CN101145750A - Multi-model integrated intelligent control method of large generator group - Google Patents

Abstract

The invention discloses a multi-model integrated intelligent controlling method of a heavy-duty generator unit. The method comprises (1) constructing a model library and a controller library, wherein the model library comprises N submodels corresponding to the N operation conditions of the generator unit, the controller library comprises N submodel fuzzy controllers, each is designed a control rule corresponding to one submodel; each submodel fuzzy controller is the controller with double inputs and double outputs, the inputs are power angle deviation amount e delta and terminal voltage deviation amount u2, and the outputs are excitation regulation amount uf and valve opening regulation amount u2; (2) acquiring the real-time operating conditions data of the generator unit, and judging the actual operating conditions value based on the variable (delta, Vt); and (3) calculating the matching degree fn of the actual operating conditions value (delta, Vt) and N submodels in the model library, which is used as a weighting coefficient wn (n is an integer) of the integration of the integrated controller. The invention has the advantages of simplicity, good reliability, good real-time performance, high control accuracy, and good practical performance.

Description

The multi-model integrated intelligent control method of Generator Set

Technical field

The present invention is mainly concerned with the control field of Generator Set, refers in particular to a kind of multi-model integrated intelligent control method of Generator Set.

Background technology

In the prior art, the security and stability problem of power system operation, promptly the dynamic security integrity problem is the major issue of power system operation, is subjected to expert's the attention of going together deeply both at home and abroad always, and has carried out a large amount of research and experimental work.Improving with the main means that improve stability of power system is to adopt novel electric power equipment and control mode.Generator Set is the key equipment in the power system operation, its effectively control help stable, safety, the economical operation of electric power system, also be a kind of effective link of improving power system transient stability.The Comprehensive Control that excitation control is regulated in conjunction with the porthole aperture is a kind of main trend of Generator Set control, for the stability and the running quality that improve electric power system significant impact is arranged, and also is main research contents of the present invention.

The excitation of generating set and porthole system are a multivariable, non-linear, cross coupling complicated system, some scholars have inquired into different generating set integrated control methods in recent years, as direct feedback linearization method, the backstepping decoupling control method, self adaptation backstepping method is based on the non-linear decentralized control method of Hamilton energy function, non-linear variable structure control method, Neural Network Adaptive Control method etc.Because these methods all depend on the Mathematical Modeling of generating set, be based upon that complex nonlinear is handled or repeatedly on the interative computation basis, be not easy to solve the practical problems such as strong nonlinearity, operation on a large scale and the continuous variation of operating mode in the generating set Comprehensive Control, and also be difficult to realize preferable decoupling zero for excitation and this coupled system of porthole.

Summary of the invention

The problem to be solved in the present invention just is: at the technical problem that prior art exists, the invention provides a kind of simple and reliable, real-time good, control precision is high, the multi-model integrated intelligent control method of the Generator Set being convenient to realize.

For solving the problems of the technologies described above, the solution that the present invention proposes is: a kind of multi-model integrated intelligent control method of Generator Set is characterized in that step is:

(1), sets up model library and controller storehouse: comprised N kind submodel in the model library, N kind operating condition corresponding to generating set, the controller storehouse is made up of N sub-model fuzzy controller, and each submodel fuzzy controller is at a kind of submodel design control law; Each submodel fuzzy controller is the controller of dual input dual output, and it is input as power angle deviation amount e _δWith terminal voltage departure e _Vt, be output as excitation regulation amount u _fWith porthole aperture regulated quantity u ₂

(2), actual condition data acquisition: gather the real-time working condition data of generating set, with variable (δ, V _t) weigh the actual condition value of generating set;

(3), the generation and the output of working control amount: calculate actual condition value (δ, V with fuzzy logic _t) with model library in the matching degree f of N submodel _n, with this as the integrated weight coefficient w of general controller _n(n=1,2 ..., N);

Set up variable (δ, V 1., at first respectively _t) membership function, be expressed as with formula (7):

${\mathrm{\μ}}_{\mathrm{\δn}}=\frac{1}{1+{\left(\frac{\mathrm{\δ}-c_{1n}}{a_{1n}}\right)}^2},$ ${\mathrm{\μ}}_{\mathrm{Vn}}=\frac{1}{1+{\left(\frac{V_t-c_{2n}}{a_{2n}}\right)}^2},$ n＝1，2，…，N (7)

C wherein _1n, c _2nRepresent variable (δ, V respectively _t) central value, and α _1n, α _2nRepresent variable (δ, V respectively _t) width;

2., according to degree of membership μ _{δ n}, μ _VnObtain the matching degree f of actual condition and submodel n _nFor:

f _n=μ _{δ n}* μ _Vn, f _n∈ [0,1] and $\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{f}_n=1---\left(8\right)$

Weight coefficient w ₁, w ₂..., w _NNumerically have

[w ₁, w ₂..., w _N] ^T=[f ₁, f ₂..., f _N] ^T, w _n∈ [0,1] and $\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{w}_n=1$ (n＝1，2，…，N)(9)

3., act on of the weighting of the working control amount of generating set for the output of submodel fuzzy controller:

${\mathrm{<figure-callout\; id="2"\; label="u"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">u</figure-callout>}}_2=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n*u_{2n},$ $u_f=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n*u_{\mathrm{fn}}---\left(10\right)$

The described idiographic flow of setting up model library and controller storehouse is:

1., primary data sample:

According to the different work condition states of the existing generator of measuring $\stackrel{\&RightArrow;}{\mathrm{\&delta;}}=\{{\mathrm{\&delta;}}_1,{\mathrm{\&delta;}}_2,...,{\mathrm{\&delta;}}_a\},$ $\stackrel{\&RightArrow;}{V_t}=\{{\mathrm{<figure-callout\; id="1"\; label="V\; t"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">V</figure-callout>}}_t,{\mathrm{<figure-callout\; id="2"\; label="V\; t"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">V</figure-callout>}}_t,...,V_{\mathrm{tb}}\},$ Choose (δ, V _t) arranging in pairs or groups obtains a * b=ab kind operating mode, as initial sample;

2., set up the IF-THEN rule according to expertise:

Under each operating mode, according to the default fuzzy control rule of expertise, adopt the fuzzy controller of dual input dual output, it is input as merit angle error amount e _δWith terminal voltage margin of error e _Vt, be output as porthole regulated quantity u ₂With excitation regulation amount u _f, fuzzy rule adopts " IF-THEN " fuzzy relation formula, all obtains 5 * 5=25 bar rule for each operating mode, that is:

rulel?IF?E _δis?PB?AND?E _Vt?is?PB?THEN?U ₂?is?NS?AND?U _f?is?NB；

rule?25?IF?E _δ?is?NB?AND?E _Vt?is?NB?THEN?U ₂?is?PB?AND?U _f?is?PS；

3., the IF-THEN rule is represented with regular matrix:

Use matrix Rule ₀Represent control law: ${\mathrm{Rule}}_{0j}={\left[\begin{array}{cccc}\mathrm{PB}& \mathrm{PB}& \mathrm{NS}& \mathrm{NB}\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \mathrm{NB}& \mathrm{NB}& \mathrm{PB}& \mathrm{PS}\end{array}\right]}_{25\&times;4,}$ (j=1,2 ..., ab), if with 1 ,-0.5,0,0.5,1} represent NB, NS, Z, PS, PB} then is equivalent to: ${\mathrm{Rule}}_{0j}^{\&prime;}={\left[\begin{array}{cccc}1& 1& 0.5& 1\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ -1& -1& 1& 0.5\end{array}\right]}_{25\&times;4;}$ With matrix Rule ₀' preceding two column skips, only stay back two row, i.e. 25 * 2 matrix, and the matrix of (25 * 2) is stretched as delegation (1 * 50), i.e. Rule _j=0.5,1 ..., 1,0.5} _{1 * 50}, this moment, the domain of object to be clustered was Rule={Rule ₁, Rule ₂..., Rule _Ab, each sample Rule _j(j=1,2 ..., ab) describe, i.e. Rule by 50 characteristic ginseng values _j={ Rule _J1, Rule _J2..., Rule _J50;

4., the standardization of regular matrix:

Adopt standardization characteristic ginseng value to be compressed to [0,1] interval:

${\mathrm{Rule}}_{\mathrm{ji}}^{\&prime;}=\frac{{\mathrm{Rule}}_{\mathrm{ji}}-{\mathrm{Rule}}_{\min }}{{\mathrm{Rule}}_{\max }-{\mathrm{Rule}}_{\min }}$ (j=1,2 ..., ab, i=1,2 ..., 50), so the object domain to be clustered after the standardization is Rule '={ Rule ₁', Rule ₂' ..., Rule _Ab';

5., demarcate:

Set up the fuzzy resembling relation R of object, the membership function mui R (Rule of R _i', Rule _j') expression Rule _i' with Rule _j' by the similarity degree of its character, R is similar matrix R=[r _Ij], r _Ij=μ _R(Rule ' _i, Rule ' _j), calculate r by the minimax method here _Ij:

$r_{\mathrm{ij}}=\left[\underset{k=1}{\overset{50}{\mathrm{\&Sigma;}}}\min ({\mathrm{Rule}}_{\mathrm{ik}}^{\&prime;},{\mathrm{Rule}}_{\mathrm{jk}}^{\&prime;})\right]/\left[\underset{k=1}{\overset{50}{\mathrm{\&Sigma;}}}\max ({\mathrm{Rule}}_{\mathrm{ik}}^{\&prime;},{\mathrm{Rule}}_{\mathrm{jk}}^{\&prime;})\right]---\left(1\right)$

6., fuzzy clustering:

According to fuzzy set theory, obtain U ₂, U _fMembership function value μ (U _2i), μ (U _Fi), then the controlled quentity controlled variable of each submodel fuzzy controller is:

$u_{2n}=\frac{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{2i}\right)\&CenterDot;U_{2i}}{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{2i}\right)},$ $u_{\mathrm{fn}}=\frac{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{fi}\right)\&CenterDot;U_{fi}}{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{\mathrm{fi}}\right)},$ n＝1，2，…，N (2)

After described step (1) finishes, learn to optimize after having set up model library and controller storehouse, the step of described study optimization is:

1., obtain N ruled surface by N controller storehouse;

To obtain model library and controller storehouse that scale is N after the cluster, each submodel fuzzy controller can be expressed as 2 ruled surfaces, i.e. U ₂About E _δWith E _VtRuled surface, U _fAbout E _δWith E _VtRuled surface, N controller storehouse is exactly 2N ruled surface, the input dimension of each ruled surface is 2, output dimension be 1;

2., 2N SVM network approaches 2N ruled surface;

Return by SVMs SVM and to approach ruled surface, fuzzy control rule is converted into SVM network weight parameter, be convenient to realize the online self study of control law by gradient algorithm; According to the controller storehouse that cluster obtains, set up learning sample collection { ((E ₁, E ₂), U) ..., ((E _1n, E _2n), U _n) _ R ²* R, the learning process that SVM returns is converted into a quadratic programming problem; Supported the vector ((E of learning outcome _1i, E _2i), U _i) and corresponding coefficient (α _i* α _i), use

Expression (E _1i, E _2i), thereby obtain regression function be:

$f\left(\stackrel{\&RightArrow;}{E}\right)=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)K\left({\stackrel{\&RightArrow;}{E},\stackrel{\&RightArrow;}{E}}_i\right)+b---\left(3\right)$

Wherein g is the support vector number, uses W _iExpression (α _i ^*-α _i), adopt the RBF kernel function: K (x _i, x _j)=exp (| x _i-x _j| ²/ σ ²), equation (6) can be expressed as:

$f\left(\stackrel{\&RightArrow;}{E}\right)\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}{W}_{i}K(\stackrel{\&RightArrow;}{E},{{\stackrel{\&RightArrow;}{E}}_i}_{})+b=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}{W}_i\exp (-{|\stackrel{\&RightArrow;}{E}-{\stackrel{\&RightArrow;}{E}}_i|}^2/{\mathrm{\&sigma;}}^2)+b---\left(4\right)$

3., whether error in judgement meets the demands;

Because the non-linear approximation capability of SVM is excellent, 2N the SVM network that step calculates in 2. can be directly used in On-line Control, and determine whether needs optimization and study according to the margin of error of On-line Control; If the error amount of merit angle and terminal voltage all one of setting in advance more among a small circle in, judge that then the control law that is obtained by expertise is more excellent, can not adjust control law; If the error amount of any one physical quantity surpasses the scope of setting in advance in merit angle and the terminal voltage, with step below adopting 4. partly algorithm be optimized;

4., become yardstick gradient optimizing algorithm;

Adopt to become yardstick gradient optimizing algorithm, the recurrence formula that calculates SVM network weight W is:

$\left.\begin{array}{c}{W}_{k+1}=W_k-H_k{E}_k\left(W_k\right)\&CenterDot;\&dtri;E_k\left(W_k\right)/{\mathrm{\&beta;}}_k\\ H_{k+1}={\mathrm{\&lambda;}}^{-1}(H_k-H_k\&dtri;E_k\&CenterDot;\&dtri;E_k^T\&CenterDot;H_k/{\mathrm{\&beta;}}_k)\\ {\mathrm{\&beta;}}_{k+1}=\mathrm{\&lambda;}+\&dtri;E_k^T\&CenterDot;H_k\&CenterDot;\&dtri;E_k\end{array}\right\}---\left(5\right)$

Here k is sampling, and 0＜λ＜1 is a forgetting factor, and H is Hessian matrix and H ₁=I (unit matrix), E _kBe error vector, be respectively δ _r-δ _kAnd V _Tr-V _Tk, hereinafter will be with E _k=δ _r-δ _kRealization for example explanation SVM parameter adjustment; ▽ E _kRepresent ▽ E (K respectively _k), ▽ E (W _k) gradient, obtain the gradient of network weight by following formula (6):

$\&dtri;E_k\left(W_k\right)=\frac{\&PartialD;J}{{\&PartialD;W}_k}=\frac{\&PartialD;J}{{\&PartialD;\mathrm{\&delta;}}_k}\&CenterDot;\frac{{\&PartialD;\mathrm{\&delta;}}_k}{{\&PartialD;u}_{2k}}\&CenterDot;\frac{{\&PartialD;u}_{2k}}{{\&PartialD;W}_k}=-({\mathrm{\&delta;}}_m-{\mathrm{\&delta;}}_k)\&CenterDot;\frac{{\&PartialD;\mathrm{\&delta;}}_k}{{\&PartialD;u}_k}\&CenterDot;{\mathrm{Ku}}_1\&CenterDot;w_n\&CenterDot;K(\stackrel{\&RightArrow;}{E},{\stackrel{\&RightArrow;}{E}}_k)---\left(6\right)$

Wherein,

Still be (E ₁, E ₂), E _k=δ _r-δ _k, _ δ _k/ u _2kWith

Obtain.

Compared with prior art, advantage of the present invention just is:

(1) set up model library and controller storehouse, control law designs at different operating modes, and specific aim is better, and can adapt to the generator 's parameter wide variation; (2) fuzzy controller does not rely on the Mathematical Modeling of generator, and decoupling zero is effective; (3) calculate the submodel matching degree based on fuzzy logic, matching degree has determined the integrated weight coefficient in the multi-model control system, helps eliminating model and switches caused vibration; (4) expertise has been merged in the control law design, can obtain preferable effect; (5) utilize SVMs (SVM) to realize the rule optimization of fuzzy controller, self-learning capability is strong.

Description of drawings

Fig. 1 is the multi-model integrated intelligent control method implementing procedure schematic diagram of Generator Set of the present invention;

Fig. 2 is the structural representation of neutron model controller of the present invention;

Fig. 3 is the interval schematic diagram of the submodel after the cluster;

Fig. 4 is the input-output ruled surface schematic diagram of FLC;

Fig. 5 is the structural representation of multi-model integrated intelligent controller;

Fig. 6 is the present invention's three machine system schematic;

Fig. 7 is the schematic diagram of generating set A terminal voltage response curve;

Fig. 8 is the schematic diagram of generating set A merit angular response curve;

Fig. 9 is the schematic diagram of generating set A active power response curve;

Figure 10 is the schematic diagram of generating set A angular speed response curve.

Embodiment

Below with reference to the drawings and specific embodiments the present invention is described in further details.

As shown in Figure 1, the multi-model integrated intelligent control method of a kind of Generator Set of the present invention the steps include:

(1), sets up model library and controller storehouse: comprised N kind submodel in the model library, N kind operating condition corresponding to generating set, the controller storehouse is made up of N sub-model fuzzy controller, and each submodel fuzzy controller is at a kind of submodel design control law; Each submodel fuzzy controller is the controller of dual input dual output, and it is input as power angle deviation amount e _δWith terminal voltage departure e _Vt, be output as excitation regulation amount u _fWith porthole aperture regulated quantity u ₂

(2), actual condition data acquisition: gather the real-time working condition data of generating set, with variable (δ, V _t) weigh the actual condition value of generating set;

(3), the generation and the output of working control amount: calculate actual condition value (δ, V with fuzzy logic _t) with model library in the matching degree f of N submodel _n, with this as the integrated weight coefficient w of general controller _n(n=1,2 ..., N).

In the present embodiment, the idiographic flow of setting up model library and controller storehouse is:

The structure flow process in model library and controller storehouse is shown in first empty frame among Fig. 1, and distributing below is described as follows:

1.1 primary data sample.

According to the different work condition states of the existing generator of measuring $\stackrel{\&RightArrow;}{\mathrm{\&delta;}}=\{{\mathrm{\&delta;}}_1,{\mathrm{\&delta;}}_2,\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{\&delta;}}_a\},$ ${\stackrel{\&RightArrow;}{V}}_t=\{{\mathrm{<figure-callout\; id="1"\; label="V\; t"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">V</figure-callout>}}_t,{\mathrm{<figure-callout\; id="2"\; label="V\; t"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">V</figure-callout>}}_t,\&CenterDot;\&CenterDot;\&CenterDot;,V_{\mathrm{tb}}\},$ Choose that (δ Vt) arranges in pairs or groups and obtains a * b=ab kind operating mode, as initial sample.

1.2 set up the IF-THEN rule according to expertise.

Under each operating mode, according to the default fuzzy control rule of expertise.Adopt the fuzzy controller (FLC) of dual input dual output shown in Figure 2, it is input as merit angle error amount e _δ, terminal voltage margin of error e _Vt, be output as porthole regulated quantity u ₂, excitation regulation amount u _fFuzzy rule adopts " IF-THEN " fuzzy relation formula, all obtains 5 * 5=25 bar rule for each operating mode, that is:

rule1?IF?E _δis?PB?AND?E _Vtis?PB?THENU ₂is?NS?ANDU _fis?NB；

......

rule25?IF?E _δis?NB?AND?E _Vt?is?NB?THEN?U ₂is?PB?AND?U _f?is?PS。

1.3 the IF-THEN rule is represented with regular matrix.

Use matrix Rule _oRepresent control law: ${\mathrm{Rule}}_{0j}={\left[\begin{array}{cccc}\mathrm{PB}& \mathrm{PB}& \mathrm{NS}& \mathrm{NB}\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \mathrm{NB}& \mathrm{NB}& \mathrm{PB}& \mathrm{PS}\end{array}\right]}_{25\&times;4,}$ (j=1,2 ..., ab), if with 1 ,-0.5,0,0.5,1} represent NB, NS, Z, PS, PB} then is equivalent to: ${\mathrm{Rule}}_{0j}^{\&prime;}={\left[\begin{array}{cccc}1& 1& 0.5& 1\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \&CenterDot;& \&CenterDot;\\ -1& -1& 1& 0.5\end{array}\right]}_{25\&times;4.,}$ Regular matrix Rule _o' in preceding two classify premise part as, be identical, and what embody regular difference is exactly back two row, i.e. conclusion part.Be convenience of calculation, with matrix Rule ₀' preceding two column skips, only stay back two row, i.e. 25 * 2 matrix, and the matrix of (25 * 2) is stretched as delegation (1 * 50), i.e. Rule _j=0.5,1 ..., 1,0.5} _{1 * 50}This moment, the domain of object to be clustered was Rule={Rule ₁, Rule ₂..., Rule _Ab, each sample Rule _j(j=1,2 ..., ab) describe, i.e. Rule by 50 characteristic ginseng values _j={ Rulej ₁, Rulej ₂..., Rule _J50.

1.4 the standardization of regular matrix.

Adopt standardization characteristic ginseng value to be compressed to [0,1] interval:

${\mathrm{Rule}}_{\mathrm{ji}}^{\&prime;}=\frac{{\mathrm{Rule}}_{\mathrm{ji}}-{\mathrm{Rule}}_{\min }}{{\mathrm{Rule}}_{\max }-{\mathrm{Rule}}_{\min }}$ (j＝1，2，…，ab，i＝1，2，…，50)。So the object domain to be clustered after the standardization be Rule '=Rule ' ₁, Rule ' ₂..., Rule ' _Ab}

1.5 demarcate.

Set up the fuzzy resembling relation R of object, the membership function mui of R _R(Rule ' _i, Rule ' _j) expression Rule ' _iWith Rule ' _jSimilarity degree by its character.R is similar matrix R=[r _Ij], r _Ij=u _R(Rule ' _i, Rule ' _j).Here calculate r by the minimax method _Ij:

$r_{\mathrm{ij}}=\left[\underset{k=1}{\overset{50}{\mathrm{\&Sigma;}}}\min ({\mathrm{Rule}}_{\mathrm{ik}}^{\&prime;},{\mathrm{Rule}}_{\mathrm{jk}}^{\&prime;})\right]/\left[\underset{k=1}{\overset{50}{\mathrm{\&Sigma;}}}\max ({\mathrm{Rule}}_{\mathrm{ik}}^{\&prime;},{\mathrm{Rule}}_{\mathrm{jk}}^{\&prime;})\right]---\left(1\right)$

1.6 fuzzy clustering.

According to fuzzy set theory, obtaining similar matrix R=[r _Ij] after, the fuzzy equivalence relation matrix R of calculating R ^*, use λ again and cut matrix and carry out fuzzy clustering.Cut the pairing λ of relation for the λ of any λ ∈ [0,1] intercepting and cut matrix, like this each common section matrix R _λ ^*Can determine the classification of a level.For arbitrary λ ∈ [0,1], divide time-like R _λ ^*In each element change such value into: all change more than or equal to the element of λ and be taken as 1, all change less than the element of λ and be taken as 0, according to R _λ ^*In 1 and 0 arranging situation can classify.

This moment, ab kind operating mode was a N kind typical condition with regard to yojan, was N submodel interval according to the combination situation of operating mode with the sample data cluster, and as shown in Figure 3, the interior data clusters of same square frame is a kind of operating mode.N kind submodel correspondence has N sub-model fuzzy controller.Obtain U according to the Mamdani fuzzy reasoning method ₂, U _fMembership function value μ (U _2i), μ (U _Fi), then the controlled quentity controlled variable of each submodel fuzzy controller is:

$u_{2n}=\frac{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{2i}\right)\&CenterDot;U_{2i}}{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{2i}\right)},$ $u_{\mathrm{fn}}=\frac{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{fi}\right)\&CenterDot;U_{fi}}{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{\mathrm{fi}}\right)},$ n＝1，2，…，N(2)

In the present embodiment, can learn to optimize to model library after setting up and controller storehouse, the idiographic flow that this study is optimized be:

2.1 obtain N ruled surface by N controller storehouse.

To obtain model library and controller storehouse that scale is N after the cluster, each submodel fuzzy controller can be expressed as 2 ruled surfaces as shown in Figure 4, i.e. U ₂About E _δWith E _VtRuled surface, U _fAbout E _δWith E _VtRuled surface.Like this, N controller storehouse is exactly 2N ruled surface, and the input dimension of each ruled surface is 2, and the output dimension is 1.

2.2 2N SVM network approaches 2N ruled surface.

Approach ruled surface by SVMs (SVM) recurrence, thereby fuzzy control rule is converted into SVM network weight parameter, be convenient to realize the online self study of control law by gradient algorithm.

According to the controller storehouse that cluster obtains, set up learning sample collection { ((E ₁, E ₂), U) ..., ((E _1n, E _2n), U _n) _ R ²* R, the learning process that SVM returns is converted into a quadratic programming problem.Supported the vector ((E of learning outcome _1i, E _2i), U _i) and corresponding coefficient (α _i ^*-α _i), use

Expression (E _1i, E _2i), thereby obtain regression function be:

$f\left(\stackrel{\&RightArrow;}{E}\right)=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)K\left({\stackrel{\&RightArrow;}{E},\stackrel{\&RightArrow;}{E}}_i\right)+b---\left(3\right)$

Wherein g is the support vector number, uses W _iExpression (α _i ^*-α _i), adopt the RBF kernel function: K (x _i, x _j)=exp (| x _i-x _j| ²/ σ ²), equation (6) can be expressed as:

$f\left(\stackrel{\&RightArrow;}{E}\right)\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}{W}_{i}K(\stackrel{\&RightArrow;}{E},{{\stackrel{\&RightArrow;}{E}}_i}_{})+b=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}{W}_i\exp (-{|\stackrel{\&RightArrow;}{E}-{\stackrel{\&RightArrow;}{E}}_i|}^2/{\mathrm{\&sigma;}}^2)+b---\left(4\right)$

2.3 whether error in judgement meets the demands.

Because the non-linear approximation capability of SVM is excellent, the 2N that 2.2 parts calculate SVM network can be directly used in On-line Control, and determine whether needs optimization and study according to the margin of error of On-line Control.If the error amount of merit angle and terminal voltage all one of setting in advance more among a small circle in, illustrate that then front 1.2 parts are more excellent by the control law that expertise obtains, can not adjust control law.If the error amount of any one physical quantity surpasses the scope of setting in advance in merit angle and the terminal voltage,, control effect preferably in the hope of reaching with adopting the algorithm of 2.4 parts to be optimized.

2.4 become yardstick gradient optimizing algorithm.

Adopt to become yardstick gradient optimizing algorithm, the recurrence formula that calculates SVM network weight W is:

$\left.\begin{array}{c}{W}_{k+1}=W_k-H_k{E}_k\left(W_k\right)\&CenterDot;\&dtri;E_k\left(W_k\right)/{\mathrm{\&beta;}}_k\\ H_{k+1}={\mathrm{\&lambda;}}^{-1}(H_k-H_k\&dtri;E_k\&CenterDot;\&dtri;E_k^T\&CenterDot;H_k/{\mathrm{\&beta;}}_k)\\ {\mathrm{\&beta;}}_{k+1}=\mathrm{\&lambda;}+\&dtri;E_k^T\&CenterDot;H_k\&CenterDot;\&dtri;E_k\end{array}\right\}---\left(5\right)$

Here k is sampling, and 0＜λ＜1 is a forgetting factor, and H is Hessian matrix and H ₁=I (unit matrix), E _kBe error vector, be respectively δ _r-δ _kAnd V _Tr-V _Tk, hereinafter will be with E _k=δ _r-δ _kRealization for example explanation SVM parameter adjustment; ▽ E _kRepresent ▽ E (K respectively _k), ▽ E (W _k) gradient. the gradient of following computing network weights:

$\&dtri;E_k\left(W_k\right)=\frac{\&PartialD;J}{{\&PartialD;W}_k}=\frac{\&PartialD;J}{{\&PartialD;\mathrm{\&delta;}}_k}\&CenterDot;\frac{{\&PartialD;\mathrm{\&delta;}}_k}{{\&PartialD;u}_{2k}}\&CenterDot;\frac{{\&PartialD;u}_{2k}}{{\&PartialD;W}_k}=-({\mathrm{\&delta;}}_m-{\mathrm{\&delta;}}_k)\&CenterDot;\frac{{\&PartialD;\mathrm{\&delta;}}_k}{{\&PartialD;u}_k}\&CenterDot;{\mathrm{<figure-callout\; id="1"\; label="Ku"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">Ku</figure-callout>}}_1\&CenterDot;w_n\&CenterDot;K(\stackrel{\&RightArrow;}{E},{\stackrel{\&RightArrow;}{E}}_k)---\left(6\right)$

Wherein,

Still be (E ₁, E ₂), Ek=δ _r-δ _k, δ _k/ u _2tWith

Calculate.Each SVM network is all optimized weights according to the online adjustment of similar gradient algorithm, and through a certain amount of study, the SVM network weight no longer changes then study optimization to be finished.

3. the enforcement of multi-model integrated Based Intelligent Control

The multi-model integrated Based Intelligent Control flow process of generating set is shown in the 3rd empty frame among Fig. 1, and its structure mainly comprises two parts in model library and controller storehouse as shown in Figure 5.Comprised N kind submodel in the model library, corresponding to the N kind operating condition of generating set, with n represent (n=1,2 ..., N) each submodel.The controller storehouse is made of N sub-model fuzzy controller (FLC), and each FLC is at a kind of submodel design control law.

Adopt variable (δ, V _t) weigh the operating mode of generating set.Adopt fuzzy logic to calculate actual condition value (δ, V _t) with model library in the matching degree f of N submodel _n, with this as the integrated weight coefficient w of general controller _n(n=1,2 ..., N).Fig. 3 has described the interval division of the submodel of the N in the model library, and dash area is the overlapping or boundary section of submodel among the figure, will belong to a certain submodel by fuzzy membership functions value decision actual condition.Set up variable (δ, V respectively _t) membership function, be expressed as with mathematical expression:

${\mathrm{\&mu;}}_{\mathrm{\&delta;n}}=\frac{1}{1+{\left(\frac{\mathrm{\&delta;}-c_{1n}}{a_{1n}}\right)}^2},$ ${\mathrm{\&mu;}}_{\mathrm{Vn}}=\frac{1}{1+{\left(\frac{V_t-c_{2n}}{a_{2n}}\right)}^2},$ n＝1，2，…，N(7)

C wherein _1n, c _2nRepresent variable (δ, V respectively _t) central value, and α _1n, α _2nRepresent variable (δ, V respectively _t) width.

According to degree of membership μ _{δ n}, μ _VnObtain the matching degree f of actual condition and submodel n _nFor:

f _n=μ _{δ n}* μ _Vn, f _n∈ [0,1] and $\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_n=1---\left(8\right)$ Weight coefficient w ₁, w ₂..., w _N[w is numerically arranged ₁, w ₂..., w _N] ^T=[f ₁, f ₂..., f _N] ^T, w _n∈ [0,1] and $\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n=1$ (n＝1，2，…，N)(9)

At this moment, act on of the weighting of the working control amount of generating set for the output of submodel fuzzy controller:

${\mathrm{<figure-callout\; id="2"\; label="u"\; filenames="A20071003576400171.png,A20071003576400181.png"\; state="\{\{state\}\}">u</figure-callout>}}_2=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n*u_{2n},$ $u_f=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n*u_{\mathrm{fn}}---\left(10\right)$

Make concrete control examples, carry out simulation study with a typical case three machine systems here, its structure comprises A, B, three units of C as shown in Figure 6, and with unit C as the reference phase place.System's actual parameter and working point are:

Unit A: impedance xd '=0.26, impedance xd=1.28, Parameter H=8, parameter D=5, initial value Td0=6.9s, parameter TB=0.15s, parameter value T μ=0.2s, initial value Pe0=0.85, initial value δ 0=46.3o, initial value Vt0=1.07;

B of Unit: impedance xd '=0.32, impedance xd=1.35, Parameter H=10, parameter D=3, initial value Td0=7.8s, parameter TB=0.2s, parameter T μ=0.25s, initial value Pe0=1.0, initial value δ 0=50.6o, initial value Vt0=1.05;

Other parameter: impedance x12=0.55, impedance x13=0.53, impedance x23=0.6, parameter Eq3 '=1.0, parameter δ 30=0o.

In order to reflect the control performance of the multi-model integrated intelligent controller that this paper designs, itself and traditional PID controller, a kind of RBF nerve network controller have been given contrast here.Unit A and B operate in the aforementioned stable working point, rise more and descend for 5% rank of generating set terminal voltage set-point respectively when t=2s and t=10s, reflect the performance of different controllers with this.Fig. 7 is that generating set A terminal voltage response curve, Fig. 8 are that generating set A merit angular response curve, Fig. 9 are that generating set A active power response curve, Figure 10 are generating set A angular speed response curve.Dotted line is the control effect of PID control among Fig. 7-10, and fine line is the control effect of RBF ANN Control, and heavy line is the control performance of multi-model integrated Based Intelligent Control.The dynamic response curve of each variable shows in the emulation experiment, multi-model integrated Based Intelligent Control has more excellent control performance, the terminal voltage of generating set A can require rapid tracing preset value according to adjusting, and each state variable of unit Vt, δ, P, ω can both enter new stable state with fast speeds.Other two kinds of controllers then need the long time just can reach stable state, and have overshoot and certain vibration.

Claims (3)

1. the multi-model integrated intelligent control method of a Generator Set is characterized in that step is:

(1), sets up model library and controller storehouse: comprised N kind submodel in the model library, N kind operating condition corresponding to generating set, the controller storehouse is made up of N sub-model fuzzy controller, and each submodel fuzzy controller is at a kind of submodel design control law; Each submodel fuzzy controller is the controller of dual input dual output, and it is input as power angle deviation amount e _δWith terminal voltage departure e _Vt, be output as excitation regulation amount u _fWith porthole aperture regulated quantity u ₂

(2), actual condition data acquisition: gather the real-time working condition data of generating set, with variable (δ, V _t) weigh the actual condition value of generating set;

(3), the generation and the output of working control amount: calculate actual condition value (δ, V with fuzzy logic _t) with model library in the matching degree f of N submodel _n, with this as the integrated weight coefficient w of general controller _n(n=1,2 ..., N);

Set up variable (δ, V 1., at first respectively _t) membership function, be expressed as with formula (7):

${\mathrm{\&mu;}}_{\mathrm{\&delta;n}}=\frac{1}{1+{\left(\frac{\mathrm{\&delta;}-c_{1n}}{a_{1n}}\right)}^2},$ ${\mathrm{\&mu;}}_{\mathrm{Vn}}=\frac{1}{1+{\left(\frac{V_t-c_{2n}}{a_{2n}}\right)}^2},$ n＝1，2，…，N (7)

C wherein _1n, c _2nRepresent variable (δ, V respectively _t) central value, and a _1n, a _2nRepresent variable (δ, V respectively _t) width;

2., according to degree of membership μ _{δ n}, μ _VnObtain the matching degree f of actual condition and submodel n _nFor:

f _n=μ _{δ n}* μ _Vn, f _n∈ [0,1] and $\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{f}_n=1---\left(8\right)$

Weight coefficient w ₁, w ₂..., w _NNumerically have

[w ₁, w ₂..., w _N] ^T=[f ₁, f ₂..., f _N] ^T, w _n∈ [0,1] and $\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n=1(n=\mathrm{1,2},...,N)---\left(9\right);$

3., act on of the weighting of the working control amount of generating set for the output of submodel fuzzy controller:

$u_2=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n*u_{2n},$ $u_f=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_n*u_{\mathrm{fn}}---\left(10\right).$

2. the multi-model integrated intelligent control method of Generator Set according to claim 1 is characterized in that the described idiographic flow of setting up model library and controller storehouse is:

1., primary data sample:

According to the different work condition states of the existing generator of measuring $\stackrel{\&RightArrow;}{\mathrm{\&delta;}}=\{{\mathrm{\&delta;}}_1,{\mathrm{\&delta;}}_2,...,{\mathrm{\&delta;}}_a\},$ ${\stackrel{\&RightArrow;}{V}}_t=\{V_{t1},V_{t2},...,V_{\mathrm{tb}}\},$ Choose (δ, V _t) arranging in pairs or groups obtains a * b=ab kind operating mode, as initial sample;

2., set up the IF-THEN rule according to expertise:

Under each operating mode, according to the default fuzzy control rule of expertise, adopt the fuzzy controller of dual input dual output, it is input as merit angle error amount e _δWith terminal voltage margin of error e _Vt, be output as porthole regulated quantity u ₂With excitation regulation amount u _f, fuzzy rule adopts " IF-THEN " fuzzy relation formula, all obtains 5 * 5=25 bar rule for each operating mode, that is:

rule1?IF?E _δ?is?PB?AND?E _Vt?is?PB?THEN?U ₂?is?NS?AND?U _f?is?NB；

......

rule25?IFE _δ?is?NB?AND?E _Vt?is?NB?THEN?U ₂?is?PB?AND?U _f?is?PS；

3., the IF-THEN rule is represented with regular matrix:

Use matrix Rule ₀Represent control law: ${\mathrm{Rule}}_{0j}={\left[\begin{array}{cccc}\mathrm{PB}& \mathrm{PB}& \mathrm{NS}& \mathrm{NB}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ \mathrm{NB}& \mathrm{NB}& \mathrm{PB}& \mathrm{PS}\end{array}\right]}_{25\&times;4},$ (j=1,2 ..., ab), if with 1 ,-0.5,0,0.5,1} represent NB, NS, Z, PS, PB} then is equivalent to: ${{\mathrm{Rule}}_{0j}}^{\&prime;}\text{=}{\left[\begin{array}{cccc}1& 1& 0.5& 1\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ -1& -1& 1& 0.5\end{array}\right]}_{25\&times;4};$ With matrix Rule ₀' preceding two column skips, only stay back two row, i.e. 25 * 2 matrix, and the matrix of (25 * 2) is stretched as delegation (1 * 50), i.e. Rule _j=0.5,1 ..., 1,0.5} _{1 * 50}, this moment, the domain of object to be clustered was Rule={Rule ₁, Rule ₂..., Rule _Ab, each sample Rule _j(j=1,2 ..., ab) describe, i.e. Rule by 50 characteristic ginseng values _j={ Rule _J1, Rule _J2..., Rule _J50;

4., the standardization of regular matrix:

Adopt standardization characteristic ginseng value to be compressed to [0,1] interval: ${{\mathrm{Rule}}_{\mathrm{ji}}}^{\&prime;}=\frac{{\mathrm{Rule}}_{\mathrm{ji}}-{\mathrm{Rule}}_{\min }}{{\mathrm{Rule}}_{\max }-{\mathrm{Rule}}_{\min }}(j=\mathrm{1,2},...,\mathrm{ab},i=\mathrm{1,2},...,50),$ So the object domain to be clustered after the standardization is Rule '={ Rule ₁', Rule ₂' ..., Rule _Ab';

5., demarcate:

Set up the fuzzy resembling relation R of object, the membership function mui of R _R(Rule _i', Rule _j') expression Rule _i' with Rule _j' by the similarity degree of its character, R is similar matrix R=[r _Ij], r _Ij=μ _R(Rule _i', Rule _j'), calculate r by the minimax method here _Ij:

$r_{\mathrm{ij}}=\left[\underset{k=1}{\overset{50}{\mathrm{\&Sigma;}}}\min ({{\mathrm{Rule}}_{\mathrm{ik}}}^{\&prime;},{{\mathrm{Rule}}_{\mathrm{jk}}}^{\&prime;})\right]/\left[\underset{k=1}{\overset{50}{\mathrm{\&Sigma;}}}\max ({{\mathrm{Rule}}_{\mathrm{ik}}}^{\&prime;},{{\mathrm{Rule}}_{\mathrm{jk}}}^{\&prime;})\right]---\left(1\right);$

6., fuzzy clustering:

According to fuzzy set theory, obtain U ₂, U _fMembership function value μ (U _2i), μ (U _Fi), then the controlled quentity controlled variable of each submodel fuzzy controller is:

$u_{2n}=\frac{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{2i}\right)\&CenterDot;U_{2i}}{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{2i}\right)},$ $u_{\mathrm{fn}}=\frac{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{\mathrm{fi}}\right)\&CenterDot;U_{\mathrm{fi}}}{\underset{i=1}{\overset{25}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(U_{\mathrm{fi}}\right)},$ n＝1，2，…，N (2)。

3. the multi-model integrated intelligent control method of Generator Set according to claim 1 and 2, it is characterized in that described step (1) finishes after, after having set up model library and controller storehouse, learn optimization, the step of described study optimization is:

1., obtain N ruled surface by N controller storehouse;

To obtain model library and controller storehouse that scale is N after the cluster, each submodel fuzzy controller can be expressed as 2 ruled surfaces, i.e. U ₂About E _δWith E _VtRuled surface, U _fAbout E _δWith E _VtRuled surface, N controller storehouse is exactly 2N ruled surface, the input dimension of each ruled surface is 2, output dimension be 1;

2., 2N SVM network approaches 2N ruled surface;

Return by SVMs SVM and to approach ruled surface, fuzzy control rule is converted into SVM network weight parameter, be convenient to realize the online self study of control law by gradient algorithm; According to the controller storehouse that cluster obtains, set up learning sample collection { ((E ₁, E ₂), U) ..., ((E _1n, E _2n), U _n) _ R ²* R, the learning process that SVM returns is converted into a quadratic programming problem; Supported the vector ((E of learning outcome _1i, E _2i), U _i) and corresponding coefficient (α _i ^*-α _i), use

Expression (E _1i, E _2i), thereby obtain regression function be:

$f\left(\stackrel{\&RightArrow;}{E}\right)=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)K(\stackrel{\&RightArrow;}{E},{\stackrel{\&RightArrow;}{E}}_i)+b---\left(3\right)$

Wherein g is the support vector number, uses W _iExpression (α _i ^*-α _i), adopt the RBF kernel function: K (x _i, x _j)=exp (| x _i-x _j| ²/ σ ²), equation (6) can be expressed as:

$f\left(\stackrel{\&RightArrow;}{E}\right)=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}{W}_{i}K(\stackrel{\&RightArrow;}{E},{\stackrel{\&RightArrow;}{E}}_i)+b=\underset{i=1}{\overset{g}{\mathrm{\&Sigma;}}}{W}_i\exp (-{|\stackrel{\&RightArrow;}{E}-{\stackrel{\&RightArrow;}{E}}_i|}^2/{\mathrm{\&sigma;}}^2)+b---\left(4\right);$

3., whether error in judgement meets the demands;

Because the non-linear approximation capability of SVM is excellent, 2N the SVM network that step calculates in 2. can be directly used in On-line Control, and determine whether needs optimization and study according to the margin of error of On-line Control; If the error amount of merit angle and terminal voltage all one of setting in advance more among a small circle in, judge that then the control law that is obtained by expertise is more excellent, can not adjust control law; If the error amount of any one physical quantity surpasses the scope of setting in advance in merit angle and the terminal voltage, with step below adopting 4. partly algorithm be optimized;

4., become yardstick gradient optimizing algorithm;

Adopt to become yardstick gradient optimizing algorithm, the recurrence formula that calculates SVM network weight W is:

$\left.\begin{array}{c}{W}_{k+1}=W_k-H_k{E}_k\left(W_k\right)\&CenterDot;\&dtri;E_k\left(W_k\right)/{\mathrm{\&beta;}}_k\\ H_{k+1}={\mathrm{\&lambda;}}^{-1}(H_k-H_k\&dtri;E_k\&CenterDot;\&dtri;E_k^T\&CenterDot;H_k/{\mathrm{\&beta;}}_k)\\ {\mathrm{\&beta;}}_{k+1}=\mathrm{\&lambda;}+\&dtri;E_k^T\&CenterDot;H_k\&CenterDot;\&dtri;E_k\end{array}\right\}---\left(5\right)$

Here k is sampling, and 0＜λ＜1 is a forgetting factor, and H is Hessian matrix and H ₁=I (unit matrix), E _kBe error vector, be respectively δ _r-δ _kAnd V _Tr-V _Tk, hereinafter will be with E _k=δ _r-δ _kRealization for example explanation SVM parameter adjustment; _ E _kExpression _ E (K respectively _k), _ E (W _k) gradient, obtain the gradient of network weight by following formula (6):

$\&dtri;E_k\left(W_k\right)=\frac{\&PartialD;J}{\&PartialD;W_k}=\frac{\&PartialD;J}{\&PartialD;{\mathrm{\&delta;}}_k}\&CenterDot;\frac{\&PartialD;{\mathrm{\&delta;}}_k}{\&PartialD;u_{2k}}\&CenterDot;\frac{\&PartialD;u_{2k}}{\&PartialD;W_k}=-({\mathrm{\&delta;}}_m-{\mathrm{\&delta;}}_k)\&CenterDot;\frac{\&PartialD;{\mathrm{\&delta;}}_k}{\&PartialD;u_k}\&CenterDot;{\mathrm{Ku}}_1\&CenterDot;w_n\&CenterDot;K(\stackrel{\&RightArrow;}{E},{\stackrel{\&RightArrow;}{E}}_k)---\left(6\right)$

Wherein,

Still be (E ₁, E ₂), E _k=δ _r-δ _k, _ δ _k/ _ u _2kWith

Obtain.

Images