CN102411728A - Transformer hot point temperature on-line monitoring method based on mixed model - Google Patents

Abstract

The invention relates to a transformer on-line soft measurement method, and provides a more economical and more reliable method in order to solve the problems of low temperature and the overheat of a transformer winding for monitoring a hot point value in real time on line. The transformer hot point temperature on-line monitoring method based on the mixed model adopts a soft measurement method for mixed modeling with a mechanism model and a neural network model, uses a radial basis function (RBF) neural network as an estimator of errors in mechanism modeling, reflects the dynamic information which the mechanism model cannot describe, estimates the errors in mechanism modeling, adds to the mechanism model, compensates the mechanism model, so that the precision of the model is greatly improved, and hot point temperature estimated values with higher precision are provided.

Description

Hot-spot temperature of transformer on-line monitoring method based on mixture model

Technical background

Power transformer is the high voltage equipment of electric energy transmitting in the electric system---its performance quality directly influences safety, the stable operation of electric system.Numerous discovering, Transformer Winding temperature, especially winding focus, become transformer safety, economical operation and serviceable life deciding factor.Most of transformer's life span termination is to have lost due insulating capacity because of it, and the winding temperature of the main factor that influence insulating capacity when to be transformer move.Winding temperature when using winding temperature gage to monitor operation helps transformer to reach the designed life of expection.So the problem of Transformer Winding cryogenic overheating is one of operation and manufacturing sector's letter major issue to be researched and solved.Power transformer insulation system in operational process is bearing the aging of chemical property always, this process be accumulation property and can cause insulation system to lose insulating property harm.For the transformer of A class B insulation, 6 ℃ of the every increases of temperature, insulation ag(e)ing rate doubles.Along with the development of electric utility, the high capacity EHV transformer has obtained developing rapidly and using.Safe and reliable to operation in order to ensure transformer, in the serviceable life of prolongation transformer, the temperature of measuring transformer winding has seemed that branch is important.At present domestic have plenty of with the transformer top-oil temperature as the switching signal that ensures the transformer safe operation; The basis for estimation of layer oil temperature, yet ensure that through the monitoring top-oil temperature this method of safe operation of transformer has many weak points as the transformer control device.This is because the high-power transformer top-oil temperature obviously lags behind the winding oil temperature.When transformer load increased fast, because heat is transmitted response speed, the transformer top-oil temperature needed can reflect through several hours the working conditions change of winding.Obviously, number can not protect the safe and reliable operation of transformer timely and effectively as the switching that ensures the transformer safe operation with the transformer top-oil temperature.

For directly measuring the internal temperature of transformer in service, need be placed on transformer inside to sensor.General employing Fibre Optical Sensor, but the used optical fiber checkout equipment price ten minutes costliness of the direct method of measurement.When the winding hotspot location was uncertain, blindly embedding optical fiber measuring point can bring sizable measuring error.Coiling hot point of transformer temperature rise estimation indirect calculation method is the classical way of coiling hot point of transformer temperature rise estimation, has certain precision.But the temperature rise of oil increases by linear from bottom to top in this method hypothesis transformer, and the lead average temperature rising of winding increases by linear from bottom to top, and these hypothesis make result of calculation and actual conditions have certain error.Therefore flexible measurement method based on mechanism model and neural network model hybrid modeling has been proposed.

Can resolve into for any real system and can describe and unknown two parts, wherein only can describe part and can describe, Here it is the said mechanism model in front with mathematics, physical model.Because in reality, people are incomplete to the understanding of real system, the simplification of being done is handled, and all makes between mechanism model and the real system and exists modeling error.It is that part is known by system, can regard external disturbance, the coefficient result of internal disturbance as.As long as therefore can estimate the depanning error, and it is added on the mechanism model, the precision that makes model is greatly improved.The characteristic that neural network identification is all makes it be suitable for doing the estimator of modeling error.Can obtain the hybrid modeling method of neural network and mechanism model based on this thought.Here neural network can adopt the multidate information that the multilayer feedforward network fails to describe with the reflection mechanism model, and it plays the effect of compensatory michanism model, and it is main neural networks compensate model that this model is called mechanism.The advantage of mechanism---neural network hybrid modeling method is the main rule (having utilized prior imformation) that had both reflected real system; Unknown disturbance or uncertain influence have been embodied again to real system; Have than the higher precision of simple use mechanism model, higher than simple neural net model establishing reliability.

Summary of the invention

To the above-mentioned defective and the deficiency of existing method, the present invention is incorporated into flexible measurement method the on-line monitoring of transformer.This method can be fast, real-time, accurately the focus of transformer is monitored.

The present invention proposes a kind of method of on-line monitoring of hot-spot temperature of transformer, and its content comprises:

(1) adopt flexible measurement method to realize the on-line monitoring of transformer.

(2) flexible measurement method is based on the flexible measurement method of mechanism model and neural network model hybrid modeling.

(3) its mechanism model can be according to the actual working environment of transformer, and concrete parameter is different with performance and adjust, revise and improve.Its neural network model can adopt suitable neural network model as the case may be.

(4) the hot(test)-spot temperature mechanism model is under the arbitrary load:

${\mathrm{\θ}}_h={\mathrm{\θ}}_a+{\mathrm{\Δ\θ}}_{\mathrm{br}}{\left(\frac{1+{\mathrm{RK}}^2}{1+R}\right)}^X+2({\mathrm{\Δ\θ}}_{\mathrm{imr}}-{\mathrm{\Δ\θ}}_{\mathrm{br}})K^y+H_g{K}^y$

Wherein, θ _hIt is hot(test)-spot temperature;

θ _aIt is environment temperature;

Δ θ _BrCooler outlet oil temperature rise under the rated condition;

Δ θ _ImrThe winding oil average temperature rising;

The y winding coefficient;

K load factor (K=I/I _H);

The index of x oil;

The R loss ratio;

Hg is the temperature difference of focus to the winding top.

(5) neural network model adopts the RBF neural network model.Also can adopt other feedforward neural network models.This paper is example with the RBF neural network model.

(6) according to claim 1,2,3; 4 is said, it is characterized in that: make the estimator of modelling by mechanism error with the RBF neural network, the multidate information that the reflection mechanism model fails to describe; Estimate the modelling by mechanism error; And it is added on the mechanism model, the compensatory michanism model greatly improves the precision of model.Its structure of block diagram is following:

(7) right 5 is said, is main with mechanism model, and the RBF neural network model is the structure of compensation mixture model, and its process flow diagram is characterized as:

(8) according to right 6, the basic variable of its selection is: load current I; θ _aEnvironment temperature; Δ θ _BrCooler outlet oil temperature rise under the rated condition; Δ θ _ImrThe winding oil average temperature rising; The y winding coefficient; K load factor (K=I/I _H); The index of X oil; The R loss ratio; Hg is the temperature difference of focus to the winding top.

(9) according to right 6, its characteristics are: the basis function of RBF network adopts Gaussian function; The learning center of basis function obtains with the K-means clustering algorithm; The LMS algorithm is adopted in the study of weights.

The K-means clustering algorithm: concrete steps are following:

The first step, the initialization cluster centre, promptly rule of thumb from training sample concentrate at random choose I sample as initial center t _i(0) (i=1,2 ..., I), iteration step number n=0 is set.

In second step, import training sample heart X at random _K

In the 3rd step, seek training sample X _KNearest from which center, promptly find i (X _K) make it satisfy i (X _K)=argmin||X _K-t _i(n) ||, i=1,2 ..., t in the I formula _IjI center of basis function when (n), being the n time iteration.

The 4th step, adjustment center formula:

The 5th step judged whether to finish all training samples and central distribution not changing, and be then to finish, otherwise n=n+1 forwarded for second step to.The t that obtains at last _i(n) be the center of the final basis function of RBF network.

The basis function variance is confirmed: ${\mathrm{\σ}}_1={\mathrm{\σ}}_2=L-{\mathrm{\σ}}_I=\frac{d_{\mathrm{Max}}}{\sqrt{2I}}$

I is the hidden unit number, d _MaxBe the ultimate range between the selected center.

Description of drawings

Accompanying drawing 1 system architecture diagram

Accompanying drawing 2 system architecture process flow diagrams

Claims (9)

1. the on-line monitoring method of a hot-spot temperature of transformer is characterized in that: adopt flexible measurement method to realize the on-line monitoring of transformer.

2. method according to claim 1 is characterized in that flexible measurement method is based on the flexible measurement method of mechanism model and neural network model hybrid modeling.

3. method according to claim 2 is characterized in that: (1) its mechanism model can be according to the actual working environment of transformer, and concrete parameter is different with performance and adjust, revise and improve.(2) its neural network model can adopt suitable neural network model as the case may be.

4. method according to claim 3 is characterized in that:

(1) the hot(test)-spot temperature mechanism model is under the multithread road transformer arbitrary load:

${\mathrm{\θ}}_h={\mathrm{\θ}}_a+{\mathrm{\Δ\θ}}_{\mathrm{br}}{\left(\frac{1+{\mathrm{RK}}^2}{1+R}\right)}^X+2({\mathrm{\Δ\θ}}_{\mathrm{imr}}-{\mathrm{\Δ\θ}}_{\mathrm{br}})K^y+H_g{K}^y$

Wherein, θ _hIt is hot(test)-spot temperature;

θ _aIt is environment temperature;

Δ θ _BrCooler outlet oil temperature rise under the rated condition;

Δ θ _ImrThe winding oil average temperature rising;

The y winding coefficient;

K load factor (K=I/I _H);

The index of x oil;

The R loss ratio;

Hg is the temperature difference of focus to the winding top.

(2) neural network model adopts the RBF neural network model.Also can adopt other feedforward neural network models.This paper is example with the RBF neural network model.

5. according to claim 1,2,3; 4 is said, it is characterized in that: make the estimator of modelling by mechanism error with the RBF neural network, the multidate information that the reflection mechanism model fails to describe; Estimate the modelling by mechanism error; And it is added on the mechanism model, the compensatory michanism model greatly improves the precision of model.

6. right 5 is said, is main with mechanism model, and the RBF neural network model is the structure of compensation mixture model.

7. according to right 6, the basic variable of its selection is: load current I; θ _aEnvironment temperature; Δ θ _BrCooler outlet oil temperature rise under the rated condition; Δ θ _ImrThe winding oil average temperature rising; The y winding coefficient; K load factor (K=I/I _H); The index of X oil; The R loss ratio; Hg is the temperature difference of focus to the winding top.

8. according to right 6, its characteristics are: the basis function of RBF network adopts Gaussian function; The learning center of basis function obtains with the K-means clustering algorithm; The LMS algorithm is adopted in the study of weights.

9. according to right 6, the algorithm end condition can precision according to actual needs be provided with, or maximum iteration time is set.

Images