CN106844972A - Transformer Winding temperature flexible measurement method based on PSO SVR - Google Patents
Transformer Winding temperature flexible measurement method based on PSO SVR Download PDFInfo
- Publication number
- CN106844972A CN106844972A CN201710053107.6A CN201710053107A CN106844972A CN 106844972 A CN106844972 A CN 106844972A CN 201710053107 A CN201710053107 A CN 201710053107A CN 106844972 A CN106844972 A CN 106844972A
- Authority
- CN
- China
- Prior art keywords
- svr
- pso
- transformer
- transformer winding
- particle
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- 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
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention relates to a kind of Transformer Winding temperature flexible measurement method based on PSO SVR, the method obtains the measured value of the auxiliary variable of setting first, using the measured value of auxiliary variable as the input vector of housebroken PSO SVR Transformer Windings temperature soft-sensing model, coiling hot point of transformer temperature is obtained so as to predict；The training of PSO SVR Transformer Winding temperature soft-sensing models is specially：1) Transformer Winding temperature survey model is set up using SVR algorithms；2) using the parameter combination of PSO algorithm optimization Transformer Winding temperature survey models, best parameter group is obtained；3) training data is obtained, training data includes multigroup auxiliary variable measured value and corresponding coiling hot point of transformer measured temperature；4) the Transformer Winding temperature survey model with best parameter group is trained using training data.Compared with prior art, the present invention has the advantages that predictive ability is good, precision of prediction is high.
Description
Technical field
The present invention relates to Winding in Power Transformer temperature indirect measuring technology, more particularly, to a kind of based on PSO-SVR's
Transformer Winding temperature flexible measurement method.
Background technology
Transformer occupies very important status in the power transmission and transformation system of power network, and it is the smooth power supply of whole power system
Feel at ease with society the basic guarantee of electricity consumption.Power transformer uses enclosed structure, is chronically at high voltage, high current and expires
Under the running status of load, it is more likely that the cause thermal damage that overheat is run and causes inside transformer element occurs.Transformer Winding
Hot(test)-spot temperature can reflect the change and potential heat event of its running status as one of most important Warm status amount of transformer
Barrier behavior, the reacting condition of its temperature goes out the corresponding information of transformer station high-voltage side bus hidden danger.Therefore to coiling hot point of transformer temperature
Research, can make operations staff more easily understand the ruuning situation of transformer, and the diagnosis and prevention to Accident of Transformer have
Very important meaning.
The prediction of current oil-filled transformer internal temperature all uses the indirect method of measurement, it by analysis draw with it is measured
Auxiliary variable that is related and easily measuring, intelligence learning and established model are carried out to it to realize the estimation to leading variable.Hard measurement
Method is only needed to measure some auxiliary variables for easily measuring, and training is passed through on the basis of thermoelectricity analogy using intelligent modeling method
To set up soft-sensing model.This model has stronger on-line correction ability, can apply well in non-linear and uncertainty
In system, therefore the soft sensor modeling of intelligence has very big development potentiality in hot spot temperature of winding hard measurement.And how to improve
The accuracy for obtaining hot-spot temperature of transformer using soft sensor modeling is then the problem that those skilled in the art need to solve.
In the prior art, IEEE Std C57.91-1995 and GB GB/T 15164-1994 recommend Transformer Winding
Hot(test)-spot temperature computation model, it ignores the actual condition influence of environment temperature and transformer station high-voltage side bus, does not account for top-oil temperature
The more slowly characteristic of speed ratio winding temperature rise speed is risen, the model needs time response when transformer load changes, past
Toward the change that cannot timely reflect hot(test)-spot temperature.
The content of the invention
The purpose of the present invention provides a kind of based on PSO-SVR for the defect for overcoming above-mentioned prior art to exist
Transformer Winding temperature flexible measurement method.
The purpose of the present invention can be achieved through the following technical solutions：
A kind of Transformer Winding temperature flexible measurement method based on PSO-SVR, the auxiliary that the method obtains setting first becomes
The measured value of amount, using the measured value of the auxiliary variable as housebroken PSO-SVR Transformer Windings temperature soft-sensing model
Input vector so that predict acquisition coiling hot point of transformer temperature.
The training of the PSO-SVR Transformer Windings temperature soft-sensing model is specially：
1) Transformer Winding temperature survey model is set up using SVR algorithms；
2) using the parameter combination of Transformer Winding temperature survey model described in PSO algorithm optimizations, optimized parameter group is obtained
Close；
3) training data is obtained, the training data includes multigroup auxiliary variable measured value and corresponding Transformer Winding heat
Point measured temperature；
4) the Transformer Winding temperature survey model with best parameter group is trained using the training data,
Obtain housebroken PSO-SVR Transformer Windings temperature soft-sensing model.
It is specially using the parameter combination of Transformer Winding temperature survey model described in PSO algorithm optimizations：
A) punishment parameter, particle maximal rate, particle total number and maximum iteration are set, the position of particle is randomly choosed
Put vector velocity；
B) whether particle is in solution space in checking iterative process, if its current location is beyond the scope of solution space, will
Its position for being re-set as previous moment；
C) fitness of current each particle is calculated according to fitness function；
D) according to the comparing of fitness value, oneself state variable and global state variable optimal at present is obtained；
E) passive attraction based on passive cogregation theoretical renewal each particle in colony
F) passive attraction item is introducedCalculate the speed and position vector for updating each particle；
G) check whether stopping criterion for iteration meets, repeat step the b)-f if being unsatisfactory for).
In the step f), the iterative formula of particle is as follows：
In formula,It is the passive attraction interfered to particle i in kth time iteration,Respectively
It is particle in the position and speed of current time and subsequent time, r1、r2And r3It is the random number between [0,1], c1、c2It is displacement
Weight factor, c3Attract item weight factor for passive, w is the inertia weight factor,It is current time individuality extreme value,For
Global extremum, β is the iteration coefficient of position vector and velocity.
Using the measured value of the auxiliary variable as the defeated of housebroken PSO-SVR Transformer Windings temperature soft-sensing model
Enter before vector and when being trained to PSO-SVR Transformer Winding temperature soft-sensing models, the survey of the auxiliary variable to obtaining
Amount is pre-processed.
Compared with prior art, the present invention has advantages below：
1st, the present invention can be set up on the basis of thermoelectricity analogy for Transformer Winding temperature hard measurement problem by training
Soft-sensing model, improves the on-line correction ability of model, it is adaptable in non-linear and uncertain system.
2nd, the present invention makes full use of the auxiliary variable that environmental factor, electricabsorption agent and Warm status factor etc. are easily measured, and chooses
Auxiliary variable including active loss, power factor, load current, environment temperature, wind speed and top-oil temperature etc., by PSO-
The soft-sensing model that SVR intelligent algorithms are set up realizes the prediction to the leading variable hot spot temperature of winding for being difficult to directly measure,
The method is also with leading and novelty in electric field.
3rd, the present invention chooses active loss, power factor, load current, environment temperature, this 6 changes of wind speed and top-oil temperature
Used as auxiliary variable, wherein active power, power factor and load current are electric parameters to amount, reflect the actual motion of transformer
Situation；The influence of oil viscosity and radiating effect to transformer fuel factor under environment temperature and wind speed reflection varying environment；Top layer oil
Hot-spot temperature of transformer computation model parameter in temperature reflection IEEE load directive/guides.Above-mentioned 6 variables reflect that transformer heat production dissipates
Heat engine is managed and Transformer Winding Temperature Rise high correlation, and auxiliary variable is easy to measurement, beneficial to winding temperature soft-sensing model iteration
Convergence.
4th, the present invention carries out parameter combination optimization using PSO algorithms to SVR models, and model prediction accuracy is high.
5th, PSO algorithms of the invention can make particle jump out locally optimal solution, and convergence speed is increased with dynamic inertia weight
Degree, until seeking the best parameter group of each auxiliary variable parameter, determines the parameter model of winding temperature, accurate prediction winding temperature
Degree.
Brief description of the drawings
Fig. 1 is PSO parameter optimizations flow chart of the present invention；
Fig. 2 is the input/output structure figure of PSO-SVR soft-sensing models of the present invention；
Fig. 3 is the hot spot temperature of winding data after normalization；
Fig. 4 is SVR final argument optimizing results；
Fig. 5 is the final training result of traditional SVR, PSO-SVR and BP neural network；
Fig. 6 is the training error of traditional SVR, PSO-SVR and BP neural network；
Fig. 7 is predicting the outcome for traditional SVR, PSO-SVR and BP neural network；
Fig. 8 is the predicated error of traditional SVR, PSO-SVR and BP neural network；
Fig. 9 is the comparison diagram of field measurement data and model prediction data.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention
Premised on implemented, give detailed implementation method and specific operating process, but protection scope of the present invention is not limited to
Following embodiments.
The present embodiment provides a kind of Transformer Winding temperature flexible measurement method based on PSO-SVR, and the method is obtained first
The measured value of the auxiliary variable of setting, unconfinement problem is translated into by penalty function method, is surveyed with reference to Transformer Winding temperature
The auxiliary variable species of amount, determines training sample and the detection data set of PSO-SVR models, by the measurement of the auxiliary variable
Value as housebroken PSO-SVR Transformer Windings temperature hard measurement mould, the input vector of type so that predict acquisition transformer around
Group hot(test)-spot temperature.This method has taken into full account that Transformer each auxiliary variable is calculated leading variable and optimized
Influence, precision of prediction is high.
First, PSO optimizes the mathematical modeling of SVR
1st, SVR algorithm models are set up：
The main purpose of support vector regression method be in order to determine one can accurately forecasting system information distribution letter
Number f (x).Meanwhile, in order to avoid the situation that over-fitting occurs, the function must be sufficiently flat.If length is the sample set of N：
(x1,y1)(x2,y2),...,(xl,yl), wherein xiRepresent input pointer and yiRepresent desired value.Seek a Nonlinear Mapping ψ
(), by the data x in samplei(i=1,2 ..., l) is mapped in higher dimensional space F one by one, thus can be by problem letter
Turn to the problem that optimal function is tried to achieve in feature space F.Nonlinear problem is carried out into linear regression with the linear function, i.e.,：
F (x)=w ψ (xi)+b i=1,2 ..., l (1)
In formula：ψ () descriptions high-dimensional feature space F, b are threshold values, and w is regression coefficient vector.After the mapping terminates, originally
A series of nonlinear datas arrived all spoiled linear function after high-dimensional feature space.Influenceing the factor of w has experience risk total
With and represent it in higher dimensional space flatness | | w | |2, by the relaxation factor ξ and ξ that introduce positive definite*, w and b can pass through
Following regular risk function is minimized to obtain:
Wherein C is the penalty for representing empiric risk, and penalty factor is a constant, and C>0, for controlling to super
Go out the punishment degree of the sample of error e.L is total sample number,It is to ensure that regular terms of the w in higher dimensional space flatness.Solve
The condition of the required satisfaction of the problem turns to treatment ε insensitive loss functions Lε(yi,f(xi)), it is defined by the formula：
Need to set up Lagrange's equation for the solution of constraints above optimization problem and be shown below：
In order to try to achieve the minimum value of L, to w, b, ξ and ξ in above formula (4)*Partial derivative is sought, and makes its value be equal to zero.I.e.：
Formula (5) is substituted into formula (4), following problem then can obtain according to duality theory：
In formula：αiWithIt is Lagrange multiplier, (i=1,2 ..., l).Derive the step, it is possible to will support to
The function regression problem of amount machine is converted into convex quadratic programming problem.Formula (6) is carried out solving the value for obtaining w：
αiWithThe solution for exactly minimizing.Thus linear regression function can be tried to achieve：
If data point is met in formula (8)Condition, then the point be just counted as the support of decision function to
Amount.Data (x according to the supporting vectori, x), just kernel function K (x can be tried to achieve in the inner product of higher dimensional space by former datai, x) such as
Under：
K(xi, x)=ψ (xi)×ψ(x) (9)
The selection of SVMs also has the form of multi-form, and this depends on the kernel function used in derivation, often
Have with kernel function：
1. RBF (K (xi, x)=exp (- | | xi-x||/2σ2))；
2. polynomial function (K (xi, x)=((xi×x)+b)d, b >=0, d is natural number)；
3. Sigmoid functions (K (xi, x)=tan (k (xi× x)+v), k ＞ 0＆v ＜ 0) and linear function etc..
Above kernel function is selected, due to being linearly can not between coiling hot point of transformer temperature and its influence factor
Divide, therefore linear function is not suitable for.In other three kinds of kernel functions, because polynomial function and Sigmoid function parameters are more,
And RBF only needs to determine a parameter, and can guarantee that training result meets requirement again, therefore, core letter of the invention
Number selection is RBF, and γ=1/2 σ is made again2, then obtain：
K(xi, x)=exp (- γ | | xi-x||) (10)
2nd, PSO algorithms are introduced and SVR parameter optimizations
Influenceing the factor of the optimizing ability of support vector regression SVR models has error penalty factor and kernel functional parameter γ.
In simulation process, it is desirable to find the parameter being best suitable for, the experience by researcher is merely able to, not direct method is available for
With reference to.Therefore, best parameter group (C, γ) is reliably found for convenience, invention introduces particle cluster algorithm optimizing side
Method.
When PSO algorithms bring into operation, a group random particles are first initialized, allow them all in solution space, each
Particle all represents a feasible solution of problem.In program operation process, per iteration once, particle all can be according to fitness function
Corresponding fitness value is obtained, for differentiating whether corresponding particle reaches optimizing solution.All particles in population all with
One vector motion, the direction of its motion and distance are all relevant with the vector.Each iteration can all occur closest to optimal solution
Particle, the direction of motion that remaining particle followed by the particle is searched for one by one, finally obtains optimal solution.Particle is more in iterative process
Newly carried out according to two extreme values, one is itself extreme value, represent the optimum position that the particle is searched out itself, it is determined
Ability and human-subject test that particle itself is found to optimal solution.Another is global extremum, represents that all particles are looked for currently
The optimal solution for arriving, embodies ability and social recognition level that whole population finds optimal solution.Particle swarm optimization algorithm is not only
Speed of searching optimization is very fast, and the characteristic of its small data optimizing also can solve the problem that the optimization of a series of complex is asked.Research understands particle
At the iteration initial stage, its optimizing performance is protruded colony optimization algorithm than remaining evolution algorithm, but with the increase of algebraically, ability
Can decline.Understand based on more than, the present invention will optimize branch using the Modified particle swarm optimization algorithm based on passive cogregation
The parameter combination (C, γ) of vector regression SVR models is held, and finds its optimal solution.
The parameter optimization carried out to SVR using PSO algorithms, can find SVM most in velocity location search model
Excellent parameter combination.A group particle is initialized in three-dimensional solution space, the particle is made up of parameter combination (C, γ), wherein sequence number i
Particle position be ui=(ui1,ui2,ui3)T, its speed is vi=(vi1,vi2,vi3)T, the individual extreme value at current time is designated as
pibest, global extremum is designated as pgbest.In each iteration, particle adjusts the direction of motion according to itself extreme value and global extremum
And speed, from the state motion of previous moment to state of lower a moment.In order to avoid traditional PS O algorithms easily lose diversity so as to
The shortcoming of locally optimal solution is absorbed in, optimizing is carried out using the Modified particle swarm optimization algorithm based on passive cogregation.
Particle cluster algorithm derives from the behavioral study looked for food to birds, when bevy is when food is searched for, in this region
Interior only serving, it is exactly to be searched around bird nearest from food at present to find food most efficient method.In optimizing
During, concept according to passive cogregation algorithm need to introduce it is passive attract item, make particle not only consider itself optimal location and
Global optimum position, also suffers from attracting the interference of item so as to be difficult to be absorbed in local optimum state.The iterative formula of particle is as follows：
In formula,It is the attraction interfered to particle in kth time iteration, can at random selects a certain in population
Particle,It is respectively particle in the position and speed of current time and subsequent time；r1、r2And r3It is
[0,1] random number between；Positive definite constant c1And c2It is the weight factor of displacement, it determines the length of Particles Moving, typically takes
Be worth is 1.5；c3Represent the passive weight factor for attracting item, it decide the speed of Particles Moving and, general value is 1；
It is current time individuality extreme value,It is global extremum, β is the iteration coefficient of position vector and velocity；ω represents inertia
Weight factor, larger ω is beneficial to global search, and less ω is then beneficial to Local Search, therefore, its value can be by following
Formula (12) is automatically adjusted：
In formula：ωmin、ωmaxIt is the minimum value and maximum of the inertia weight factor, NmaxIt is the total iterations of colony, N is
Current particle iterations, inertia weight is taken to 0.4 from 0.9.It is worth noting that：Because the SVR parameter combinations of optimization
The value of (C, γ) differs greatly, not on the same order of magnitude.Therefore, it is necessary to be multiplied by before particle rapidity in initialization procedure
Corresponding coefficient.In order to be able to directly reflect support vector regression performance, the fitness function of selection is root-mean-square error
(RMSE), its formula is following (13)：
In formula, m is total sample number, yiWithRepresent respectively be i-th training sample measured value and predicted value.
As shown in figure 1, improving step of the particle cluster algorithm optimization to parameter (C, γ) optimizing of support vector regression SVR models
It is rapid as follows：
1) c is set1, c2And particle maximal rate, particle number N, maximum iteration Nmax, randomly choose the position of particle
Put vector velocity；
2) whether i-th particle is checked also in the middle of solution space, if its current locationBeyond the scope of solution space, then will
Its position for being re-set as previous moment
3) fitness of current each particle is calculated by fitness function, fitness function is root-mean-square error RMSE；
4) according to the comparing of fitness value, oneself state variable and global state variable optimal at present is found.By than
Compared with pibestFitness value and object function, if object function is more excellent, using current location updateSimilarly, such as
Fruit object function is not only better thanFitness value, be also advantageous overFitness value, then use current locationUpdate
5) R is updatedi, individuality R is passively attracted for each particle randomly chooses one in colony in each step of iterationi；
6) calculating speed vector position coordinates, the speed and position vector of each particle are updated by computing formula；
7) check whether stopping criterion for iteration meets, the repeat step 2 if being unsatisfactory for) -6).
2nd, the Transformer Winding temperature soft-sensing model based on PSO-SVR
1st, the selection of state variable
The selection of state variable is the basis of soft sensor modeling, and good auxiliary variable will improve the precision of soft-sensing model,
It is easy to train and learns, it is more convenient to be that industrial process brings.Therefore, the state of hot spot temperature of winding soft-sensing model is selected
Transformer heat production radiating should be carefully studied during variable, from finding the associated factor of influence winding temperature rise in practice.It is first
First, by the environmental condition of transformer (including environment temperature θa, wind speed etc.) as input quantity because transformer is in varying environment temperature
Under degree, oil viscosity and its radiating effect are differed certainly, therefore environmental variance influences whether the thermal characteristics of transformer.Secondly, root
The hot-spot temperature of transformer computation model recommended according to IEEE load directive/guides can be seen that top-oil temperature can also influence coiling hotspot temperature
Degree, therefore it is seen as one of input quantity.Finally, due to the active loss of transformer, power factor and load current are electric
Parameter, can to a certain extent react the operation conditions of transformer, therefore also receive as among input quantity.In sum, select altogether
6 characteristic quantities such as the active loss of transformer, power factor, load current, environment temperature, wind speed, top-oil temperature are used as defeated
Enter vector.Its input/output structure figure is as shown in Figure 2.
After being determined state variable, so that it may data are trained and soft-sensing model is established.The present embodiment is received altogether
In Ji Liaomou cities 110kV certain transformer station's daily sheet transformer data of 4 days as PSO-SVR models inputoutput data collection.
These data are divided into two groups：First group is the data of first 3 days, and for training generation soft-sensing model, the data of the 4th day are made
It it is second group, for verifying the accuracy and reliability of model.In training process, improved passive cogregation particle cluster algorithm can be
Support vector regression SVR searches out optimal parameter combination (C, γ), then brings it back into and can be based in SVR algorithms
The Transformer Winding temperature soft-sensing model of selected auxiliary variable.The soft-sensing model utilizes environmental factor, electricabsorption agent and heat
The auxiliary variable that status consideration etc. is easily measured, the soft-sensing model set up by PSO-SVR intelligent algorithms is realized straight to being difficult to
The prediction of the leading variable hot spot temperature of winding for measuring is connect, this application process is also with leading and novelty in electric field
's.
2nd, the pretreatment of data and Performance Evaluating Indexes
Because the state variable for obtaining is more, and some data are even on the different orders of magnitude, it is therefore desirable to right
Data are pre-processed.First according to error of whether ating fault in Analysis on Mechanism data, because data are collected from transformer station
Come, be not in generally phenomenon of the failure, therefore data are all available, without rejecting.Next to that the normalization of data, this
Be conducive on Data Integration a to dimension, so that model more facilitates when fitting is restrained, it is also possible to increase model
Generalization Capability.Fig. 3 is the hot spot temperature of winding data after normalization.
It is exactly error analysis to also have a most important part during soft sensor modeling, and this is conducive to the pre- of testing model
Effect is surveyed, and model can be adjusted as feedback quantity makes its performance more excellent.In order to more accurately reflect that it is predicted
Effect, employs four Performance Evaluating Indexes as error analysis function in this chapter, they are respectively：Maximum definitely percent error
MPE, mean absolute error MAE, average absolute percent error MAPE and coefficient R2, expression formula is distinguished as follows：
3rd, sample calculation analysis
1st, data source and simulation parameter
Certain transformer station's daily sheet 29 days to 2015 Septembers of August in 2015 are total on the 01st during the present embodiment have collected certain city 110kV
The transformer data of 4 days, and by active loss, power factor, load current, environment temperature, wind speed and top-oil temperature six
Auxiliary variable as PSO-SVR models input, hot spot temperature of winding as its export.The interval of data acquisition is 1h, four days
96 groups altogether, first 72 groups are used as training set, and 24 groups are used as test set afterwards.Table 1 is part input and output sample.
Table 1
2nd, soft-sensing model predicts the outcome analysis
After coiling hot point of transformer temperature soft-sensing model based on PSO-SVR is established, in order to verify its accuracy,
Start to be trained the observed temperature data of large-scale power transformer with checking, while employing traditional supporting vector
Homing method SVR and BP neural network are trained to data, predict the outcome to protrude this finally by three kinds of methods are compared
The superior part of the PSO-SVR soft-sensing models of invention.This process programming realization in Matlab softwares.
During model sets up SVR parameter optimizations, in order to reach best prediction effect, it is necessary to adjust PSO algorithms
Parameter, the ability of optimizing can be given full play to.It is final as follows to its parameter setting：Local search ability c1=1.5, entirely
Office search capability c2=2.0, the inertia weight factor ω value are 0.4~0.9, and population maximum quantity is 20, maximum evolution quantity
It is 200.The result of final iteration optimizing is as shown in figure 4, train best parameter group (ε, C, the γ)=(6.253e- for obtaining
005,98.5953,2.8928), fitness function RMSE=0.049362%.
The Optimal error penalty factor and kernel functional parameter γ that PSO is sought bring SVR models into and can obtain coiling hotspot temperature
Degree soft-sensing model.It is the accuracy of prominent this model experiment results, with identical data, using traditional support vector regression
Method SVR and BP neural network method predict the hot(test)-spot temperature of winding.The final training of three kinds of models and predict the outcome such as Fig. 5
And the training error and predicated error that Fig. 7, Fig. 6 and Fig. 8 are three kinds of models.
Be can be seen that from the training result of Fig. 5 and first three days of Fig. 6, in first three day data training stage, BP neural network and
PSO-SVR soft-sensing models can it is accurate fitting actual temperature curve, its prediction temperature error values 0.5 DEG C with
Interior, the two has stronger learning ability.And the result of tradition SVR model predictions is less than satisfactory, error maximum can reach
1.7 DEG C, the selection of parameter influences its fitting effect.But during the 4th day test data, the performance of BP neural network
But it is not so good as PSO-SVR soft-sensing models, error is significantly increased, particularly when hot spot temperature of winding change is violent, its is pre-
Survey max value of error and can reach 1.085 DEG C.Illustrate that BP neural network inside capability of fitting is very strong, but it is pre- for external data
Survey result and be but not so good as PSO-SVR.Because easily there is the phenomenon of over-fitting in neutral net, its precision of prediction under small data sample
It is also to be strengthened.And PSO-SVR algorithms prediction effect is more preferably, error is controlled within 0.3897 DEG C always, the soft-sensing model
With more preferable predictive ability.Traditional SVR methods reflect the variable quantity of hot spot temperature of winding to a certain extent, but its
Training error of the training error more than BP neural network.It is also seen that PSO-SVR not only has ratio in the training stage from figure
The fitting effect of pseudo neural network, and its predicated error in test set is minimum, its precision of prediction in three kinds of models
It is better.
To check the prediction effect of above-mentioned model, performance evaluation is carried out to above-mentioned model according to evaluation index, as a result such as table 2
It is shown.
Table 2
As can be seen from Table 2, during the winding temperature prediction of identical training set data input, the soft surveys of PSO-SVR
The maximum of amount model definitely percent error MPE=1.04%, MAE=0.016 DEG C of mean absolute error and average absolute percentage are missed
Difference MAPE=0.043% is minimum.The coefficient correlation each predicted the outcome under three kinds of schemes, PSO- are given in table simultaneously
The coefficient R of SVR soft-sensing models2=0.967 is maximum, and this explanation PSO-SVR temperature soft-sensing model is to coiling hotspot temperature
The precision of prediction of degree is higher, and generalization ability is stronger.
In order to more intuitively find out the prediction effect of PSO-SVR models, Fig. 9 depicts field measurement data and model prediction
The comparison diagram of data.Coordinate points are substantially all on the straight line of y=x as can be seen from Figure, this explanation predicted value and measured value
Data are very close to of substantially equal.In sum, the results show is of the invention based on PSO-SVR soft-sensing models can
Accurately predict the hot spot temperature of winding of transformer.
By the result to three kinds of prediction schemes and performance evaluation, whether SVR, BP neural network or PSO-
SVR flexible measurement methods, the increase of the sample of training data is all conducive to improving the regression accuracy of model.But in identical training sample
Under conditions of this input, this method has more preferable generalization and accuracy, and this is oil-immersed power transformer coiling hotspot temperature
Degree prediction provides a kind of new method, also for Transformer's Condition Monitoring and failure predication provide certain technical support and reason
By guidance, human cost and financial resources cost can be to a certain extent saved.
Preferred embodiment of the invention described in detail above.It should be appreciated that one of ordinary skill in the art without
Need creative work just can make many modifications and variations with design of the invention.Therefore, all technologies in the art
Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Technical scheme, all should be in the protection domain being defined in the patent claims.
Claims (5)
1. a kind of Transformer Winding temperature flexible measurement method based on PSO-SVR, it is characterised in that the method is set first
Auxiliary variable measured value, the measured value of the auxiliary variable is soft as housebroken PSO-SVR Transformer Windings temperature
The input vector of measurement model, so as to predict acquisition coiling hot point of transformer temperature.
2. the Transformer Winding temperature flexible measurement method based on PSO-SVR according to claim 1, it is characterised in that institute
The training for stating PSO-SVR Transformer Winding temperature soft-sensing models is specially：
1) Transformer Winding temperature survey model is set up using SVR algorithms；
2) using the parameter combination of Transformer Winding temperature survey model described in PSO algorithm optimizations, best parameter group is obtained；
3) training data is obtained, the training data includes multigroup auxiliary variable measured value and corresponding coiling hot point of transformer temperature
Degree measured value；
4) the Transformer Winding temperature survey model with best parameter group is trained using the training data, is obtained
Housebroken PSO-SVR Transformer Windings temperature soft-sensing model.
3. the Transformer Winding temperature flexible measurement method based on PSO-SVR according to claim 1, it is characterised in that choosing
Taking the auxiliary variable includes active loss, power factor, load current, environment temperature, wind speed and top-oil temperature.
4. the Transformer Winding temperature flexible measurement method based on PSO-SVR according to claim 1, it is characterised in that profit
It is specially with the parameter combination of Transformer Winding temperature survey model described in PSO algorithm optimizations：
A) punishment parameter, particle maximal rate, particle total number and maximum iteration are set, the position arrow of particle is randomly choosed
Amount and velocity；
B) whether particle is in solution space in checking iterative process, if its current location is beyond the scope of solution space, its is heavy
The new position for being set to previous moment；
C) fitness of current each particle is calculated according to fitness function；
D) according to the comparing of fitness value, oneself state variable and global state variable optimal at present is obtained；
E) passive attraction based on passive cogregation theoretical renewal each particle in colony
F) passive attraction item is introducedCalculate the speed and position vector for updating each particle；
G) check whether stopping criterion for iteration meets, repeat step the b)-f if being unsatisfactory for).
5. the Transformer Winding temperature flexible measurement method based on PSO-SVR according to claim 4, it is characterised in that institute
State in step f), the iterative formula of particle is as follows：
In formula,It is the passive attraction interfered to particle i in kth time iteration,It is respectively grain
Son is in the position and speed of current time and subsequent time, r1、r2And r3It is the random number between [0,1], c1、c2It is displacement weight
The factor, c3Attract item weight factor for passive, ω is the inertia weight factor,It is current time individuality extreme value,It is the overall situation
Extreme value, β is the iteration coefficient of position vector and velocity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710053107.6A CN106844972A (en) | 2017-01-22 | 2017-01-22 | Transformer Winding temperature flexible measurement method based on PSO SVR |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710053107.6A CN106844972A (en) | 2017-01-22 | 2017-01-22 | Transformer Winding temperature flexible measurement method based on PSO SVR |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106844972A true CN106844972A (en) | 2017-06-13 |
Family
ID=59120432
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710053107.6A Pending CN106844972A (en) | 2017-01-22 | 2017-01-22 | Transformer Winding temperature flexible measurement method based on PSO SVR |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106844972A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108038795A (en) * | 2017-12-05 | 2018-05-15 | 武汉大学 | Hot-spot temperature of transformer inversion method and system based on streamline and support vector machines |
| CN108089077A (en) * | 2017-12-05 | 2018-05-29 | 武汉大学 | Characteristic quantity choosing method and system applied to transformer hot spot inverting |
| CN108120521A (en) * | 2017-12-08 | 2018-06-05 | 囯网河北省电力有限公司电力科学研究院 | Coiling hot point of transformer temperature predicting method and terminal device |
| CN108520111A (en) * | 2018-03-06 | 2018-09-11 | 宁波大学 | A kind of flexible measurement method based on orthogonal component optimal selection and optimum regression |
| CN108844624A (en) * | 2018-06-01 | 2018-11-20 | 北京科技大学 | A kind of SLM process laser power monitor method based on temperature field |
| CN109063729A (en) * | 2018-06-20 | 2018-12-21 | 上海电力学院 | A kind of Multisensor Image Fusion Scheme based on PSO-NSCT |
| CN110869728A (en) * | 2017-07-19 | 2020-03-06 | 林德股份公司 | Method for determining the stress level in a material of a process engineering device |
| CN110991011A (en) * | 2019-11-14 | 2020-04-10 | 广东电网有限责任公司 | Transformer real-time hot spot temperature prediction method |
| CN111401657A (en) * | 2020-04-02 | 2020-07-10 | 国网冀北电力有限公司唐山供电公司 | Transformer hot spot temperature time sequence prediction method based on data mining algorithm |
| CN112381123A (en) * | 2020-10-30 | 2021-02-19 | 鄂尔多斯应用技术学院 | Coal mining machine fault prediction method based on time series prediction method |
| CN112752960A (en) * | 2018-09-21 | 2021-05-04 | 赖茵豪森机械制造公司 | Determining a characteristic temperature of an electrical or electronic system |
| WO2022069149A1 (en) * | 2020-09-30 | 2022-04-07 | Maschinenfabrik Reinhausen Gmbh | Device and system for indirectly determining the temperature of a power transformer |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101447048A (en) * | 2008-12-30 | 2009-06-03 | 上海发电设备成套设计研究院 | Method for predicting life of transformer insulation and management system thereof |
| CN103838955A (en) * | 2013-12-09 | 2014-06-04 | 浙江大学 | Optimum soft measurement instrument and method for optimum mixing in propylene polymerization production process |
| WO2015027127A1 (en) * | 2013-08-23 | 2015-02-26 | Abb Inc. | Oil-immersed transformed thermal monitoring and prediction system |
| CN104459089A (en) * | 2014-12-12 | 2015-03-25 | 东北大学 | Soft measurement method of freeness of high-consistency pulp grinding system |
| CN105182161A (en) * | 2015-09-23 | 2015-12-23 | 国网山东莒县供电公司 | Transformer monitoring system and method |
-
2017
- 2017-01-22 CN CN201710053107.6A patent/CN106844972A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101447048A (en) * | 2008-12-30 | 2009-06-03 | 上海发电设备成套设计研究院 | Method for predicting life of transformer insulation and management system thereof |
| WO2015027127A1 (en) * | 2013-08-23 | 2015-02-26 | Abb Inc. | Oil-immersed transformed thermal monitoring and prediction system |
| CN103838955A (en) * | 2013-12-09 | 2014-06-04 | 浙江大学 | Optimum soft measurement instrument and method for optimum mixing in propylene polymerization production process |
| CN104459089A (en) * | 2014-12-12 | 2015-03-25 | 东北大学 | Soft measurement method of freeness of high-consistency pulp grinding system |
| CN105182161A (en) * | 2015-09-23 | 2015-12-23 | 国网山东莒县供电公司 | Transformer monitoring system and method |
Non-Patent Citations (1)
| Title |
|---|
| 杨芳勋 等: "基于改进粒子群算法的ICPT配电***规划", 《华中科技大学学报(自然科学版) 》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110869728A (en) * | 2017-07-19 | 2020-03-06 | 林德股份公司 | Method for determining the stress level in a material of a process engineering device |
| CN108089077A (en) * | 2017-12-05 | 2018-05-29 | 武汉大学 | Characteristic quantity choosing method and system applied to transformer hot spot inverting |
| CN108089077B (en) * | 2017-12-05 | 2020-05-19 | 武汉大学 | Characteristic quantity selection method and system applied to transformer hotspot inversion |
| CN108038795A (en) * | 2017-12-05 | 2018-05-15 | 武汉大学 | Hot-spot temperature of transformer inversion method and system based on streamline and support vector machines |
| CN108120521A (en) * | 2017-12-08 | 2018-06-05 | 囯网河北省电力有限公司电力科学研究院 | Coiling hot point of transformer temperature predicting method and terminal device |
| CN108120521B (en) * | 2017-12-08 | 2020-06-26 | 囯网河北省电力有限公司电力科学研究院 | Transformer winding hot spot temperature prediction method and terminal equipment |
| CN108520111A (en) * | 2018-03-06 | 2018-09-11 | 宁波大学 | A kind of flexible measurement method based on orthogonal component optimal selection and optimum regression |
| CN108520111B (en) * | 2018-03-06 | 2021-11-30 | 宁波大学 | Soft measurement method based on optimal selection and optimal regression of orthogonal components |
| CN108844624B (en) * | 2018-06-01 | 2020-01-21 | 北京科技大学 | SLM process laser power monitoring method based on temperature field |
| CN108844624A (en) * | 2018-06-01 | 2018-11-20 | 北京科技大学 | A kind of SLM process laser power monitor method based on temperature field |
| CN109063729A (en) * | 2018-06-20 | 2018-12-21 | 上海电力学院 | A kind of Multisensor Image Fusion Scheme based on PSO-NSCT |
| CN112752960A (en) * | 2018-09-21 | 2021-05-04 | 赖茵豪森机械制造公司 | Determining a characteristic temperature of an electrical or electronic system |
| CN112752960B (en) * | 2018-09-21 | 2024-05-17 | 赖茵豪森机械制造公司 | Determining characteristic temperatures of electrical or electronic systems |
| CN110991011A (en) * | 2019-11-14 | 2020-04-10 | 广东电网有限责任公司 | Transformer real-time hot spot temperature prediction method |
| CN110991011B (en) * | 2019-11-14 | 2023-05-02 | 广东电网有限责任公司 | Transformer real-time hot spot temperature prediction method |
| CN111401657B (en) * | 2020-04-02 | 2022-07-19 | 国网冀北电力有限公司唐山供电公司 | Transformer hot spot temperature time sequence prediction method based on support vector regression algorithm |
| CN111401657A (en) * | 2020-04-02 | 2020-07-10 | 国网冀北电力有限公司唐山供电公司 | Transformer hot spot temperature time sequence prediction method based on data mining algorithm |
| WO2022069149A1 (en) * | 2020-09-30 | 2022-04-07 | Maschinenfabrik Reinhausen Gmbh | Device and system for indirectly determining the temperature of a power transformer |
| CN112381123A (en) * | 2020-10-30 | 2021-02-19 | 鄂尔多斯应用技术学院 | Coal mining machine fault prediction method based on time series prediction method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106844972A (en) | Transformer Winding temperature flexible measurement method based on PSO SVR | |
| Ge et al. | A hybrid model for short-term PV output forecasting based on PCA-GWO-GRNN | |
| Zhang et al. | A comprehensive wind speed prediction system based on Monte Carlo and artificial intelligence algorithms | |
| Homod et al. | A novel hybrid modelling structure fabricated by using Takagi-Sugeno fuzzy to forecast HVAC systems energy demand in real-time for Basra city | |
| CN103218675A (en) | Short-term load prediction method based on clustering and sliding window | |
| CN104636985A (en) | Method for predicting radio disturbance of electric transmission line by using improved BP (back propagation) neural network | |
| Kavitha et al. | Improved harris hawks optimization with hybrid deep learning based heating and cooling load prediction on residential buildings | |
| Islam et al. | Non-linear autoregressive neural network approach for inside air temperature prediction of a pillar cooler | |
| Zeng et al. | Short-term load forecasting of smart grid systems by combination of general regression neural network and least squares-support vector machine algorithm optimized by harmony search algorithm method | |
| Cuevas et al. | An optimization algorithm guided by a machine learning approach | |
| Kiran et al. | Tree-seed programming for modelling of Turkey electricity energy demand | |
| Ren et al. | Power load forecasting based on support vector machine and particle swarm optimization | |
| CN116526473A (en) | Particle swarm optimization LSTM-based electrothermal load prediction method | |
| Wang et al. | The improved least square support vector machine based on wolf pack algorithm and data inconsistency rate for cost prediction of substation projects | |
| Li et al. | Improved sparrow search algorithm with the extreme learning machine and its application for prediction | |
| Zhang et al. | Short-Term Load Forecasting Based on DBO-LSTM Model | |
| Zhao et al. | Annual Energy Consumption Forecasting Based on PSOCA‐GRNN Model | |
| Ling et al. | Predicting the temperature dynamics of scaled model and real-world IoT-enabled smart homes | |
| Dan et al. | Application of machine learning in forecasting energy usage of building design | |
| Ding et al. | A statistical upscaling approach of region wind power forecasting based on combination model | |
| CN115759458A (en) | Load prediction method based on comprehensive energy data processing and multi-task deep learning | |
| Ramsami et al. | Neural Network Frameworks for Electricity Forecasting in Mauritius and Rodrigues Islands | |
| Li et al. | Residential Photovoltaic Power Forecasting Considering Division of Weather Type Index Interval | |
| Huang et al. | Short-term load forecasting based on similar day approach and intelligent algorithm using analytic hierarchy process | |
| Elahe et al. | An adaptive and parallel forecasting strategy for short-term power load based on second learning of error trend |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170613 |
|
| RJ01 | Rejection of invention patent application after publication |