaptability.

Claims (3)

1. a distributed power generation island detection method with automatic measure on line ability, it is characterised in that comprise the following steps:

1) microgrid SCADA system is utilized to obtain original sample online；

2) utilize the on-line talking method containing micro-bunch, adaptively original sample is carried out re-sampling；

3) calculating the weight of each sample set according to multi-categorizer model and select the superior and eliminate the inferior, superseded classification accuracy is less than Set the sample set of threshold value, it is thus achieved that preferably sample set；

4) according to preferred sample set, on-line training generates sorter model；

5) sorter model utilizing above-mentioned automatic measure on line to obtain, updates the classification used by the detection of real-time isolated island in an asynchronous manner Device model；

Described utilize microgrid SCADA system to obtain original sample online particularly as follows:

If sample set is:

E={ (x_i,C(x_i), i=1 ... N (1)

N is total sample number, x_iIt is characterized vector, C (x_i) { 0,1} is the class label of this sample to ∈, and 0 is non-isolated island, and 1 is isolated island；

Using frequency, voltage magnitude, current amplitude, voltage-to-current phase angle difference, active power and reactive power as characteristic vector, It is designated as:

In formula (1), characteristic vector x_i6 eigenvalues reported by each distributed generation system DG, and class label C (x_i) from PCC On off state obtains, and in SCADA system, when being realized the network pair of low cost by simple network agreement SNTP, transmits x_iWith C (x_i) report all contain time tag, with C (x_i) timestamp on the basis of, to x_iCarry out linear interpolation, i.e. obtain the same time The sample instance in cross section.

A kind of distributed power generation island detection method with automatic measure on line ability the most according to claim 1, it is special Levy and be, described utilize the on-line talking method containing micro-bunch, adaptively original sample is carried out re-sampling particularly as follows:

Step 1, under off-line state, carry out cluster analysis, if Q is bunch number set, produce C₁...C_Q；For each bunch of C_q, Form micro-bunch of corresponding M_q；

When step 2, on-line operation, when new point?After moment arrives, calculateWith each bunch of C_qThe distance of barycenter

(1) ifThen create new bunch of C_new, turn to step 3；

(2) otherwise, by pointIt is merged into a bunch C_pIn, update bunch C_pMicro-bunch of corresponding M_p, return step 2；

Step 3, generalIncrease in training sample set, complete once to sample；Meanwhile, find out there is minimum CF1^tValue bunch, Abandoned, thus the sum of holding bunch is Q, returned step 2；

Wherein parameter is defined as follows:

Bunch point that C is tieed up by a series of dThe set constituted, the timestamp that each point is corresponding isDefining its micro-bunch of M is (2d+3) the first ancestral tieed up:

Wherein,For d dimensional vector, the value of its pth item is For d dimensional vector, the value of its pth item isN is the number at bunch midpoint；

The barycenter of bunch CIt is defined as:

The maximum boundary of bunch C is defined as τ σ, and wherein, τ ＞ 0 is the distance factor；σ be in bunch C each point to barycenter C_cDistance mean square Root error, its computing formula is:

$\mathrm{\σ}=\sqrt{\underset{p=1}{\overset{d}{\Σ}}(\frac{CF{2}^p}{n}-{\left(\frac{CF{1}^p}{n}\right)}^2)}---\left(5\right).$

A kind of distributed power generation island detection method with automatic measure on line ability the most according to claim 2, it is special Levy and be, update the sorter model used by the detection of real-time isolated island in an asynchronous manner, specific as follows:

501) setting data window and comprise k sub-sample set the most simultaneously, each subsample collection generates a grader M_i；

502) after cycle T, PCC on off state Tong Bu with the on-line sampling data block report that SCADA system is sent is received；

503) M is used₁,M₂,...,M_kK grader to new subsample collection c_k+1Carry out classification prediction respectively, according on off state Report can be classified error rate, with Error (M_i) represent M_iClassification error rate, wherein i=1,2 ..., k, M_k+1Intersect Verify its error rate is Error (M_k+1), Error (M_i) grader weight more than 0.5 is assigned to 0, calculates remaining weight:

${\widehat{\mathrm{\ω}}}_i={\log }_{10}\left(\frac{1-Error\left(M_i\right)}{Error\left(M_i\right)}\right)---\left(6\right)$

504) M is rejected₁,M₂,...,M_k+1The grader that middle weight is minimum, if there being two or more weight minimum and identical, Then reject time grader the earliest, the subsample of other k grader is integrated into preferred sample set；

505) grader using the training of preferred sample set to generate updates the grader for real-time grading prediction in an asynchronous manner Model；

506) circulation step 502) to 505).