CN106570526A - Classifier integration method for power transmission and transformation primary device load curve mining - Google Patents

Classifier integration method for power transmission and transformation primary device load curve mining Download PDF

Info

Publication number
CN106570526A
CN106570526A CN201610958899.7A CN201610958899A CN106570526A CN 106570526 A CN106570526 A CN 106570526A CN 201610958899 A CN201610958899 A CN 201610958899A CN 106570526 A CN106570526 A CN 106570526A
Authority
CN
China
Prior art keywords
dimensional
classification
classification results
dimentional
tuple
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.)
Pending
Application number
CN201610958899.7A
Other languages
Chinese (zh)
Inventor
庄池杰
胡军
张斌
段炼
尹立群
郭丽娟
张玉波
罗怿
曾嵘
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.)
Tsinghua University
Electric Power Research Institute of Guangxi Power Grid Co Ltd
Original Assignee
Tsinghua University
Electric Power Research Institute of Guangxi Power Grid Co Ltd
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 Tsinghua University, Electric Power Research Institute of Guangxi Power Grid Co Ltd filed Critical Tsinghua University
Priority to CN201610958899.7A priority Critical patent/CN106570526A/en
Publication of CN106570526A publication Critical patent/CN106570526A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

A classifier integration method for power transmission and transformation primary device load curve mining comprises the following steps: performing k times of iteration sampling with replacement on a data set D to get k training sets Di; machine-learning the training sets Di to get classification models M and corresponding classifiers; inputting the data set D to the classification models M to get classification results Q; and classifying the classification results to determine the nature of tuples. The beneficial effect lies in that an integrated classifier model built through the method can effectively improve the accuracy of equipment load abnormal data identification. Based on the power transmission and transformation primary device load abnormality sorting result obtained by the method, the focus of inspection is put on the top-ranked devices. Thus, the operation and maintenance efficiency of devices is improved effectively, and cost is saved for electric power companies.

Description

For the categorizer integration method that power transmission and transformation main equipment load curve is excavated
Technical field
The present invention relates to power system big data excavation applications, particularly a kind of power transmission and transformation main equipment load curve is extremely right The excavation of elephant.
Background technology
By the excavation to power transmission and transformation main equipment load curve, many effective informations, a kind of effect therein can be obtained It is exactly power transmission and transformation main equipment load anomalous data identification.Due to main equipment faults itself, load harvester failure, communicate different Often, the anthropic factor such as data storage problem, stealing electricity, may all cause power transmission and transformation main equipment load data to occur abnormal.For Utilities Electric Co., at present conventional processing method is that this not only needs substantial amounts of human resourcess but also increase based on site inspection mostly Operation cost.By effective identification of power transmission and transformation main equipment load abnormal data, accurately judge to cause the various of data exception Factor, can realize causing the quick positioning of data exception link, improve equipment O＆M efficiency, save Utilities Electric Co.'s spending.
The content of the invention
The invention aims to solve the above problems, devise a kind of for the excavation of power transmission and transformation main equipment load curve Categorizer integration method.Specific design scheme is：
A kind of categorizer integration method excavated for power transmission and transformation main equipment load curve, integrated step is：
Step one, comprising p tuple dj(j=1,2, in data set D p), k time is carried out to data set D repeatedly For sampling with replacement, k training set D is obtainedi(i=1,2, k)；
Step 2, to k training set DiMachine learning is carried out, the learning method of the machine learning is n kinds, is obtained multiple Grader Mn, (n=1,2,3), each grader MnInclude k disaggregated model Mni；
Step 3, with iteration input mode, by each tuple d in data set DjIt is input into each disaggregated model MnqIn, obtain Obtain p*k*n output result djni；
Step 4, according to output result djni, with iteration input mode, to each tuple djBinary is carried out in the first dimension Classification, each tuple obtains k*n one-dimensional classification results Qjni；
Step 5, according to one-dimensional classification results Qjni, with iteration input mode, to each tuple djCarry out in the second dimension Binary classification, each tuple obtains n two-dimentional classification results Qjn；
Step 6, according to two-dimentional classification results Qjn, to each tuple djBinary classification is carried out in third dimension, each tuple Obtain three-dimensional categorisation result Qj；
Step 7, according to three-dimensional categorisation result Qj, judge each tuple d in data set DjClassification results.
In the step 4, first dimension is tuple dimension, described the step of the first dimension carries out binary classification For：
Step 4-1, number of iterations j=1, i=1, n=1 are taken, by tuple d1Input disaggregated model M11In, obtain a kind of element Classification results d111；
Disaggregated model M11The binary element for including is classification element r, classification element w, when the output valve of d1 is normal, is sentenced Disconnected to classify as classification element r, when the output valve of d1 is abnormal, judgement classifies as classification element w, the classification results d111To return Class result d111rWith categorization results d111wIn one kind.
Step 4-2, number of iterations j=(j=1)+1, i=1, n=1 is made, tuple d2 is input in disaggregated model M11 and carries out Binary classification,
Obtain a kind of element classification result d211, the classification results d211For categorization results d211rWith categorization results d211wIn It is a kind of；
Step 4-3, repeat step step 4-1, step 4-2p time, obtain p classification results dj11；
Step 4-4, number of iterations i=(i=1)+1 is taken, repeat step four-1, step 4-2 obtains p*k categorization results dj1i；
Step 4-5, number of iterations n=(n=1)+1 is taken, repeat step four-1, step 4-2, step 4-4 obtain n*p*k Individual categorization results djni。
In the step 5, second dimension is disaggregated model dimension, described to carry out binary classification in the second dimension Step is：
Step 5-1 is by j values are identical, the one-dimensional classification results Q of n value identicalsjniDivide in one-dimensional sorted group F, Obtain p*n one-dimensional sorted group Fjn, each one-dimensional sorted group FjnInclude k one-dimensional classification results Qjni；
Step 5-2, make number of iterations j=1, n=1, take one-dimensional sorted group F11, in one-dimensional sorted group F11In, make number of iterations i =1, judge the one-dimensional classification results Q111Species, if the one-dimensional classification results Q111For Q111r, then Q is made111=1, if described One-dimensional classification results Q111For Q111w, then the Q is made111=-1,
Step 5-3, make number of iterations j=1, take one-dimensional sorted group F11, in one-dimensional sorted group F11In, make number of iterations i=(i =1)+1, judge the one-dimensional classification results Q112Species, if the one-dimensional classification results Q112For Q112r, then Q is made112=1, if The one-dimensional classification results Q112For Q112w, then the Q is made112=-1,
Step 5-4, repeat step five-3 obtains one-dimensional sorted group F11In k one-dimensional classification results value, then The value of k one-dimensional classification results is sued for peace, two-dimentional classification value ∑ Q is obtained11If, two-dimentional classification value ∑ Q11>0, then judge One-dimensional sorted group F11In all one-dimensional classification results Qj1iIt is two-dimentional classification results Qj1rIf, two-dimentional classification value ∑ Q11<0, then sentence Disconnected one-dimensional sorted group F11In all one-dimensional classification results Qj1iIt is two-dimentional classification results Qj1w,
Step 5-5, make j=(j=1)+1, n=1 repeat steps five-2, step 5-3, step 5-4, and acquisition p is two-dimentional Classification results Qj1；
Step 5-6, make n=(n=1)+1, repeat step five-5 obtain p*n two-dimentional classification results Qjn.
In the step 6, the third dimension be grader dimension, the step that binary classification is carried out in third dimension Suddenly it is：
Step 6-1, by the identical two-dimentional classification results Q of j valuesjnDivide in sorted group F, obtain p two-dimentional sorted group Fj, each two-dimentional sorted group FjInclude n two-dimentional classification results Qjn；
Step 6-2, make number of iterations j=1 take two-dimentional sorted group F1, in two-dimentional sorted group F1In, number of iterations n=1 is made, judge The two-dimentional classification results Q11Species, if the two-dimentional classification results Q11For Q11r, then Q is made11=1, if the two dimension classification knot Fruit Q11For Q11w, then the Q is made11=-1,
Step 6-3, make number of iterations j=1, take two-dimentional sorted group F1, in two-dimentional sorted group F1In, make number of iterations n=(n= 1)+1, judge the two-dimentional classification results Q12Species, if the two-dimentional classification results Q12For Q12r, then Q is made12=1, if described two Dimension classification results Q12For Q12w, then the Q is made12=-1；
Step 6-4, repeat step six-3 obtains two-dimentional sorted group F1In n two-dimentional classification results value, then The value of n two-dimentional classification results is sued for peace, three-dimensional classification value ∑ Q is obtained1If, three-dimensional classification value ∑ Q1>0, then judge two Dimension sorted group F1In all two-dimentional classification results QjnIt is three-dimensional categorisation result QjrIf, three-dimensional classification value ∑ Q1<0, then judge two Dimension sorted group F1In all two-dimentional classification results QjnIt is three-dimensional categorisation result Qjw；
Step 6-5, make j=(j=1)+1, and repeat step five-2, step 5-3, step 5-4 obtain p three-dimensional categorisation As a result Qj, each tuple djCorrespond to a j value identical three-dimensional categorisation result.
If three-dimensional categorisation result is Qjr, then judge and its j value identical tuple djFor normal tuple, if three-dimensional categorisation knot Fruit is Qjw, then j values identical tuple d therewith is judgedjFor abnormal tuple.
The sample mode of the data set D is sampled for playback, and training set DiIn tuple d quantity be p, unit Group quantity of the d in training set Di is 0,1, it is multiple in one kind.
The disaggregated model MjIncluding decision-tree model, its classifying step is：
By training set DiRegard a node N as；
Analyzing and training collection DiVariable partitioning scheme, determine multiple partitioning scheme N；
Two optimum partitioning schemes are selected, node N is divided into into N1、N2；
By the dividing method to node N, to node N1、N2Split；
Repeat previous step, to training set DiMulti-stage division is carried out, decision tree is formed；
With remaining K-1 training set DiDecision tree is tested；
Carrying out node secateurs and node according to assay increases, and determines decision tree.
The disaggregated model MjIncluding Bayesian Classification Model, its expression formula is：
Wherein, dH、dXIn two linear properties variables, P (dH) it is dHPrior probability, P (dH|dX) it is dHPosteriority it is general Rate；P(dX) it is dXPrior probability, P (dX|dH) it is dXPosterior probability.
The disaggregated model MjIncluding supporting vector machine model,
The supporting vector machine model includes linear separability supporting vector machine model, linearly inseparable support vector machine mould Type,
The optimization problem of the linear separability supporting vector machine model can be solved by the method for convex quadratic programming；
The linearly inseparable supporting vector machine model can be mapped to the input data of lower dimensional space by kernel function High bit space, and then the optimal separating hyper plane in new described high bit space.
The kernel function includes many Xu formulas kernel functions, gaussian radial basis function and S-shaped kernel function.
The grader collection excavated for power transmission and transformation main equipment load curve obtained by the above-mentioned technical proposal of the present invention Into method, its advantage is：
The integrated classifier model set up by this method, can effectively be lifted power transmission and transformation main equipment load abnormal data and be distinguished The accuracy of knowledge.The power transmission and transformation main equipment load abnormality degree obtained by using this method is sorted, and equipment in the top is entered Row emphasis is verified, and can effectively improve equipment O＆M efficiency, saves Utilities Electric Co.'s spending.
Description of the drawings
Fig. 1 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Receiver operating curve's comparison diagram that collection statistical result passes through Decision-Tree Classifier Model statistical result with data set；
Fig. 2 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Collection statistical result is with data set by the detection number of Decision-Tree Classifier Model statistical result and recall ratio comparison diagram；
Fig. 3 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Receiver operating curve's comparison diagram that collection statistical result passes through Bayesian Classification Model statistical result with data set；
Fig. 4 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Collection statistical result is with data set by the detection number of Bayesian Classification Model statistical result and recall ratio comparison diagram；
Fig. 5 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Receiver operating curve's comparison diagram that collection statistical result passes through support vector cassification modeling statistics result with data set；
Fig. 6 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Collection statistical result is with data set by the detection number of support vector cassification modeling statistics result and recall ratio comparison diagram；
Fig. 7 is all main equipments that the present invention is predicted based on the pack sorter model group that supporting vector machine model is obtained The scatterplot of the extremely doubtful probability of the load data of object；
Fig. 8 is the distribution situation of the abnormal main equipment object of load data in the present invention；
In figure, ctree, Decision-Tree Classifier Model；Nb, Bayesian Classification Model；Svm, support vector cassification model； CtreeBag, the pack sorter model group obtained based on Decision-Tree Classifier Model；NbBag, is obtained based on Bayesian Classification Model The pack sorter model group for obtaining；SvmBag, based on the pack sorter model group that supporting vector machine model is obtained.
Specific embodiment
The present invention is specifically described below in conjunction with the accompanying drawings.
A kind of categorizer integration method excavated for power transmission and transformation main equipment load curve, integrated step is：
Step one, comprising p tuple dj(j=1,2, in data set D p), k time is carried out to data set D repeatedly For sampling with replacement, k training set D is obtainedi(i=1,2, k)；
Step 2, to k training set DiMachine learning is carried out, the learning method of the machine learning is n kinds, is obtained multiple Grader Mn, (n=1,2,3), each grader MnInclude k disaggregated model Mni；
Step 3, with iteration input mode, by each tuple d in data set DjIt is input into each disaggregated model MnqIn, obtain Obtain p*k*n output result djni；
Step 4, according to output result djni, with iteration input mode, to each tuple djBinary is carried out in the first dimension Classification, each tuple obtains k*n one-dimensional classification results Qjni；
Step 5, according to one-dimensional classification results Qjni, with iteration input mode, to each tuple djCarry out in the second dimension Binary classification, each tuple obtains n two-dimentional classification results Qjn；
Step 6, according to two-dimentional classification results Qjn, to each tuple djBinary classification is carried out in third dimension, each tuple Obtain three-dimensional categorisation result Qj；
Step 7, according to three-dimensional categorisation result Qj, judge each tuple d in data set DjClassification results.
In the step 4, first dimension is tuple dimension, described the step of the first dimension carries out binary classification For：
Step 4-1, number of iterations j=1, i=1, n=1 are taken, by tuple d1Input disaggregated model M11In, obtain a kind of element Classification results d111；
Disaggregated model M11The binary element for including is classification element r, classification element w, when the output valve of d1 is normal, is sentenced Disconnected to classify as classification element r, when the output valve of d1 is abnormal, judgement classifies as classification element w, the classification results d111To return Class result d111rWith categorization results d111wIn one kind.
Step 4-2, number of iterations j=(j=1)+1, i=1, n=1 is made, tuple d2 is input in disaggregated model M11 and carries out Binary classification,
Obtain a kind of element classification result d211, the classification results d211For categorization results d211rWith categorization results d211wIn It is a kind of；
Step 4-3, repeat step step 4-1, step 4-2p time, obtain p classification results dj11；
Step 4-4, number of iterations i=(i=1)+1 is taken, repeat step four-1, step 4-2 obtains p*k categorization results dj1i；
Step 4-5, number of iterations n=(n=1)+1 is taken, repeat step four-1, step 4-2, step 4-4 obtain n*p*k Individual categorization results djni。
In the step 5, second dimension is disaggregated model dimension, described to carry out binary classification in the second dimension Step is：
Step 5-1 is by j values are identical, the one-dimensional classification results Q of n value identicalsjniDivide in one-dimensional sorted group F, Obtain p*n one-dimensional sorted group Fjn, each one-dimensional sorted group FjnInclude k one-dimensional classification results Qjni；
Step 5-2, make number of iterations j=1, n=1, take one-dimensional sorted group F11, in one-dimensional sorted group F11In, make number of iterations i =1, judge the one-dimensional classification results Q111Species, if the one-dimensional classification results Q111For Q111r, then Q is made111=1, if described One-dimensional classification results Q111For Q111w, then the Q is made111=-1,
Step 5-3, make number of iterations j=1, take one-dimensional sorted group F11, in one-dimensional sorted group F11In, make number of iterations i=(i =1)+1, judge the one-dimensional classification results Q112Species, if the one-dimensional classification results Q112For Q112r, then Q is made112=1, if The one-dimensional classification results Q112For Q112w, then the Q is made112=-1,
Step 5-4, repeat step five-3 obtains one-dimensional sorted group F11In k one-dimensional classification results value, then The value of k one-dimensional classification results is sued for peace, two-dimentional classification value ∑ Q is obtained11If, two-dimentional classification value ∑ Q11>0, then judge One-dimensional sorted group F11In all one-dimensional classification results Qj1iIt is two-dimentional classification results Qj1rIf, two-dimentional classification value ∑ Q11<0, then sentence Disconnected one-dimensional sorted group F11In all one-dimensional classification results Qj1iIt is two-dimentional classification results Qj1w,
Step 5-5, make j=(j=1)+1, n=1 repeat steps five-2, step 5-3, step 5-4, and acquisition p is two-dimentional Classification results Qj1；
Step 5-6, make n=(n=1)+1, repeat step five-5 obtain p*n two-dimentional classification results Qjn.
In the step 6, the third dimension be grader dimension, the step that binary classification is carried out in third dimension Suddenly it is：
Step 6-1, by the identical two-dimentional classification results Q of j valuesjnDivide in sorted group F, obtain p two-dimentional sorted group Fj, each two-dimentional sorted group FjInclude n two-dimentional classification results Qjn；
Step 6-2, make number of iterations j=1 take two-dimentional sorted group F1, in two-dimentional sorted group F1In, number of iterations n=1 is made, judge The two-dimentional classification results Q11Species, if the two-dimentional classification results Q11For Q11r, then Q is made11=1, if the two dimension classification knot Fruit Q11For Q11w, then the Q is made11=-1,
Step 6-3, make number of iterations j=1, take two-dimentional sorted group F1, in two-dimentional sorted group F1In, make number of iterations n=(n= 1)+1, judge the two-dimentional classification results Q12Species, if the two-dimentional classification results Q12For Q12r, then Q is made12=1, if described two Dimension classification results Q12For Q12w, then the Q is made12=-1；
Step 6-4, repeat step six-3 obtains two-dimentional sorted group F1In n two-dimentional classification results value, then The value of n two-dimentional classification results is sued for peace, three-dimensional classification value ∑ Q is obtained1If, three-dimensional classification value ∑ Q1>0, then judge two Dimension sorted group F1In all two-dimentional classification results QjnIt is three-dimensional categorisation result QjrIf, three-dimensional classification value ∑ Q1<0, then judge two Dimension sorted group F1In all two-dimentional classification results QjnIt is three-dimensional categorisation result Qjw；
Step 6-5, make j=(j=1)+1, and repeat step five-2, step 5-3, step 5-4 obtain p three-dimensional categorisation As a result Qj, each tuple djCorrespond to a j value identical three-dimensional categorisation result.
If three-dimensional categorisation result is Qjr, then judge and its j value identical tuple djFor normal tuple, if three-dimensional categorisation knot Fruit is Qjw, then j values identical tuple d therewith is judgedjFor abnormal tuple.
The sample mode of the data set D is sampled for playback, and training set DiIn tuple d quantity be p, unit Group quantity of the d in training set Di is 0,1, it is multiple in one kind.
The disaggregated model MjIncluding decision-tree model, its classifying step is：
By training set DiRegard a node N as；
Analyzing and training collection DiVariable partitioning scheme, determine multiple partitioning scheme N；
Two optimum partitioning schemes are selected, node N is divided into into N1、N2；
By the dividing method to node N, to node N1、N2Split；
Repeat previous step, to training set DiMulti-stage division is carried out, decision tree is formed；
With remaining K-1 training set DiDecision tree is tested；
Carrying out node secateurs and node according to assay increases, and determines decision tree.
The disaggregated model MjIncluding Bayesian Classification Model, its expression formula is：
Wherein, dH、dXIn two linear properties variables, P (dH) it is dHPrior probability, P (dH|dX) it is dHPosteriority it is general Rate；P(dX) it is dXPrior probability, P (dX|dH) it is dXPosterior probability.
The disaggregated model MjIncluding supporting vector machine model,
The supporting vector machine model includes linear separability supporting vector machine model, linearly inseparable support vector machine mould Type,
The optimization problem of the linear separability supporting vector machine model can be solved by the method for convex quadratic programming；
The linearly inseparable supporting vector machine model can be mapped to the input data of lower dimensional space by kernel function High bit space, and then the optimal separating hyper plane in new described high bit space.
The kernel function includes many Xu formulas kernel functions, gaussian radial basis function and S-shaped kernel function.
So that data set is 6200 power transmission and transformation main equipment object load datas of 18 months as an example, sample frequency is 30 points Clock.Due to the abnormity of this paper primary study equipment long-term load data, therefore the unit of time studied takes one month, i.e., to original Beginning data set is processed, and calculates the monthly average load of each device object to reflect its data characteristic modes, i.e., 30 day.Therefore this One has 111600 load curves in example.6200 power transmission and transforming equipment objects are right comprising 6123 normal subjects and 77 exceptions As exception object ratio is 1.24%.The input of model is raw data set, is output as machine utilization data exception degree and doubtful Probability sorting.
Embodiment 1
Step one, feature extraction is carried out to all load curves in this example, the property set D of homologous thread is obtained, wherein common 11160 tuples d；
K sampling with replacement is carried out to set D, 31 different training sets D are obtainedi(i=1,2,31).By In adopting sampling with replacement, some tuples d of D may not D againiMiddle appearance, and part tuple is likely to occur repeatedly；
31 training sets D of step 2 pairiDecision tree machine learning is carried out, 31 disaggregated model M are obtained1i, by 31 moulds of classifying Type M1iClassify as grader M1；
To 31 training sets DiNaive Bayes machine learning is carried out, 31 disaggregated model M are obtained2i, by 31 disaggregated model M2i Classify as grader M2；
To 31 training sets DiVector machine machine learning is supported, 31 disaggregated model M are obtained3i, by 31 moulds of classifying Type M3iClassify as grader M3,
After above-mentioned machine learning, disaggregated model 3*31=91 is set up.
11160 tuples d in data set D are input into into 91 disaggregated models, 111160*91 output result is obtained, often 91 one-dimensional classification results Q of individual tuple d correspondence1ni。
Embodiment 2
Take first tuple d1ni, by its 91 one-dimensional classification results according to grader M1, grader M2, grader M3It is divided into Three groups, grader M1Include 31 one-dimensional classification Q11i(i=1,2,31), grader M2Include 31 one-dimensional point Class Q12i(i=1,2,31), grader M3Include 31 one-dimensional classification Q13i(i=1,2,31).
Judge 31 one-dimensional classification results Q in first tuple Classifier M1j1i, to first one-dimensional classification results Q111 Judged, if one-dimensional classification results Q111Show tuple d1Numerical value is normal, it is determined that one-dimensional classification results Q111Numerical value be 1, if one-dimensional classification results Q111Show the numerical exception of tuple d1, it is determined that one-dimensional classification results Q111Numerical value be -1, so Afterwards by same procedure to first tuple Classifier M1In second one-dimensional classification results Q112Judged, until by 31 Individual one-dimensional classification results Q11iAll carry out judgement to finish, 31 judged results are added, because 31 is odd number, its result is only possible to It is greater than 0 or less than 0, when it is more than 0, first tuple d1It is normal, when its result is less than 0, then judges first Tuple d1It is abnormal, an i.e. two-dimentional classification results Q of first tuple11。
Above-mentioned steps, determine that 11160 tuples pass through M1The result that grader is obtained is normal or abnormal, obtains every The two-dimentional classification results Q of individual tuple1n。
With same method, first tuple is classified using M2 graders, obtain second of first tuple Two-dimentional classification results Q12；
With same method, first tuple is classified using M3 graders, obtain the 3rd of first tuple Two-dimentional classification results Q13；
If two-dimentional classification results Q11Classification results explanation numerical value be normal, then two-dimentional classification results Q11=1, it is no Then two-dimentional classification results Q11=-1；
If two-dimentional classification results Q12Classification results explanation numerical value be normal, then two-dimentional classification results Q12=1, it is no Then two-dimentional classification results Q12=-1；
If two-dimentional classification results Q13Classification results explanation numerical value be normal, then two-dimentional classification results Q13=1, it is no Then two-dimentional classification results Q13=-1；
By two-dimentional classification results Q11、Q12、Q13Numerical value be added, obtain three-dimensional classification value ∑ Q1, the three-dimensional classification value ∑Q1Value be -3, -2, -1,1,2,3.
As three-dimensional classification value ∑ Q1When taking 1,2,3, then judge tuple d1 for normal tuple, i.e. Q1r, when three-dimensional classification value ∑Q1When taking -1, -2, -3, then judge tuple d1 for abnormal tuple, i.e. Q1w。
Embodiment 3
To the equal repeat the above steps of 11160 tuples, you can in judging 11160 tuples, whether each tuple is abnormal. Then three-dimensional categorization results are found out for QjwTuple, determine the category such as time, main equipment object, sampling time belonging to the tuple Property, it is possible to find out 77 warping apparatus objects.
Embodiment 4
Fig. 1 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Receiver operating curve's comparison diagram that collection statistical result passes through Decision-Tree Classifier Model statistical result with data set；Fig. 2 is It is of the present invention for power transmission and transformation main equipment load curve excavate categorizer integration method obtain data set statistical result with Detection number and recall ratio comparison diagram that data set passes through Decision-Tree Classifier Model statistical result, as shown in Figure 1 and Figure 2, based on decision-making Tree classification model obtains grader, then exports desired value by importing data to grader, and legacy data collection and desired value are entered The method of row matching, can effectively lift the classification performance of decision tree, improve classification accuracy.
Embodiment 5
Fig. 3 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Receiver operating curve's comparison diagram that collection statistical result passes through Bayesian Classification Model statistical result with data set；Fig. 4 is It is of the present invention for power transmission and transformation main equipment load curve excavate categorizer integration method obtain data set statistical result with Data set is by the detection number of Bayesian Classification Model statistical result and recall ratio comparison diagram；As shown in Figure 3, Figure 4, based on pattra leaves This disaggregated model obtains grader, then exports desired value by importing data to grader, and legacy data collection and desired value are entered The method of row matching, can effectively lift Bayesian classification performance, improve classification accuracy.
Embodiment 6
Fig. 5 is the data obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention Receiver operating curve's comparison diagram that collection statistical result passes through support vector cassification modeling statistics result with data set；Figure 6 is the data set statistical result obtained for the categorizer integration method that power transmission and transformation main equipment load curve is excavated of the present invention With data set by the detection number of support vector cassification modeling statistics result and recall ratio comparison diagram；Fig. 7 is that the present invention is based on The load data of all main equipment objects that the pack sorter model group that supporting vector machine model is obtained is predicted is extremely doubtful The scatterplot of probability, as illustrated in figs. 5-7, based on support vector cassification model grader is obtained, then by importing data to point Class device exports desired value, and the method that legacy data collection is matched with desired value can effectively lift dividing for support vector machine Class performance, improves classification accuracy.
Embodiment 7
Fig. 8 is the distribution situation of the abnormal main equipment object of load data in the present invention, as shown in figure 8, following in figure Scatterplot represents that original input data concentrates the position of abnormal main equipment object, it is seen that abnormal main equipment pair as if random distribution 's.Scatterplot above represents that exception is main to be set after the numerical value by doubtful probability is to the rearrangement of all main equipment objects in figure The distribution situation of standby object.It can be seen that most exception main equipment objects can be concentrated on into new data set by abnormality degree sequence Front portion, front 1000 main equipment objects include 65 exception objects, and then 5200 main equipment objects only include 12 exceptions Object.Emphasis verification can be carried out to the main equipment object of doubtful data exception according to the sequence, effectively improve equipment fortune inspection Efficiency.
Above-mentioned technical proposal only embodies the optimal technical scheme of technical solution of the present invention, those skilled in the art Some of which part is made some variation embody the present invention principle, belong to protection scope of the present invention it It is interior.

Claims (10)

1. a kind of categorizer integration method excavated for power transmission and transformation main equipment load curve, it is characterised in that integrated step is：
Step one, comprising p tuple dj(j=1,2, in data set D p), k iteration is carried out to data set D and is put Pumpback sample, obtains k training set Di(i=1,2, k)；
Step 2, to k training set DiMachine learning is carried out, the learning method of the machine learning is n kinds, obtains multiple classification Device Mn, (n=1,2,3), each grader MnInclude k disaggregated model Mni；
Step 3, with iteration input mode, by each tuple d in data set DjIt is input into each disaggregated model MnqIn, obtain p* K*n output result djni；
Step 4, according to output result djni, with iteration input mode, to each tuple djBinary classification is carried out in the first dimension, Each tuple obtains k*n one-dimensional classification results Qjni；
Step 5, according to one-dimensional classification results Qjni, with iteration input mode, to each tuple djBinary point is carried out in the second dimension Class, each tuple obtains n two-dimentional classification results Qjn；
Step 6, according to two-dimentional classification results Qjn, to each tuple djBinary classification is carried out in third dimension, each tuple is obtained One three-dimensional categorisation result Qj；
Step 7, according to three-dimensional categorisation result Qj, judge each tuple d in data set DjClassification results.
2., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 1, it is special Levy and be, in the step 4, first dimension is tuple dimension, described the step of the first dimension carries out binary classification For：
Step 4-1, number of iterations j=1, i=1, n=1 are taken, by tuple d1Input disaggregated model M11In, obtain a kind of element classification As a result d111；
Disaggregated model M11The binary element for including is classification element r, classification element w, when the output valve of d1 is normal, judges to return Class is classification element r, and when the output valve of d1 is abnormal, judgement classifies as classification element w, the classification results d111To sort out knot Fruit d111rWith categorization results d111wIn one kind.
Step 4-2, number of iterations j=(j=1)+1, i=1, n=1 is made, by tuple d2Input disaggregated model M11In carry out binary point Class,
Obtain a kind of element classification result d211, the classification results d211For categorization results d211rWith categorization results d211wIn one Kind；
Step 4-3, repeat step step 4-1, step 4-2p time, obtain p classification results dj11；
Step 4-4, number of iterations i=(i=1)+1 is taken, repeat step four-1, step 4-2 obtains p*k categorization results dj1i；
Step 4-5, number of iterations n=(n=1)+1 is taken, repeat step four-1, step 4-2, step 4-4 obtain n*p*k and return Class result djni。
3., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 2, it is special Levy and be,
In the step 5, second dimension is disaggregated model dimension, described the step of the second dimension carries out binary classification For：
Step 5-1 is by j values are identical, the one-dimensional classification results Q of n value identicalsjniDivide in one-dimensional sorted group F, obtain P*n one-dimensional sorted group Fjn, each one-dimensional sorted group FjnInclude k one-dimensional classification results Qjni；
Step 5-2, make number of iterations j=1, n=1, take one-dimensional sorted group F11, in one-dimensional sorted group F11In, number of iterations i=1 is made, Judge the one-dimensional classification results Q111Species, if the one-dimensional classification results Q111For Q111r, then Q is made111=1, if described one-dimensional Classification results Q111For Q111w, then the Q is made111=-1,
Step 5-3, make number of iterations j=1, take one-dimensional sorted group F11, in one-dimensional sorted group F11In, make number of iterations i=(i=1)+ 1, judge the one-dimensional classification results Q112Species, if the one-dimensional classification results Q112For Q112r, then Q is made112=1, if described one Dimension classification results Q112For Q112w, then the Q is made112=-1,
Step 5-4, repeat step five-3 obtains one-dimensional sorted group F11In k one-dimensional classification results value, then to k The value of one-dimensional classification results is sued for peace, and obtains two-dimentional classification value ∑ Q11If, two-dimentional classification value ∑ Q11>0, then judge one-dimensional Sorted group F11In all one-dimensional classification results Qj1iIt is two-dimentional classification results Qj1rIf, two-dimentional classification value ∑ Q11<0, then judge one Dimension sorted group F11In all one-dimensional classification results Qj1iIt is two-dimentional classification results Qj1w,
Step 5-5, make j=(j=1)+1, n=1 repeat steps five-2, step 5-3, step 5-4, obtain p two dimension classification As a result Qj1；
Step 5-6, make n=(n=1)+1, repeat step five-5 obtain p*n two-dimentional classification results Qjn.
4., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 3, it is special Levy and be, in the step 6, the third dimension is grader dimension, described the step of third dimension carries out binary classification For：
Step 6-1, by the identical two-dimentional classification results Q of j valuesjnDivide in sorted group F, obtain p two-dimentional sorted group Fj, Each two-dimentional sorted group FjInclude n two-dimentional classification results Qjn；
Step 6-2, make number of iterations j=1 take two-dimentional sorted group F1, in two-dimentional sorted group F1In, make number of iterations n=1, judge this two Dimension classification results Q11Species, if the two-dimentional classification results Q11For Q11r, then Q is made11=1, if the two-dimentional classification results Q11 For Q11w, then the Q is made11=-1,
Step 6-3, make number of iterations j=1, take two-dimentional sorted group F1, in two-dimentional sorted group F1In, number of iterations n=(n=1)+1 is made, Judge the two-dimentional classification results Q12Species, if the two-dimentional classification results Q12For Q12r, then Q is made12=1, if the two dimension point Class result Q12For Q12w, then the Q is made12=-1；
Step 6-4, repeat step six-3 obtains two-dimentional sorted group F1In n two-dimentional classification results value, then to n The value of two-dimentional classification results is sued for peace, and obtains three-dimensional classification value ∑ Q1If, three-dimensional classification value ∑ Q1>0, then judge two dimension point Class group F1In all two-dimentional classification results QjnIt is three-dimensional categorisation result QjrIf, three-dimensional classification value ∑ Q1<0, then judge two dimension point Class group F1In all two-dimentional classification results QjnIt is three-dimensional categorisation result Qjw；
Step 6-5, make j=(j=1)+1, and repeat step five-2, step 5-3, step 5-4 obtain p three-dimensional categorisation result Qj, each tuple djCorrespond to a j value identical three-dimensional categorisation result.
5., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 4, it is special Levy and be, if three-dimensional categorisation result is Qjr, then judge and its j value identical tuple djFor normal tuple, if three-dimensional categorisation knot Fruit is Qjw, then j values identical tuple d therewith is judgedjFor abnormal tuple.
6., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 1, it is special Levy and be, the sample mode of the data set D is sampled for playback, and training set DiIn tuple d quantity be p, unit Group quantity of the d in training set Di is 0,1, it is multiple in one kind.
7., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 1, it is special Levy and be, the disaggregated model MjIncluding decision-tree model, its classifying step is：
By training set DiRegard a node N as；
Analyzing and training collection DiVariable partitioning scheme, determine multiple partitioning scheme N；
Two optimum partitioning schemes are selected, node N is divided into into N1、N2；
By the dividing method to node N, to node N1、N2Split；
Repeat previous step, to training set DiMulti-stage division is carried out, decision tree is formed；
With remaining K-1 training set DiDecision tree is tested；
Carrying out node secateurs and node according to assay increases, and determines decision tree.
8., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 1, it is special Levy and be, the disaggregated model MjIncluding Bayesian Classification Model, its expression formula is：
P ( d H | d X ) = P ( d X Td H ) P ( d H ) P ( d X )
Wherein, dH、dXIn two linear properties variables, P (dH) it is dHPrior probability, P (dH|dX) it is dHPosterior probability；P (dX) it is dXPrior probability, P (dX|dH) it is dXPosterior probability.
9., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 1, it is special Levy and be, the disaggregated model MjIncluding supporting vector machine model,
The supporting vector machine model includes linear separability supporting vector machine model, linearly inseparable supporting vector machine model,
The optimization problem of the linear separability supporting vector machine model can be solved by the method for convex quadratic programming；
The input data of lower dimensional space can be mapped to a high position by the linearly inseparable supporting vector machine model by kernel function Space, and then the optimal separating hyper plane in new described high bit space.
10., according to the categorizer integration method excavated for power transmission and transformation main equipment load curve described in claim 5, it is special Levy and be, the kernel function includes many Xu formulas kernel functions, gaussian radial basis function and S-shaped kernel function.
CN201610958899.7A 2016-10-27 2016-10-27 Classifier integration method for power transmission and transformation primary device load curve mining Pending CN106570526A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610958899.7A CN106570526A (en) 2016-10-27 2016-10-27 Classifier integration method for power transmission and transformation primary device load curve mining

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610958899.7A CN106570526A (en) 2016-10-27 2016-10-27 Classifier integration method for power transmission and transformation primary device load curve mining

Publications (1)

Publication Number Publication Date
CN106570526A true CN106570526A (en) 2017-04-19

Family

ID=58535782

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610958899.7A Pending CN106570526A (en) 2016-10-27 2016-10-27 Classifier integration method for power transmission and transformation primary device load curve mining

Country Status (1)

Country Link
CN (1) CN106570526A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034241A (en) * 2018-07-24 2018-12-18 南京千智电气科技有限公司 Load cluster control method and system based on support vector machines
CN109272058A (en) * 2018-11-27 2019-01-25 电子科技大学中山学院 Integrated power load curve clustering method
CN111751650A (en) * 2020-07-06 2020-10-09 重庆大学 Non-invasive household electric equipment on-line monitoring system and fault identification method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034241A (en) * 2018-07-24 2018-12-18 南京千智电气科技有限公司 Load cluster control method and system based on support vector machines
CN109272058A (en) * 2018-11-27 2019-01-25 电子科技大学中山学院 Integrated power load curve clustering method
CN111751650A (en) * 2020-07-06 2020-10-09 重庆大学 Non-invasive household electric equipment on-line monitoring system and fault identification method
CN111751650B (en) * 2020-07-06 2021-06-22 重庆大学 Non-invasive household electric equipment on-line monitoring system and fault identification method

Similar Documents

Publication Publication Date Title
Xiao et al. Feature-selection-based dynamic transfer ensemble model for customer churn prediction
CN107944480A (en) A kind of enterprises &#39; industry sorting technique
CN105373606A (en) Unbalanced data sampling method in improved C4.5 decision tree algorithm
Vijaya et al. Computing efficient features using rough set theory combined with ensemble classification techniques to improve the customer churn prediction in telecommunication sector
CN100595780C (en) Handwriting digital automatic identification method based on module neural network SN9701 rectangular array
Pineda-Bautista et al. General framework for class-specific feature selection
CN102521656A (en) Integrated transfer learning method for classification of unbalance samples
CN101464964A (en) Pattern recognition method capable of holding vectorial machine for equipment fault diagnosis
CN101819601A (en) Method for automatically classifying academic documents
Pardeshi et al. Improved k-medoids clustering based on cluster validity index and object density
Lawi et al. Ensemble GradientBoost for increasing classification accuracy of credit scoring
CN104598586A (en) Large-scale text classifying method
CN103400190A (en) Integrated framework method for optimizing extremity learning machine by using genetic algorithm
CN106570526A (en) Classifier integration method for power transmission and transformation primary device load curve mining
Hall et al. An overview of machine learning with SAS® enterprise miner™
CN112767106B (en) Automatic auditing method, system, computer readable storage medium and auditing equipment
CN105045913A (en) Text classification method based on WordNet and latent semantic analysis
Hebert Predicting rare failure events using classification trees on large scale manufacturing data with complex interactions
CN106156163A (en) File classification method and device
CN101295362A (en) Combination supporting vector machine and pattern classification method of neighbor method
CN104699819A (en) Sememe classification method and device
Hirst Machine learning for Hilbert series
Patel et al. A reduced error pruning technique for improving accuracy of decision tree learning
CN109858541A (en) A kind of specific data self-adapting detecting method based on data integration
Kumar et al. Classification of faults in Web applications using machine learning

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20170419

RJ01 Rejection of invention patent application after publication