CN109165504A - A kind of electric system false data attack recognition method generating network based on confrontation - Google Patents
A kind of electric system false data attack recognition method generating network based on confrontation Download PDFInfo
- Publication number
- CN109165504A CN109165504A CN201810982950.7A CN201810982950A CN109165504A CN 109165504 A CN109165504 A CN 109165504A CN 201810982950 A CN201810982950 A CN 201810982950A CN 109165504 A CN109165504 A CN 109165504A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- measurement
- residual error
- follows
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/50—Monitoring users, programs or devices to maintain the integrity of platforms, e.g. of processors, firmware or operating systems
- G06F21/55—Detecting local intrusion or implementing counter-measures
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/048—Activation functions
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Security & Cryptography (AREA)
- Computer Hardware Design (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The present invention discloses a kind of electric system false data attack (False Data Injection Attack that network is generated based on confrontation, FDIA) recognition methods, the following steps are included: measuring Step 1: obtaining residual error measurement, estimator and substantial amount by the pretreatment based on state estimation, the detected data of FDIA is given over to；Step 2: carrying out filtering out bad data based on two norm threshold tests to residual error measurement；Differentiate Step 3: whether there is with the discrimination model for generating network based on confrontation to FDIA；Step 4: being carried out orientation problem data with based on the generation model generation positioning residual error data for generating network is fought to positioning two norm threshold tests of residual error data progress and being filtered out.The present invention simplifies assumed condition too strong in conventional method, and model training meets industry and require the practical of the detection of metric data containing FDIA independent of large-scale FDIA abnormal data sample.
Description
Technical field
The invention belongs to Operation of Electric Systems security maintenance technical fields, and in particular to a kind of to generate network based on confrontation
Electric system false data attack recognition method.
Background technique
For a long time, whether the often power quality reliable, power equipment whether up to standard that traditional power grid is most paid attention to is safe
Economy, and cause the ignorance to electrical power system network safety.However nowadays, under the propulsion of Demand-side reform, modern electric
The composition in market is far richer compared with traditional power grid, and a large amount of communication apparatus is introduced therewith, and the power grid in China is also by traditional electricity
Gradually transition and upgrade becomes an information physical catenet throughout the country to net.The network security problem of electric system is carried on the back herein
Become more severe under scape.False data injection attacks (False Data Injection Attack, FDIA) are many electric power
Most typically in system network safety problem, a kind of problem of greatest concern.
The accuracy that FDIA is mainly estimated in the power system by the telemetry intelligence (TELINT) in the system of distorting come collapse state
With the purpose of reliability.Currently, the detection method of FDIA is carried out abnormality detection to the telemetry intelligence (TELINT) for being sent into Energy Management System,
It is wanted in measurement information with determining with the presence or absence of abnormal data.In conventional operation, the detection of abnormal data is examined by bad data
Examining system is realized.However external current research is it has already been indicated that traditional raw data detection can not detect to contain FDIA
The telemetry distorted needs an individual module to detect it containing the telemetry that FDIA is distorted.
The detection of electric system FDIA worldwide still falls within knotty problem at present, although having existed needle both at home and abroad
To the tens of pieces of the research paper of the problem, but industry there is no the maturation for FDIA so far and the solution being used widely is done
Method.The work of the researcher of successive dynasties FDIA detection in the literature can be broadly divided into two classes: 1) utilizing the mathematics sides such as probabilistic model
Method detects FDIA 2) using the method comprising artificial intelligence detect FDIA.Why first kind method cannot obtain preferably
Application be because they are usually required based on stronger assumed condition, such as: the state variable in system follows specific distribution,
FDIA be influenced by director measurement Power system state estimation etc..Too strong hypothesis often makes these methods
It is difficult to move even one step in true work.Deficiency existing for second class method is that they are extremely trained dependent on huge FDIA
Sample, reality in exceptional sample scale be often up to less than these methods training condition.These two types of deficiencies make existing document
In method seldom have realistic meaning.
Summary of the invention
It is difficult to detected problem the technical problem to be solved by the present invention is to electric system metric data containing FDIA, provides
A kind of electric system false data attack recognition method generating network based on confrontation, to hypothesis too strong in previous conventional method
Condition is simplified, and model training meets independent of large-scale FDIA abnormal data sample industrially to amount containing FDIA
The practical of measured data detection requires.
The technical problems to be solved by the invention are achieved by the following technical programs:
A kind of electric system false data attack recognition method generating network based on confrontation, comprising the following steps:
Step 1: linear state estimation model is established, it is collected everywhere in the whole network based on remote-measuring equipment in electric system
Telemetering amount carries out state estimation and obtains estimated state amount；It is pushed away, is obtained according to linear state estimation model is counter with estimated state amount
To the corresponding estimation measurement of estimated state amount；Estimation measurement is made the difference with substantial amount measurement, obtains residual error measurement；Residual error
Measurement, estimator and substantial amount measurement give over to the detected data of FDIA；
Step 2: being filtered out in data by carrying out the two norm threshold tests based on given threshold value size to residual error measurement
The bad data generated by Physical Network operation troubles, device measuring error, communication system noise makes only to retain true number in system
Accordingly and FDIA distort after false data；
Step 3: establishing confrontation generates network model, and net is generated to confrontation with the health data in electric system history
Network model is trained, and is obtained mature parameters of electric power system discrimination model and is generated model；The residual error that will retain in step 1
Measurement, estimator and substantial amount measurement are sent into discrimination model and are differentiated, obtain differentiating result；If differentiating, result is not
Then detecting there are false data terminates, and otherwise enters step four；
Step 4: will be judged as making the difference in the presence of abnormal data with the generation data for generating model, positioning residual error is obtained；
Positioning residual error data asked based on two norm threshold tests of given threshold value size with determining problem data position and filtering out
Inscribe data.
The step 1 specifically:
Establish linear state estimation model are as follows:
Z=Hx+e
Wherein, z be telemetering amount, H be representative topology Jacobian matrix, e be state estimation during inevitably generate miss
Difference and disturbance；
The relationship of quantity of state and measurement are as follows:
Qk=0
Pk-m=bk-m(θk-θm)
Qk-m=0
Wherein, PkActive power, Q are injected for node kkReactive power, P are injected for node kk-mIt is k node between m node
The active power of route, Qk-mFor k node to the reactive power of route between m node, Bk-mThe imaginary part of (k, m) in admittance battle array,
bk-mFor k node to the susceptance of route between m node, θkWith θmThe polar coordinates voltage phase angle of respectively node k and node m；
The objective function of noise least square is established to carry out state estimation, objective function expression formula are as follows:
Min J (x)=[z-Hx]TR-1[z-Hx]
Wherein, R is weight matrix；
Decline principle according to gradient and acquire estimated state amount:
Measurement is carried out according to linear relationship to be back-calculated to obtain estimation measurement, Extrapolation after obtaining estimated state amount
Follow formula are as follows:
Pass through formulaResidual error measurement Δ z is obtained, obtained residual error measurement, estimator and substantial amount are measured
Give over to the detected data of FDIA.
Number is filtered out by carrying out the two norm threshold tests based on given threshold value size to residual error measurement described in step 2
The bad data generated in by Physical Network operation troubles, device measuring error, communication system noise, detection follow formula are as follows:
Wherein, τ1For constant threshold.
The step 3 specifically:
It includes discrimination model and generation model, their objective function that the confrontation, which generates network model, are as follows:
Wherein, formula above is the objective function for generating model in optimization, and formula below is discrimination model excellent
Objective function when change；The competition that the core of confrontation production network is to introduce discrimination model and generate between model comes so that two
A model can obtain significant progress and optimization in training, and competition manifestation is wished to make and correctly be sentenced for discrimination model
Disconnected, and generate model and wish to deceive discrimination model, can two models around " accurate judgement " this problem expansion alternately game；
The output result D of discrimination model represents discrimination model to the trusting degree of input data authenticity, and generating model output result G is
Generate the simulation truthful data of model production；The optimization training for generating model is realized by deceiving discrimination model, therefore G
Objective function wish the judgement mistake as far as possible of D, i.e. D (G (z)) is as big as possible；And the optimization training of discrimination model is to pass through
Promote what its own judgement can not realized, therefore the objective function of D wishes that the judgement of D is as correct as possible, i.e. D (G (z)) to the greatest extent may be used
Can be small, while D (x) is as big as possible when facing truthful data；E is binary matrix, and specific value is according to feeding data source
It determines, to guarantee that the optimization process of separate sources data is independent of each other.
The generation model is all made of with discrimination model containing the reverse transmittance nerve network for intersecting Entropy principle, includes three layers
Neuron, two linear plus sigmoid activation primitive mappings, the number of each layer neuron according to grid nodes number to be checked and
Fixed, first layer is all input layer, and the last layer is output layer, and intermediate one layer is then hidden layer, and the calculating of forward-propagating follows public affairs
Formula are as follows:
Wherein, anFor n-th layer neuron output as a result, ωnFor weight of the n-th layer neuron into n+1 layers of mapping, bn
For biasing of the n-th layer neuron into n+1 layers of mapping, r is the input of neural network, and s is the output of neural network, and σ is represented
Sigmoid activation primitive, calculation method follow formula are as follows:
Wherein, e is natural Exponents；
It is based on intersecting Entropy principle with the Direct mapping result of neural network and given optimization aim and seeks output layer error, meter
Calculation follows formula are as follows:
Again by asking local derviation to obtain updating gradient output layer error, whole network is reversely updated using gradient decline principle
Weight and biasing, calculation method follow formula are as follows:
Wherein, y is preset optimization aim, and C is output layer error, ρiFor Gradient learning rate；
Mature discrimination model can be obtained by backpropagation training and generate model, the residual error that will retain in step 1
Measurement, estimator and substantial amount measurement be sent into mature discrimination model can be obtained FDIA presence or absence differentiation as a result, if
Differentiate that result is that there is no FDIA to distort, differentiation terminates, and distorts then enter the 4th step if it exists.
A pair positioning residual error data described in step 4 is carried out the two norm threshold tests based on given threshold value size and is asked with determination
Topic Data Position simultaneously filters out problem data, and detection follows formula are as follows:
Wherein, ztotalFor the residual error measurement of data to be tested, estimator and substantial amount measurement summation,To generate
The summation that residual error measurement, estimator and the substantial amount for the data to be tested that model generates measure, τ2For constant threshold.
Compared with prior art, the invention has the following advantages:
(1) present invention solves the problems, such as that traditional raw data detection can not detect data containing FDIA.
(2) present invention be not based on very strong scenario, can be applicable in multiple types FDIA initiation the problem of, with compared with
Good industrial applicibility.
(3) present invention in the training to model independent of large-scale abnormal data training sample, but based on just
Normal sample data overcomes sample shortage instantly, it is difficult to the problem of maintaining model training.
Detailed description of the invention
Fig. 1 is π type equivalent circuit model figure；
Fig. 2 is that confrontation generates network model figure；
Fig. 3 is neural network back-propagation process figure；
Fig. 4 is 118 node systems training performance figure；
Fig. 5 is 118 node system abnormal point positioning result figures.
Specific embodiment
In order to which the above objects, features and advantages of the present invention is more clearly understood, below with reference to the specific implementation of model
Technical solution of the present invention is further described in detail in form.
A kind of electric system false data attack recognition method generating network based on confrontation, comprising the following steps:
Step 1: linear state estimation model is established, using the quantity of state of electric system as uncertain collection, with electric system
In telemetering amount, Jacobian matrix and noise modeling；Estimated state amount can be solved by minimizing noise and counter push away estimated state
Measure corresponding estimation measurement；Estimation measurement is made the difference with substantial amount measurement, obtains residual error measurement；Residual error measurement is estimated
Metering gives over to the detected data of FDIA with substantial amount measurement.It is specific as follows:
π type equivalent circuit model figure according to figure 1 is visible: telemetering amount z=[P in electric systemk,Qk,Pk-m,Qk-m,
Vk]TWith estimator x=[θk,Vk]TThere are non-linear relations.Based on the assumption that: 1) the per unit value voltage of node with 1 very close；
2) resistance, receive be 0, i.e. the loss of circuit is 0, phase difference close to 0, non-linear relationship can be simplified is abstracted as it is linear
Relational expression, as shown in formula (1).
Z=Hx+e (1)
Wherein: H is the Jacobian matrix of representative topology, and e is inevitably generates error and disturbance during state estimation.
Generally there are relationships in such as (2) for quantity of state and measurement in nonlinear state Eq model:
The form in formula (3) is being obtained after above-mentioned hypothesis simplifies:
Wherein: PkActive power, Q are injected for node kkReactive power, P are injected for node kk-mIt is k node between m node
The active power of route, Qk-mFor k node to the reactive power of route between m node, Gk-mThe real part of (k, m) in admittance battle array,
Bk-mThe imaginary part of (k, m), g in admittance battle arrayk-mFor k node to the conductance of route between m node, bk-mFor k node to m node it
Between route susceptance, y is susceptance over the ground, VkWith VmThe respectively voltage magnitude of node k and node m, θkWith θmRespectively node k
With the polar coordinates voltage phase angle of node m.
In order to enable state estimation is accurate as far as possible, it should which reduction noise as far as possible establishes the objective function of noise least square
Carry out state estimation, objective function is such as shown in (4).
Min J (x)=[z-Hx]TR-1[z-Hx] (4)
Wherein: R is weight matrix.
Estimated state amount (5) can be acquired by declining principle according to gradient:
Measurement can be carried out according to linear relationship being back-calculated to obtain estimation measurement after obtaining estimated state amount, it is counter to push away meter
Calculation follows formula (6).
Thus residual error measurement can be obtained by formula (7).The residual error that the present invention will be collected in this step be generated
Measurement, estimator and substantial amount measurement give over to the detected data of FDIA.
Wherein: Δ z is residual error measurement.
Step 2: being filtered out in data by carrying out the two norm threshold tests based on given threshold value size to residual error measurement
The bad data generated by Physical Network operation troubles, device measuring error, communication system noise makes only to retain true number in system
Accordingly and FDIA distort after false data.It is specific as follows:
Generally, due to which Physical Network operation troubles, device measuring error, communication system noise make electric system telemetering amount
Measured data can always have contaminated bad data, and in order to reduce the workload and difficulty of false data differentiation, the present invention is first
The bad data in metric data is filtered out first with raw data detection, only leaves normal data and false data in next step
False data arbiter detection.
Bad data is filtered out using two norm threshold test of residual error, detection follows formula (8).
Wherein: τ1For constant threshold.When the establishment of (12) formula then shows to need in detected data with the presence of bad data
It filters out worst error amount in measurement and re-starts detection until detection passes through.
The step for of the invention, can greatly improve the efficiency of subsequent detection, reduce the workload of false data identification
And difficulty, to improve the accuracy rate of false data differentiation.
Step 3: establish confrontation generate network model, the confrontation generate network model include discrimination model and generate model,
Discrimination model and generation model are symmetrical one group of three-layer neural network；With the health data in electric system history to antibiosis
It is trained at network model, obtains mature discrimination model and generates model；By the residual error measurement retained in step 1, estimate
It measures to be sent into mature discrimination model with substantial amount measurement and be differentiated, obtain differentiating result；If differentiating, result is not deposit
Then detecting in false data terminates, and otherwise enters step four.It is specific as follows:
As shown in Fig. 2, building model includes a generation model and a discrimination model.Generation model is for generating and very
The close false data of real data, discrimination model are used to judge that the source for the data being sent into be to generate model or authentic specimen.
When discrimination model differentiates it is correct when, illustrate to generate model capability insufficient, thus optimize and generate model, it is on the contrary then optimize differentiation mould
Type.By generating the confrontation and competition of model and discrimination model, two kinds of models can be improved.Finally, due to training data
It is the health data in successive dynasties, whether the data that mature discrimination model can determine feeding are completely healthy, either with or without presence
FDIA is distorted.And the data very similar with healthy sample can be produced by generating model then.
The model can be realized by establishing the objective function such as formula (9).Formula above wherein is that generator G is optimizing
When objective function, following formula is objective function of the arbiter D in optimization.
Model is generated to be all made of with discrimination model containing the reverse transmittance nerve network for intersecting Entropy principle.According to Fig.3, raw
It include three layers of neuron, two linear plus sigmoid activation primitive mappings at model and discrimination model.Each layer neuron
Number is depending on grid nodes number to be checked.Their first layer is all input layer, and the last layer is output layer, and intermediate one layer then
It is hidden layer.The calculating of their forward-propagatings follows formula (10).
Wherein: anFor n-th layer neuron output as a result, ωnFor n-th layer neuron to n+1 layers mapping in weight,
bnFor biasing of the n-th layer neuron into n+1 layers of mapping, r is the input of neural network, and s is the output of neural network, and σ is represented
Sigmoid activation primitive, calculation method follow formula (11).
Wherein: e is natural Exponents.
According to Fig. 3, it is based on intersecting Entropy principle with the Direct mapping result of neural network and given optimization aim and seeks output layer
Error, calculating follow formula (12).It is anti-using gradient decline principle again by asking local derviation to obtain updating gradient output layer error
To the weight and biasing for updating whole network, calculation method follows formula (13).Maturation can be obtained by backpropagation training
Discrimination model and generation model.
Wherein: y is preset optimization aim, and C is output layer error, ρiFor Gradient learning rate.
The residual error measurement of the data to be tested of step 1, estimator and substantial amount are measured and are sent into the mature differentiation of training
The differentiation result of FDIA presence or absence can be obtained in model.Differentiating if differentiating that result is to distort there is no FDIA terminates, if depositing
Then enter the 4th step distorting.
Step 4: will be judged as making the difference in the presence of abnormal data with the generation data for generating model, positioning residual error is obtained.
Positioning residual error data asked based on two norm threshold tests of given threshold value size with determining problem data position and filtering out
Inscribe data.Its detection follows formula (14).
Wherein: ztotalFor the residual error measurement of data to be tested, estimator and substantial amount measurement summation,It makes a living
The summation that residual error measurement, estimator and the substantial amount of the data to be tested generated at model measure, τ2For constant threshold.
Filter out position when the establishment of (14) formula will then filter out in measurement worst error amount and record, re-start detection until
Detection passes through.
Instance analysis
Be illustrated respectively from following four part: healthy sample acquisition, confrontation generate the training of network, differentiate mould
The accuracy situation of the differentiation accuracy situation of type, abnormal point location.
One, healthy sample acquisition
The sample of example is derived from the metric data of the IEEE118 node system with high international endorsement degree.It executes following
Operate the state estimation health sample that can get 118 node system of single group: by disconnecting the portion in 118 node system metric data
Separated time road simulates the electric system topologies change under normal operation, can be obtained under present topology according to step 1
118 node systems state estimation health sample.It can get 118 groups to 118 node system Data duplication aforesaid operations 118 times
Training sample set.
Two, confrontation generates the training of network
The gap between data and truthful data is mainly generated to assess training by observation in the training process
Quality.As described above, generation model and discrimination model are grown up jointly in confrontation, generate between data and truthful data
Gap reflects the training for generating model, also reflects the training of discrimination model simultaneously.In 118 groups of training samples
It randomly selects in one group of sample input model as generation training, iterates four all ages.Table in the training of 118 node systems
Now it is presented in Fig. 4.Since algebra is excessive, generation data are extracted every 100 generations and are compared with truthful data, the flat of the present age is calculated
Equal error and error variance.Dotted line and solid line two lines respectively represent the mean error and error variance of successive dynasties generation model in figure
Tendency.As can be seen that training process is steady, work well.Average on the way down to close to zero point, final mean error is down to not
To 0.01, and error variance is up to the ten negative biquadratic order of magnitude.Normalized data are mainly distributed between [0.3,0.7],
0.01 less than 0.3 5%, it is seen that 118 node systems are trained successfully.
Three, the differentiation accuracy situation of discrimination model
Table 1 is the confusion matrix of 118 node system discrimination models.93 groups of health samples are usurped with 25 groups based on FDIA principle
The exceptional sample random alignment changed, the detection sample set as discrimination model.Testing result is visualized by confusion matrix.93
The healthy sample of group differentiates that accuracy rate is 100%, and 25 groups of exceptional samples differentiate that accuracys rate are 96%, and comprehensive accuracy rate is up to 99.15%.
Discrimination model is accurately higher.
Table 1
Four, the accuracy situation of abnormal point location
Abnormal point location is carried out to 118 data volumes (linear measurement P) for having FDIA to invade using trained model.
In the case where there is 25 points to attack (25/118) by FDIA, abnormal point positioning result figure is as shown in Figure 5.Detection judges in figure
" 1 ", which represents, has exception, and " 0 " represents without exception.Circle represents the truth of test sample, and starlike point represents positioning result,
Circle is overlapped then positioning with point correctly.If result is shown in, locating accuracy is higher than 90%.
Claims (5)
1. a kind of electric system false data attack recognition method for generating network based on confrontation, which is characterized in that including following
Step:
Step 1: linear state estimation model is established, using the quantity of state of electric system as uncertain collection, in electric system
Telemetering amount, Jacobian matrix and noise modeling；By minimizing, noise can solve estimated state amount and counter estimate counts quantity of state pair
The estimation measurement answered；Estimation measurement is made the difference with substantial amount measurement, obtains residual error measurement；Residual error measurement, estimator
The detected data of FDIA is given over to substantial amount measurement；
Step 2: being filtered out in data by carrying out the two norm threshold tests based on given threshold value size to residual error measurement by object
Manage the bad data that net operation troubles, device measuring error, communication system noise generate, make only to retain in system truthful data with
And FDIA distort after false data；
Step 3: establishing confrontation generates network model, which generates network model and includes discrimination model and generate model, differentiate
Model and generation model are symmetrical one group of three-layer neural network；Net is generated to confrontation with the health data in electric system history
Network model is trained, and is obtained mature discrimination model and is generated model；By the residual error measurement retained in step 1, estimator
It is sent into mature discrimination model and is differentiated with substantial amount measurement, obtain differentiating result；If differentiating, result is that there is no void
False data, which then detects, to be terminated, and otherwise enters step four；
Step 4: will be judged as making the difference in the presence of abnormal data with the generation data for generating model, positioning residual error is obtained；To fixed
Position residual error data is carried out based on two norm threshold tests of given threshold value size to determine problem data position and filter out problem number
According to.
2. the method according to claim 1, wherein the step 1 specifically:
Establish linear state estimation model are as follows:
Z=Hx+e
Wherein, z be telemetering amount, H be representative topology Jacobian matrix, e be state estimation during inevitably generate error with
Disturbance；
The relationship of quantity of state and measurement are as follows:
Qk=0
Pk-m=bk-m(θk-θm)
Qk-m=0
Wherein, PkActive power, Q are injected for node kkReactive power, P are injected for node kk-mFor k node to route between m node
Active power, Qk-mFor k node to the reactive power of route between m node, Bk-mThe imaginary part of (k, m), b in admittance battle arrayk-mFor
Susceptance of the k node to route between m node, θkWith θmThe polar coordinates voltage phase angle of respectively node k and node m；
The objective function of noise least square is established to carry out state estimation, objective function expression formula are as follows:
Min J (x)=[z-Hx]TR-1[z-Hx]
Wherein, R is weight matrix；
Decline principle according to gradient and acquire estimated state amount:
Measurement is carried out according to linear relationship to be back-calculated to obtain estimation measurement after obtaining estimated state amount, Extrapolation follows
Formula are as follows:
Pass through formulaResidual error measurement Δ z is obtained, the measurement of obtained residual error measurement, estimator and substantial amount is stayed
Make the detected data of FDIA.
3. according to the method described in claim 2, it is characterized in that, by carrying out being based on giving to residual error measurement described in step 2
The two norm threshold tests for determining threshold size filter out in data is made an uproar by Physical Network operation troubles, device measuring error, communication system
The bad data that sound generates, detection follow formula are as follows:
Wherein, τ1For constant threshold.
4. according to the method described in claim 3, it is characterized in that, step 3 specifically:
It includes discrimination model and generation model, their objective function that the confrontation, which generates network model, are as follows:
Wherein, formula above is the objective function for generating model in optimization, and formula below is discrimination model in optimization
Objective function；The output result D of discrimination model represents discrimination model to the trusting degree of input data authenticity, generates model
Output result G is the simulation truthful data for generating model production；E is binary matrix, and specific value is according to feeding data source
It determines, to guarantee that the optimization process of separate sources data is independent of each other；
The generation model is all made of with discrimination model containing the reverse transmittance nerve network for intersecting Entropy principle, includes three layers of nerve
Member, two linear plus sigmoid activation primitive mappings, the number of each layer neuron is depending on grid nodes number to be checked, and the
One layer is all input layer, and the last layer is output layer, and intermediate one layer is then hidden layer, and the calculating of forward-propagating follows formula are as follows:
Wherein, anFor n-th layer neuron output as a result, ωnFor weight of the n-th layer neuron into n+1 layers of mapping, bnIt is n-th
Biasing of the layer neuron into n+1 layers of mapping, r are the input of neural network, and s is the output of neural network, and σ represents sigmoid
Activation primitive, calculation method follow formula are as follows:
Wherein, e is natural Exponents；
Output layer error is sought based on Entropy principle is intersected with given optimization aim with the Direct mapping result of neural network, calculating is abided by
Follow formula are as follows:
Again by asking local derviation to obtain updating gradient output layer error, the weight of whole network is reversely updated using gradient decline principle
With biasing, calculation method follows formula are as follows:
Wherein, y is preset optimization aim, and C is output layer error, ρiFor Gradient learning rate；
Mature discrimination model can be obtained by backpropagation training and generate model, the residual error retained in step 1 is measured
Mature discrimination model is sent into amount, estimator and substantial amount measurement can be obtained the differentiation of FDIA presence or absence as a result, if differentiating
As a result for there is no FDIA to distort, differentiation terminates, and distorts then enter the 4th step if it exists.
5. according to the method described in claim 4, it is characterized in that, a pair positioning residual error data described in step 4 is carried out based on given
For two norm threshold tests of threshold size to determine problem data position and filter out problem data, detection follows formula are as follows:
Wherein, ztotalFor the residual error measurement of data to be tested, estimator and substantial amount measurement summation,To generate model
The summation that residual error measurement, estimator and the substantial amount of the data to be tested of generation measure, τ2For constant threshold.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810982950.7A CN109165504B (en) | 2018-08-27 | 2018-08-27 | Power system false data attack identification method based on anti-generation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810982950.7A CN109165504B (en) | 2018-08-27 | 2018-08-27 | Power system false data attack identification method based on anti-generation network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109165504A true CN109165504A (en) | 2019-01-08 |
| CN109165504B CN109165504B (en) | 2021-05-07 |
Family
ID=64896801
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810982950.7A Active CN109165504B (en) | 2018-08-27 | 2018-08-27 | Power system false data attack identification method based on anti-generation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109165504B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110035090A (en) * | 2019-05-10 | 2019-07-19 | 燕山大学 | A kind of smart grid false data detection method for injection attack |
| CN110245302A (en) * | 2019-05-24 | 2019-09-17 | 阿里巴巴集团控股有限公司 | The strategy-generating method and device and electronic equipment of fraud case for identification |
| CN110365647A (en) * | 2019-06-13 | 2019-10-22 | 广东工业大学 | A kind of false data detection method for injection attack based on PCA and BP neural network |
| CN110796237A (en) * | 2019-10-28 | 2020-02-14 | 宁夏吉虎科技有限公司 | Method and device for detecting attack resistance of deep neural network |
| CN110830514A (en) * | 2019-12-12 | 2020-02-21 | 四川大学 | Detection method for collusion-based false data injection attack of smart power grid |
| CN110852597A (en) * | 2019-11-07 | 2020-02-28 | 东南大学 | Electricity consumption peak period resident load ratio calculation method based on generation of countermeasure network |
| CN110995761A (en) * | 2019-12-19 | 2020-04-10 | 长沙理工大学 | Method and device for detecting false data injection attack and readable storage medium |
| CN112241532A (en) * | 2020-09-17 | 2021-01-19 | 北京科技大学 | Method for generating and detecting malignant confrontation sample based on jacobian matrix |
| CN112861461A (en) * | 2021-03-05 | 2021-05-28 | 北京华大九天科技股份有限公司 | Abnormity detection method and device for circuit simulation model |
| CN113268729A (en) * | 2021-05-01 | 2021-08-17 | 群智未来人工智能科技研究院(无锡)有限公司 | Smart grid attack positioning method based on convolutional neural network |
| CN114330486A (en) * | 2021-11-18 | 2022-04-12 | 河海大学 | Power system bad data identification method based on improved Wasserstein GAN |
| CN114978586A (en) * | 2022-04-12 | 2022-08-30 | 东北电力大学 | Power grid attack detection method and system based on attack genes and electronic equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127047A (en) * | 2016-06-24 | 2016-11-16 | 河海大学 | A kind of power system malicious data detection method based on Jensen Shannon distance |
| CN106375344A (en) * | 2016-11-02 | 2017-02-01 | 华北电力大学（保定） | Intelligent grid load integrity attack detection method for cloud storage |
| US20180176249A1 (en) * | 2016-12-21 | 2018-06-21 | Abb Inc. | System and method for detecting false data injection in electrical substations |
-
2018
- 2018-08-27 CN CN201810982950.7A patent/CN109165504B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127047A (en) * | 2016-06-24 | 2016-11-16 | 河海大学 | A kind of power system malicious data detection method based on Jensen Shannon distance |
| CN106375344A (en) * | 2016-11-02 | 2017-02-01 | 华北电力大学（保定） | Intelligent grid load integrity attack detection method for cloud storage |
| US20180176249A1 (en) * | 2016-12-21 | 2018-06-21 | Abb Inc. | System and method for detecting false data injection in electrical substations |
Non-Patent Citations (5)
| Title |
|---|
| MD. ASHFAQUR RAHMAN等: "False data injection attacks with incomplete information against smart power grids", 《2012 IEEE GLOBAL COMMUNICATIONS CONFERENCE》 * |
| XUAN LIU等: "Modeling of Local False Data Injection Attacks With Reduced Network Information", 《IEEE TRANSACTIONS ON SMART GRID》 * |
| 何耀等: "基于扩展卡尔曼滤波的虚假数据攻击检测方法", 《中国电力》 * |
| 李雪等: "基于快速回归算法的虚假数据攻击构造新方法", 《仪器仪表学报》 * |
| 王先培等: "输电网虚假数据攻击研究综述", 《电网技术》 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110035090A (en) * | 2019-05-10 | 2019-07-19 | 燕山大学 | A kind of smart grid false data detection method for injection attack |
| CN110035090B (en) * | 2019-05-10 | 2020-09-15 | 燕山大学 | False data injection attack detection method for smart grid |
| CN110245302A (en) * | 2019-05-24 | 2019-09-17 | 阿里巴巴集团控股有限公司 | The strategy-generating method and device and electronic equipment of fraud case for identification |
| CN110245302B (en) * | 2019-05-24 | 2023-08-08 | 创新先进技术有限公司 | Policy generation method and device for identifying fraudulent cases and electronic equipment |
| CN110365647A (en) * | 2019-06-13 | 2019-10-22 | 广东工业大学 | A kind of false data detection method for injection attack based on PCA and BP neural network |
| CN110365647B (en) * | 2019-06-13 | 2021-09-14 | 广东工业大学 | False data injection attack detection method based on PCA and BP neural network |
| CN110796237A (en) * | 2019-10-28 | 2020-02-14 | 宁夏吉虎科技有限公司 | Method and device for detecting attack resistance of deep neural network |
| CN110796237B (en) * | 2019-10-28 | 2023-04-07 | 宁夏吉虎科技有限公司 | Method and device for detecting attack resistance of deep neural network |
| CN110852597A (en) * | 2019-11-07 | 2020-02-28 | 东南大学 | Electricity consumption peak period resident load ratio calculation method based on generation of countermeasure network |
| CN110852597B (en) * | 2019-11-07 | 2022-01-28 | 国网江西省电力有限公司电力科学研究院 | Electricity consumption peak period resident load ratio calculation method based on generation of countermeasure network |
| CN110830514B (en) * | 2019-12-12 | 2021-06-22 | 四川大学 | Detection method for collusion-based false data injection attack of smart power grid |
| CN110830514A (en) * | 2019-12-12 | 2020-02-21 | 四川大学 | Detection method for collusion-based false data injection attack of smart power grid |
| CN110995761B (en) * | 2019-12-19 | 2021-07-13 | 长沙理工大学 | Method and device for detecting false data injection attack and readable storage medium |
| CN110995761A (en) * | 2019-12-19 | 2020-04-10 | 长沙理工大学 | Method and device for detecting false data injection attack and readable storage medium |
| CN112241532A (en) * | 2020-09-17 | 2021-01-19 | 北京科技大学 | Method for generating and detecting malignant confrontation sample based on jacobian matrix |
| CN112241532B (en) * | 2020-09-17 | 2024-02-20 | 北京科技大学 | Method for generating and detecting malignant countermeasure sample based on jacobian matrix |
| CN112861461A (en) * | 2021-03-05 | 2021-05-28 | 北京华大九天科技股份有限公司 | Abnormity detection method and device for circuit simulation model |
| CN112861461B (en) * | 2021-03-05 | 2022-05-17 | 北京华大九天科技股份有限公司 | Abnormity detection method and device for circuit simulation model |
| CN113268729A (en) * | 2021-05-01 | 2021-08-17 | 群智未来人工智能科技研究院(无锡)有限公司 | Smart grid attack positioning method based on convolutional neural network |
| CN114330486A (en) * | 2021-11-18 | 2022-04-12 | 河海大学 | Power system bad data identification method based on improved Wasserstein GAN |
| CN114978586A (en) * | 2022-04-12 | 2022-08-30 | 东北电力大学 | Power grid attack detection method and system based on attack genes and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109165504B (en) | 2021-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109165504A (en) | A kind of electric system false data attack recognition method generating network based on confrontation | |
| CN110415215A (en) | Intelligent detecting method based on figure neural network | |
| CN109767054A (en) | Efficiency cloud appraisal procedure and edge efficiency gateway based on deep neural network algorithm | |
| WO2022021726A1 (en) | Pmu-based power system state estimation performance evaluation method | |
| CN106443566B (en) | The electric energy metering device error detection method and system deduced based on big data | |
| CN109816031A (en) | A kind of Transformer State Assessment clustering method based on the unbalanced measurement of data | |
| CN109766952A (en) | Photovoltaic array fault detection method based on Partial Least Squares and extreme learning machine | |
| CN105866725A (en) | Method for fault classification of smart electric meter based on cluster analysis and cloud model | |
| CN109597396B (en) | A kind of distribution transforming on-line fault diagnosis method based on high amount of traffic and transfer learning | |
| CN110289613A (en) | The identification of distribution net topology and line parameter circuit value discrimination method based on sensitivity matrix | |
| CN110231447A (en) | The method, apparatus and terminal device of water quality abnormality detection | |
| CN102364490B (en) | Automatic synchronization recognition method based on hierarchical analyzing model | |
| CN109376366A (en) | A kind of analysis method and device of exception line loss influence factor | |
| CN114720819A (en) | Fault section binary positioning method based on self-checking learning | |
| CN112202630A (en) | Network quality abnormity detection method and device based on unsupervised model | |
| CN114330486A (en) | Power system bad data identification method based on improved Wasserstein GAN | |
| CN110889207A (en) | System combination model credibility intelligent evaluation method based on deep learning | |
| CN116466408B (en) | Artificial neural network superbedrock identification method based on aeromagnetic data | |
| CN108845217A (en) | Bushing shell for transformer monitoring method based on FUZZY NETWORK fusion | |
| CN114779029A (en) | CVT internal insulation online evaluation method fusing group redundancy association and structural parameters | |
| CN113420803A (en) | Multi-detector combined fire alarm determination method suitable for transformer substation | |
| Ren et al. | Fault Diagnosis of UHVDC Transmission System Based on Gated Recurrent Unit | |
| Jiang et al. | Distribution transformers connectivity recognition based on multi-layer perceptron classifier | |
| CN113009107A (en) | Method for detecting unfrozen water content of frozen soil | |
| Dobrić et al. | Non-Technical losses detection in power system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |