CN103410660B - Wind-power generating variable pitch Learning Control Method based on support vector machine - Google Patents

Abstract

The present invention is directed to wind generator system variable pitch control problem, it is proposed that a kind of sliding formwork variable pitch control method based on support vector machine.Study control is divided into two steps, and the first step uses conventional sliding mode controller (SMC) to be controlled, and SVM SMC controller obtains structure and the preliminary parameters of controller by support vector machine study.When study arrives to a certain degree and the approximate error of SMC is less than SVM SMC a threshold value, pitch-variable system is switched to SVM SMC and controls.Second step uses discovery mechanism, and actual controlled quentity controlled variable is made up of plus the random disturbance of the normal distribution that average is zero output of SVM SMC controller, obtains learning sample according to estimated performance index, carries out real-time optimization by on-line learning algorithm to controlling parameter.The present invention not only have that sliding formwork controls anti-interference, to advantages such as running parameter strong robustness and speed are fast, while guaranteed output output is stable near rated value, it is achieved the steady regulation of propeller pitch angle, alleviate the abrasion of unit fatigue strength and inter-module.

Description

Wind-power generating variable pitch Learning Control Method based on support vector machine

Technical field

The present invention relates to the control method of a kind of technical field of wind power generation, specifically, relate to a kind of based on propping up Hold the award setting method of vector machine.

Background technology

Variable-pitch system of wind turbine generator passes through pitch controller, completes the control at vanepiston angle, more than incision wind speed Time in the range of below rated wind speed, keep wind energy conversion system propeller pitch angle constant, make wind energy conversion system operate in by changing motor speed Good tip speed ratio gets off to realize maximal wind-power tracking control；Time more than rated wind speed to cut-out wind speed, rotating speed is made to maintain volume Determine near rotating speed, make generating set output keep power invariability by regulation propeller pitch angle, when wind speed is more than cut-out wind speed, carry out Stoppage protection.

Due to randomness, the time variation of Wind turbines parameter of wind speed, it is negative that tower shadow, wind shear, driftage revolution etc. cause Carry disturbance, the reciprocating action of feather regulation oar, drive the inertial element that big quality impeller loads so that variable-pitch control system There is the features such as parametrical nonlinearity, parameter time varying, hysteresis quality, cause the instability of Wind turbines output.Additionally, along with The increase of wind power generating set single-machine capacity, the wind energy conversion system fatigue failure that unbalanced load causes becomes wind-powered electricity generation operation expense Main source.How to reduce wind energy conversion system fatigue failure and abrasion, the service life extending each parts is control field technology people Member is presently required consideration and solves the technical problem that.

Summary of the invention

Technical problem: the invention provides a kind of sliding formwork variable pitch control method based on Online SVM, comprehensively Sliding formwork controls and the advantage of support vector machine, utilizes the Online SVM algorithm to realize controller self-learning function, effectively Ground weakens to be buffeted, it is adaptable to this kind of indeterminate of wind-power generating variable pitch system is unable to estimate, model and be difficult to accurately set up and transport The control problem that during row, inside and outside disturbance is stronger.

Technical scheme: in order to overcome the problems referred to above, controls sliding formwork and Online SVM algorithm combines, it is not necessary to Line identification objects model, makes up the deficiency of traditional method so that faster system response, strong robustness, has good dynamic quality, Ensure Wind turbines normally, efficiently and reliably run.

A kind of wind-power generating variable pitch Learning Control Method based on support vector machine, uses based on support vector machine Sliding formwork controls, with wind speed wind direction sensor collection about wind speed and direction data signal, when wind velocity signal overrate and meet During wind energy conversion system service condition, then start the feather regulation of blower fan, measure unit output, try to achieve sliding formwork according to power deviation Input variable；Described Learning Control Method is realized by variable-pitch control system, and including two benches, described two benches is:

Stage 1: control early stage use sliding mode controller output feather object is controlled, and online acquisition each adopt The input of the sliding mode controller in sample moment, output and power deviation form data pair, the input of described sliding mode controller and output It is sliding variable s respectively_iWith reference pitch angle beta_i ^*, described data are to being (s_i, β_i ^*)；Through the assessment cycle set and through logarithm According to adding the training sample set D supporting vector learning controller after carrying out Effective judgement；

Described sliding mode controller is called for short SMC controller；Described support vector learning controller is called for short SVM-SMC controller；

SVM-SMC controller passes through support vector machine learning algorithm simultaneously, enters training sample data according to object function Row study obtains structure and the preliminary parameters of SVM-SMC controller；

By the learning functionality matching Equivalent Sliding Mode control action of support vector machine, obtain the controlled quentity controlled variable of feather object, I.e. with reference to propeller pitch angle

$U=f\left(x\right)=\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\θ}}_{i}K(x_i,x)+b$

Wherein, b is biasing；θ_i=α_i-α^* _i, α_iAnd α^* _iFor Lagrange multiplier；K(x_i, x) it is kernel function.

Stage 2: when study reaches to a certain degree and SVM-SMC controller is little to the approximate error of conventional sliding mode controller When a threshold value, the input of variable-pitch control system is switched to the output of SVM-SMC controller, enters the self study stage；Described The self study stage is SVM-SMC controller parameter real time correction algorithm based on Online SVM, new number of training Constantly it is sequentially generated alternately with controlled device by SVM-SMC controller according to needs, newly-increased data is carried out effectively simultaneously Property judge, according to newly obtained training sample set D, utilize on-line learning algorithm that control structure and parameter are carried out real-time school Just, its step is as follows:

Step1 receives new data, and structure new data is to (s_c,β_c)；Described s_cAnd β_cIt is sliding variable and propeller pitch angle respectively；

Step2 judges data validity after the assessment cycle set, if effectively, adds training sample set D；No Then abandon data and turn Step1；

Step3 uses batch type or increment type support vector machine training algorithm to carry out on-line training；

Step4 calculates SVM-SMC controller output U_svm；

Step5 is with U_svmFor average, calculate the disturbed value for exploring by normal distribution；

Step6 adds disturbed value as propeller pitch angle reference value using the output of SVM-SMC controller, controls feather object System；

Step7 waits that online data updates, and turns Step1；

Constantly new data is joined support vector machine training sample set D, by increase in each iterative process Sample learns, on-line tuning SVM-SMC controller parameter, it is achieved continuing to optimize of system.

Acquisition methods and the Effective judgement method of described training sample data be,

The quality of training sample data determines the quality of the SVM-SMC controller parameter that study obtains, it is desirable to training sample Notebook data can reflect controlled device Changing Pattern during controlling, and system can be made again to have good response characteristic；

γ=(1-δ) E{ (P-P_rated)}/E{|Δv|³}+δE{(Δβ)²}≤ζ

Wherein, γ is the effective threshold value of learning data, P and P_ratedBeing respectively real output and rated power, Δ β is oar Elongation changes, the absolute value that | Δ v | changes for wind speed；P and P_ratedIt is perunit value；E{ } it is to set assessment cycle T_vBracket The average statistical of interior correlated variables, T_vFor the integral multiple in sampling period, herein refer to E{ } correlated variables in bracket, ζ is The threshold value of definition；γ expression formula Section 1 (1-δ) E{ (P-P_rated)}/E{|Δv|³Represent that desired output power is attached at rated value Near fluctuation is minimum, because changed power and wind speed are changing into cube relation, therefore introduces wind speed change cube absolute value；The Binomial δ E{ (Δ β)²Representing that the change of expectation propeller pitch angle minimizes, the balance between these two is regulated by weight coefficient δ, only When the power of wind power system exports and pitch change meets formula threshold condition, just by new data to adding training sample set D； Controlling preliminary stage 1, the target of SVM-SMC controller primarily focus on the rejection of output and controller architecture and The quick obtaining of parameter, sets δ as less value, and ζ is constant；Enter the stage 2 controlled, owing to now actual change oar performs The change oar signal of mechanism is determined by SVM-SMC controller and disturbed value, is tuned up by weight coefficient δ, and ζ is according to the statistical average of γ Value constantly reduces, and reduces the fatigue load that blade vibration produces while ensureing power limitation control.

Described SVM-SMC controller parameter real time correction algorithm based on Online SVM can utilize computing above As a result, computation complexity is reduced, it is achieved the new samples study under less time cost；Online SVM instruction can be improved Practicing the time, and can improve the adaptation ability of SVM-SMC controller, when making ambient conditions change, model can become accordingly Change.

Described wind-power generating variable pitch Learning Control Method based on support vector machine also includes the exploration machine used System；

Normal distribution is a kind of important continuous distribution, and its distribution curve is that peak is centrally located, and both sides are gradually reduced, Curve that is symmetrical, that do not intersect with transverse axis, the density function of distribution is,

$f\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}{e}^{-{(x-u)}^2/\left(2{\mathrm{\σ}}^2\right)}$

$\mathrm{\σ}=\frac{k_{\exp }}{1+1/\mathrm{\λ}}$

Wherein, u is average, and σ is population standard deviation, i.e. explores scope；

Described discovery mechanism is, if SVM-SMC controller controls effect preferably, namely controls recruitment evaluation value and meets Threshold values condition, then reduce exploration scope σ；As controlled effect, can suitably increase exploration scope σ, but total exploration scope limits Make system stability little within the scope of, be used for exploring by normal distribution calculation perturbation value, with SVM-SMC controller output add Feather object, as propeller pitch angle reference value, is controlled by disturbed value, by with system interaction, voluntarily explore make control strategy Continue to optimize, thus obtain globally optimal solution.

Beneficial effect

Award setting method of the present invention not only has the anti-interference, to running parameter strong robustness and speed of sliding formwork control The advantages such as degree is fast, and multiple optimized Control Mode can be learnt, propeller pitch angle steadily regulates, and reduces torque oscillation and cabin shakes Swing, not only optimize output, and the noise effectively reduced, improve the force-bearing situation of blade and complete machine, alleviate unit Fatigue strength and the abrasion of inter-module.

Accompanying drawing explanation

Fig. 1 is sliding formwork variable-pitch control system structure chart based on Online SVM.

Detailed description of the invention:

The sliding formwork variable pitch control method based on Online SVM that the present invention proposes combines accompanying drawing and is embodied as Details are as follows for mode:

The present invention uses wind speed wind direction sensor collection about data signals such as wind speed, and the state being sent in cabin is adopted Collection PLC system (or special processor) processes, and passes to main control computer by Profibus-DP bus, it is judged that unit Running status.When wind velocity signal overrate and when meeting wind energy conversion system service condition, then start the feather regulation of blower fan；Survey Amount unit output, tries to achieve sliding formwork input variable s according to power deviation, and control system is divided into two benches to learn, and controls System construction drawing is as shown in Figure 1.

First study stage (having teacher learning) uses sliding mode controller to be controlled feather object.At this moment SVM- SMC controller is in the study stage, by input (the sliding variable s of sliding mode controller_i) and output (reference pitch angle beta_i ^*) after conjunction Form the training sample (s of SVM-SMC controller_i, β_i ^*), add training sample set D.

From alleviate mechanical load and firm power export two in terms of definition judge the effective threshold condition of learning data:

γ=(1-δ) E{ (P-P_rated)}/E{|Δv|³}+δE{(Δβ)²}≤ζ

Wherein, P and P_ratedBeing respectively real output and rated power, Δ β is propeller pitch angle change, and | Δ v | is for wind speed The absolute value of change, P and P_ratedIt is perunit value.E{ } it is to set assessment cycle T_vThe average statistical of correlated variables in bracket, T_vFor the integral multiple in sampling period, herein refer to E{ } correlated variables in bracket, ζ is the threshold value of definition, and δ is less than 1 Weight coefficient.

Only when the power of wind power system exports and pitch change meets formula threshold condition, just by new data to adding sample This set D.Control the early stage first stage, the target of SVM-SMC controller primarily focus on output rejection and Controller architecture and the quick obtaining of parameter, therefore set δ as less value, and ζ be constant.Enter the second stage controlled, by Determined by SVM-SMC controller and disturbance quantity in the now actual change oar signal becoming oar actuator, weight coefficient δ is adjusted Greatly, and ζ constantly reduces according to the assembly average of γ.The fatigue that blade vibration produces is reduced while ensureing power limitation control Load.

SVM-SMC controller passes through support vector machine learning algorithm simultaneously, to sample data according to object function Acquistion is to the structure of SVM-SMC controller and preliminary parameters.Controlled by the learning functionality matching Equivalent Sliding Mode of support vector machine Effect, obtains the input controlled quentity controlled variable of feather object, i.e. with reference to propeller pitch angle:

$U=f\left(x\right)=\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\θ}}_{i}K(x_i,x)+b$

Wherein, b is biasing；θ_i=α_i-α^* _i, α_iAnd α^* _iFor Lagrange multiplier.Polynomial kernel function, radially base can be used Function (RBF) kernel function, Sigmoid kernel function etc., use RBF RBF kernel function here, Its center is for supporting vector, and σ is the width of gaussian kernel function.

When study reach to a certain degree and SVM-SMC controller to the approximate error of conventional sliding mode controller less than one During threshold value, pitch-variable system is switched to SVM-SMC controller and controls, and enters for the second study stage, i.e. self study stage.New instruction Practice sample to need constantly to be sequentially generated alternately with controlled device by controller, newly-increased sample data is carried out effectively simultaneously Property judge, owing to the now actual change oar signal becoming oar actuator is directly given by SVM-SMC controller, by weight coefficient δ tunes up, and reduces the fatigue load that blade vibration produces while ensureing power limitation control.

The sample increased in each iterative process can be learnt, before utilization once by Online SVM learning method The operation result of iteration, it is possible to reduce computation complexity, it is achieved the new samples study under less time cost.Therefore basis Newly obtained sample set, utilizes Online SVM algorithm that controller architecture and parameter are carried out real time correction.The present invention adopts Sliding formwork variable pitch control based on increment type Online SVM to realize process as follows:

Step1 receives new data, constructs new samples (s_c,β_c), s here_cAnd β_cIt is sliding variable and propeller pitch angle respectively；

Step2 judges data validity after the assessment cycle set, if effectively, adding training sample set；Otherwise lose Abandon data and turn Step1；

Step3 uses algorithm of support vector machine to carry out on-line training；

Step4 calculates SVM-SMC controller output U_svm；

Step5 is with U_svmFor average, calculate the disturbed value for exploring by normal distribution；

Using the output of SVM-SMC controller, Step6 adds that feather object, as propeller pitch angle reference value, is controlled by disturbed value System；

Step7 uses and successively decreases forgetting algorithm, delete in training set required for the sample forgotten；

Step8 waits that online data updates, and turns Step1.

Above-mentioned algorithm can carry out there is " selective " or " directional " on-line training, is ensureing that online support vector is non- On the basis of linear approximation ability, meet the systematic training requirement under different phase, reduce on-line training process amount of calculation simultaneously. For allow the controller to by with system interaction, explore voluntarily and make control strategy continue to optimize, thus obtain globally optimal solution, this Bright discovery mechanism is joined in control strategy.Normal distribution is a kind of important continuous distribution, and its distribution curve is peak Centrally located, both sides are gradually reduced, symmetrical, the curve not intersected with transverse axis.The density function of distribution:

$f\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}{e}^{-{(x-u)}^2/\left(2{\mathrm{\σ}}^2\right)}$

$\mathrm{\σ}=\frac{k_{\exp }}{1+1/\mathrm{\λ}}$

Wherein u is average, and σ is that population standard deviation i.e. explores scope.If it is preferable that SVM-SMC controller controls effect, also It is exactly to control recruitment evaluation value to meet threshold values condition, then reduces exploration scope σ, as controlled effect, can suitably increase exploration Scope σ, but total exploration scope be limited in system stability little within the scope of.In the present invention, the disturbance calculated by normal distribution Value is used for exploring, and using the output of SVM-SMC controller plus disturbed value as propeller pitch angle reference value, controls feather object System.

The above-mentioned preferable realization implementing the simply present invention, certainly, the present invention also can have other various embodiments, In the case of without departing substantially from present invention spirit and essence thereof, those of ordinary skill in the art are various when making according to the present invention Corresponding change and deformation, but these changes accordingly and deformation all should belong to the scope of the claims of the present invention.

Claims (3)

1. a wind-power generating variable pitch Learning Control Method based on support vector machine, it is characterised in that: described self study Control method uses sliding formwork based on support vector machine to control, and believes about wind speed and direction data with wind speed wind direction sensor collection Number, when wind velocity signal overrate and when meeting wind energy conversion system service condition, then start the feather regulation of blower fan, measure unit Output, tries to achieve sliding formwork input variable according to power deviation；Described Learning Control Method is realized by variable-pitch control system, Including two benches, described two benches is:

Stage 1: control early stage and use sliding mode controller output that feather object is controlled, and during each sampling of online acquisition The input of sliding mode controller, output and the power deviation carved form data pair, and the input of described sliding mode controller and output are respectively It is sliding variable s_iWith reference pitch angle beta_i ^*, described data are to being (s_i, β_i ^*)；Through the assessment cycle set and through data are entered The training sample set D supporting vector learning controller is added after row Effective judgement；

Described sliding mode controller is called for short SMC controller；Described support vector learning controller is called for short SVM-SMC controller；

SVM-SMC controller passes through support vector machine learning algorithm simultaneously, to training sample data according to object function Acquistion is to the structure of SVM-SMC controller and preliminary parameters；

By the learning functionality matching Equivalent Sliding Mode control action of support vector machine, obtain the controlled quentity controlled variable of feather object, i.e. join Examine propeller pitch angle

$U=f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\θ}}_{i}K(x_i,x)+b$

Wherein, b is biasing；θ_i=α_i-α^* _i, α_iAnd α^* _iFor Lagrange multiplier；K(x_i, x) it is kernel function；

Stage 2: when study reach to a certain degree and SVM-SMC controller to the approximate error of conventional sliding mode controller less than one During individual threshold value, the input of variable-pitch control system is switched to the output of SVM-SMC controller, enters the self study stage；Described self-study The habit stage is SVM-SMC controller parameter real time correction algorithm based on Online SVM, and new training sample data need SVM-SMC controller to be passed through is constantly mutual with controlled device and is sequentially generated, and newly-increased data is carried out effectiveness simultaneously and sentences Disconnected, according to newly obtained training sample set D, utilize on-line learning algorithm that control structure and parameter are carried out real time correction, its Step is as follows:

Step1 receives new data, constructs new data pair, and this new data is to being (s_c,β_c)；Described s_cAnd β_cBe respectively sliding variable and Propeller pitch angle；

Step2 judges data validity after the assessment cycle set, if effectively, adds training sample set D；Otherwise lose Abandon data and turn Step1；

Step3 uses batch type or increment type support vector machine training algorithm to carry out on-line training；

Step4 calculates SVM-SMC controller output U_svm；

Step5 is with U_svmFor average, calculate the disturbed value for exploring by normal distribution；

Step6 adds disturbed value as propeller pitch angle reference value using the output of SVM-SMC controller, is controlled feather object；

Step7 waits that online data updates, and turns Step1；

Constantly new data is joined support vector machine training sample set D, by the sample increased in each iterative process Learn, on-line tuning SVM-SMC controller parameter, it is achieved continuing to optimize of system.

Wind-power generating variable pitch Learning Control Method based on support vector machine the most according to claim 1, its feature It is: acquisition methods and the Effective judgement method of described training sample data be,

The quality of training sample data determines the quality of the SVM-SMC controller parameter that study obtains, it is desirable to number of training According to controlled device Changing Pattern during controlling can be reflected, system can be made again to have good response characteristic；

γ=(1-δ) E{ (P-P_rated)}/E{|Δv|³}+δE{(Δβ)²}≤ζ

Wherein, γ is the effective threshold value of learning data, P and P_ratedBeing respectively real output and rated power, Δ β is propeller pitch angle Change, the absolute value that | Δ v | changes for wind speed；P and P_ratedIt is perunit value；E{ } it is to set assessment cycle T_vBracket internal phase Close the average statistical of variable, T_vFor the integral multiple in sampling period, herein refer to E{ } correlated variables in bracket, ζ is definition Threshold value；γ expression formula Section 1Represent that desired output power is near rated value Fluctuation minimum, because changed power and wind speed are changing into cube relation, therefore introduces wind speed change cube absolute value；Section 2 δE{(Δβ)²Representing that the change of expectation propeller pitch angle minimizes, the balance between these two is regulated by weight coefficient δ, only works as wind When the power output of electricity system and pitch change meet formula threshold condition, just by new data to adding training sample set D；In control Preliminary stage 1 processed, the target of SVM-SMC controller primarily focuses on the rejection of output and controller architecture and parameter Quick obtaining, set δ as less value, and ζ be constant；Enter the stage 2 controlled, due to now actual change oar actuator Change oar signal determined by SVM-SMC controller and disturbed value, weight coefficient δ is tuned up, and ζ is according to the assembly average of γ not Break and reduce, while ensureing power limitation control, reduce the fatigue load that blade vibration produces.

Wind-power generating variable pitch Learning Control Method based on support vector machine the most according to claim 1, its feature It is: described SVM-SMC controller parameter real time correction algorithm based on Online SVM can utilize computing knot above Really, computation complexity is reduced, it is achieved the new samples study under less time cost；Online SVM training can be improved Time, and the adaptation ability of SVM-SMC controller can be improved, when making ambient conditions change, model can change accordingly.