CN106597178A - ANFIS digital pre-distorting method of relay protection test device LPA - Google Patents
ANFIS digital pre-distorting method of relay protection test device LPA Download PDFInfo
- Publication number
- CN106597178A CN106597178A CN201710035791.5A CN201710035791A CN106597178A CN 106597178 A CN106597178 A CN 106597178A CN 201710035791 A CN201710035791 A CN 201710035791A CN 106597178 A CN106597178 A CN 106597178A
- Authority
- CN
- China
- Prior art keywords
- anfis
- lpa
- abnormal
- relay protection
- current
- 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
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/10—Numerical modelling
Abstract
The invention discloses an ANFIS (Adaptive Neuro-Fuzzy Inference System) digital pre-distorting method of a relay protection test device LPA (Linear Power Amplifier). An ANFIS after training and learning approaches a nonlinear inverse digital model of a relay protection test device LPA on a compact set, and replaces the inverse digital model to produce an instruction voltage/current corresponding to an ideal output voltage/current to realize digital pre-distorting. The method considers all possible nonlinearity of the LPA from an analog-to-digital converter to each amplifier stage to a load on the whole, and no longer requires rigorous precision of each amplifier stage. A quasi-online mode is adopted, digital pre-distorting can be realized on most of existing microprocessors, and the cost performance is improved. With the help of the adaptive ability of ANFIS, the method can adapt to loads of different features and values. The method not only can meet the precision requirement of fault waveform recurrence, but also can make full use of the limited bandwidth and output capacity of LPA.
Description
Technical field
The invention belongs to technical field of relay protection, more particularly to a kind of test device for relay protection linear power amplifier
(LPA) the pre- abnormal method of ANFIS numerals.
Background technology
Test device for relay protection has become the indispensable instrument of verification relay protection and automatic safety device of adjusting.For
The action details of relay protection and automatic safety device under power system true fault, modern test can be meticulously analyzed comprehensively
It is required that test device for relay protection should be able to relatively accurately export various false voltage/current waveforms, Wave data can derive from soft
Part is emulated or fault oscillograph.On the one hand, because relay protection and automatic safety device are connected to the secondary of potential and current transformers
Side, therefore the horizontal summation current transformer of secondary voltage that voltage and current must be amplified to voltage transformer by test device for relay protection
Secondary current level, the voltage and current at " scene " could be provided for tested device.Voltage transformer secondary rated voltage one
As be 100V (virtual value, root mean square value, RMS);To grassroot project, it is reduce loss, current transformer
Secondary rated current typically selects 1A (RMS).Under fault transient, the primitive period component of false voltage is up to the 5 of rated voltage
Again, 30 times up to rated current of the primitive period component of fault current, it is considered to aperiodic component harmonic, total voltage and total electricity
Stream can also be higher.Therefore voltage and current at least should be able to be amplified to test device for relay protection 500V (RMS) and 30A (RMS), i.e.,
Voltage dynamic range is at least 0~5 (perunit value, per unit, pu), and current dynamic range is at least 0~30 (pu).The opposing party
Face, is the test for meeting fault location device, harmonic source monitoring positioning device, active filter and SVC etc.,
It is required that the bandwidth of output waveform is at least 0~10kHz.According to sampling thheorem, it is desirable to the frequency band of test device for relay protection
Width is at least 0~20kHz.To sum up, test device for relay protection of new generation requires to realize while wide dynamic range is realized
Broadband, this is always the technological difficulties of test device for relay protection Designing power amplifier, causes when fault waveform reappears
Wave distortion.Because current design criteria was formulated before 10 years, there are fault waveform producing contents to be added, therefore existing product is thought to receive
Operation constraint, it is difficult for fear of realizing, it is often empty to claim faulty wave recurrence function.To overcome the technological difficulties, most relay protections
Test device is tended to adopt LPA, it is believed that the control of LPA is easy compared with switching power amplifier, and opened loop control precision even can
Match in excellence or beauty with the closed-loop control precision of switching power amplifier.This is the misunderstanding that LPA " linear " is caused to designer.Actually
" linear " of LPA is a relative concept, and no LPA can be completely eliminated non-linear.Reason is to obtain sufficiently large putting
Big multiple, LPA need to be made up of the cascade of some amplifier stages.Each amplifier stage can be considered a LPA, typically by a differential amplifier
Realize.The non-linear of one differential amplifier is easily controlled, however, after several cascade, the non-linear of each can be by step by step
Amplify simultaneously superposed on one another, eventually become one it is uppity non-linear.In theory, be reduce cascade after it is non-linear, each
The hardware of amplifier stage must connect circuit and must also optimize between sophisticated design and adjustment, and power supply and level, can adopt following methods:①
Use dynamic switch power supply instead to be connected with LPA, by specific modulation reduce Traditional DC power-supply fluctuation cause it is non-linear;2. it is every
Individual amplifier stage design simulation feed forward circuit, corrects the non-linear of each amplifier stage in advance;3. by optimize and revise peak compensation line and
Connection capacitor and switch bias realize the linearisation of LPA, and improve the linearity using backward distortion compensation;4. cross-couplings are adopted
Double deference to and two linearisation measures of source-degeneration resistance realize LPA High Linear response.High line acquired by these methods
Property degree is, to increase the complexity of hardware as cost, and thus may to bring the reduction of reliability, and the transplantation of method
Property is poor.It is different from hardware above amendment or compensation method, it is to be no longer limited to improve each portions of LPA the characteristics of digital pre- abnormal method
The precision of part, normally only needs to change software, and need not change hardware circuit, you can realize the lifting of precision.It is current to realize numeral
Pre- abnormal method mainly has:1. LPA is modeled using memoryless multinomial, polynomial coefficient is true by recurrent least square method
It is fixed, it is pre- abnormal by model realization numeral;2. LPA is modeled using Wiener wave filter, and using some piecewise linear functions to non-
Linearly estimated, it is pre- abnormal that both are implemented in combination with numeral;3. using the linearization technique that can be combined with iteration noise cancellation technology, eliminate
The impact of additive noise, makes the pre- abnormal device of numeral be independent of the accurate model of LPA.These pre- abnormal methods of numeral have preferable transplantation
Property, but set up on the basis of the linear math instrument such as multinomial, linear filter and piecewise linear function, for estimating non-thread
Property, it is possible in theory, and can be limited by exponent number in practicality, precision is difficult to improve.Additionally, in estimation or adjusting parameter and coefficient
When, numerical stability is difficult to ensure that sometimes, may cause system unstability.In view of being limited by rated capacity, the output of LPA is limited
It from from the point of view of functional analysis, can be defined as a unknown nonlinear in compact set in a limited range by system
Functional.Theoretical proof, ANFIS Jing finite number of time training and study after, any non-linear letter that infinitely can be approached in compact set
Number.Therefore the more aforementioned linear math instruments of ANFIS are more suitable for Nonlinear Modeling.Additionally, provable, when adopting gradient descent method
When the hybrid learning algorithm combined with method of least square replaces single gradient descent method, there would not be numerical stability and ask
Topic.
The content of the invention
It is an object of the invention to a kind of pre- abnormal method of ANFIS numerals of relay protection test test device LPA is proposed, with
Solve test device for relay protection of new generation to require wide band technical problem is realized while wide dynamic range is realized.
What the pre- abnormal method of ANFIS numerals of relay protection test test device LPA of the present invention was realized in:Utilize
ANFIS removes the nonlinear inverse mathematical model of test device for relay protection LPA approached in compact set after training and learning,
Replace inverse number word model to produce the command voltage/electric current corresponding with desired output voltage/electric current, realize that numeral is pre- abnormal.
Further, the realization of the pre- abnormal method of ANFIS numerals of relay protection test test device LPA includes following step
Suddenly:
1) the input/output data of LPA inverse number word models are trained into number to the input/output data pair as ANFIS
According to;
2) type and number of membership function are chosen;
3) generate the initiating structure of ANFIS;
4) training parameter is selected, i.e., maximum training batch, initial step length and target error;
5) training is completed by hybrid learning algorithm, generate the pre- abnormal device of ANFIS numerals;
6) pre- abnormal command voltage/electric current is produced with the pre- abnormal device of desired output voltage/electric current Jing ANFIS numerals;
7) the amplification output of voltage/current is realized with pre- abnormal command voltage/current control LPA;
8) back production the RMSE of actual output voltage/electric current and desired output voltage/electric current is calculated, confirms that ANFIS is digital
The performance of pre- abnormal device.
Further, the establishment step of the LPA mathematical models is as follows:
1) by actual test wiring, test device for relay protection is connected with tested relay protection and automatic safety device;
2) sinusoidal voltage and sinusoidal current that LPA output frequencies are 10kHz are controlled, smooth adjustment its amplitude is allowed to from 0 increase
Arrive greatly maximum;
3) sampling, the voltage of record gradual change, current amplitude;
4) amplitude by the amplitude of command signal with output voltage, electric current is arranged in input/output data correspondingly
It is right.
Further, the formulation step of the hybrid learning algorithm is as follows:
1) it is learning target to choose minimum RMSE:
In formula:dmExport for preferable, | | | |2Represent 2- norms;
2) coefficient of backward network is adjusted by following learning algorithm:
In formula:K=0,1,2 ... represents kth time adjustment, β>0 is learning rate;
3) will Jing after 1 adjustmentIt is temporarily fixed, next adjust the parameter of feedforward network, different membership functions
There are different parameters, without loss of generality, ifThere are 2 parameters aijAnd bij, its learning algorithm is:
In formula:L=0,1,2 ... represents the l time adjustment, γ>0 is learning rate;
4) useaijAnd b (l+1)ij(l+1) matrix A is constituted with input data, use αjComposition unknown vector α, uses
dmPreferable output vector d of composition, obtains overdetermination matrix equation A α=d;With Least Square Method optimal solution α, a most young waiter in a wineshop or an inn can be obtained
Take advantage of solution:
αLS=(ATA)-1ATd。
Another object of the present invention is to provide a kind of relay protection test dress using the pre- abnormal method of ANFIS numerals
Put.
The ANFIS pre- abnormal methods of numeral of test device for relay protection LPA proposed by the present invention, overall thinking LPA
From digital to analog converter to each amplifier stage (including DC source) again to all possible non-linear, the essence to each amplifier stage of load
Degree is no longer made excessive demands so that only need to change software, you can traditional relay protection test device is upgraded to device of new generation, can not only
Meet the required precision of fault waveform reproduction, also make limited bandwidth and output capacity be fully used;Meanwhile, also make
Universal LPA is in use to device of new generation from traditional relay protection test device and is possibly realized, because universal LPA technology maturations,
Reliability is high and cost is relatively low, therefore can keep the cost performance of device of new generation.The mathematical model of LPA can be filled by relay protection test
Put and self testing record before output false voltage/electric current, and the training and study for the pre- abnormal device of ANFIS numerals leaves enough
Time, this is quasi- online mode;Using quasi- online mode, the pre- abnormal device of ANFIS numerals can be realized on original microprocessor, therefore
High speed microprocessor need not be changed to meet requirement of real-time control, the cost performance of device is thus further improved.Additionally, by means of
The adaptive ability of ANFIS, the present invention can also adapt to load (i.e. different relay protections and the safety of different qualities and numerical value automatically
The input characteristics of automaton).
Description of the drawings
Fig. 1 is the realization stream of the pre- abnormal method of ANFIS numerals of test device for relay protection LPA provided in an embodiment of the present invention
Cheng Tu.
Fig. 2 is the generic structure diagram of ANFIS provided in an embodiment of the present invention.
Fig. 3 is the positive digital model curve figure of current mode LPA provided in an embodiment of the present invention.
Fig. 4 is the inverse number word model curve chart of current mode LPA provided in an embodiment of the present invention.
Fig. 5 is simulated fault current curve diagram provided in an embodiment of the present invention.
Fig. 6 (a) is training provided in an embodiment of the present invention and the shape graph for learning front membership function.
Fig. 6 (b) is the shape graph of membership function after training provided in an embodiment of the present invention and study.
Fig. 7 is simulation example provided in an embodiment of the present invention without pre- abnormal actual output current curve chart.
Fig. 8 is the pre- abnormal instruction current curve chart of simulation example provided in an embodiment of the present invention.
Fig. 9 is the true fault current curve diagram of fault oscillograph record provided in an embodiment of the present invention.
Figure 10 is true fault electric current provided in an embodiment of the present invention without pre- abnormal actual output current curve chart.
Figure 11 is the pre- abnormal instruction current curve chart of true fault electric current provided in an embodiment of the present invention.
Specific embodiment
To make the objects, technical solutions and advantages of the present invention become apparent from understanding, below in conjunction with specific embodiment, to this
It is bright to be described in detail.It is noted that particular embodiments described herein is not intended to limit this only to explain the present invention
It is bright.
The invention discloses a kind of test device for relay protection linear power amplifier (linear power
Amplifier, LPA) adaptive neural network-fuzzy inference system (adaptive neuro-fuzzy inference
System, ANFIS) the pre- abnormal method of numeral.The present invention using ANFIS by training and study after, go to approach in compact set after
The nonlinear inverse mathematical model of electric protection test device LPA, replaces inverse number word model to produce relative with desired output voltage/electric current
Command voltage/the electric current answered, realizes that numeral is pre- abnormal.Overall thinking of the present invention LPA is from digital to analog converter to each amplifier stage
(including DC source) arrives all possible non-linear of load again, and the precision of each amplifier stage is no longer made excessive demands.This makes universal
LPA is in use to device of new generation from traditional relay protection test device and is possibly realized, because of universal LPA technology maturations, reliability
It is high and cost is relatively low, therefore the cost performance of device of new generation can be kept.Using quasi- online mode, the present invention can be original in major part
Realize on microprocessor, therefore high speed microprocessor need not be changed to meet requirement of real-time control, thus further improve device
Cost performance.By means of the adaptive ability of ANFIS, the present invention can also adapt to different qualities automatically and the load of numerical value is (i.e. different
Relay protection and the input characteristics of automatic safety device).The present invention can not only meet the required precision of fault waveform reproduction, may be used also
The bandwidth and output capacity for making LPA limited is fully used so that only need to change software, you can by traditional relay protection
Test device upgrades to device of new generation.
The present invention using ANFIS by training and study after, the test device for relay protection LPA's for going to approach in compact set
Nonlinear inverse mathematical model, so as to replace inverse number word model to produce the command voltage/electricity corresponding with desired output voltage/electric current
Stream, is achieved that numeral is pre- abnormal.
Below in conjunction with the accompanying drawings the application principle of the present invention is described in detail.
As shown in figure 1, the pre- abnormal method of ANFIS numerals of test device for relay protection LPA provided in an embodiment of the present invention
Realization is comprised the following steps:
S101:By the output of linear power amplifier (linear power amplifier, LPA) inverse number word model/defeated
Enter data to as adaptive neural network-fuzzy inference system (adaptive neuro-fuzzy inference system,
ANFIS input/output data pair), i.e. training data;
S102:Choose the type and number of membership function;
S103:Generate the initiating structure of ANFIS;
S104:Training parameter is selected, i.e., maximum training batch, initial step length and target error;
S105:Training is completed by hybrid learning algorithm, the pre- abnormal device of ANFIS numerals is generated;
S106:Pre- abnormal command voltage/electric current is produced with the pre- abnormal device of desired output voltage/electric current Jing ANFIS numerals;
S107:The amplification output of voltage/current is realized with pre- abnormal command voltage/current control LPA;
S108:Back production simultaneously calculates the root-mean-square error (root of actual output voltage/electric current and desired output voltage/electric current
Mean square error, RMSE), confirm the effectiveness of the pre- abnormal method of ANFIS numerals.
Below in conjunction with the accompanying drawings the application principle of the present invention is further described.
1st, design the universal architecture of ANFIS.
Because Takagi-Sugeno (T-S) fuzzy model is easy to analyze and calculates, therefore setting up ANFIS in T-S fuzzy models more
On the basis of.If input vector x=[x1,x2,…,xM]T, wherein xiFor Fuzzy Linguistic Variable, dimensions of the M for x, T represent that vector turns
Put.If:
In formula:For xiJ-th fuzzy language value (fuzzy set), NiFor fuzzy language value
Number.DefinitionMembership function beThen T-S fuzzy rules are:
then yj=pj0+pj1x1+…+pjMxM;
In formula:J is expanded, j=1,2 ..., N,pj0,pj1,…,pjMFor xiLinear combination is
Number.By defined above, the universal architecture of ANFIS can be designed as shown in Fig. 2 which is by feedforward network and backward network two parts group
Into.
Feedforward network is constituted by 4 layers for being matched with the former piece of T-S fuzzy rules.1st layer is input layer, directly will be each
xiNext layer is sent to, therefore the layer has M node.2nd layer 1 fuzzy language value of each node on behalf, function is to pass throughCalculate xiIt is rightDegree of membership, the layer hasIndividual node.The 1 of 3rd layer of each 1 article of fuzzy rule of node on behalf
Individual former piece, function are the relevance grades for calculating every fuzzy rule:
Or
The layer has Q node.4th layer of function is the weight for calculating normalized i.e. every fuzzy rule of relevance grade:
The layer also has Q node.
Backward network is constituted by 3 layers for being matched with the consequent of T-S fuzzy rules.1st layer is input layer, directly will be each
xiIt is sent to next layer, wherein x0=1 is used to produce y in following formulamjConstant termThe layer can arrange N≤M as needed
Node.If ANFIS has R output, the output of every fuzzy rule of the 2nd layer of calculating:
In formula:M=1,2 ..., R, the layer have R × N number of node.3rd layer is output layer, and function is that the output to the 2nd layer is entered
Row is comprehensive:
The layer has R node.
2nd, formulate hybrid learning algorithm.
The pre- abnormal method of ANFIS numerals is using the hybrid learning algorithm with good numerical stability and fast convergence.Choosing
Minimum RMSE is taken for learning target:
In formula:dmExport for preferable, | | | |2Represent 2- norms.It is theoretical by back propagation, first by following learning algorithm
Adjust the coefficient of backward network:
In formula:K=0,1,2 ... represents kth time adjustment, β>0 is learning rate.Jing after 1 adjustment,Temporarily will fix,
Next adjust the parameter of feedforward network.Different membership functions has different parameters, without loss of generality, it is assumed thatHave
2 parameters aijAnd bij, its learning algorithm is:
In formula:L=0,1,2 ... represents the l time adjustment, γ>0 is learning rate.Under the classical gradient of 3 formula of above composition
Drop method.Now, useaijAnd b (l+1)ij(l+1) matrix A is constituted with input data, use αjComposition unknown vector α, uses
dmPreferable output vector d of composition, then obtain overdetermination matrix equation A α=d.With Least Square Method optimal solution α, can obtain minimum
Two take advantage of solution:
αLS=(ATA)-1ATd;
Thus just realize hybrid learning algorithm.
3rd, set up the mathematical model of LPA.
Mathematical model is different from mathematical model.The former is obtained by testing, and using input/output data to representing, is most pasted
Nearly practical situation;The latter is obtained by theoretical derivation by introducing certain approximate condition, can be represented with mathematical formulae, but may be with
There is error in practical situation.To obtain high accuracy, the mathematical model of LPA should be adopted.
The mathematical model of LPA is set up by following steps:
1) by actual test wiring, test device for relay protection is connected with tested relay protection and automatic safety device;
2) sinusoidal voltage and sinusoidal current that LPA output frequencies are 10kHz are controlled, smooth adjustment its amplitude is allowed to from 0 increase
It is big that to maximum, (voltage isElectric current is);
3) sampling, the voltage of record gradual change, current amplitude;
4) amplitude by the amplitude of command signal with output voltage, electric current is arranged in input/output data correspondingly
It is right.
When one timing of output waveform amplitude, the non-linear of LPA can increase with the increase of output waveform frequency, and failure is electric
The frequency of pressure/current harmonics component is usually no more than 10kHz, therefore mathematical model compatibility below the 10kHz measured under 10kHz
Situation.
Simply can exchange in view of the control method of voltage-type LPA and current mode LPA, and current mode LPA it is non-linear compared with
Significantly, current precision control is difficult compared with voltage accuracy control, therefore only illustrates by taking current mode LPA as an example below for voltage-type LPA.
The mathematical model of certain current mode LPA obtained by above flow process can with input data as transverse axis, output data is plotted as the longitudinal axis
Curve as shown in Figure 3.It can be seen that, although calculate by rated current 30A (RMS), maximum output current is up to 42A, but reality
Linearity range has been compressed to -30~30A because non-linear, and positive curve and negative sense curve are not origin symmetries, be have lost near
The resource of 10A.The mathematical model of current mode LPA, from from the point of view of function, is common " input → output " mapping, i.e., just reflects
Penetrate, therefore referred to as positive number word model.
For positive number word model, when the amplification of instruction current is more than 30 times, output current will just produce non-linear
Distortion.Obviously, in the ideal situation, i.e. in the state of LPA remains linear, when amplification reaches 40 times of (specified amplifications
Multiple) when, should not also there is non-linear distortion.To obtain desired output current, can be inverse mapping by positive mapping transformation, i.e., it is " defeated
Go out → be input into " mapping, correspondingly by positive digital model conversion be inverse number word model.Tried to achieve and preferable output electricity by inverse number word model
The instruction current can be obtained desired output current by positive number word model by the corresponding instruction current of stream.Inverse number word model
Can be by input/output data be changed into input/output data to obtaining to simple switched position, the sign on curve is exactly
By the horizontally and vertically exchange of Fig. 3, as shown in Figure 4.
4th, the pre- abnormal device of ANFIS numerals is realized by training and study.
Using ANFIS after training and learning, the test device for relay protection LPA's for going to approach in compact set is non-linear
Inverse number word model, so as to replace inverse number word model to produce the command voltage/electric current corresponding with desired output voltage/electric current, just
Realize numeral pre- abnormal.Why it is referred to as numeral pre- abnormal, is because that the instruction current corresponding with desired output current is different from
Former instruction current, sees to seem that former instruction current there occurs distortion in appearance, and is occur before into current mode LPA in advance.
Thus, the instruction current corresponding with desired output current is also referred to as pre- abnormal instruction current.By pre- abnormal instruction current input current
Type LPA will obtain desired output current.
The training of the ANFIS pre- abnormal devices of numeral, study and use (i.e. the realization of the pre- abnormal method of ANFIS numerals) step as follows:
1) the input/output data of LPA inverse number word models are trained into number to the input/output data pair as ANFIS
According to;
2) type and number of membership function are chosen;
3) generate the initiating structure of ANFIS;
4) training parameter is selected, i.e., maximum training batch, initial step length and target error;
5) training is completed by hybrid learning algorithm, generate the pre- abnormal device of ANFIS numerals;
6) pre- abnormal command voltage/electric current is produced with the pre- abnormal device of desired output voltage/electric current Jing ANFIS numerals;
7) the amplification output of voltage/current is realized with pre- abnormal command voltage/current control LPA;
8) back production the RMSE of actual output voltage/electric current and desired output voltage/electric current is calculated, confirms that ANFIS is digital
The effectiveness of pre- abnormal method.
The mathematical model of LPA self testing can be recorded before output false voltage/electric current by test device for relay protection,
And the training and study for the pre- abnormal device of ANFIS numerals leaves enough time, this is quasi- online mode.Using quasi- online mode,
The pre- abnormal device of ANFIS numerals can be realized on most of original microprocessor, without the need for upgrading.
5th, the parameter of the pre- abnormal device of preferred ANFIS numerals.
Performance to make the pre- abnormal device of ANFIS numerals is optimal, the type and number of membership function, maximum training batch
Secondary, initial step length and target error need Jing emulation experiments to carry out preferably.Below with simulated fault electric current shown in Fig. 5 as example pair
The parameter of the pre- abnormal device of ANFIS numerals is carried out preferably, and the simulated fault electric current contains the up to harmonic component of 9.75kHz, can be used for
Traveling wave is tested.
1) type of membership function.Because simulated fault electric current be it is ambipolar, i.e., it is positive and negative, therefore should select bilateral
Membership function, should not select monolateral membership function, and otherwise training will error.Secondly, in bilateral membership function, also not
Linear section of membership function can be selected, because ANFIS is not intended to " being/non-(1/0) " judgement, but is used for output waveform
The control of precision, therefore should be using continuous curved membership function.By upper, membership function (should be divided in double sigmoid functions
Difference function and Product function), joint Gaussian function, single Gaussian function, select in generalized bell function and Π shape functions.Shown in following table
Fix for remaining parameter (membership function number 6, maximum training batch 100, initial step length 0.1, target error 0), using not
During same type membership function, the RMSE of actual output current and desired output current.It can be seen that, generalized bell function is corresponding
RMSE is minimum, therefore selects the type of membership function to be generalized bell function.
2) number of membership function.Following table show remaining parameter and fixes (generalized bell membership degree function, maximum training
Batch 100, initial step length 0.1, target error 0), different membership function number when, actual output current and preferable output electricity
The RMSE of stream.It can be seen that, when just starting (2~3), RMSE constantly declines with the increase of membership function number, illustrates to increase ANFIS
The complexity of structure is conducive to improving precision;But when membership function number increases to after 4, RMSE not only no longer declines anti-
And gradually increase, illustrating to occur in that " over-fitting " phenomenon, this needs is avoided.In addition, when membership function number is 3
RMSE is minimum, therefore membership function number should be selected to be 3.Only 3 membership functions, the structure of ANFIS are very simple, training and
Learning time greatly shortens, highly beneficial to real-time control.
3) maximum training batch.Following table show remaining parameter and fixes (generalized bell membership degree function, number 3, initial step
It is long by 0.1, target error 0), different maximum training batches when, the RMSE of actual output current and desired output current.It can be seen that, and
Non- training batch is the bigger the better, and when training batch is reached after 100 times, is further added by training batch, and the difference of RMSE is obvious,
After illustrating training 100 times, ANFIS reliable conveyances.This is also demonstrated using after hybrid learning algorithm, and ANFIS has quick receipts
Holding back property.Training time is more short better, therefore selects maximum training batch to be 100 optimums.
4) initial step length.Following table show remaining parameter and fixes (generalized bell membership degree function, number 3, maximum training batch
Secondary 100, target error 0), different initial step length when, the RMSE of actual output current and desired output current.It can be seen that, it is initial to walk
Length is too big or too little all bad, and optimal values between 0.01~0.1, are easy, and can use initial step length is 0.1.
5) target error.Using in, it is desirable to which target error RMSE is the smaller the better, therefore target error should be set as 0, and in practice
It is only capable of being close to 0, it is impossible to equal to 0.
To sum up, the preferred parameter of the pre- abnormal device of ANFIS numerals is:Generalized bell membership degree function, number 3, maximum training batch
Secondary 100, initial step length 0.1, target error 0.
6th, embodiment.
The application effect of the present invention is described in detail with reference to emulation and test example.
1) emulation embodiment.
The simulation experiment result using example shown in Fig. 5 during preferred parameter is given herein.It is pre- that Fig. 6 show ANFIS numerals
Before and after abnormal device is trained and is learnt to the inverse number word model of current mode LPA, the situation of change of membership function.It can be seen that, generalized bell
Membership function is stretched in bilateral scope or compresses continuously gentle compared with other type membership functions, therefore precision is higher.
By the simulated fault electric current of example divided by preferable amplification 40, obtain normalized instruction current and (be input into electricity
Stream perunit value), passed through the positive number word model of current mode LPA, be obtained without the pre- abnormal actual output current of numeral, such as scheme
Shown in 7.Contrast Fig. 5 it is visible, the positive waveform of the actual output current occur in that cut push up and burr is floating, illustrate because of times magnification
Number more than 30 times and contains higher hamonic wave, current mode LPA it is non-linear very significantly, cause amplitude distortion and frequency distortion.
Quantitative Analysis show that actual output current is up to 11.0199 with the RMSE of desired output current.
To eliminate the non-linear distortion for causing of current mode LPA, it would be desirable to which output current is obtained by the pre- abnormal device of ANFIS numerals
Pre- abnormal instruction current, as shown in Figure 8.Jing numerals are obtained final product by the positive number word model of current mode LPA by pre- abnormal instruction current pre-
Abnormal actual output current, it is almost identical with Fig. 5 because of which, only it is difficult to tell difference from visual angle, therefore no longer provides song
Line chart.It is computed, at this moment actual output current only has 0.0311 with the RMSE of desired output current, quantitatively illustrates that both are non-
Very close to.
The simulation experiment result shows that the pre- abnormal method of ANFIS numerals preferably eliminates the non-linear of current mode LPA, it is ensured that
The precision of actual output current, makes the non-linear section of current mode LPA also be utilized.
2) test example.
The mathematical model of lifted current mode LPA of the present invention derives from a universal test device for relay protection product, its skill
Art parameter is:Rated capacity 2kVA, 0~30A of output-current rating (RMS);When output current is less than 0.2A (RMS), definitely by mistake
Difference is less than 0.4mA (RMS);At 0.2~30A (RMS), relative error is less than 0.2% to output current;Rated output frequency 0~
2.5kHz;Specified amplification 40.It may be noted that these technical parameters are only capable of meeting current design criteria《DL/T 995-2006 relays are protected
Shield and power grid security automaton inspection procedure》The requirement of the traditional test of regulation.And before the code is published on 10 years, lay particular emphasis on
Power frequency measures examination, needs to be supplemented current new test content and requirement.Therefore above technical parameter be to power frequency amount i.e. 50Hz just
For string electric current.So surface is seen, precision is very high, but this is only the precision under 50Hz sinusoidal currents, if using which
The fault current of output harmonic wave complicated component, amplification more than 30 times, then gained precision will be far from each other.Additionally, smart above
Degree index (absolute error and relative error) is only capable of the virtual value of simply estimation of sinusoidal electric current, it is impossible to is deep into inside waveform and goes
Point-to-point error is assessed, the fault current of the harmony wave component of the DC component containing non-standing is not applied to completely.Therefore,
The present invention with can count and point-to-point error RMSE as the target training ANFIS pre- abnormal devices of numeral, while also actual with RMSE assessments
The precision of output current.
ANFIS numerals pre- abnormal method is applied to current mode LPA of above-mentioned test device for relay protection below, by test
Verify its effectiveness.Test is from the true fault electric current recorded by fault oscillograph shown in Fig. 9.By the true fault
Electric current obtains normalized instruction current (i.e. input current perunit value), with its control electric current type divided by preferable amplification 40
LPA amplifies and reappears, and obtains without the pre- abnormal actual output current of numeral, as shown in Figure 10.Fig. 9 is visible for contrast, the reality output
There is obvious distortion in failure of the current section, this is because the harmony of the DC component containing non-standing wave component in faulty section,
Present current mode LPA significant non-linear, amplification changes therewith.Quantitative Analysis show, the actual output current with
The RMSE of desired output current is 10.7955.
Desired output current is obtained into pre- abnormal instruction current by the pre- abnormal device of ANFIS numerals below, as shown in figure 11.By pre-
Abnormal instruction current control electric current type LPA amplifies and reappears, and obtains the pre- abnormal actual output current of Jing numerals, because its with Fig. 9 almost phase
Together, therefore similar to emulation embodiment, no longer provide curve chart.It is computed, the RMSE of actual output current and desired output current
Only 0.0372, the precision of quantitative explanation actual output current is very high, and the pre- abnormal device of ANFIS numerals is same for actual device
Effectively.
Finally by emulation embodiment and test example Jing numeral it is pre- it is abnormal after RMSE do a contrast, its value is respectively
0.0311 and 0.0372, it is seen that both are relatively.Although this explanation emulation and test are respectively adopted model and actual dress
Put and carry out for different fault currents, but its precision has preferable concordance.
It is noted that emulation embodiment provided above and test example are only used for the application effect for stating the present invention, not
To limit the present invention, all any modification, equivalent and improvement made within the spirit and principles in the present invention etc., all should
It is included within protection scope of the present invention.
Claims (6)
1. a kind of ANFIS pre- abnormal methods of numeral of test device for relay protection LPA, it is characterised in that the relay protection test
The pre- abnormal method of ANFIS numerals of device LPA goes the relay protection approached in compact set using ANFIS after training and learning
The nonlinear inverse mathematical model of test device LPA, replaces inverse number word model to produce the finger corresponding with desired output voltage/electric current
Voltage/current is made, realizes that numeral is pre- abnormal.
2. ANFIS pre- abnormal methods of numeral of test device for relay protection LPA as claimed in claim 1, it is characterised in that described
The realization of the pre- abnormal method of ANFIS numerals of test device for relay protection LPA is comprised the following steps:
1) using the input/output data of LPA inverse number word models to the input/output data pair as ANFIS, i.e. training data;
2) type and number of membership function are chosen;
3) generate the initiating structure of ANFIS;
4) training parameter is selected, i.e., maximum training batch, initial step length and target error;
5) training is completed by hybrid learning algorithm, generate the pre- abnormal device of ANFIS numerals;
6) pre- abnormal command voltage/electric current is produced with the pre- abnormal device of desired output voltage/electric current Jing ANFIS numerals;
7) the amplification output of voltage/current is realized with pre- abnormal command voltage/current control LPA;
8) back production the RMSE of actual output voltage/electric current and desired output voltage/electric current is calculated, confirms that ANFIS numerals are pre- abnormal
The effectiveness of method.
3. ANFIS pre- abnormal methods of numeral of test device for relay protection LPA as claimed in claim 2, it is characterised in that described
The establishment step of LPA mathematical models is as follows:
1) by actual test wiring, test device for relay protection is connected with tested relay protection and automatic safety device;
2) sinusoidal voltage and sinusoidal current that LPA output frequencies are 10kHz are controlled, smooth adjustment its amplitude is allowed to from 0 increase to
Maximum;
3) sampling, the voltage of record gradual change, current amplitude;
4) amplitude by the amplitude of command signal with output voltage, electric current is arranged in input/output data pair correspondingly.
4. ANFIS pre- abnormal methods of numeral of test device for relay protection LPA as claimed in claim 2, it is characterised in that described
The formulation step of hybrid learning algorithm is as follows:
1) it is learning target to choose minimum RMSE:
In formula:dmExport for preferable, | | | |2Represent 2- norms;
2) coefficient of backward network is adjusted by following learning algorithm:
In formula:K=0,1,2 ... represents kth time adjustment, β>0 is learning rate;
3) will Jing after 1 adjustmentIt is temporarily fixed, the parameter of feedforward network is next adjusted, different membership functions have difference
Parameter, without loss of generality, ifThere are 2 parameters aijAnd bij, its learning algorithm is:
In formula:L=0,1,2 ... represents the l time adjustment, γ>0 is learning rate;
4) useaijAnd b (l+1)ij(l+1) matrix A is constituted with input data, use αjComposition unknown vector α, uses dmGroup
Into preferable output vector d, overdetermination matrix equation A α=d is obtained;With Least Square Method optimal solution α, least square solution can be obtained:
αLS=(ATA)-1ATd。
5. a kind of pre- abnormal method of ANFIS numerals of relay protection test LPA described in utilization claim 1~4 any one is set up
The ANFIS pre- abnormal devices of numeral.
6. a kind of test device for relay protection for being provided with the pre- abnormal device of the numeral of ANFIS described in claim 5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710035791.5A CN106597178B (en) | 2017-01-18 | 2017-01-18 | A kind of pre- abnormal method of ANFIS number of test device for relay protection LPA |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710035791.5A CN106597178B (en) | 2017-01-18 | 2017-01-18 | A kind of pre- abnormal method of ANFIS number of test device for relay protection LPA |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106597178A true CN106597178A (en) | 2017-04-26 |
CN106597178B CN106597178B (en) | 2019-07-19 |
Family
ID=58586491
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710035791.5A Active CN106597178B (en) | 2017-01-18 | 2017-01-18 | A kind of pre- abnormal method of ANFIS number of test device for relay protection LPA |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106597178B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106990312A (en) * | 2017-05-25 | 2017-07-28 | 贵州电网有限责任公司电力科学研究院 | The test device and method of testing of a kind of digital relay protection device |
CN108107294A (en) * | 2017-12-27 | 2018-06-01 | 国网冀北电力有限公司张家口供电公司 | A kind of microcomputer protective relay pilot system based on measurement module |
CN108195704A (en) * | 2018-03-05 | 2018-06-22 | 重庆水利电力职业技术学院 | A kind of grinding parameter test bed component automatically |
CN110108957A (en) * | 2019-05-07 | 2019-08-09 | 武汉理工大学 | A kind of tractor electric fault diagnosis method based on structured analysis method |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20040103027A (en) * | 2003-05-30 | 2004-12-08 | 국방과학연구소 | Method for estimating an inductance profile of srm using adaptive neuro-fuzzy inference system |
CN103001591A (en) * | 2012-07-16 | 2013-03-27 | 北京邮电大学 | Method for eliminating nonlinear distortion |
CN104155574A (en) * | 2014-07-31 | 2014-11-19 | 国网湖北省电力公司武汉供电公司 | Power distribution network fault classification method based on adaptive neuro-fuzzy inference system |
CN104536435A (en) * | 2014-12-18 | 2015-04-22 | 中国科学院电工研究所 | Online diagnosis method for linear control system network |
CN105912806A (en) * | 2016-04-28 | 2016-08-31 | 湖南大学 | Small-overlap collision safety airbag control method based on adaptive neural fuzzy inference |
CN106291354A (en) * | 2016-07-18 | 2017-01-04 | 河海大学 | Electric automobile permagnetic synchronous motor Fault Classification based on ANFIS |
-
2017
- 2017-01-18 CN CN201710035791.5A patent/CN106597178B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20040103027A (en) * | 2003-05-30 | 2004-12-08 | 국방과학연구소 | Method for estimating an inductance profile of srm using adaptive neuro-fuzzy inference system |
CN103001591A (en) * | 2012-07-16 | 2013-03-27 | 北京邮电大学 | Method for eliminating nonlinear distortion |
CN104155574A (en) * | 2014-07-31 | 2014-11-19 | 国网湖北省电力公司武汉供电公司 | Power distribution network fault classification method based on adaptive neuro-fuzzy inference system |
CN104536435A (en) * | 2014-12-18 | 2015-04-22 | 中国科学院电工研究所 | Online diagnosis method for linear control system network |
CN105912806A (en) * | 2016-04-28 | 2016-08-31 | 湖南大学 | Small-overlap collision safety airbag control method based on adaptive neural fuzzy inference |
CN106291354A (en) * | 2016-07-18 | 2017-01-04 | 河海大学 | Electric automobile permagnetic synchronous motor Fault Classification based on ANFIS |
Non-Patent Citations (1)
Title |
---|
陈庆霆: "RF功放的神经网络模型研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106990312A (en) * | 2017-05-25 | 2017-07-28 | 贵州电网有限责任公司电力科学研究院 | The test device and method of testing of a kind of digital relay protection device |
CN106990312B (en) * | 2017-05-25 | 2023-06-16 | 贵州电网有限责任公司电力科学研究院 | Testing device and testing method for digital relay protection device |
CN108107294A (en) * | 2017-12-27 | 2018-06-01 | 国网冀北电力有限公司张家口供电公司 | A kind of microcomputer protective relay pilot system based on measurement module |
CN108107294B (en) * | 2017-12-27 | 2020-02-07 | 国网冀北电力有限公司张家口供电公司 | Microcomputer relay protection test system based on measurement modularization |
CN108195704A (en) * | 2018-03-05 | 2018-06-22 | 重庆水利电力职业技术学院 | A kind of grinding parameter test bed component automatically |
CN110108957A (en) * | 2019-05-07 | 2019-08-09 | 武汉理工大学 | A kind of tractor electric fault diagnosis method based on structured analysis method |
Also Published As
Publication number | Publication date |
---|---|
CN106597178B (en) | 2019-07-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhang et al. | A novel hybrid model for wind speed prediction based on VMD and neural network considering atmospheric uncertainties | |
CN106597178A (en) | ANFIS digital pre-distorting method of relay protection test device LPA | |
Roinila et al. | Hardware-in-the-loop methods for real-time frequency-response measurements of on-board power distribution systems | |
Erenturk | ANFIS-based compensation algorithm for current-transformer saturation effects | |
Tremblay et al. | Contribution to stability analysis of power hardware‐in‐the‐loop simulators | |
CN112100962B (en) | Electromagnetic transient simulation method and system of MMC | |
Garanayak et al. | Fast and accurate measurement of harmonic parameters employing hybrid adaptive linear neural network and filtered‐x least mean square algorithm | |
Lee et al. | An enhanced Lagrangian neural network for the ELD problems with piecewise quadratic cost functions and nonlinear constraints | |
Tremblay et al. | A multi-time-step transmission line interface for power hardware-in-the-loop simulators | |
Maciejewski et al. | Application of the waveform relaxation technique to the co-simulation of power converter controller and electrical circuit models | |
Klimas et al. | Application of shallow neural networks in electric arc furnace modeling | |
Monfared et al. | Prediction of state-of-charge effects on lead-acid battery characteristics using neural network parameter modifier | |
Rupasinghe et al. | Multi‐rate co‐simulation of power system transients using dynamic phasor and EMT solvers | |
Bahoura et al. | FPGA-implementation of an adaptive neural network for RF power amplifier modeling | |
Mahdian Dehkordi et al. | Robust reliable fault tolerant control of islanded microgrids using augmented backstepping control | |
Siegers et al. | Stability and accuracy considerations in the design and implementation of a kilowatt-scale DC power hardware-in-the-loop platform | |
CN106681168A (en) | Decoupling method for MMC converter valve digital-analog hybrid simulation interface network | |
Parang et al. | Residualisation‐based model order reduction in power networks with penetration of photovoltaic resources | |
Akçay et al. | Identification of power transformer models from frequency response data: A case study | |
Romay et al. | Transient stability assessment considering hard limits on dynamic states | |
CN105515423A (en) | Energy balance control method for modular multi-level converter bridge arms | |
CN112348384B (en) | Method for rapidly evaluating reliability of power distribution network planning scheme based on data driving | |
Oleskovicz et al. | Estimation of harmonic currents injected by nonlinear loads for a distorted power supply scenario using artificial neural networks | |
Byczkowska-Lipińska et al. | Adaptation of discrete macromodels for calculations in MATLAB/Simulink environment | |
Luo et al. | Implementation and value of power hardware in the loop testing bed for wind turbines integrated into grid |
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 |