CN110458230A - A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method - Google Patents
A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method Download PDFInfo
- Publication number
- CN110458230A CN110458230A CN201910740107.2A CN201910740107A CN110458230A CN 110458230 A CN110458230 A CN 110458230A CN 201910740107 A CN201910740107 A CN 201910740107A CN 110458230 A CN110458230 A CN 110458230A
- Authority
- CN
- China
- Prior art keywords
- data
- point
- voltage
- value
- adopting
- 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
Links
Classifications
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/251—Fusion techniques of input or preprocessed data
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a kind of distribution transformings based on the fusion of more criterions with adopting data exception discriminating method, comprising: to for statistical analysis with data breakpoint, abnormal point and scene actual operating data situation is adopted；The examination that four kinds of methods such as prototype clustering procedure, Density Clustering method, probability density method, deep learning method carry out exceptional value is respectively adopted, " 4 take 2 " verification result is carried out to four kinds of models, i.e. four kinds of models think that point to be determined is abnormal point there are two model, then point to be determined is abnormal point.The difficulty that faces is big when the present invention solves conventional machines learning method processing mass data, low efficiency, the problems such as real-time is not high.
Description
Technical field
The invention belongs to electric system switching data processing technology fields, and in particular to a kind of matching based on the fusion of more criterions
Become with adopting data exception discriminating method.
Background technique
With computer, communication, the extensive use of sensing technology, the continuous propulsion of distribution operation monitoring business and a large amount of prisons
The deployment of metering device is surveyed, distribute-electricity transformer district monitoring obtains magnanimity operation data, user power utilization data and device status data, right
These data are analyzed, are excavated, extracted and promotion service quality, expand electricity with processing, realization distribute-electricity transformer district the safe and economic operation
Measuring electricity charges becomes distribution facing challenges.In particular, in the magnanimity electric network data that distribute-electricity transformer district monitoring obtains
In the presence of about 10% abnormal data, it is necessary to being analyzed with adopting the quality of data, abnormal data is screened, thus to carry out monitoring
Operation business provides reliable, accurate, effective data supporting.Have with the main reason for time series abnormal data occurs is adopted:
(1) metering device failure: metering device includes terminal, mutual inductor, terminal box, meter, and failure, which is likely to be present in, appoints
In what link.Such as: shelf depreciation or completely electric discharge is presented in the corona of mutual inductor, causes data collection inaccurate；It connects
Wire box exception of the continuous data as caused by poor contact etc..
(2) signal of communication is poor: some areas use 3G signal, and signal is caused to cut in and out, and the transmission of partial period data is lost
It loses.Meanwhile large buildings also can generate shielding to signal of communication, influence to communicate.
(3) collector failure: collector realizes the data summarization of all devices and distribution in control range, realizes to intelligence
The effect of ammeter control command transmission.In low-voltage customer, collector is separated with metering device, and the control of each collector is multiple
Intelligent electric meter.When communication or ontology failure occur for collector, all intelligent electric meter data are adopted in entire acquisition range
Collection failure.
(4) human factor: mainly unreasonable electricity consumption, so that ammeter is in overload state and power stealing stealing for a long time
Behavior, this is all caused with the appearance for adopting time series abnormal data.
With the quality for adopting the quality of data, it is largely fixed the quality of modal analysis results.Therefore, analysis is being established
Before model to adopt exceptional value present in data carry out detection screen be improve the quality of data important channel.It commonly uses at present different
Normal point detecting method mainly has:
(1) statistical method: it is used for outlier detection earliest, is generally divided into based on the assumption that the method for inspection and based on model
Method.Since real data Mining Problems majority need to find abnormal point, but the consistency check of the overwhelming majority in hyperspace
It is only applicable to single inspection by attributes；Simultaneously as must be known by data distribution model before this method, so that this method has very big limitation
Property.
(2) based on apart from rejecting outliers method: its distance function and parameter are not easy to select, and can only detect global different
Chang Dian, and cannot detect local outlier.
(3) be based on density anomaly value detection method: it is capable of detecting when global and local abnormal point, but calculate it is complicated,
It is cumbersome, it is not suitable for high dimensional data occasion.
(4) based on cluster rejecting outliers method: it can find class and abnormal point simultaneously, but general efficiency is lower, needle
It is stronger to property.
(5) it is based on machine learning rejecting outliers method: artificial neural network (artificial neural can be divided
Networks, ANN) and support vector machines (support vector machines, SVM) two major classes.ANN is small-scale in processing
There is good application effect, but lower to large-scale data scene efficiency in problem, it is difficult to preferably solve the problems, such as parameter training,
And training process easily falls into local optimum, model structure and weight setting is improper can also seriously affect model accuracy.SVM has
Better generalization ability, but will face a severe challenge in processing Massive Sample, and modeling is more complex, there is one in practical applications
Determine difficulty.
The electric current of intelligent electric meter, voltage, active power, the abnormal data in reactive capability curve have directly reacted intelligent electricity
The operating status of table, such abnormal data belong to the measurement abnormal point and user power utilization abnormal point of table note aspect.Intelligent electric meter
Break down is frequently not to cause moment, but inferior health operating status is in a period of time before the failure.In this shape
Under state, the abnormal data on curve is more hidden, not easily passs through basic norm to distinguish.Due to being seriously affected with adopting the quality of data
The credibility of the segment analysis result such as operation centre, and the quality of data is drastically influenced with abnormal data is adopted.In addition, with number is adopted
It is general at present excessively dead in the presence of being arranged with abnormal data examination rule is adopted according to the problems such as there are breakpoint, phase shortage, abnormal high low values
The deficiency of plate needs pointedly to improve exceptional value decision rule, screens accuracy to improve rejecting outliers.
The research hotspot of data mining and deep learning theory as computer field instantly, can be carried out effectively height
Training set is divided into small lot data in advance and calculated, mentioned by dimension, complicated, nonlinear problem analysis and processing, deep learning
High training effectiveness.Therefore, in comparison, deep learning is more suitable for carrying out the time serieses magnanimity such as Current Voltage with adopting number
According to the detection and examination of exceptional value, it is existing when handling mass data that conventional machines learning method can be solved using deep learning
Committed memory it is high, operation processing speed is slow and is difficult to the defects of handling high dimensional feature data.
Summary of the invention
The purpose of the present invention is to provide a kind of distribution transformings based on the fusion of more criterions with data exception discriminating method is adopted, respectively
The examination of exceptional value is carried out using four kinds of methods such as prototype clustering procedure, Density Clustering method, probability density method, deep learning method,
" 4 take 2 " verification result is carried out to four kinds of models, the difficulty faced when solving conventional machines learning method processing mass data
Greatly, low efficiency, the problems such as real-time is not high.
In order to achieve the above objectives, The technical solution adopted by the invention is as follows:
A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method, comprising:
Distribution transforming is obtained with adopting initial data；
The distribution transforming is pre-processed with initial data is adopted；
Noise spot is added in pretreated distribution transforming at random with adopting in initial data, formed containing abnormal point with adopting data sequence
Column；
Four kinds of prototype clustering procedure, Density Clustering method, probability density method and deep approach of learning models are respectively adopted to containing abnormal
Point carries out abnormal point examination with adopting data sequence；
Determine distribution transforming with adopting abnormal data；The distribution transforming screens knot with the abnormal point that abnormal data is any two kinds of models is adopted
The intersection of fruit, the union for the intersection for then taking all combination of two to determine.
Further, the acquisition distribution transforming is with adopting initial data, comprising:
Three-phase current, three-phase voltage and active power initial data, acquisition time are acquired based on metering device is operated normally
Between be divided into 15min.
It is further, described that the distribution transforming is pre-processed with initial data is adopted, comprising:
Missing values processing and the obvious exceptional value of rejecting are carried out with initial data is adopted to distribution transforming；
It is described that missing values processing is carried out with initial data is adopted to distribution transforming, comprising: individual to being lacked in continuous time data set
Data, the data lacked using linear interpolation method polishing；To mass data is lacked in continuous time data set, the section is directly rejected
Data；
It is described to reject obvious exceptional value and refer to, will be shown as in three-phase current, three-phase voltage and active power initial data-
9999 data are rejected.
Further, it when the distribution transforming is with adopting initial data and large-scale data missing occur, selects before shortage of data
Curve calculated.
Further, the noise spot Normal Distribution.
Further, described that abnormal point examination is carried out with adopting data sequence to containing abnormal point using prototype clustering procedure, it wraps
It includes:
Determine cluster attribute；Include: choose measuring point voltage actual value, measuring point to be checked and former point voltage change to be checked and
Measuring point to be checked and cluster attribute of the preceding two o'clock voltage change as voltage time sequence, choose measuring point current actual value to be checked,
Measuring point to be checked and former point current variation value and measuring point to be checked and cluster of the preceding two o'clock current variation value as current time sequence
Attribute chooses measuring point active power actual value, measuring point to be checked and former point active power changing value to be checked and measuring point to be checked with before
Cluster attribute of the two o'clock active power changing value as active power time series；
According to the cluster attribute, using k-means algorithm, time series to be detected is gathered for 4 classes, and determination is all kinds of
Mass center；
According to all kinds of mass centers, each measuring point to be checked is calculated to the distance of nearest cluster centre and each measuring point to be checked to recently
The relative distance of cluster centre；
The relative distance of each measuring point to be checked to nearest cluster centre is made comparisons with given threshold value；If certain measuring point to be checked arrives
The relative distance of nearest cluster centre is greater than given threshold value, then the measuring point to be checked is outlier, i.e. abnormal point.
Further, described that abnormal point examination is carried out with adopting data sequence to containing abnormal point using Density Clustering method, it wraps
It includes:
The flat distribution map of voltage-to-current is drawn respectively, and electric current-active power flat distribution map and voltage-are active
The flat distribution map of power；
Point on the flat distribution map is clustered, comprising: if between two points in the flat distribution map
Distance be no more than setting maximum distance, then be divided into one kind；Wherein, the point on voltage-to-current flat distribution map is certain phase
Current time sequence and voltage time sequence；Point on electric current-active power flat distribution map be certain phase current time series and
Active power time series；When point on voltage-active power flat distribution map is certain phase voltage time series and active power
Between sequence；
It is abnormal point that loop iteration, which is found out and is not belonging to the point of any class,.
Further, described that abnormal point examination is carried out with adopting data sequence to containing abnormal point using probability density method, it wraps
It includes:
Determine mode input and model output；The mode input are as follows: be to be determined electric current for current time sequence
With the changing value of former point electric current；It is the changing value of to be determined voltage and former point voltage for voltage time sequence；For
Active power time series is the changing value of to be determined voltage and former point voltage；The model output are as follows: changing value is normal
Range；
It is fitted the probability distribution of voltage, electric current and active power data respectively using kernel density function, and it is close to obtain probability
Spend function；
For any numerical value d to be detected, occurred to the carry out integral calculation of probability density function [d ,+∞) numerical value model
The probability enclosed, and and threshold value comparison, the i.e. probability whether be lower than the corresponding probability 0.003 of 3 σ；If so, the numerical value to be detected is
Abnormal point.
Further, described that abnormal point Zhen is carried out with adopting data sequence to containing addition abnormal point using deep approach of learning
Not, comprising:
The deep learning mould based on shot and long term memory network completed using the training of electric current, voltage and power time series
Type predicts non-incoming current, voltage or power data, the error of comparison prediction value and true value；If predicted value deviates true
Real value is more than the threshold value of setting, then predicted value is abnormal point；
The deep learning model of the shot and long term memory network are as follows: after completing forward calculation, calculated using error back propagation
Method is updated adjustment to model parameter, comprising:
The neuron weighting input net of t moment shot and long term memory networkf,t, neti,t, netc′,t, neto,tAre as follows:
Wherein, Wox、Wfx、Wix、Wcx、Woh、Wfh、Wih、WchIndicate weight, ht-1It is previous moment LSTM output, xtIt is current
Moment input, bf、bi、bo、bcRespectively forget door, input door, output door and current time input unit
Biasing；
The neuron error item δ of t moment shot and long term memory networkf,t, δi,t, δc′,t, δo,tAre as follows:
Wherein, E is prediction error；
When error is propagated along time reversal, the error term δ at t-1 momentt-1Are as follows:
Wherein,For Jacobian matrix；
When error by current l layer back transfer to l-1 layers when, l-1 layers of errorAre as follows:
Finally, weight W is obtainedoh、Wfh、Wih、WchAre as follows:
Wherein, Woh,t, Wfh,t, Wih,t, Wch,tThe weight of t moment is respectively indicated, subscript T indicates transposition；
Weight Wox、Wfx、Wix、WcxAre as follows:
bf、bi、bo、bcAre as follows:
Wherein, bo,t, bf,t, bi,t, bc,tRespectively indicate the error term of t moment.
Compared with prior art, technical solution of the present invention has the advantages that
(1) deep learning method used in the method for the present invention, deep learning algorithm can handle mass data, and carry out height
Imitate comprehensive feature learning, reduce it is artificial carry out feature learning it is inefficient with it is incomplete, to make the extensive energy of the feature learnt
Power is stronger；
(2) LSTM neural network used in deep learning algorithm is shot and long term memory network, is a kind of time circulation mind
Through network, it is suitable for being spaced and postpone relatively long critical event in processing and predicted time sequence.Metering device is transported extremely
The main task of row state recognition model is the time series data for identifying metering device and obtaining, and LSTM neural network is handling this
There is good performance in a problem；
(3) the operation data sample set quantity in ammeter acquisition is few or time series characteristic is unobvious, the mould of LSTM
When type is performed poor, often the effect than LSTM is good for the effect of three kinds of algorithms of other in model.More criterions fusion of the invention
Algorithm can cope with difference in the case where abnormal operating condition identification, Generalization Capability is more preferable, and accuracy rate is higher；
(4) the method for the present invention uses " 4 take 2 " method cross validation, improves abnormal point and screens recall rate, reduces exception
The False Rate that point is screened, improves the abnormal point discrimination capabilities of model, provides precise information guarantee for operation detection business, power grid
Relevant departments staff can overhaul and check according to recognition result, solve potential existing for abnormal ammeter ask as early as possible
Topic, improves the security reliability of operation of power networks.
Detailed description of the invention
Fig. 1 is that shot and long term memory network respectively forms partial structure diagram in the embodiment of the present invention；
Fig. 2 is depth shot and long term memory network structural framing in the embodiment of the present invention；
Fig. 3 is that prototype clusters voltage analysis result under the normally distributed error that standard deviation is 6 in the embodiment of the present invention；
Fig. 4 is the normally distributed error lower density cluster analysis result for being 6 in standard deviation in the embodiment of the present invention；Fig. 4 (a)
For power voltage plane distribution, Fig. 4 (b) is voltage and current plane distribution；
Fig. 5 is the normally distributed error lower probability distributional analysis result for being 6 in standard deviation in the embodiment of the present invention；
Fig. 6 is that LSTM prediction result and voltage are true under the normally distributed error that standard deviation is 6 in the embodiment of the present invention
Value；
Fig. 7 is that LSTM predicts error under the normally distributed error that standard deviation is 6 in the embodiment of the present invention；
Fig. 8 be in the embodiment of the present invention under the normally distributed error that standard deviation is 8 prototype cluster analysis result；
Fig. 9 is the normally distributed error lower density cluster analysis result for being 8 in standard deviation in the embodiment of the present invention；Fig. 9 (a)
For power voltage plane distribution, Fig. 9 (b) is voltage and current plane distribution；
Figure 10 is that voltage true value and LSTM prediction are tied under the normally distributed error that standard deviation is 8 in the embodiment of the present invention
Fruit；
Figure 11 be in the embodiment of the present invention under the normally distributed error that standard deviation is 8 LSTM model predictive error；
Figure 12 is 49932 ammeter electric currents and current variation value curve in the embodiment of the present invention；
Figure 13 is 29047 ammeter power and power change values curve in the embodiment of the present invention；
Figure 14 is 45000 ammeter rejecting outliers curves in the embodiment of the present invention；Figure 14 (a) is power and power change values
Curve；Figure 14 (b) is ammeter electric current and current variation value curve；
Figure 15 is 29047 ammeter electric currents and power change values curve in the embodiment of the present invention；
Figure 16 is 64258 ammeter electric currents and power change values curve in the embodiment of the present invention.
Specific embodiment
The invention will be further described below.Following embodiment is only used for clearly illustrating technical side of the invention
Case, and not intended to limit the protection scope of the present invention.
The present invention provides a kind of distribution transforming based on the fusion of more criterions with adopting data exception discriminating method, specific as follows:
1) with adopt data acquisition
There are each phase current, voltage, active power, reactive power and meter reading electricity with real-time data acquisition content is adopted mainly
Amount.Wherein, first four every 15min measurements are primary, generate 96 data points daily, meter reading electricity measures once daily.In reality
In the operation monitoring business and relevant recording, checking, and charging work of enterprise, for every electric current, voltage, active power and meter reading electricity
The quality of data is more demanding.In contrast, the quality of data of reactive power is required relatively low.And meter reading electricity with it is active
Power data is closely related, and active power data exception then means that meter reading electricity data is abnormal.So main right in the present invention
It is analyzed with the electric current, voltage and active power data for adopting data.Practical with adopting in data collection, it is primarily present with adopting number
The problem of according to breakpoint and abnormal point.
2) with data analysis is adopted, abnormal data is determined
Certain normal distribution transforming in city, distribute-electricity transformer district, Jiangning District and failure switching data are collected, wherein normal switching data and failure are matched
Parameter is according to being August in 2017 1 day to August three-phase current on the 31st, three-phase voltage, positive active power data, time interval
Scale is 15min.Brief analysis is carried out to provided data entirety overview, Main Conclusions is as follows:
A) voltage data: in whole voltage datas of offer, there are distribution transforming totally 8208 that data acquire, always acquire data point
Number is 13,555,680, and total missing data (NULL) number is 885,129, and the ratio that missing data accounts for total amount of data is
6.53%.Wherein: the voltage data for having 6001 distribution transformings to acquire is complete, no missing data.The Denver 10kV small town 7# change, five cells
The public change of 5#, rivers and mountains house property #3 become, Yu Jianyuan #2 becomes, modern times 21st century city #7 become, lake mountain honor residence of a high official 3# beauty becomes, the garden Tan Qiaobei
23 changes, the overlooking change #1 in the various schools of thinkers lake area International Garden B, the overlooking change #3 in the various schools of thinkers lake area International Garden B, the various schools of thinkers lake area International Garden B face
Street becomes the missing data ratio maximum that 2# becomes (totally 10), is 97.92%.
In the data of no missing, having 2400 data is -9999, belongs to obvious abnormal data, proportion is
0.02%.Specially China Oil and Food Import and Export Corporation's house property bright cloud occupies #1 distribution substation #1 main transformer, #2 main transformer, #3 main transformer and #4 main transformer in 28 days-August of August
Collected voltage data value on the 30th, is shown as -9999V.
B) current data: in whole current datas of offer, there are distribution transforming totally 8271 that data acquire, total acquisition data
Points are 84,205,344, and total missing data (NULL) number is 4,390,384, and total proportion that lacks is 5.21%.
Wherein, the current data for having 4464 distribution transformings to acquire is complete, no missing data.The Denver 10kV small town 7# change, China Oil and Food Import and Export Corporation's house property bright cloud
Occupy #1 distribution substation #1 main transformer, China Oil and Food Import and Export Corporation's house property bright cloud occupies #1 distribution substation #2 main transformer, China Oil and Food Import and Export Corporation's house property bright cloud occupies #1 distribution substation #3 main transformer, justice
Black commodity city #5 becomes, modern times 21st century city #7 become, the public change of five cell 5#, Wuyi oasis #2 case becomes, rivers and mountains house property #3 becomes,
The change of lake mountain honor residence of a high official 3# U.S., the change of the garden Tan Qiaobei 23, Yu Jianyuan #2 change, the overlooking change #1 in the various schools of thinkers lake area International Garden B, various schools of thinkers lake world flower
The garden area B is overlooking to become #3, the overlooking missing data ratio maximum for becoming 2# and becoming (totally 15) in the various schools of thinkers lake area International Garden B, is
97.92%.
In the data of no missing, having 2195 data is -9999, belongs to obvious abnormal data, proportion is
0.003%, the bis- collected current data of station August part of Zhang Qiaoyang is concentrated on, these abnormal current values are all greater than 700A.
C) active power data: in whole active power data of offer, there is distribution transforming totally 8153 that data acquire, always
Acquiring data points is 59,650,944, and total missing data (NULL) number is 1,748,960, and total proportion that lacks is
2.93%.Wherein, the active power data for having 4111 distribution transformings to acquire are complete, no missing data.There is the missing number of 122 distribution transformings
It is maximum side by side according to ratio, and be 97.92%.
On the basis of data aggregate analysis, raw data set is pre-processed, raw data set is lacked
Value processing, hence it is evident that the rejecting of exceptional value.Two kinds of situations are divided into for the processing of missing values: first is that in continuous time data set
Individual data is lacked, is handled at this time using common linear interpolation method.Second is that a large amount of for being lacked in continuous time data set
Data can not carry out interpolation processing, directly reject this partial data at this time.The rejecting of obvious exceptional value refers to voltage, electric current
It is rejected with the data for being shown as -9999 in active power data.
Using prototype clustering procedure, Density Clustering method, probability density method and deep learning theory totally four kinds of models to voltage, electricity
Stream and active power data carry out rejecting outliers and screen, and target is that exceptional value data point is found out from initial data, from
And the quality of data is improved, effective data supporting is provided for other related services.In testing, it takes before breaking down 1-3 months
A phase current, A phase voltage, active power curves carry out outlier detection.When there is large-scale data missing, data are selected
Curve before missing is calculated.
3) with adopt abnormal data screen process
For the performance for verifying model, certain noise is artificially added on the basis of initial data first, test model is
It is no these Interference Detections to be come out.On this basis, by prototype clustering procedure, Density Clustering method, probability density method, deep learning
The models such as method are used for real data, and the detection of exceptional value is carried out to voltage, electric current and active power initial data.Specific implementation
Process are as follows:
31) based on electric current, the voltage, active data for operating normally metering device acquisition, increase in initial data different
The random noise and interference of degree, these noise Normal Distributions, normal distribution noise spot is added in initial data at random,
Initial data is modeled as abnormal point after noise is added, and forms the time series containing abnormal point.Use prototype clustering procedure, density
Four kinds of clustering procedure, probability density method, deep learning method models test interference with noise, testing accuracy.In test, most
Whole rejecting outliers result are as follows: any two kinds and model above testing result are abnormal point as final determining exception
Point.
32) identical as step 31) implementation process, the data before breaking down to known fault ammeter carry out abnormal point Zhen
Not, it is screened using the abnormal point that four kinds of models carry out multi-angle to the data such as electric current, voltage, active respectively.
33) abnormal point of classification is analyzed, finds rule and common ground.
4) based on more criterions fusion use adopt abnormal data examination
It is different that four kinds of methods progress such as prototype clustering procedure, Density Clustering method, probability density method, deep learning method are respectively adopted
The examination of constant value, and the exceptional value judgement precision of more each model.
41) prototype clustering procedure: the output of prototype clustering procedure rejecting outliers mode input is as shown in table 1.
The output of 1 prototype clustering procedure rejecting outliers mode input of table
Exceptional value based on prototype clustering procedure, which is screened, realizes step are as follows:
A) cluster attribute is chosen；Such as: when carrying out exceptional value examination to voltage time sequence, voltage actual value, voltage being become
Change value comprehensively considers voltage value height with pace of change to exceptional value collective effect as cluster attribute.Similarly, to electric current and function
When rate time series carries out exceptional value examination, cluster attribute is respectively current actual value, current variation value and power actual value
And power change values.
B) k-means algorithm is used, sample is gathered for 4 classes, and determines all kinds of mass centers.
C) distance of each point to be determined of calculating to nearest cluster centre.
D) relative distance of each point to be determined of calculating to nearest cluster centre.
E) it makes comparisons with given threshold value.Threshold value is determined according to each distribute-electricity transformer district voltage characteristic.If point to be determined arrives
The relative distance of nearest cluster centre is greater than the threshold value, it is believed that the point to be determined is outlier.
By above-mentioned steps, it can be deduced that the rejecting outliers result based on prototype clustering procedure.
42) Density Clustering method: Density Clustering method assumes that classification can be determined by the tightness degree of sample distribution, can incite somebody to action
Sample is divided into cluster sampling classification and discrete sample noise spot.Specific step is as follows:
A) it is accounted for by voltage, electric current, the relationship of power three data between any two, draws voltage-to-current respectively
Two-dimensional surface distribution map, the flat distribution map of current-power and voltage-power flat distribution map.
B) one maximum distance d is set, when two points distance is more than d in flat distribution map, to think them not be that density can
It reaches, i.e., they are not belonging to same category；
C) loop iteration finds out a series of reachable sample points of all density, these points are divided into one kind.Remaining does not belong to
In the point of any class be noise.
Density Clustering method rejecting outliers model is as shown in table 2.
The output of 2 Density Clustering method rejecting outliers mode input of table
43) probability density method: the data exception value detection method based on deviation is main to judge exception according to " 3 σ " criterion
Value.If data Normal Distribution, under " 3 σ " criterion, exceptional value is defined as in measured value with mean deviation being more than 3 times
The value of standard deviation.Under normal distribution hypothesis, the value probability of occurrence except 3 σ of distance average isThis
Probability belongs to extremely a other small probability event.In formula, σ is that initial data normal distribution standard is poor.
For test data r1,r2,…,rnElectric current, power or the voltage time sequence for indicating input, take arithmetic mean of instantaneous value:
Wherein, n is the number of samples in electric current, power or contact potential series.
And residual error valueFind out root-mean-square-deviation are as follows:
Then exceptional value distinguishing rule are as follows: ifThe value is abnormal data；IfThen riFor normal number
According to.
But to the voltage, electric current and power data of actual measurement on site, it is difficult to judge its probability distribution in advance, and
Normal distribution is disobeyed under normal circumstances.To, judge exceptional value using " 3 σ " criterion there are errors it is larger, be difficult to completely retouch
State the probability distribution of voltage, electric current and power.
Probability density method carries out exceptional value and screens mode input output as indicated at 3.
The output of 3 probability density method rejecting outliers mode input of table
Rejecting outliers method specific implementation flow based on probability density method is as follows:
A) voltage, electric current and power data are obtained；
When b) carrying out exceptional value examination to the different times sequence such as voltage, electric current and power, distinguished using kernel density function
It is fitted the probability distribution of voltage, electric current and power data, and obtains probability density function；
C) for any numerical value d to be detected, to the carry out integral calculation of probability density function can calculate appearance [d ,+∞)
The probability of numberical range, and and threshold value comparison, the i.e. probability whether be lower than the corresponding probability 0.003 of 3 σ；If so, the point is different
Chang Dian；
D) abnormal data is judged whether it is according to comparison result.
By above-mentioned steps, it can be deduced that the rejecting outliers result based on probability density method.
44) deep learning algorithm model
Exceptional value based on deep learning algorithm screens model according to statistical method principle, by the depth of training completion
It practises model and scientific and reasonable prediction, the error of comparison prediction value and true value is carried out to following electric current, voltage or power data.
If predicted value much deviates true value, which is outlier.If predicted value fluctuates near true value, the point tolerance
Belong to normal random error, namely illustrates that the point is normal point.Shot and long term memory network has at good long-term sequence
Reason ability can be realized the preservation and control of remote information, be conducive to provide accurate electric current, voltage and power prediction value.
Therefore, the present invention divides on the basis of analyzing shot and long term memory network (long short-term memory, LSTM) basic principle
It is other that shot and long term memory network prediction model is established to electric current, voltage and power, it realizes to electric current, voltage and power time series
Exceptional value is screened.The output of deep learning rejecting outliers mode input is as shown in table 4.
The output of 4 deep learning method rejecting outliers mode input of table
Traditional artificial neural network model, input layer are connect entirely with hidden layer, hidden layer with output interlayer neuron, and each
It is connectionless between layer neuron.It individually isolated the front and back moment is had ignored to the mode of each sample process input number however, this
According to relevance, it is poor to certain long-term sequence problem throughputs such as natural language processing, machine translation.Recurrent neural
Network (recur-rent neural network, RNN) is a kind of network structure important in deep learning field, typical
It is characterized between neuron not only thering is internal feedback connection, also containing feedforward connection.RNN is easy to appear gradient in the training process
It disappears and gradient explosion issues, causes RNN that can not capture the influence that remote output exports current time, it is wide to limit it
General application and development.
It is illustrated with reference to Fig. 1 LSTM structure each section calculating process, wherein the concrete meaning of each input/output variable are as follows: xt
For mode input, 96 points of history of electric current, voltage and power are indicated；In Fig. 1 (e), otFor the out gate of LSTM, indicate to be predicted
Electric current, voltage or the power at moment；E is prediction error, i.e. the difference of the predicted value and actual value of model output, for judging whether
For abnormal point.Remaining variables are the intermediate variable and parameter of model.
The training algorithm of shot and long term memory network: after completing forward calculation, that is, error backpropagation algorithm can be used to mould
Shape parameter is updated adjustment.LSTM needs learn totally 4 groups of parameter, it may be assumed that WfAnd bf、WiAnd bi、WoAnd bo, WcAnd bc.For convenient for
It derives, by weight matrix Wf、Wi、Wo、WcIt is written as two separated matrixes: Wfh、Wfx、Wih、Wix、Woh、Wox、Wch、Wcx。ht-1Before being
The output of one moment LSTM, xtIt is current time input, bf、bi、bo、bcRespectively forget door, input door, out gate knot
The biasing of structure and current time input unit.
Define the error term δ of t momenttIt is loss function to the derivative of output valve, i.e.,Meanwhile defining each mind
It is respectively as follows: through member weighting input and its error term
When error is propagated along time reversal, the error term δ at t-1 moment is calculatedt-1Are as follows:
In formula:For Jacobian matrix.
Current time location mode ctBy previous moment location mode ct-1By element multiplied by forgetting door ft, and current input list
First state c 'tBy element multiplied by input gate itTwo parts composition.Due to ot、ft、it、c′tFor ht-1Function, utilize total derivative public
Shi Ke get:
Further it can be obtained:
SymbolIt indicates to press element multiplication.
Formula (7) are substituted into (6), can be obtained:
By δo,t、δf,t、δi,t、δc′,tDefinition, it is known that:
Error by current l layer back transfer to l-1 layers when, define l-1 layers of errorForThat is error
Derivative of the function to l-1 layers of weighting input.Due toAndIt is all xtLetter
Number, is obtained using total derivative formula:
To obtain Woh、Wfh、Wih、WchEach parameter gradients are as follows:
Wox、Wfx、Wix、WcxGradient calculation formula are as follows:
bf、bi、bo、bcGradient calculation formula are as follows:
Forward calculation and error backpropagation algorithm based on LSTM can construct depth LSTM network frame as shown in Figure 2
Frame.
It 55) is abnormal point as final exceptional value by two kinds of models rejecting outliers result any in four kinds of models
Testing result.
Embodiment
The embodiment of the present invention is primarily based on the electric current for operating normally metering device acquisition, voltage, active data, in original number
Increase different degrees of random noise and interference according to middle, form abnormal point, interference and noise are carried out using above-mentioned four kinds of models
Test, testing accuracy.In test, final rejecting outliers result is the intersection of four kinds of model inspection results.By being arranged not
With the random error of degree, can test above-mentioned four kinds of models effectively be detected these outliers, so as to authentication
The validity of method.It is specific as follows:
1) it tests 1: choosing the ammeter that number is 15661, time range is on May 3rd, 2017 to May 31 total 2785
Point, wherein A phase voltage mean value 228.891V, maximum value 232.8V, minimum value 221.9V.The random mean value that generates is 0, standard deviation
For the normally distributed error of 6 (A phase voltages), and these interference are put into voltage original time series at random.Table 5 is with chance error
Poor size and addition point.
Table 5 manually adds the electrical voltage point (small noise disturbance) of noise
A) exceptional value based on prototype clustering procedure screens test: model parameter setting are as follows: cluster classification is 4 classes, exceptional value
It is 500 that point judgment criterion threshold value trial, which is set as 2.75, cluster maximum cycle,.Distance function uses Euclidean distance:
Certain is put into voltage actual value, certain point with former point voltage change, certain point and preceding two o'clock voltage change as poly-
Generic attribute comprehensively considers the height of voltage value and the size of pace of change, and cluster result is as shown in figure 3, the correct points of detection are 3
Point, ID 64,372,2192.
B) exceptional value based on Density Clustering method screens test: model parameter setting are as follows: maximum distance is set as 0.5, sample
This point normalization range is (0,4), a kind of other minimum number of samples is 5, distance function is Euclidean distance calculation method.Test
As a result as shown in Fig. 4 (a) and Fig. 4 (b), abnormal ID is 372,663,995,997,2192.
C) exceptional value based on probability density method screens test: Fig. 5 is voltage change probability density curve, from Fig. 5
Voltage change integrated distribution be can be seen that near 0, the basic Normal Distribution of voltage change, in conjunction with its probability density
Function can obtain probability when a certain value occurs in voltage change.According to " 3 σ " criterion, it is assumed that abnormal voltage changing value occurs
One thousandth probability below is very small, and can calculate voltage change endpoint value at this time is -1.7516 and 1.7075.Recognize
Are as follows: normal voltage changing value range [- 1.7516,1.7075].What it is not in this range is abnormal voltage variation.
Table 6 is the rejecting outliers result obtained using probability density method.
6 probability density method exception search results (small noise disturbance) of table
D) exceptional value based on deep learning algorithm screens test: being shifted to an earlier date using LSTM deep learning algorithm to voltage
The prediction of a bit.Model parameter setting are as follows: four layers of Recognition with Recurrent Neural Network, including input layer (96 × 1 sequence inputting), one layer of LSTM
Layer (8 node), one layer of common hidden layer (4 node), one layer of output layer (1 node).Input and output: use at nearest 96 points of history
The current value (sampling interval 15min) of data prediction subsequent time.Model optimization is RMSProp (under the stochastic gradient with momentum
Algorithm drops), the number of iterations 400, training lot number 512 (training sample has more than 2000 altogether, and an iteration is about divided into 4-5 batches),
It is 5% that verifying, which collects shared ratio, in training sample.Objective function is the mean square error MSE of model output value and true value.
Prediction uses preceding 96 day data, and ID+96 is practical moment value, and Fig. 6 is prediction result.
Fig. 7 be LSTM model error, i.e., model predication value subtract true value obtain curve, abnormal ID be 372,866,998,
2192、2193。
Four kinds of method effective anomaly points summarize as shown in table 7.
The abnormal search results (small noise disturbance) of 7 four kinds of detection methods of table
Conclusion: using the method for inspection of dual crossing, 3 abnormal points be detected in the abnormal point that 5 are arranged.
It is as shown in table 8 to omit point analysis.
The missed point (small noise disturbance) of 8 abnormality detection of table
| Serial number | Add ID | U original value (V) | Random error (V) | U exception point value (V) |
| 3 | 663 | 229.3 | -1.4623 | 227.8377 |
| 4 | 1163 | 230.6 | 1.294 | 231.894 |
From data, the error of No. 3 abnormal points is 0.63%, and the error 0.56% of No. 5 abnormal points, error is smaller.A phase
Average voltage 228.891V, maximum value 232.8V, minimum value 221.9V.And the average change value of voltage is 0.4083V (the latter
Voltage value subtracts previous voltage value), maximum+6.8V, minimum -6.7V.Therefore, the voltage after the addition of the two random errors
Value deviates initial value very little, is not easy to detect.
2) test 2: change standard deviation is 8 (A phase voltages).Random error and addition point are as shown in table 9.
Table 9 manually adds the electrical voltage point (big noise disturbance) of noise
| Serial number | Add ID | U original value (V) | Random error (V) | U exception point value (V) |
| 1 | 64 | 227.9 | -0.9325 | 226.9675 |
| 2 | 372 | 228.2 | -5.0017 | 223.1983 |
| 3 | 663 | 229.3 | 9.4597 | 238.7597 |
| 4 | 1163 | 230.6 | -1.5734 | 229.0266 |
| 5 | 2192 | 228.6 | -4.1232 | 224.4768 |
A) the outlier detection method based on cluster is used: comprehensive using voltage actual value, voltage change as cluster attribute
The height and the size of pace of change for considering voltage value are closed, as a result as shown in figure 8, the correct points of detection are, abnormal ID is at 3 points
372、663、2192。
B) use Density Clustering rejecting outliers method, as a result as shown in Fig. 9 (a) and (b), abnormal ID is 663,994,
995、997。
C) probability density algorithm is used:
In conjunction with its probability density function, probability when a certain value occurs in voltage change is obtained: assuming that there is one thousandth
Probability below is very small, that is, thinks: normal voltage changing value range is [- 1.9255,1.7397], is not in this range
Exceptional value.Table 10 is that probability density method exceptional value screens result.
The abnormal search results (big noise disturbance) of 10 probability density method of table
| Failure ID | Active power value (kW) | Voltage value (V) | Current value (A) |
| 120 | 0.0512 | 231.2 | 0.061 |
| 372 | 0.0953 | 228.3 | 0.187 |
| 373 | 0.1035 | 223.1983 | 0.115 |
| 663 | 0.1092 | 229.4 | 0.183 |
| 664 | 0.0805 | 238.7597 | 0.126 |
| 828 | 0.0419 | 226.5 | 0.07 |
| 866 | 0.072 | 232.6 | 0.109 |
| 977 | 0.0407 | 228 | 0.065 |
| 998 | 0.0422 | 222.2 | 0.067 |
| 1163 | 0.0419 | 230.8 | 0.085 |
| 1302 | 0.0824 | 228.1 | 0.12 |
| 1314 | 0.0574 | 227.2 | 0.073 |
| 1315 | 0.0639 | 225.4 | 0.132 |
| 1780 | 0.0709 | 228.6 | 0.14 |
| 2160 | 0.0512 | 228.8 | 0.096 |
| 2192 | 0.1183 | 228.4 | 0.196 |
D) use LSTM algorithm: prediction uses preceding 96 day data, and ID+96 is practical moment value.Figure 10 is voltage true value
And LSTM prediction result, Figure 11 be LSTM model error, i.e., model predication value subtract true value obtain curve, abnormal ID be 372,
663、866、2192、2768。
Four kinds of method effective anomaly points summarize as shown in table 11.
The abnormal search results (big noise disturbance) of 11 4 kinds of detection methods of table
Test result: using the method for inspection of dual crossing, 3 abnormal points be detected in 5 abnormal points.
It is as shown in table 12 to omit point analysis.
The missed point (big noise disturbance) of 12 abnormality detection of table
| Serial number | Add ID | U original value (V) | U random error (V) | U exception point value (V) |
| 1 | 64 | 227.9 | -0.9325 | 226.9675 |
| 4 | 1163 | 230.6 | -1.5734 | 229.0266 |
From data, the error of No. 1 abnormal point is 0.41%, and the error 0.68% of No. 5 abnormal points, error is smaller.A phase
Average voltage 228.891V, maximum value 232.8V, minimum value 221.9V.And the average change value of voltage is 0.4083V (the latter
Voltage value subtracts previous voltage value), maximum+6.8V, minimum -6.7V.Therefore, the voltage after the addition of the two random errors
Value deviates initial value very little, is not easy to detect.
3) electric current and power measurements of practical ammeter
Rejecting outliers research is carried out to voltage, electric current and power data, obtains test electric current and power abnormal point, test
Ammeter includes known fault ammeter and does not find failure ammeter.
(1) the excessive exception of current changing rate
The ammeter current curve for being 49932 to number is analyzed, May 7 current curve and current variation value curve
As shown in figure 12.It can be seen from the figure that current variation value is larger at the 37th point, it is determined as outlier.
(2) the excessive exception of power variation rate
Figure 13 is power and power change values of 29047 ammeter of number May 3, it can be seen from the figure that the 62nd,
Power change values are larger at 63 and 64 points, are determined as outlier.
Figure 14 is that exceptional value of 45000 ammeter of number May 22 screens power, electric current and its changing value curve.From figure
As can be seen that power change values are larger at the 46th point in 14 (a), belong to obvious burr point, numerical value is exceptional value.And at this time
Current variation value in Figure 14 (b) is smaller, does not reach changing value discrimination threshold, therefore the 46th curent change is in normal model
In enclosing, it is not judged as exception.To sum up, it is determined as power exception at the 46th point.
(3) current power correlation sexual abnormality
Figure 15 is A phase power curve and A phase current curve of 29047 ammeter of number May 2, it can be seen from the figure that
At 22,34,55 and 60 etc. four points, there is significant exception in electric current and power dependency, that is, there is following two situation:
1) when electric current obviously rises or maintains the larger value, power is presented abnormal decline phenomenon or maintains lower value horizontal；
2) when electric current is decreased obviously or maintains lower value, power is presented abnormal rise phenomenon or maintains the larger value horizontal.
Similarly such as Figure 16, the ammeter that number is 64258, in the A phase current and A phase power curve on May 6 the figure institute
Show.The 34,35th of this day points out that there are correlation abnormal phenomenon.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the technical principles of the invention, several improvement and deformations can also be made, these improvement and deformations
Also it should be regarded as protection scope of the present invention.
Claims (9)
1. a kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method characterized by comprising
Distribution transforming is obtained with adopting initial data；
The distribution transforming is pre-processed with initial data is adopted；
Noise spot is added in pretreated distribution transforming at random with adopting in initial data, formed containing abnormal point with adopting data sequence；
Four kinds of prototype clustering procedure, Density Clustering method, probability density method and deep approach of learning models are respectively adopted to containing abnormal point
Abnormal point examination is carried out with data sequence is adopted；
Determine distribution transforming with adopting abnormal data；The distribution transforming screens result with the abnormal point that abnormal data is any two kinds of models is adopted
Intersection, the union for the intersection for then taking all combination of two to determine.
2. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, the acquisition distribution transforming is with adopting initial data, comprising:
Three-phase current, three-phase voltage and active power initial data, acquisition time interval are acquired based on metering device is operated normally
For 15min.
3. a kind of distribution transforming based on the fusion of more criterions according to claim 2 is with adopting data exception discriminating method, feature
It is, it is described that the distribution transforming is pre-processed with initial data is adopted, comprising:
Missing values processing and the obvious exceptional value of rejecting are carried out with initial data is adopted to distribution transforming；
It is described that missing values processing is carried out with initial data is adopted to distribution transforming, comprising: to lacking individual data in continuous time data set,
The data lacked using linear interpolation method polishing；To mass data is lacked in continuous time data set, the segment data is directly rejected；
The obvious exceptional value of rejecting refers to, will be shown as -9999 in three-phase current, three-phase voltage and active power initial data
Data reject.
4. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, when the distribution transforming is with adopting initial data and large-scale data missing occur, the curve before selecting shortage of data is counted
It calculates.
5. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, the noise spot Normal Distribution.
6. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, it is described that abnormal point examination is carried out with adopting data sequence to containing abnormal point using prototype clustering procedure, comprising:
Determine cluster attribute；It include: to choose measuring point voltage actual value, measuring point to be checked and former point voltage change to be checked and to be checked
Measuring point and cluster attribute of the preceding two o'clock voltage change as voltage time sequence choose measuring point current actual value to be checked, to be checked
Measuring point and former point current variation value and measuring point to be checked and cluster attribute of the preceding two o'clock current variation value as current time sequence,
Choosing measuring point active power actual value, measuring point to be checked and former point active power changing value to be checked and measuring point to be checked and preceding two o'clock has
Cluster attribute of the function power change values as active power time series；
According to the cluster attribute, using k-means algorithm, time series to be detected is gathered for 4 classes, and determine all kinds of matter
The heart；
According to all kinds of mass centers, calculates each measuring point to be checked and clustered to the distance of nearest cluster centre and each measuring point to be checked to nearest
The relative distance at center；
The relative distance of each measuring point to be checked to nearest cluster centre is made comparisons with given threshold value；If certain measuring point to be checked is to recently
The relative distance of cluster centre is greater than given threshold value, then the measuring point to be checked is outlier, i.e. abnormal point.
7. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, it is described that abnormal point examination is carried out with adopting data sequence to containing abnormal point using Density Clustering method, comprising:
The flat distribution map of voltage-to-current, electric current-active power flat distribution map and voltage-active power are drawn respectively
Flat distribution map；
Point on the flat distribution map is clustered, comprising: if between two points in the flat distribution map away from
From the maximum distance for being no more than setting, then one kind is divided into；Wherein, the point on voltage-to-current flat distribution map is certain phase current
Time series and voltage time sequence；Point on electric current-active power flat distribution map is certain phase current time series and active
Power time series；Point on voltage-active power flat distribution map is certain phase voltage time series and active power time sequence
Column；
It is abnormal point that loop iteration, which is found out and is not belonging to the point of any class,.
8. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, it is described that abnormal point examination is carried out with adopting data sequence to containing abnormal point using probability density method, comprising:
Determine mode input and model output；The mode input are as follows: be to be determined electric current with before for current time sequence
The changing value of some electric currents；It is the changing value of to be determined voltage and former point voltage for voltage time sequence；For active
Power time series are the changing value of to be determined voltage and former point voltage；The model output are as follows: changing value normal range (NR)；
It is fitted the probability distribution of voltage, electric current and active power data respectively using kernel density function, and obtains probability density letter
Number；
For any numerical value d to be detected, occurred to the carry out integral calculation of probability density function [d ,+∞) numberical range
Probability, and and threshold value comparison, the i.e. probability whether be lower than the corresponding probability 0.003 of 3 σ；If so, the numerical value to be detected is abnormal
Point.
9. a kind of distribution transforming based on the fusion of more criterions according to claim 1 is with adopting data exception discriminating method, feature
It is, it is described that abnormal point examination is carried out with adopting data sequence to containing addition abnormal point using deep approach of learning, comprising:
The deep learning model based on shot and long term memory network completed using the training of electric current, voltage and power time series, it is right
Non- incoming current, voltage or power data predicted, the error of comparison prediction value and true value；If it is super that predicted value deviates true value
The threshold value of setting is crossed, then predicted value is abnormal point；
The deep learning model of the shot and long term memory network are as follows: after completing forward calculation, using error backpropagation algorithm pair
Model parameter is updated adjustment, comprising:
The neuron weighting input net of t moment shot and long term memory networkf,t, neti,t, netc′,t, neto,tAre as follows:
Wherein, Wox、Wfx、Wix、Wcx、Woh、Wfh、Wih、WchIndicate weight, ht-1It is previous moment LSTM output, xtIt is current time
Input, bf、bi、bo、bcRespectively forget the inclined of door, input door, output door and current time input unit
It sets；
The neuron error item δ of t moment shot and long term memory networkf,t, δi,t, δc′,t, δo,tAre as follows:
Wherein, E is prediction error；
When error is propagated along time reversal, the error term δ at t-1 momentt-1Are as follows:
Wherein,For Jacobian matrix；
When error by current l layer back transfer to l-1 layers when, l-1 layers of errorAre as follows:
Finally, weight W is obtainedoh、Wfh、Wih、WchAre as follows:
Wherein, Woh,t, Wfh,t, Wih,t, Wch,tThe weight of t moment is respectively indicated, subscript T indicates transposition；
Weight Wox、Wfx、Wix、WcxAre as follows:
bf、bi、bo、bcAre as follows:
Wherein, bo,t, bf,t, bi,t, bc,tRespectively indicate the error term of t moment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910740107.2A CN110458230A (en) | 2019-08-12 | 2019-08-12 | A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910740107.2A CN110458230A (en) | 2019-08-12 | 2019-08-12 | A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110458230A true CN110458230A (en) | 2019-11-15 |
Family
ID=68485974
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910740107.2A Pending CN110458230A (en) | 2019-08-12 | 2019-08-12 | A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110458230A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110807014A (en) * | 2019-09-24 | 2020-02-18 | 国网北京市电力公司 | Cross validation based station data anomaly discrimination method and device |
| CN111241656A (en) * | 2019-12-28 | 2020-06-05 | 国网江西省电力有限公司电力科学研究院 | Distribution transformer outlet voltage abnormal point detection algorithm |
| CN111984925A (en) * | 2020-07-29 | 2020-11-24 | 江苏方天电力技术有限公司 | Circuit abnormity positioning method based on loop impedance, storage medium and computing equipment |
| CN112084229A (en) * | 2020-07-27 | 2020-12-15 | 北京市燃气集团有限责任公司 | Method and device for identifying abnormal gas consumption behaviors of town gas users |
| CN112464848A (en) * | 2020-12-07 | 2021-03-09 | 国网四川省电力公司电力科学研究院 | Information flow abnormal data monitoring method and device based on density space clustering |
| CN112669879A (en) * | 2020-12-24 | 2021-04-16 | 山东大学 | Air conditioner indoor unit noise anomaly detection method based on time-frequency domain deep learning algorithm |
| CN112819373A (en) * | 2021-02-25 | 2021-05-18 | 云南电网有限责任公司电力科学研究院 | Distribution network voltage abnormal data detection method and device |
| CN112819088A (en) * | 2021-02-20 | 2021-05-18 | 苏州安极能新能源发展有限公司 | Anomaly detection algorithm based on power data |
| CN112818052A (en) * | 2021-02-25 | 2021-05-18 | 云南电网有限责任公司电力科学研究院 | Abnormal voltage data detection method and device |
| CN113076945A (en) * | 2021-03-17 | 2021-07-06 | 华夏芯（北京）通用处理器技术有限公司 | Camera direct-reading meter reading instrument abnormal point removing method based on improved RANSAC |
| CN113434496A (en) * | 2021-07-15 | 2021-09-24 | 广东电网有限责任公司 | Distribution transformer weight overload real-time monitoring system method, system and computer medium |
| CN113554117A (en) * | 2021-08-16 | 2021-10-26 | 中国南方电网有限责任公司 | Abnormal load data identification method and electronic equipment |
| CN113762507A (en) * | 2021-08-24 | 2021-12-07 | 浙江中辰城市应急服务管理有限公司 | Semi-supervised deep learning arc voltage anomaly detection method based on phase space reconstruction |
| CN114997253A (en) * | 2021-02-23 | 2022-09-02 | 哈尔滨工业大学 | Intelligent state anomaly detection method, monitoring system and monitoring method for satellite constellation |
| CN116150239A (en) * | 2022-12-16 | 2023-05-23 | 彭州华润燃气有限公司 | Data mining method for gas stealing behavior |
| WO2023246070A1 (en) * | 2022-06-22 | 2023-12-28 | 中国移动通信集团广东有限公司 | Abnormal power consumption station detection method and apparatus, electronic device, and readable storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101561878A (en) * | 2009-05-31 | 2009-10-21 | 河海大学 | Unsupervised anomaly detection method and system based on improved CURE clustering algorithm |
| CN103901880A (en) * | 2014-04-01 | 2014-07-02 | 浙江大学 | Industrial process fault detection method based on multiple classifiers and D-S evidence fusion |
| CN108536741A (en) * | 2018-03-09 | 2018-09-14 | 国网江苏省电力公司无锡供电公司 | A kind of energy method for monitoring abnormality |
-
2019
- 2019-08-12 CN CN201910740107.2A patent/CN110458230A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101561878A (en) * | 2009-05-31 | 2009-10-21 | 河海大学 | Unsupervised anomaly detection method and system based on improved CURE clustering algorithm |
| CN103901880A (en) * | 2014-04-01 | 2014-07-02 | 浙江大学 | Industrial process fault detection method based on multiple classifiers and D-S evidence fusion |
| CN108536741A (en) * | 2018-03-09 | 2018-09-14 | 国网江苏省电力公司无锡供电公司 | A kind of energy method for monitoring abnormality |
Non-Patent Citations (1)
| Title |
|---|
| 罗慧等: "基于长短期记忆网络的智能用电数据甄别方法", 《广东电力》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110807014A (en) * | 2019-09-24 | 2020-02-18 | 国网北京市电力公司 | Cross validation based station data anomaly discrimination method and device |
| CN111241656B (en) * | 2019-12-28 | 2023-03-14 | 国网江西省电力有限公司电力科学研究院 | Distribution transformer outlet voltage abnormal point detection algorithm |
| CN111241656A (en) * | 2019-12-28 | 2020-06-05 | 国网江西省电力有限公司电力科学研究院 | Distribution transformer outlet voltage abnormal point detection algorithm |
| CN112084229A (en) * | 2020-07-27 | 2020-12-15 | 北京市燃气集团有限责任公司 | Method and device for identifying abnormal gas consumption behaviors of town gas users |
| CN111984925A (en) * | 2020-07-29 | 2020-11-24 | 江苏方天电力技术有限公司 | Circuit abnormity positioning method based on loop impedance, storage medium and computing equipment |
| CN111984925B (en) * | 2020-07-29 | 2024-03-12 | 江苏方天电力技术有限公司 | Circuit abnormality positioning method based on loop impedance, storage medium and computing device |
| CN112464848A (en) * | 2020-12-07 | 2021-03-09 | 国网四川省电力公司电力科学研究院 | Information flow abnormal data monitoring method and device based on density space clustering |
| CN112464848B (en) * | 2020-12-07 | 2023-04-07 | 国网四川省电力公司电力科学研究院 | Information flow abnormal data monitoring method and device based on density space clustering |
| CN112669879A (en) * | 2020-12-24 | 2021-04-16 | 山东大学 | Air conditioner indoor unit noise anomaly detection method based on time-frequency domain deep learning algorithm |
| CN112669879B (en) * | 2020-12-24 | 2022-06-03 | 山东大学 | Air conditioner indoor unit noise anomaly detection method based on time-frequency domain deep learning algorithm |
| CN112819088A (en) * | 2021-02-20 | 2021-05-18 | 苏州安极能新能源发展有限公司 | Anomaly detection algorithm based on power data |
| CN114997253A (en) * | 2021-02-23 | 2022-09-02 | 哈尔滨工业大学 | Intelligent state anomaly detection method, monitoring system and monitoring method for satellite constellation |
| CN112819373A (en) * | 2021-02-25 | 2021-05-18 | 云南电网有限责任公司电力科学研究院 | Distribution network voltage abnormal data detection method and device |
| CN112818052A (en) * | 2021-02-25 | 2021-05-18 | 云南电网有限责任公司电力科学研究院 | Abnormal voltage data detection method and device |
| CN113076945A (en) * | 2021-03-17 | 2021-07-06 | 华夏芯（北京）通用处理器技术有限公司 | Camera direct-reading meter reading instrument abnormal point removing method based on improved RANSAC |
| CN113076945B (en) * | 2021-03-17 | 2024-05-14 | 华夏芯（北京）通用处理器技术有限公司 | Method for eliminating abnormal points of camera direct-reading meter reading instrument based on improved RANSAC |
| CN113434496A (en) * | 2021-07-15 | 2021-09-24 | 广东电网有限责任公司 | Distribution transformer weight overload real-time monitoring system method, system and computer medium |
| CN113554117A (en) * | 2021-08-16 | 2021-10-26 | 中国南方电网有限责任公司 | Abnormal load data identification method and electronic equipment |
| CN113762507A (en) * | 2021-08-24 | 2021-12-07 | 浙江中辰城市应急服务管理有限公司 | Semi-supervised deep learning arc voltage anomaly detection method based on phase space reconstruction |
| CN113762507B (en) * | 2021-08-24 | 2023-12-29 | 浙江中辰城市应急服务管理有限公司 | Semi-supervised deep learning arc voltage anomaly detection method based on phase space reconstruction |
| WO2023246070A1 (en) * | 2022-06-22 | 2023-12-28 | 中国移动通信集团广东有限公司 | Abnormal power consumption station detection method and apparatus, electronic device, and readable storage medium |
| CN116150239B (en) * | 2022-12-16 | 2023-09-22 | 彭州华润燃气有限公司 | Data mining method for gas stealing behavior |
| CN116150239A (en) * | 2022-12-16 | 2023-05-23 | 彭州华润燃气有限公司 | Data mining method for gas stealing behavior |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110458230A (en) | A kind of distribution transforming based on the fusion of more criterions is with adopting data exception discriminating method | |
| CN110097297B (en) | Multi-dimensional electricity stealing situation intelligent sensing method, system, equipment and medium | |
| CN110223196B (en) | Anti-electricity-stealing analysis method based on typical industry feature library and anti-electricity-stealing sample library | |
| Yan et al. | Performance analysis of electricity theft detection for the smart grid: An overview | |
| León et al. | Variability and trend-based generalized rule induction model to NTL detection in power companies | |
| Xia et al. | Detection methods in smart meters for electricity thefts: A survey | |
| CN110458195A (en) | A kind of abnormal data discriminating method based on the fusion of more criterions | |
| Bian et al. | Abnormal detection of electricity consumption of user based on particle swarm optimization and long short term memory with the attention mechanism | |
| Gu et al. | Electricity theft detection in AMI with low false positive rate based on deep learning and evolutionary algorithm | |
| CN109376944A (en) | The construction method and device of intelligent electric meter prediction model | |
| CN110879377A (en) | Metering device fault tracing method based on deep belief network | |
| Lin et al. | Electricity theft detection in power consumption data based on adaptive tuning recurrent neural network | |
| Javaid et al. | Using GANCNN and ERNET for detection of non technical losses to secure smart grids | |
| CN114240041A (en) | Lean line loss analysis method and system for distribution network distribution area | |
| Fang et al. | A machine learning-based detection framework against intermittent electricity theft attack | |
| Saeed et al. | Electricity theft detection in power utilities using bagged CHAID-based classification trees | |
| CN115905319B (en) | Automatic identification method and system for abnormal electricity fees of massive users | |
| CN115730962A (en) | Big data-based electric power marketing inspection analysis system and method | |
| CN109636057A (en) | Based on the link prediction technique and storage medium for improving artificial immune system | |
| CN114066219A (en) | Electricity stealing analysis method for intelligently identifying electricity utilization abnormal points under incidence matrix | |
| Yang et al. | Electricity theft detection based on ReliefF feature selection algorithm and BP neural network | |
| Zhang et al. | Multi-feature fusion based anomaly electro-data detection in smart grid | |
| Yadav et al. | The detection of non-technical losses and electricity theft by smart meter data and Artificial Intelligence in the context of electric distribution utilities: A comprehensive review | |
| CN114154617A (en) | Low-voltage resident user abnormal electricity utilization identification method and system based on VFL | |
| Shijian et al. | Energy efficiency evaluation system based on ladder-based comprehensive assessment method |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191115 |