CN106569095A - Power grid fault diagnosis system based on weighted average dependence classifier - Google Patents

Abstract

The invention provides a power grid fault diagnosis system based on a weighted average dependence classifier; the system comprises the following units: an electric power grid history data acquisition module used for obtaining history data from an electric power grid history database, wherein the obtained data comprises a symptom variable set and a fault class variable set; a model building module used for building a weighted average dependence classifier prediction model and a weighted average dependence classifier forming module, wherein the weighted average dependence classifier prediction model automatically learns classifier parameters through the history data, thus forming a trained weighted average dependence classifier; a fault diagnosis module using the trained weighted average dependence classifier to estimate test data in a fault diagnosis process, thus finally obtaining the corresponding fault diagnosis result.

Description

A kind of electric network failure diagnosis system that grader is relied on based on weighted average

Technical field

The invention belongs to electrical network field, and in particular to a kind of electric network failure diagnosis system that grader is relied on based on weighted average System.

Background technology

Modern power network has the characteristics of electrical network scale is big, and topology is complicated.When the grid collapses, control centre can receive A large amount of abundant fault messages, this is provided for the fault diagnosis for the purpose of quick positioning failure region, identification fault element Precondition.Early stage electrical network diagnostic system has protection, chopper tripping malfunction or communication line mistake in fault message Fault-tolerance is difficult to ensure that during etc. uncertain factor.If electrical grid failure timely, does not accurately detect trouble point, can be right Economy, the people's livelihood produce great adverse effect.Therefore, timely and effectively electric network fault detection is highly important.

Early stage relies on the method for expertise detection electric network fault and has been difficult to ensure that current extensive, high complexity electrical network Stability.Therefore, in large complicated electrical network, intelligent diagnostics are widely applied, and are just included in these intelligent diagnosing methods Data classification algorithm in machine learning field.

NB Algorithm (naive Bayes, abbreviation NB) is the supervised learning algorithm based on Bayes rule, and it abides by Bayesian assumption, also referred to as naive Bayesian conditional independence assumption are followed, the hypothesis greatly simplify the Bayes of the algorithm Network structure.Therefore, NB is efficient in the model training stage or in test phase.It is this efficiently substantially from Conditional independence assumption.And this conditional independence assumption is generally disagreed with real data cases in practice, the classification essence of NB Degree has therefore suffered from affecting.In order to improve the nicety of grading of NB, many research worker are by discharging strict conditional independence assumption Propose some new sorting algorithms.Follow-up, GI Webb et al. propose the AODE algorithms for improving naive Bayesian.With NB is compared, and AODE considers impact of the appearance of common property value to test data probability distribution for test data, is come with this Improve nicety of grading, at the same time also increase the complexity of the Bayesian network of AODE models, its training time complexity and Testing time complexity is accordingly increased compared with NB.

However, AODE is not accounted in addition to class label, the actively impact of the relation pair data distribution between common property.This Sample, data classification are accurately just affected.Therefore, this area needs that a kind of accuracy is higher and computation complexity increases less Electric network failure diagnosis method.

It is contemplated that to for any attribute is to Ai, Aj, if the dependence between them is stronger, to P (A_j|A_i) or Person P (A_i|A_j) estimated value it is just more confident.Therefore, we on the basis of AODE expect to consider the relation between attribute Adjustment probability distribution, improves the nicety of grading of AODE.In addition, expertise can be acted on our algorithm, adjustment category Relation between property pair, improves the classification degree of accuracy of algorithm with this.

The content of the invention

It is an object of the invention to solve a difficult problem present in above-mentioned prior art, there is provided a kind of weighted average relies on classification The electric network failure diagnosis system of device, the relatively low problem of effectively solving Large-scale power networks fault diagnosis accuracy rate, and keep good appearance Wrong ability.

The present invention is achieved by the following technical solutions：

A kind of electric network failure diagnosis system based on weighted average-dependence grader, it is characterised in that：Including：

Electric power network historical data acquisition module：Historical data is obtained from electric power network historical data base, including Symptom variables collection and failure classes variables set；

Model construction module：Build weighted average and rely on grader forecast model；

Weighted average relies on grader generation module：The weighted average relies on grader forecast model by the history Automatically study, to classifier parameters, is formed and has trained the weighted average-dependence grader for completing data；

Fault diagnosis module：When carrying out fault diagnosis, to test data using described in trained the weighted average for completing-according to Bad grader estimated, finally gives corresponding fault diagnosis result.

The weighted average of the model construction module-dependence grader forecast model is as follows：

Wherein, t is test data, and Y is failure classes variables set, values of the y for failure class variable, p (y, x_i) for corresponding therefore Barrier class variable y and symptom variables x_iJoint probability, i=1,2 ..., m, m are symptom variables number and F (x_i)≥g,F(x_i) table Show x_iFrequency in training data；In addition, p (x_j|y,x_i) for symptom variables x_jWith symptom variables x_iWith the condition of failure variable y Probability, j=1,2 ..., m, w_ijFor symptom variables x_iWith symptom variables x_jWeight.

The value of the g is 30.

The symptom variables x_iWith symptom variables x_jWeight w_ijIt is as follows：

Wherein, p (x_i, x_j, it is y) x_iAnd x_jAnd the joint probability of y, p (x_i| y) with p (x_j| y) it is respectively x_iAnd x_jCondition Probability, i=1,2 ..., m and j=2,3 ..., m and j >=i.

Rely in grader generation module in the weighted average, described classifier parameters include that symptom variables joint is general Rate p (x_i,x_j,y),p(x_i, y) and failure variable Probability p (y), specially：

Wherein n is the training examples number in the case of given y values, n_iValue and failure classes value for i-th symptom variables It is determined that in the case of training examples number, n_ijValue and failure classes value for i-th, j symptom variables is trained in the case of determining The number of sample, and the base value that c () is, the frequency that F () is.

The test data in the fault diagnosis module includes the property value stated by symptom variables collection.

Compared with prior art, the invention has the beneficial effects as follows：

1, the inventive method improves the accuracy rate of electric network failure diagnosis, and is not affected by attribute missing values, with appearance Wrong ability；

2, the present invention is a kind of based on weighting for, the characteristics of existing electric power networks scale is big, electric network composition is complicated, devising Average one relies on grader, carries out fault diagnosis to electrical network by the grader, improves the learning capacity of grader itself.

Description of the drawings

Fig. 1 is AODE grader structure charts；

Fig. 2 is fault diagnosis protocol procedures figure；

Fig. 3 is WAODE grader structure charts；

Fig. 4 is the historical data schematic diagram obtained in embodiment；

Fig. 5 is failure variable and symptom variables cartogram in example；

Fig. 6 is using the fault diagnosis accuracy rate cartogram under different forecast models when historical data is lost.

Specific embodiment

Below in conjunction with the accompanying drawings the present invention is described in further detail：

For the disadvantages described above of prior art, the present invention proposes a kind of electrical network for relying on grader based on weighted average and examines Disconnected system, the relatively low problem of effectively solving Large-scale power networks fault diagnosis accuracy rate.

The present invention as shown in Fig. 2 including：

Electric power network historical data acquisition module：Historical data is obtained from electric power network historical data base, including Symptom variables collection and failure classes variables set；

Model construction module：Build weighted average and rely on grader forecast model；

Weighted average relies on grader generation module：The weighted average relies on grader forecast model by the history Automatically study, to classifier parameters, is formed and has trained the weighted average for completing to rely on grader data；

Fault diagnosis module：When carrying out fault diagnosis, to test data using described in trained the weighted average for completing according to Bad grader estimated, finally gives corresponding fault diagnosis result.

The test data in the fault diagnosis module includes the property value stated by symptom variables collection, the weighting Average one relies on grader by training, provides the classification results of the test data, i.e. fault type.So-called test data can To think to be made up of, n=1,2,3,4 ... n bars test sample (test instance), grader once tests a test Instance, and classification results are given to the test instance.

The present invention assesses the classification performance of grader using classification accuracy.

Hypothesis has 100 test samples, and to 70 samples, grader classification accuracy is exactly 70% to grader point.

General is so description in paper：Accurate rate is by " 10 folding cross validation " on " 95% trusts interval " Mode estimate.

Grader provides classification results for a certain test sample, and this " result " is exactly the diagnosis knot that the grader is given Really.

WAODE sorter models structure as shown in figure 3, wherein Y is class node, i.e. failure classes variables set (failure variables set As shown in Figure 4, c1-c4), point to all of attribute A₁, A₂..., A_i..., A_m, i.e., (Ai represents symptom variables to symptom variables collection The attribute concentrated, it is substantially a stochastic variable.If as shown in figure 5, i=2, A2 are a category of symptom variables collection Property (being called stochastic variable).According to the restriction in the present invention to symptom set attribute value, the value of A2 can be 1, alternatively 0), W_ijIt is Attribute A_iWith A_jBetween weight, attribute A_iPoint to other all properties but do not include class node.

Described symptom variables collection, specially：Symptom variables value be nominal property value (as the A1-A7 in Fig. 4 this 7 attributes, they can only be considered as nominal attribute (mutually distinguishing with numerical attribute), if there is value type from value in { 0,1 } Value, be both needed to sliding-model control (generally can using the unsupervised attribute filter Discretize of Weka softwares it is discrete fall it is all Continuous property value), and the missing values of certain symptom variables are only marked (can generally use the unsupervised attribute of Weka softwares Missing values are substituted for by filter ReplaceMissingValues " * ") and do not do other special handlings.

In model construction module, the weighted average one relies on grader forecast model C_WAODESpecially：

Wherein, t is test data, and Y is failure classes variables set, values of the y for failure class variable, p (y, x_i) for corresponding therefore Barrier class variable y and symptom variables x_iJoint probability, i=1,2 ..., m, m are symptom variables number and F (x_i)≥g,F(x_i) table Show x_iFrequency in training data, g are typically set at 30.In addition, p (x_j|y,x_i) for symptom variables x_jWith symptom variables x_iWith The conditional probability of failure variable y, j=1,2 ..., m.Symbol w_ijFor symptom variables x_iWith symptom variables x_jWeight.

The symptom variables x_iWith symptom variables x_jWeight w_ijSpecially：

Wherein, m be symptom variables collection variable number, p (x_i, x_j, it is y) x_iAnd x_jAnd the joint probability of y, p (x_i| y) and p (x_j| y) it is respectively x_iAnd x_jConditional probability, i=1,2 ..., m and j=2,3 ..., m and j >=i.

Rely in grader generation module in weighted average, described classifier parameters include symptom variables joint probability p (x_i,x_j,y),p(x_i, y) and failure variable Probability p (y), specially：

Wherein n is the training examples number in the case of given y values, n_iValue and failure classes value for i-th symptom variables It is determined that in the case of training examples number, n_ijValue and failure classes value for i-th, j symptom variables is trained in the case of determining The number of sample, and the base value that c () is, the frequency that F () is.

In embodiment, the present invention have chosen 400 cases from electrical network historical data from the present invention, and summarize some diseases Shape variables set is attribute, and some failure classes variables sets are class (symptom only correspond to a fault value), details such as Fig. 4 institutes Show.The corresponding relation of these symptoms and failure is as shown in Figure 5.A1, A2, A3, A4, A5, A6, A7 in Fig. 5, these values are represented for 0 Chopper is not conjugated or protector is not operating, is 1 expression chopper by closing displacement to disconnect or protection act.

400 record electrical network historical datas shown in Fig. 5 are sent into WAODE models as test and training data carries out event Barrier diagnosis, and and two kinds of graders of NB, AODE (shown in Fig. 1) carry out the comparison in nicety of grading.In figure 6, when loss attribute When value gradually increases, the nicety of grading of three graders is gradually reduced, but in the same state, the nicety of grading of WAODE is all the time Better than NB and AODE, illustrating WAODE has good fault-tolerant ability and learning capacity.

Above-mentioned technical proposal is one embodiment of the present invention, for those skilled in the art, at this On the basis of disclosure of the invention application process and principle, it is easy to make various types of improvement or deformation, this is not limited solely to The method described by above-mentioned specific embodiment is invented, therefore previously described mode is simply preferred, and do not had and limit The meaning of property.

Claims (6)

1. it is a kind of based on weighted average rely on grader electric network failure diagnosis system, it is characterised in that：Including：

Electric power network historical data acquisition module：Historical data is obtained from electric power network historical data base, including symptom Variables set and failure classes variables set；

Model construction module：Build weighted average and rely on grader forecast model；

Weighted average relies on grader generation module：The weighted average-dependence grader forecast model is by the history number According to study automatically to classifier parameters, formation has trained the weighted average for completing to rely on grader；

Fault diagnosis module：When carrying out fault diagnosis, using described in, train to test data the weighted average for completing to rely on divided Class device estimated, finally gives corresponding fault diagnosis result.

2. it is according to claim 1 based on weighted average rely on grader electric network failure diagnosis system, it is characterised in that： It is as follows that the weighted average of the model construction module relies on grader forecast model：：

Wherein, t is test data, and Y is failure classes variables set, values of the y for failure class variable, p (y, x_i) become for corresponding failure classes Amount y and symptom variables x_iJoint probability, i=1,2 ..., m, m are symptom variables number and F (x_i)≥g,F(x_i) represent x_i Frequency in training data；In addition, p (x_j|y,x_i) for symptom variables x_jWith symptom variables x_iWith the conditional probability of failure variable y, J=1,2 ..., m, w_ijFor symptom variables x_iWith symptom variables x_jWeight.

3. it is according to claim 2 based on weighted average rely on grader electric network failure diagnosis system, it is characterised in that： The value of the g is 30.

4. it is according to claim 3 based on weighted average rely on grader electric network failure diagnosis system, it is characterised in that： The symptom variables x_iWith symptom variables x_jWeight w_ijIt is as follows：

$w_{ij}=\frac{\underset{i,j,y}{\Σ}p(x_i,x_j,y)\log \frac{p(x_i,x_j|y)}{p\left(x_i\right|y)p\left(x_j\right|y)}}{\underset{i=1}{\overset{m}{\Σ}}\underset{j=2,j>i}{\overset{m}{\Σ}}\underset{i,j,y}{\Σ}p(x_i,x_j,y)\log \frac{p(x_i,x_j|y)}{p\left(x_i\right|y)p\left(x_j\right|y)}}m$

Wherein, p (x_i, x_j, it is y) x_iAnd x_jAnd the joint probability of y, p (x_i| y) with p (x_j| y) it is respectively x_iAnd x_jConditional probability, I=1,2 ..., m and j=2,3 ..., m and j >=i.

5. it is according to claim 4 based on weighted average rely on grader electric network failure diagnosis system, it is characterised in that： In the weighted average-dependence grader generation module, described classifier parameters include symptom variables joint probability p (x_i, x_j,y),p(x_i, y) and failure variable Probability p (y), specially：

$p\left(y\right)=\frac{F\left(y\right)+1}{n+c\left(Y\right)}$

$p(x_i,y)=\frac{F(x_i,y)+1}{n_i+c(attri\_i)\×c\left(Y\right)}$

$p(x_i,x_j,y)=\frac{F(x_i,x_j,y)+1}{n_{ij}+c(attri\_i)\×c(attri\_j)\×c\left(Y\right)}$

Wherein n is the training examples number in the case of given y values, n_iWhat the value and failure classes value for i-th symptom variables determined In the case of training examples number, n_ijTraining examples in the case of value and failure classes value determination for i-th, j symptom variables Number, and the base value that c () is, the frequency that F () is.

6. it is according to claim 5 based on weighted average rely on grader electric network failure diagnosis system, it is characterised in that： The test data in the fault diagnosis module includes the property value stated by symptom variables collection.