CN106874963A - A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology - Google Patents
A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology Download PDFInfo
- Publication number
- CN106874963A CN106874963A CN201710159906.1A CN201710159906A CN106874963A CN 106874963 A CN106874963 A CN 106874963A CN 201710159906 A CN201710159906 A CN 201710159906A CN 106874963 A CN106874963 A CN 106874963A
- Authority
- CN
- China
- Prior art keywords
- attribute
- property
- decision table
- discretization
- distribution network
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/043—Architecture, e.g. interconnection topology based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Physiology (AREA)
- Genetics & Genomics (AREA)
- Automation & Control Theory (AREA)
- Computational Mathematics (AREA)
- Fuzzy Systems (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of Fault Diagnosis Method for Distribution Networks based on big data technology and system, be it is a kind of data volume it is big, dimension is more, in the power distribution network more than data class, the mechanism of failure and rapid rush-repair present in existing big data information diagnosis power distribution network promptly and accurately, the system can be utilized mainly includes three parts：Data discrete device, attribute reduction device, sample training device.Rough set theory is used for the present invention pretreatment of neural metwork training data, first with the theoretical calculation reduction of rough set and create-rule, unnecessary conditional attribute is deleted during yojan, it is favorably improved the learning efficiency of network, maintain the approximate classification error rate of relatively low stabilization, it is last into rule carry out high-level abstract expression as the inference machine of power distribution network, can well ensure the safe operation of power distribution network.
Description
Technical field
The present invention relates to a kind of Fault Diagnosis Method for Distribution Networks based on big data technology and system, for solving power distribution network
The problem of on-line fault diagnosis, belongs to Distributed Calculation software field.
Background technology
With the fast development of modern electric, the continuous increase of power distribution network scale, the inevitable property of distribution network failure is again
So that improving Fault Diagnosis of Distribution Network rate turns into one of key technical index for measurement power supply reliability.Distribution network failure
Diagnosis is the important process of distribution network operation, but failure cause, phenomenon of the failure, failure process are intricate, to improve distribution
The rapidity of fault diagnosis in net, domestic and foreign scholars propose the various faults such as fuzzy theory, genetic algorithm, artificial neural network
Diagnostic method, when fault diagnosis be based on information it is correct, it is complete when, the result that these methods can be more satisfied with,
But disturbed and many uncertain factors such as loss information because the information in power distribution network is present, the big data in power distribution network is big
Amount, at a high speed, changeable information, it is that the amount and complexity of data develops into the product in certain stage, to data computing capability,
The aspects such as the operational efficiency of parser propose requirement higher, and the above method has some limitations.
A Single Point of Faliure as processed not in time in power distribution network, it will causes the extension of failure, or even can cause personnel
With the massive losses of property, the generation of failure how is reduced in power distribution network, and can in time be processed after failure generation, rapidly
Resume production, it has also become ensure the key point of power distribution network safe operation.The height reliability of distribution network system and maintainability
Closely related fault diagnosis technology seems increasingly important, the existing faulty tree diagnosis of method for diagnosing faults, fault mode
Method of identification, the diagnostic method based on expert system, the fault diagnosis based on neutral net, the fault diagnosis based on rough set etc.,
But all there are some drawbacks in these methods, it is impossible to timely and effectively diagnose online failure, it is necessary to propose in fault diagnosis
A kind of strong robustness, the Fault Diagnosis Method for Distribution Networks for adapting to various rough sledding, help dispatcher to recognize failure rapidly, protect
Demonstrate,prove the safe and stable operation of power distribution network.Therefore, scientific and efficient diagnosis to the failure in power distribution network is supplied improving power distribution network
Electric reliability and power distribution network service quality are significant.
The content of the invention
The technical problems to be solved by the invention are to provide a kind of Fault Diagnosis Method for Distribution Networks based on big data technology
And system solves troubleshooting issue under power distribution network big data environment, the present invention is a kind of tactic method, by using this
Method can cause that the failure in power distribution network obtains quick diagnosis, it is ensured that the safe and stable operation of power distribution network.
The present invention uses following technical scheme to solve above-mentioned technical problem：
On the one hand, the present invention provides a kind of Fault Diagnosis Method for Distribution Networks based on big data technology, comprises the following steps that：
Step 1：Data to being collected in power distribution network carry out the extraction of continuous property, and the continuous property for extracting is made
It is original decision table, for carrying out network training；
Step 2：Sliding-model control is carried out to the continuous property in the original decision table of formation in step 1, after discretization
Continuous property as initial decision table；
Step 3：Using the method for Dynamic Reduct Based, structural environment attribute reduction function, to the condition category in initial decision table
Property carries out yojan, the minimal condition property set of formation；
Step 4, using the power distribution network data set corresponding to minimal condition property set in step 3 as new training sample set；
Step 5：New training sample set pair BP neural network according to being formed in step 4 is trained study；
Step 6：The BP neural network that will be tested in step 5 is used as Fault Diagnosis of Distribution Network device, so as to realize online event
Barrier diagnosis.
Used as further prioritization scheme of the invention, continuous property includes conditional attribute and decision attribute values in step 1.
The continuous property in original decision table carried out as further prioritization scheme of the invention, in step 2 discrete
Change is processed, specially：
2.1) the connection attribute value set for defining training sample set is w, orderUsing similar matrix
Continuous property in original decision table is arranged as the form of matrix, wherein, adiabatic index function It is the l in original decision table1Generic attribute,It is the kth in original decision table1Generic attribute；
2.2) collection for defining different element composition in similar matrix H is combined into K, to any q ∈ K successively using q- from
Relation formula Lq={ (s, t) | (L (s, t) >=q } calculate in original decision table between any two continuous property from pass
System, wherein, s, t are the continuous property of training sample concentration, and L (s, t) is the fuzzy relation of continuous property s and t, q ∈
[0,1]；Cluster result A1 is obtained with net-making method, the continuous property in original decision table is carried out just from Automated generalization by q-
Step discretization, forms preliminary discretization property set；
2.3) reservation degree function is constructedProperty set to preliminary discretization carries out picking for redundancy
Remove, wherein, WcX () is the reservation collection of decision attribute x, U is the property set of preliminary discretization, RcX () represents decision attribute x to bar
The reservation degree of part attribute c；
2.4) structural classification control functionTo step 2.3) the middle discretization rejected after redundancy
Property set carries out clustering processing, wherein, nlTo reject the number of l generic attributes in the discretization property set after redundancy, m is rejecting
Total class number of attribute in discretization property set after redundancy, y is to reject in the discretization property set after redundancy in l generic attributes
The heart,Be reject redundancy after discretization property set in attribute class center, r be reject redundancy after discretization property set in
Attribute number,It is the distance in the discretization property set after rejecting redundancy between l generic attributes and kth generic attribute center；
2.4) using genetic algorithm for solving object function ming (x)；
2.5) comprehensive function Q is definedq=n1Rc(x)+n2G (x), wherein, n1And n2It is weight coefficient, makes Q=0, if Qq-Q
＜ 0, then be transferred to step 2.2)；If Qq- Q ＞ 0, then make A=A1 and Q=Qq, then it is transferred to step 2.2)；
2.6) at the end of the connection attribute collection in original decision table is discrete, Q is takenqMaximum as cluster result A, and
The all kinds of of cluster result are encoded.
As further prioritization scheme of the invention, step 2.4) it is middle using genetic algorithm for solving object function ming (x),
Idiographic flow is as follows：
A) using step 2.3) in reject the discretization property set after redundancy as initial population, and initialize；
B) fitness function is utilizedThe quality of individuality in population is evaluated, wherein, N is
It is the range of choice of each point in discretization property set conditional Attribute class, C(a,b)It is that a classes conditional attribute is fitted to b class conditional attributes
Response return value, TbIt is the fitness desired value of b class conditional attributes；
C) selection that the size according to fitness value in b) is selected the superior and eliminated the inferior to the individuality in population；
D) individuality for selecting fitness value larger from current population using predetermined probability breeds filial generation as parent；
E) gene intersection, gene mutation and restructuring are carried out to current sub- population at individual；
F) population heredity of future generation, function to achieve the objective are enteredOptimal value approach.
As further prioritization scheme of the invention, using the method for Dynamic Reduct Based, structural environment attribute reduction in step 3
Function Si=m1Li+m2Ni, wherein, m1And m2It is importance degree weight, and m1+m2=1；LiIt is relation on attributes importance degree,Wherein, n is the number of initial decision table conditional attribute, NiIt is attributive character importance degree,| * | represents the data amount check included in attribute set, and D is the property set of initial decision table, and G is special data
Levy constant, MivIt is conditional attribute TiAnd TvBetween dependence, posj(D) represent jth class conditional attribute collection in initial decision table
In the important series of feature, j=1 ..., n.
On the other hand, the present invention also provides a kind of Fault Diagnosis of Distribution Network system based on big data technology, including data
Discretizer, attribute reduction device, sample training device, wherein, data discrete device, for being carried out to the data collected in power distribution network
Continuous property carries out sliding-model control, forms initial decision table；Attribute reduction device, for the condition category in initial decision table
Property carry out yojan, form minimal condition property set；Sample training device, for according to the corresponding distribution netting index of minimal condition property set
Learn according to BP neural network, the BP neural network that output training is completed is used as Fault Diagnosis of Distribution Network device.
The present invention uses above technical scheme compared with prior art, with following technique effect：The inventive method is a kind of
Based on the Fault Diagnosis Method for Distribution Networks of big data technology, failure diagnoses problem in time in being mainly used in solving power distribution network, passes through
Using in the present invention propose method can be in current power distribution network big data, using rough set theory and BP neural network
The method being combined is diagnosed in time to the failure in power distribution network, so as to ensure the safe and stable operation of power distribution network well.
Brief description of the drawings
Fig. 1 is Fault Diagnosis of Distribution Network system construction drawing.
Fig. 2 is reference architecture schematic diagram.
Fig. 3 is the schematic flow sheet of the inventive method.
Specific embodiment
Technical scheme is described in further detail below in conjunction with the accompanying drawings：
Fault Diagnosis of Distribution Network system based on big data technology mainly needs to consider two problems of aspect：(1) how from
The correlation of knowledge, and the study by there is supervision are extracted in fuzzy, uncertain, incomplete information, by training dataset
Conjunction carries out classification merger, excavates the online failure cause of power distribution network presence.(2) how the nonlinear spy of distribution network failure is directed to
Point, finds a kind of with very strong learning ability, and the method for adaptability and robustness is classified and built to distribution network failure
Found a simplified fault diagnosis system.
The method of the present invention is a kind of method of tactic, will not known in power distribution network by using rough set theory, no
Complete information yojan removal redundant attributes, simplified rule are learnt by the training of BP neural network, the god for training
It is used in power distribution network as diagnostic reasoning machine through network, so as to solve the problems, such as the on-line fault diagnosis of power distribution network.
First, architecture
Fig. 1 gives the Fault Diagnosis of Distribution Network system construction drawing based on big data technology, and it mainly includes three parts：
Data discrete device, attribute reduction device, sample training device.Data discrete device in Fig. 1 will be obscured in system, uncertain original number
According to Discretization for Continuous Attribute；Attribute reduction device is used for removing the information abstraction rule and reduction rules of redundancy；Sample training device
For simplified rule to be trained into study.
Specific introduction is given below：
Data discrete device：Data discrete device is mainly used in the sliding-model control to the big data connection attribute in distribution, this
To carry out the attribute reduction of data in invention with rough set theory, but rough set theory is one kind is entered based on discretization data
The method of row treatment, the attribute discretization of continuous data directly influences its treatment effect, and reservation degree letter is proposed in the present invention
Number RcX () and the method for classification control function g (x), realizes the cluster analysis function of Automatic feature recognition.Separately, it is right in this patent
Implementing for data discrete device do not do any limitation.
Attribute reduction device：Under conditions of keeping knowledge classification ability constant, carrying out attribute reduction to information table can be with letter
Change the complexity of information system, attribute reduction device is deleted unnecessary conditional attribute and (deleted mainly by structure attribute reduced function
A certain row in except information table), the row that cancellation is repeated, and the redundancy of attribute of each decision rule is eliminated, number is realized with this
According to the Dynamic Reduct Based of attribute.
Sample training device：Sample training device is mainly directed to the nonlinear feature of distribution network failure with BP neural network technology
The characteristics of very strong learning ability, adaptability, robustness for having with neutral net, with BP neural network in power distribution network
Failure classified, set up a simplified fault diagnosis system.Is trained to simplified rule with BP neural network
Practise, and using the BP neural network for training as distribution network system diagnostic reasoning machine.
2nd, method flow
1st, data discrete device
By continuous attribute discretization, it is critical only that the selection of breakpoint, thus attribute Discretization be summed up as
What selection breakpoint carries out rational partition problem to the space that conditional attribute is constituted.Rough set can only be grasped to discrete attribute
Make, and the quality of attribute discretization result directly affects last analysis, so the problem that rough set faces is how will
Continuous attribute discretization, constructs reservation degree function R in this patentcX () makes discrete result reflect original data as far as possible
Information, reduces the loss of data message, while again without the data message of redundancy as far as possible.In view of making in discrete classification
Each point as far as possible together, and makes significantly separate between class in class, and other structural classification control function g (x) comes in this patent
The effect of reaction classification, therefore discrete results are by the two function co- controllings.
(1) reservation degree function is：
In formula, c is conditional attribute, and x decision attributes, D is the property set of training sample set, RcX () is reservation degree of the x to c,
I.e.：
WcX () is the reservation collection of x, W is the reservation subset of training sample set, and D is property set (including the bar of training sample set
Part attribute and decision attribute), V is the conditional attribute reservation degree collection of training sample set；
Define comprehensive function Qq=n1Rc(x)+n2G (x), wherein, n2And n1It is reservation degree function and control function of classifying
Weight coefficient, rational distribution weight is crucial；
(2) classification control function is：
In formula, m is data classification number, and y is the center of l class conditional attribute values,It is all conditional attribute value centers, r
It is the sample data number by reservation degree function discretization, n is the number of kth class data,It is kth class data and l classes
Between class distance.Wherein the molecule of g (x) represents the distance between class and class, and denominator represents the distance of element in class, so g (x)
Value is smaller, and it is more reasonable to classify.
Departure process in the present invention is as follows：
A) connection attribute (including conditional attribute and decision attribute) of training sample set carried out one by one in this patent discrete
Change.It is w that the connection attribute of definition training sample set first integrates (including conditional attribute and decision attribute).And setUsing phase
Like matrixThe form of matrix is arranged as to training sample set continuous property, wherein using adiabatic index functionAs similar function, in formulaFor training sample concentrates l1Generic attribute,For training sample is concentrated
Kth1Generic attribute.
B) comprehensive function Q is definedq=n1Rc(x)+n2G (x), wherein n1And n2It is reservation degree function and control function of classifying
Weight coefficient, rational distribution weight is crucial.
C) it is the fuzzy relation on D × D to set L, and fuzzy relation meets： And set Lq={ (s, t) | (L (s, t) >=q } is the q- of L from relation,
Represent the company concentrated for training sample from relation, wherein s, t, w, c, z between training sample set any two continuous property
Continuous attribute, L (s, t) is the fuzzy relation of connection attribute s and t, the wherein arbitrary value of q ∈ [0,1].For similar matrix H, if its
All different integrating for element composition are K, and q- is made successively from relation i.e. L to q ∈ Kq。
D) q is obtained from relation G successivelyq, cluster result A1 is obtained according to net-making method, calculate QPValue.By q- from relation
The preliminary discretization of the continuous property of training sample set is formed preliminary discretization property set by treatment, and discrete result is as far as possible
The original data message of reflection, reduces the loss of data message as far as possible.
If e) Qp- Q ＜ 0, then be transferred to b) step, next q is calculated, if Qp- Q ＞ 0, then make A=A1, Q=Qp, enter back into
B) step.
F) at the end of the connection attribute collection that training sample is concentrated is discrete, Q is takenqMaximum as cluster result A, be used in combination
1,2,3 ... it is all kinds of to cluster result to encode.
(3) in order that reaching preferable Clustering Effect in training sample set connection attribute departure process, constructed in the present invention
Classification control function g (x) carries out clustering processing, distance, denominator generation between the molecules present Attribute class and class of control function of classifying
The distance between in Table Properties class, so target function value more subclassification is more reasonable.
Thought in the present invention using genetic algorithm solves the minimum value of object function g (x), and main working process is as follows：
A) input training sample set is used as initial population.
B) by client initialization population.
C) fitness function is utilized, i.e.,：
The quality of individuality in population is evaluated, wherein N is the range of choice of each point in original sample Attribute class, C(a,b)It is a classes
Attribute data is to the fitness return value of b generic attributes, TbIt is the fitness desired value of b generic attributes.According to the size of fitness value
The selection selected the superior and eliminated the inferior to the individuality in population, the probability that the bigger individuality of fitness value is chosen to remain is bigger.
D) retain the optimum individual in per generation heredity, i.e., select from current population larger individual of fitness value with certain probability
Body breeds filial generation, the effective convergence for ensureing genetic algorithm as parent.
E) gene intersection, gene mutation and restructuring are carried out to current sub- population at individual.
F) population heredity of future generation, function to achieve the objective are enteredOptimal value is approached.nlFor
The number of l classes, m is training sample set class number, and y is the center of l generic attributes,It is the class center of training sample set attribute, r
It is training sample set data amount check,It is l classes in kth class data and the class distance between centers of kth, the molecules present of object function
Distance between Attribute class and class, denominator the distance between is represented in Attribute class, so target function value more subclassification is more reasonable.Profit
The thought final output approached with genetic algorithm optimal value makes significantly separate between class, and each point is as far as possible together most in class
Excellent decision table.
2nd, attribute reduction device
Attribute reduction refers to logarithm while retaining key message in the case where nicety of grading is not reduced in rough set
According to carrying out abbreviation and try to achieve the minimum expression of knowledge, concept simple mode is disclosed, and the dependence pass between recognizing and assessing data
System.Yojan causes that identical decision-making is drawn by less condition, therefore just can obtain same precision with less judgement
As a result, the process and method of yojan seem especially important.The process of attribute reduction is exactly the process for deleting redundancy condition attribute, is looked for
The least reduction for going out a decision table is a np hard problem.
In the present invention using the method construct attribute reduction function of Dynamic Reduct Based, data qualification attribute TiImportance degree by
2 dimensions of conditional attribute relation and conditional attribute feature are constituted, and are expressed as Si.Conditional attribute relation importance degree reflects single condition
Influence of the attribute to other conditional attributes of whole system.If 1 conditional attribute is related to other conditional attributes, i.e. attribute Tj's
Depend on TiWhen, it is believed that task TiThere is relation importance degree higher, it is believed that both attributes can not be differentiated, a model can be classified as
Farmland；When conditional attribute collection B is independent, there is B=J-e so that ind (J)=ind (B), then B is called the yojan of A, wherein e
It is a conditional attribute collection of sample, attribute reduction function is represented with following formula, i.e.,：
Si=miLi+m2Ni
In formula：m1And m2It is importance degree weight, m1+m2=1；LiIt is relation on attributes importance degree.NiIt is attributive character importance degree.
The relation importance degree of attribute is represented with the dependence between data attribute, matrix, M is set toij=0 one represents attribute TiAnd TjIt
Between there is no dependence；Mij=1 represents conditional attribute TjNeed to rely on Ti.Therefore, the relation importance degree of attribute is：
The characteristic importance of attribute is：
0≤N in formulai≤1.| * | represents the data amount check included in attribute set, and D is the property set of training sample set, G
It is data characteristics constant, MivIt is attribute TiAnd TvBetween dependence, n for initial decision table conditional attribute number, posA
(D) feature important series of the jth class conditional attribute collection in initial decision table, j=1 ..., n are represented.
3rd, sample training device
BP neural network main composition data sample training aids, BP neural network is made up of three layers of neuron, i.e., input layer,
Hidden layer, output layer.Same layer neuron is mutually not attached in network, is connected with each other between different neuronal layers, BP neural network
One layer of state of neuron under the influence of the state of each layer of neuron in forward-propagating, sample input quantity is input into from input layer
By hidden layer, last incoming output layer.When the output layer of forward-propagating is not input into preferably, that is, it is transferred to nerve net
The backpropagation of network, is now input to input layer by error signal, constantly adjusts the threshold value and weights of each layer, makes error gradually
Diminish, data sample finally makes weight convergence in smallest point by repetition training.
(1) the forward-propagating process of BP neural network：
If o-th data sample input vector is Yo=(yo1,yo2,......yon), obtain v-th node of hidden layer
Output pvI.e.：
F is Sigmiod functions in formula, i.e.,U is input layer, hivIt is input layer to hidden layer
Connection weight, xkvIt is b-th input of sample, θvIt is v-th neuron threshold value of input layer.D-th node of output layer
Output HrI.e.：
S is hidden layer neuron number w in formularmIt is the connection weight of hidden layer to output layer, θrIt is r-th nerve of hidden layer
The threshold value of unit.
(2) back-propagation process of BP neural network：
H is provided with to learning sample (Ag,Bg), (g=1,2...... p), actually export Bg' and require BgError letter
Number is：
BP algorithm will be along DgNegative gradient direction constantly change weights and make the network convergence, each knots modification be：
0 ＜ η ＜ 1 in formula, η is learning efficiency.For the connection weight of hidden layer output quantity
Input layer is to the connection weight of hidden layer：
A kind of main working process of the Fault Diagnosis Method for Distribution Networks based on big data technology of the present invention is：
(1) continuous property of the data that will be collected in power distribution network as system original sample, and by original sample
In data it is random be divided into two parts：A part as training sample set, as original decision table, for carrying out network instruction
Practice；Another part as test sample collection, for testing the network after training.
(2) sliding-model control is carried out to the continuous property in original decision table, reservation degree function R is proposed in the present inventionc
X () and the method for classification control function g (x), control function g (x) will be classified as target with the thought of Gene hepatitis B vaccine
Function, and optimization approximation process is carried out to object function, realize the cluster analysis function of Automatic feature recognition.
(3) conditional attribute and decision attribute values after discretization forms initial decision table, the every a line wherein in decision table
One object of description, an attribute of each of which row corresponding objects.With the method construct attribute reduction function of Dynamic Reduct Based
Si=miLi+m2NiMake decision table ensure not containing unnecessary attribute under the conditions of classification is correct, form minimal condition property set.
(4) structure of neutral net is determined with training sample, BP neural network is carried out with the learning sample after yojan
Learning training and output category result；The BP neural network trained with test sample set pair is tested, the BP for having tested
Neutral net reduces the Diagnostic Time of neutral net as diagnostor in power distribution network, realizes on-line fault diagnosis.
A kind of Fault Diagnosis Method for Distribution Networks based on big data technology of the invention comprising the step of be：
Step 1：The continuous property of the data that will be collected in power distribution network as system original sample, and by original sample
Data in this it is random be divided into two parts：A part as training sample set, as original decision table, for carrying out network instruction
Practice；Another part as test sample collection, for testing the network after training.Into step 2；
Step 2：Continuous property sliding-model control.Continuous property in the original decision table of formation in step 1 is entered
Row sliding-model control, proposes reservation degree function R in the present inventioncX () and the method for classification control function g (x), realizes that feature is automatic
The cluster analysis function of identification.Into step 3；
Step 3：Conditional attribute after discretization and decision attribute values are formed into initial decision table, it is wherein every in decision table
A line describes an object, an attribute of each of which row corresponding objects.Into step 4；
Step 4：With the method for Dynamic Reduct Based, structure attribute reduced function, i.e.,：
Si=miLi+m2Ni
In formula：m1And m2It is importance degree weight, m1+m2=1；LiIt is relation on attributes importance degree.NiIt is attributive character importance degree.
Conditional attribute to decision table carries out yojan, makes decision table ensure not containing unnecessary attribute under the conditions of classification is correct, is formed
Minimal condition property set.Into step 5；
Step 5：The minimal condition property set that step 4 is formed constitutes new training sample, now only contains in sample and has an impact
The essential condition attribute of classification, it will not be necessary to which the conditional attribute has been removed.Into step 6；
Step 6：Learning training is carried out to BP neural network and is exported according to the new training sample set formed in step 5
Classification results.Into step 7；
Step 7：Tested with the BP neural network in the test sample set pair step 6 in step 1, and having tested
Neutral net realizes on-line fault diagnosis as diagnostor in power distribution network.
In the present invention：
Data discrete device is by using reservation degree function Rc(x) and the method for classification control function g (x), it is ensured that in bar
In the case that part attribute and decision attribute are constant, suitable segmentation point set is searched out, the space that conditional attribute is constituted is drawn
Point, using the method, the indistinguishable relation between sample that can be expressed by guarantee information system again can accurately be to reality
Big data in example is classified.By using data discrete device so that the big data attribute in power distribution network has automatic identification
Cluster analysis the characteristics of.
Attribute reduction device passes through structure attribute reduced function, so that the data in power distribution network are according to the relation between attribute
Importance degree and attributive character realize Dynamic Attribute Reduction.Each attribute of data is not in power distribution network big data information system
Of equal importance, or even some attributes are redundancies, the attribute reduction method based on importance degree is built upon decision-making in this patent
On the core collection of table, the conditional attribute of interpolation data successively, until meeting the attribute of added attribute set between dependency degree and
Untill the indistinguishability of attributive character reaches minimum, and the non-core attributes in data are deleted on the basis of this, until sample set
In attribute all meet it is indispensable untill.
Describe for convenience, it will be assumed that have following application example：
Big data in power distribution network has the characteristics of data volume is big, dimension is more, data class is more, to user, company and society
Meeting economy has huge value, if occurring in that failure in power distribution network, causes it to be not normally functioning, with rough set and BP
The thought that neutral net is combined is diagnosed in time to failure, construction reservation degree function Rc(x) and classification control function g (x)
The sliding-model control of data connection attribute is realized, with the method construct attribute reduction function S of Dynamic Reduct Basedi=miLi+m2NiLogarithm
Yojan is carried out according to attribute, study is trained to simplified rule with BP neural network, a simplification is set up by the method
Fault diagnosis system, realize the inline diagnosis of distribution network failure
Its specific embodiment is：
(1) data of current network have complicated relation and need to excavate, and in most cases have requirement of real-time,
The mass data that will be collected in current power distribution network first carries out characteristics extraction, as the initial data of system, forms initial
Decision table.
(2) because rough set can only be operated to discrete attribute, therefore using reservation degree function RcX () and classification are controlled
The method of function g (x) processed, sliding-model control is carried out and with Gene hepatitis B vaccine by the continuous property in initial decision table
Thought will classify control function g (x) as object function, and carry out optimization approximation process to object function.
(3) in the attribute reduction of rough set theory, using the method construct attribute reduction function S of Dynamic Reduct Basedi=miLi
+m2NiMake decision table ensure not containing unnecessary attribute under the conditions of classification is correct, minimal condition property set is formed, so as to complete
Classification merger to new data acquisition system.Because the data mining algorithm of rough set theory is advantageously implemented executed in parallel, can be with
The efficiency of data mining is greatly improved, therefore it is combined excavation of the realization to implicit information with BP neural network.
The structure of neutral net is determined with training sample, makes the structure of neutral net simpler, it is easy to understand,
Reduce the training time of network.Learning training and output category result are carried out to BP neural network with the learning sample after yojan,
Using the BP neural network for training as diagnostor in power distribution network, on-line fault diagnosis are realized.
The above, the only specific embodiment in the present invention, but protection scope of the present invention is not limited thereto, and appoints
What be familiar with the people of the technology disclosed herein technical scope in, it will be appreciated that the conversion or replacement expected, should all cover
It is of the invention include within the scope of, therefore, protection scope of the present invention should be defined by the protection domain of claims.
Claims (6)
1. a kind of Fault Diagnosis Method for Distribution Networks based on big data technology, it is characterised in that comprise the following steps that：
Step 1：Data to being collected in power distribution network carry out the extraction of continuous property, and the continuous property for extracting is used as original
Beginning decision table, for carrying out network training；
Step 2：Sliding-model control is carried out to the continuous property in the original decision table of formation in step 1, the company after discretization
Continuous property value is used as initial decision table；
Step 3：Using the method for Dynamic Reduct Based, structural environment attribute reduction function enters to the conditional attribute in initial decision table
Row yojan, the minimal condition property set of formation；
Step 4, using the power distribution network data set corresponding to minimal condition property set in step 3 as new training sample set；
Step 5：New training sample set pair BP neural network according to being formed in step 4 is trained study；
Step 6：The BP neural network of completion will be trained in step 5 as Fault Diagnosis of Distribution Network device, so as to realize online failure
Diagnosis.
2. a kind of Fault Diagnosis Method for Distribution Networks based on big data technology according to claim 1, it is characterised in that step
Continuous property includes conditional attribute and decision attribute values in rapid 1.
3. a kind of Fault Diagnosis Method for Distribution Networks based on big data technology according to claim 1, it is characterised in that step
Sliding-model control is carried out to the continuous property in original decision table in rapid 2, specially：
2.1) the connection attribute value set for defining training sample set is w, orderUsing similar matrixBy original
Continuous property in beginning decision table is arranged as the form of matrix, wherein, adiabatic index function
It is the l in original decision table1Generic attribute,It is the kth in original decision table1Generic attribute；
2.2) collection of element composition different in definition similar matrix H is combined into K, to any q ∈ K successively using q- from relation
Formula Lq={ (s, t) | (L (s, t) >=q } calculate in original decision table between any two continuous property from relation, its
In, s, t are the continuous property of training sample concentration, and L (s, t) is the fuzzy relation of continuous property s and t, q ∈ [0,1]；
Cluster result A1 is obtained with net-making method, is carried out the continuous property in original decision table from Automated generalization by q- preliminary discrete
Change, form preliminary discretization property set；
2.3) reservation degree function is constructedProperty set to preliminary discretization carries out the rejecting of redundancy, its
In, WcX () is the reservation collection of decision attribute x, U is the property set of preliminary discretization, RcX () represents decision attribute x to conditional attribute
The reservation degree of c；
2.4) structural classification control functionTo step 2.3) the middle discretization attribute rejected after redundancy
Collection carries out clustering processing, wherein, nlTo reject the number of l generic attributes in the discretization property set after redundancy, m is rejecting redundancy
Total class number of attribute in discretization property set afterwards, y is the center for rejecting l generic attributes in the discretization property set after redundancy,To reject the class center of attribute in the discretization property set after redundancy, r is to reject the category in the discretization property set after redundancy
Property number,It is the distance in the discretization property set after rejecting redundancy between l generic attributes and kth generic attribute center；
2.4) using genetic algorithm for solving object function ming (x)；
2.5) comprehensive function Q is definedq=n1Rc(x)+n2G (x), wherein n1And n2It is reservation degree function and the power of classification control function
Weight coefficient, makes Q=0, if Qq- Q ＜ 0, then be transferred to step 2.2)；If Qq- Q ＞ 0, then make A=A1 and Q=Qq, then it is transferred to step
2.2)；
2.6) at the end of the connection attribute collection in original decision table is discrete, Q is takenqMaximum as cluster result A, and to poly-
Class result is all kinds of to be encoded.
4. a kind of Fault Diagnosis Method for Distribution Networks based on big data technology according to claim 3, it is characterised in that step
Rapid 2.4) middle using genetic algorithm for solving object function ming (x), idiographic flow is as follows：
A) using step 2.3) in reject the discretization property set after redundancy as initial population, and initialize；
B) fitness function is utilizedThe quality of individuality in population is evaluated, wherein, N is discrete
Change the range of choice of each point in property set conditional Attribute class, C(a,b)It is that a classes conditional attribute is returned to the fitness of b class conditional attributes
Return value, TbIt is the fitness desired value of b class conditional attributes；
C) selection that the size according to fitness value in b) is selected the superior and eliminated the inferior to the individuality in population；
D) individuality for selecting fitness value larger from current population using predetermined probability breeds filial generation as parent；
E) gene intersection, gene mutation and restructuring are carried out to current sub- population at individual；
F) population heredity of future generation, function to achieve the objective are enteredOptimal value approach.
5. a kind of Fault Diagnosis Method for Distribution Networks based on big data technology according to claim 1, it is characterised in that step
Using the method for Dynamic Reduct Based, structural environment attribute reduction function S in rapid 3i=m1Li+m2Ni, wherein, m1And m2For importance degree is weighed
Weight, and m1+m2=1；LiIt is conditional attribute relation importance degree,N is initial decision table conditional
The number of attribute, NiIt is conditional attribute characteristic importance,| * | represents the number included in conditional attribute set
According to number, D is the conditional attribute collection of initial decision table, and G is data characteristics constant, MivIt is conditional attribute TiAnd TvBetween dependence
Relation, posj(D) feature important series of the jth class conditional attribute collection in initial decision table, j=1 ..., n are represented.
6. a kind of Fault Diagnosis of Distribution Network system based on big data technology, it is characterised in that including data discrete device, attribute about
Simple device, sample training device, wherein, data discrete device, for entering to carrying out continuous property to the data collected in power distribution network
Row sliding-model control, forms initial decision table；Attribute reduction device, for carrying out yojan to the conditional attribute in initial decision table,
Form minimal condition property set；Sample training device, for neural to BP according to the corresponding distribution network data of minimal condition property set
Network is learnt, and the BP neural network that output training is completed is used as Fault Diagnosis of Distribution Network device.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710159906.1A CN106874963B (en) | 2017-03-17 | 2017-03-17 | A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710159906.1A CN106874963B (en) | 2017-03-17 | 2017-03-17 | A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106874963A true CN106874963A (en) | 2017-06-20 |
| CN106874963B CN106874963B (en) | 2018-11-02 |
Family
ID=59171980
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710159906.1A Active CN106874963B (en) | 2017-03-17 | 2017-03-17 | A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106874963B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107301296A (en) * | 2017-06-27 | 2017-10-27 | 西安电子科技大学 | Circuit breaker failure influence factor method for qualitative analysis based on data |
| CN107609769A (en) * | 2017-09-06 | 2018-01-19 | 重庆邮电大学 | A kind of intelligent distribution network fault early warning method based on failure gene table |
| CN109284818A (en) * | 2018-09-07 | 2019-01-29 | 北方爆破科技有限公司 | A kind of blasting vibration control prediction technique based on accident tree and genetic algorithm |
| CN110674468A (en) * | 2019-09-25 | 2020-01-10 | 郑州轻工业学院 | Quantitative analysis method for spun yarn breakage factor based on improved rough set algorithm |
| CN110688809A (en) * | 2019-09-05 | 2020-01-14 | 西安理工大学 | Box transformer substation fault diagnosis method based on VPRS-RBF neural network |
| WO2020125166A1 (en) * | 2018-12-20 | 2020-06-25 | 北京科东电力控制系统有限责任公司 | Power grid anticipated fault set prediction method and apparatus, and electronic device and storage medium |
| CN115408196A (en) * | 2022-10-31 | 2022-11-29 | 国网四川省电力公司电力科学研究院 | High fault tolerance power grid fault diagnosis method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050266395A1 (en) * | 2003-09-10 | 2005-12-01 | Bioimagene, Inc. | Method and system for morphology based mitosis identification and classification of digital images |
| CN104298873A (en) * | 2014-10-10 | 2015-01-21 | 浙江大学 | Attribute reduction method and mental state assessment method on the basis of genetic algorithm and rough set |
| CN105445646A (en) * | 2015-12-14 | 2016-03-30 | 国家电网公司 | Testing method of analog-digital circuit fault diagnosis based on neural network expert system |
-
2017
- 2017-03-17 CN CN201710159906.1A patent/CN106874963B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050266395A1 (en) * | 2003-09-10 | 2005-12-01 | Bioimagene, Inc. | Method and system for morphology based mitosis identification and classification of digital images |
| CN104298873A (en) * | 2014-10-10 | 2015-01-21 | 浙江大学 | Attribute reduction method and mental state assessment method on the basis of genetic algorithm and rough set |
| CN105445646A (en) * | 2015-12-14 | 2016-03-30 | 国家电网公司 | Testing method of analog-digital circuit fault diagnosis based on neural network expert system |
Non-Patent Citations (3)
| Title |
|---|
| SOUAHLIA SEIFEDDINE等: "《Power Transformer Fault Diagnosis Based on Dissolved Gas Analysis by Artificial Neural Network》", 《2012 FIRST INTERNATIONAL CONFERENCE ON RENEWABLE ENERGIES AND VEHICULAR TECHNOLOGY》 * |
| 邓松等: "电力大数据智能化高效分析挖掘技术框架", 《电子测量与仪器学报》 * |
| 邓武等: "粗糙集、神经网络和专家系统模型用于电力系统故障诊断", 《高电压技术》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107301296A (en) * | 2017-06-27 | 2017-10-27 | 西安电子科技大学 | Circuit breaker failure influence factor method for qualitative analysis based on data |
| CN107301296B (en) * | 2017-06-27 | 2020-09-04 | 西安电子科技大学 | Data-based qualitative analysis method for circuit breaker fault influence factors |
| CN107609769A (en) * | 2017-09-06 | 2018-01-19 | 重庆邮电大学 | A kind of intelligent distribution network fault early warning method based on failure gene table |
| CN107609769B (en) * | 2017-09-06 | 2020-07-31 | 重庆邮电大学 | Intelligent power distribution network fault early warning method based on fault gene table |
| CN109284818A (en) * | 2018-09-07 | 2019-01-29 | 北方爆破科技有限公司 | A kind of blasting vibration control prediction technique based on accident tree and genetic algorithm |
| WO2020125166A1 (en) * | 2018-12-20 | 2020-06-25 | 北京科东电力控制系统有限责任公司 | Power grid anticipated fault set prediction method and apparatus, and electronic device and storage medium |
| CN110688809A (en) * | 2019-09-05 | 2020-01-14 | 西安理工大学 | Box transformer substation fault diagnosis method based on VPRS-RBF neural network |
| CN110674468A (en) * | 2019-09-25 | 2020-01-10 | 郑州轻工业学院 | Quantitative analysis method for spun yarn breakage factor based on improved rough set algorithm |
| CN110674468B (en) * | 2019-09-25 | 2022-09-27 | 郑州轻工业学院 | Quantitative analysis method for spun yarn breakage factor based on improved rough set algorithm |
| CN115408196A (en) * | 2022-10-31 | 2022-11-29 | 国网四川省电力公司电力科学研究院 | High fault tolerance power grid fault diagnosis method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106874963B (en) | 2018-11-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106874963B (en) | A kind of Fault Diagnosis Method for Distribution Networks and system based on big data technology | |
| CN101063643B (en) | Intelligent diagnostic method for airplane functional failure and system thereof | |
| CN105487526B (en) | A kind of Fast RVM sewage treatment method for diagnosing faults | |
| CN108536123A (en) | The method for diagnosing faults of the train control on board equipment of the long neural network of memory network combination in short-term | |
| CN108062572A (en) | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on DdAE deep learning models | |
| CN106570513A (en) | Fault diagnosis method and apparatus for big data network system | |
| CN111292195A (en) | Risk account identification method and device | |
| CN109949929A (en) | A kind of assistant diagnosis system based on the extensive case history of deep learning | |
| CN112732921B (en) | False user comment detection method and system | |
| CN109685653A (en) | A method of fusion deepness belief network and the monitoring of the credit risk of isolated forest algorithm | |
| CN112217674B (en) | Alarm root cause identification method based on causal network mining and graph attention network | |
| CN109101584A (en) | A kind of sentence classification improved method combining deep learning with mathematical analysis | |
| CN111046961B (en) | Fault classification method based on bidirectional long-time and short-time memory unit and capsule network | |
| CN114756686A (en) | Knowledge reasoning and fault diagnosis method based on knowledge graph | |
| CN109522917A (en) | A method of fusion forecasting is stacked based on multi-model | |
| CN104732067A (en) | Industrial process modeling forecasting method oriented at flow object | |
| CN110222733A (en) | The high-precision multistage neural-network classification method of one kind and system | |
| CN110008332A (en) | The method and device of trunk word is extracted by intensified learning | |
| CN117151222A (en) | Domain knowledge guided emergency case entity attribute and relation extraction method thereof, electronic equipment and storage medium | |
| CN104850862B (en) | A kind of sorting technique based on unit cost income sensitivity decision tree | |
| Vasilev et al. | Decision Tree Extraction using Trained Neural Network. | |
| CN113095501A (en) | Deep reinforcement learning-based unbalanced classification decision tree generation method | |
| CN115796635A (en) | Bank digital transformation maturity evaluation system based on big data and machine learning | |
| Fukumi et al. | Rule extraction from neural networks trained using evolutionary algorithms with deterministic mutation | |
| Hassan et al. | Rough neural classifier system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |