CN101788789A - Nonlinear predictive control method of unit plant based on chaos and hybrid optimization algorithm - Google Patents

Abstract

The invention discloses a nonlinear predictive control method of a unit plant based on a chaos and hybrid optimization algorithm, belonging to the control field of thermal process. In the method, a neural network is used as a model prediction measure, the chaos and hybrid optimization algorithm based on SkewTent mapping is used as a rolling optimization strategy, and then the a predictive object is combined to control a steam valve so that the stability of a system is improved and the rapidity of the load regulation is also strengthened. Furthermore, the method solves the problems of nonlinear identification of a load control system for a plant furnace, large on-line rolling optimization calculated amount, slow constriction and the like of a load control system for a plant furnace, improves the real-time performance, the robustness and the dynamic performance of control and develops a new field for the application of the chaos optimization in the control of the thermal process.

Description

Monoblock non-linear predication control method based on the chaotic mixing optimized Algorithm

Technical field

The present invention relates to thermal process control field, relate in particular to monoblock non-linear predication control method based on the chaotic mixing optimized Algorithm.

Background technology

Modernization large-scale thermal power machine group is commonly high parameter, high capacity unit unit, often adopts monoblock to coordinate control.Thermal power plant's monoblock is made up of boiler, steam turbine, generator, electrical network and load equipment.Wherein, machine stove master control system is the core of control system, and its main task is that the realization unit is being kept main vapour pressure in allowed limits in the response external load demand fast.Thermal power plant's monoblock load control system be a typical nonlinear, the time become, the multivariate complication system.Because when coupling between machine stove control loop and peaking operation load fluctuation are big, the reasons such as nonlinear characteristic that machine stove object shows make that present classical control system is difficult to obtain satisfied effect.

PREDICTIVE CONTROL adopts control strategies such as multi-step prediction, rolling optimization and feedback compensation, have effective, the strong robustness of control, to the less demanding advantage of model accuracy, can effectively overcome the uncertainty of process, non-linear and relevance, and can easily the processing procedure controlled variable and manipulated variable in various constraints.This basic idea is to predict afterwards earlier to control that control action has stronger foresight.But the control effect that basic predictive control algorithm is difficult to obtain for non-linear stronger system, thereby necessary research is based on the prediction and the optimized Algorithm of nonlinear model.

The chaos optimization algorithm of the overwhelming majority all is based on the chaos optimization of Chaos Variable at present, and its thought all is to utilize the method for similar carrier wave that selected Chaos Variable is mapped to the optimization variable space linearly, utilizes the ergodicity of Chaos Variable to search for then.The process of utilizing chaos to be optimized is divided into coarse search and two stages of fine searching.Chaos optimization method is few to the requirement condition of objective function, higher execution efficient is arranged, utilize the ergodicity characteristics of Chaos Search, can also avoid being absorbed in locally optimal solution, but because the shortcomings such as unevenness of probability distribution, make it slow, i.e. the fine searching scarce capacity in optimum solution neighborhood speed of convergence.

Pattern search method (Pattern Search) is also referred to as the Hooke-Jeeves method as a kind of direct search method, originates from five sixties of last century, and is used to find the solution in the actual optimum problem.Though the pattern search method has obtained paying close attention to widely and using, its theory is also not really perfect.Up to 1997, the convergence that Torczon has finished the unconstrained problem pattern search algorithm proved, and has provided the framework of pattern search algorithm.After this, Lewis and the Torczon convergence of having delivered the pattern search algorithm of square boundary constraint and equality constraint proves.Pattern search method does not require the information of any target function gradient.More traditional optimized Algorithm of searching for the optimization point with use gradient or higher derivative information is opposite, and pattern search algorithm is searched for the series of points around the current point, seeks those points that target function value is lower than current point value.

Therefore, the present invention introduces pattern search, in conjunction with the chaos optimization algorithm, for the shortcoming that solves chaos optimization method provides method.

Summary of the invention

The objective of the invention is the problem that exists at the present monoblock nonlinear prediction control described in the background technology, proposed monoblock non-linear predication control method based on the chaotic mixing optimized Algorithm.

It is characterized in that, may further comprise the steps:

Step 1: monoblock is carried out modelling by mechanism;

Step 2: initialization neural network identifier, initialization network weight, given prediction length and Chaos Variable;

Step 3: the off-line training neural network identifier according to the described modelling by mechanism of step 1, generates the monoblock forecast model f based on neural network _NN

Step 4: above-mentioned forecast model is dropped in line traffic control, keeps network weight constant,, calculate the output of forecast model and carry out recursion, obtain the multi-step prediction result with the input of control variable as forecast model; Wherein, utilize iterative algorithm to obtain N step forecast model forward:

$\hat{y}(k+1|k)=f_{\mathrm{NN}}[y\left(k\right),...,y(k-n+1),u\left(k\right),...,u(k-m+1)];$

.

.

.

$\hat{y}(k+N|k)=f_{\mathrm{NN}}[\hat{y}(k+N-1|k),...,\hat{y}(k+N-n),$

u(k+N-1)，…，u(k+N-m)]

Wherein:

The model of predicting constantly at k for model is in k+j output valve constantly, j=1 ..., N in iterative process, uses the constantly later predicted value of k

The predicted value before the moment to k and k replaces with its actual value, promptly

$\hat{y}(k+i-j)=y(k+i-j),i-j\≤0,$

Wherein,

u(k+j)＝u(k)，i＝1，…，j-1；

Step 5: utilize the optimizing of rolling of chaotic mixing optimized Algorithm, determine the control increment of algorithm, more current optimum control amount is acted on controlled device;

The control increment of above-mentioned algorithm Δ u (k+j-1), j=1 ..., m} is next definite for minimum by the value that makes optimization criterion J,

$J=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{[\hat{y}(k+j)-w(k+j)]}^2+\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\λ}\left(j\right){\left[\mathrm{\Δu}(k+j-1)\right]}^2,$

Wherein, n and m are respectively maximum predicted length and control length; λ (j) generally gets constant λ for the control weighting factor, and w (k+j) is a k+j reference locus value constantly;

Step 6: online correction neural network identifier network weight;

Step 7: return step 4, enter next control cycle.

The described chaotic mixing optimized Algorithm of step 5 is based on the chaotic mixing optimized Algorithm of Skew Tent mapping.May further comprise the steps:

Starting condition: establishing a class optimization problem is:

minf(x _i)，x _imin≤x _i≤x _imax，

i＝1，2，…，n.

Wherein, f (x _i) be any optimization aim function;

1) chooses n the initial value x that fine difference is arranged in (0,1) interval _i ⁰(i=1,2 ..., n),, make current optimum Chaos Variable as the initial value of Chaos Variable

Step-up error scope ε＞0; The maximum iteration time M of chaos optimization coarse search and pattern search is set ₁, M ₂And binary search number of times M ₃

2) with x _i ⁰As the initial point of chaos iteration, utilize Skew Tent mapping to generate chaos sequence, and foundation and optimization range [x _Imin, x _Imax] between corresponding relation, iteration is carried out M ₁The optimizing of inferior coarse search chaos, if Then obtain current optimum point and optimum solution

3) carry out M ₂Inferior pattern search fine searching fast seeking, if

Then

4) check stopping criterion for iteration is if twice Function Optimization value is poor

Or t＞M ₃(t=1,2 ..., M ₃), then algorithm finishes, and obtains x _i ^*And f ^*Otherwise make x _i ^*Add microvariations, Economy β ∈ (0,1), t=t+1 changes step 2 over to).As shown in Figure 1.

Application mode search procedure of the present invention improves the efficient and the precision of chaos optimization fine searching, a kind of hybrid optimization algorithm of searching for based on the chaos optimization binding pattern of Skew Tent mapping has been proposed, global optimization ability with chaos optimization algorithm has the local fast seeking performance of pattern search method again.In addition, because algorithm is taked Skew Tent to shine upon to produce chaos sequence, overcome to a certain extent because the low searching efficiency problem that the traversal unevenness that Logistic shines upon causes.

Description of drawings

Fig. 1: the basic flow sheet of chaotic mixing optimized Algorithm;

Fig. 2: boiler kernel model structure;

Fig. 3: coal pulverizer dynamically reaches the water-cooling wall dynamic model;

Fig. 4: 500MW monoblock dynamic model structure;

Fig. 5: the input signal that is used for identification;

Fig. 6: the comparison of actual output of system and model output;

Fig. 7: non-linear prediction control system structure;

Fig. 8: the simulation result during the load step disturbance.

Embodiment

Below in conjunction with accompanying drawing, preferred embodiment is elaborated.Should be emphasized that following explanation only is exemplary, rather than in order to limit the scope of the invention and to use.

The invention provides a kind of monoblock non-linear predication control method based on the chaotic mixing optimized Algorithm.

The modeling object that present embodiment is chosen is the steam turbine power generation unit of refreshing two 500MW of power plant in Shanxi, its specified main steam flow D _TeBe 1650t/h, specified drum pressure P _DeBe 18.97MPa, specified main steam pressure P _TeBe 16.18MPa.The dum boiler monoblock is the multi-variable system of a complexity, generally can be reduced to an object with dual input dual output.

Step 1: monoblock is carried out modelling by mechanism.

Shown in Figure 2 is the kernel model of boiler, comprises two input quantities and two output quantities.Its input quantity is respectively effective caloric receptivity and valve opening instruction; Output quantity is respectively main steam flow and main steam pressure.Each coefficient implication is: K ₁Heat storage coefficient C for drum _dInverse; K ₃Heat storage coefficient C for main steam line _tInverse; K ₂, K ₄, K ₆, K ₆Be inner coefficient, be used for the internal balance of realistic model.The nominal parameter of unit drum pressure is P _De, be set to the initial value of integrator 1; The nominal parameter of main steam pressure is P _Te, be set to the initial value of integrator 2; The nominal parameter of main steam flow is D _TeSuppose effective caloric receptivity D of boiler _qWith steam turbine valve aperture instruction μ all standardization, then D _qCan be described as with μ

With

This model has embodied two energy equilibrium of unit and two nonlinear relationships:

One, two of the boiler dynamic model kinds of relevant energy balance relations:

1. drum pressure P _dThe effective caloric receptivity D that has reflected boiler _qWith drum outlet steam heating amount D _kBalance;

2. main steam pressure P _tReflected drum outlet steam heating amount D _kWith main steam thermal value D _tBalance.

Two, the nonlinear characteristic that exists in the boiler dynamic process mainly is reflected in two aspects:

1. drum pressure P _dWith main steam pressure P _tPressure fall with drum outlet steam flow D _kBetween exist square root relationship;

2. main steam flow D _tWith turbine governor valve flow area F and main steam pressure P _tThe proportional relation of product.

Model parameter is tried to achieve by the turbine governor valve upset test, and the parameter occurrence is as follows.

K ₁＝1.0/(27.7678×0.9)，

K ₂＝2.4082，

K ₃＝1.0/(27.7678×0.1)，

K ₄＝4.0224，

K ₅＝0.2486，

K ₆＝101.9778.

In order more intactly to reappear the boiler object, need derive the effectively dynamic model of caloric receptivity of boiler.Usually, stove internal combustion and diabatic process can be reduced to the dynamic and dynamic two parts of water-cooling wall of coal pulverizer, as shown in Figure 3.

The coal pulverizer dynamic model is:

The water-cooling wall dynamic model is:

By with Fig. 2 and Fig. 3 combination, can further obtain the model structure of whole 500MW monoblock, as shown in Figure 4.Wherein, the boiler kernel model includes nonlinear element.Main steam pressure is a controlled volume before the electromotive power output of unit and the machine; Main inlet throttle-stop valve control valve opening and fuel quantity are controlled quentity controlled variables, require to regulate according to valve opening instruction and quantity combusted instruction B respectively.

Step 2: initialization neural network identifier, initialization network weight, given prediction length and Chaos Variable.

Determine that neural network structure is (4-6-1) structure, its prediction length is that 3, two Chaos Variable get 0.2 and 0.21 respectively.

Step 3: the off-line training neural network identifier according to the described modelling by mechanism of step 1, generates the monoblock forecast model f based on neural network _NN

The discrete system of supposing a NARMAX model description is:

y(k)＝f[y(k-1)，…，y(k-n)，u(k-1)，…，u(k-m)]+ε(k)

In the formula: f () is a continuous complex nonlinear function; M and n are respectively the order of output y (k) and input u (k); ε (k) is a zero mean noise.This model is normally unknown, and present embodiment adopts as drag is the identification model of forecast model:

$\hat{y}\left(k\right)=f_{\mathrm{NN}}[y(k-1),...,y(k-n),u(k-1),...,u(k-m)]+\mathrm{\ϵ}\left(k\right)$

With neural network 500MW monoblock dynamic model shown in Figure 4 is carried out identification.

Fig. 5 is the input signal (quantity combusted instruction and valve opening instruction) that is used for identification, and Fig. 6 is the identification result of electromotive power output and main vapour pressure, and error is respectively 1.0513 * 10 ^-5With 1.7361 * 10 ^-3As can be seen from Figure 6, the dynamic perfromance that the neural network model of being built can well the predicting unit unit, wherein network weight adopts w ₁=rands (4,6) and w ₂=rands (6,1) function picked at random.

Step 4: above-mentioned forecast model is dropped in line traffic control, as shown in Figure 7, keeps network weight constant,, calculate the output of forecast model and carry out recursion, obtain the multi-step prediction result with the input of control variable as forecast model.

In the present embodiment, control variable u ₁=80; u ₂=85, be output as 450MW.

Wherein, utilize iterative algorithm to obtain N step forecast model forward:

$\hat{y}(k+1|k)=f_{\mathrm{NN}}[y\left(k\right),...,y(k-n+1),u\left(k\right),...,u(k-m+1)];$

.

.

.

$\hat{y}(k+N|k)=f_{\mathrm{NN}}[\hat{y}(k+N-1|k),...,\hat{y}(k+N-n),$

u(k+N-1)，…，u(k+N-m)]

Wherein:

The model of predicting constantly at k for model is in k+j output valve constantly, j=1 ..., N in iterative process, uses the constantly later predicted value of k

The predicted value before the moment to k and k replaces with its actual value, promptly

$\hat{y}(k+i-j)=y(k+i-j),i-j\≤0,$

In recursive process, because current time u (k+j-1) ..., u (k) the unknown is in this order

u(k+j)＝u(k)，i＝1，…，j-1；

Step 5: utilize the optimizing of rolling of chaotic mixing optimized Algorithm, determine the control increment of algorithm, more current optimum control amount is acted on controlled device.

The control increment of above-mentioned algorithm Δ u (k+j-1), j=1 ..., m} is next definite for minimum by the value that makes optimization criterion J,

$J=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{[\hat{y}(k+j)-w(k+j)]}^2+\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\λ}\left(j\right){\left[\mathrm{\Δu}(k+j-1)\right]}^2,$

Wherein, n=3; M=2; λ (j)=0.1; The initial value of w is 450.

1) choosing two in (0,1) interval has several 0.2 and 0.21 of fine difference, as the initial value of Chaos Variable, makes current optimum Chaos Variable

Step-up error scope ε＞0; The maximum iteration time M of chaos optimization coarse search and pattern search is set ₁, M ₂And binary search number of times M ₃Be respectively 50,30 and 5;

2) with x _i ⁰As the initial point of chaos iteration, utilize Skew Tent mapping to generate chaos sequence, and foundation and optimization range [x _Imin, x _Imax] between corresponding relation, iteration is carried out M ₁The optimizing of inferior coarse search chaos, if

Then obtain current optimum point And optimum solution

Promptly

f ^*=450;

3) carry out M ₂Inferior pattern search fine searching fast seeking, if

Then

4) check stopping criterion for iteration is if twice Function Optimization value is poor

Or t＞M ₃(t=1,2 ..., M ₃), then algorithm finishes, and obtains x _i ^*And f ^*Otherwise make x _i ^*Add microvariations, Economy β=0.1, t=t+1 changes step 2 over to).

Step 6: according to the feedback information of above-mentioned control procedure, forecast model is further revised the network weight of neural network identifier.

Step 7: return step 4, enter next control cycle.

Verify the validity of present embodiment nonlinear prediction control strategy by an emulation experiment.

Test: unit is in the fixed pressure operation mode, and the instruction of will loading drops to 425MW from the 450MW step.Simulation result during the load step disturbance as shown in Figure 8.As seen the very fast and non-overshoot of the tracking velocity of power, the main vapour pressure amplitude of variation is little, and transit time is short.Be that the nonlinear prediction control strategy based on the chaotic mixing optimized Algorithm of the present invention has better controlling performance.

The above; only for the preferable embodiment of the present invention, but protection scope of the present invention is not limited thereto, and is familiar with those skilled in the art in the technical scope of the present invention's exposure; the variation that can expect easily or replacement all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (2)

1. based on the monoblock non-linear predication control method of chaotic mixing optimized Algorithm, it is characterized in that, may further comprise the steps:

Step 1: monoblock is carried out modelling by mechanism;

Step 2: initialization neural network identifier, initialization network weight, given prediction length and Chaos Variable;

Step 3: the off-line training neural network identifier according to the described modelling by mechanism of step 1, generates the monoblock forecast model f based on neural network _NN

Step 4: above-mentioned forecast model is dropped in line traffic control, keeps network weight constant,, calculate the output of forecast model and carry out recursion, obtain the multi-step prediction result with the input of control variable as forecast model; Wherein, utilize iterative algorithm to obtain N step forecast model forward:

$\hat{y}(k+1|k)=f_{\mathrm{NN}}[y\left(k\right),...,y(k-n+1),u\left(k\right),...,u(k-m+1)];$

；

$\hat{y}(k+N|k)=f_{\mathrm{NN}}[\hat{y}(k+N-1|k),...,\hat{y}(k+N-n),$

$u(k+N-1),...,u(k+N-m)]$

Wherein: The model of predicting constantly at k for model is in k+j output valve constantly, j=1 ..., N in iterative process, uses the constantly later predicted value of k

The predicted value before the moment to k and k replaces with its actual value, promptly

$\hat{y}(k+i-j)=y(k+i-j),i-j\≤0,$

Wherein,

u(k+j)＝u(k)，i＝1，…，j-1；

Step 5: utilize the optimizing of rolling of chaotic mixing optimized Algorithm, determine the control increment of algorithm, more current optimum control amount is acted on controlled device;

The control increment of above-mentioned algorithm Δ u (k+j-1), j=1 ..., m} is next definite for minimum by the value that makes optimization criterion J,

$J=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{[\hat{y}(k+j)-w(k+j)]}^2+\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\λ}\left(j\right){\left[\mathrm{\Δu}(k+j-1)\right]}^2,$

Wherein, n and m are respectively maximum predicted length and control length; λ (j) generally gets constant λ for the control weighting factor, and w (k+j) is a k+j reference locus value constantly;

Step 6: online correction neural network identifier network weight;

Step 7: return step 4, enter next control cycle.

2. the monoblock non-linear predication control method based on the chaotic mixing optimized Algorithm according to claim 1, it is characterized in that, the described chaotic mixing optimized Algorithm of step 5 is based on the chaotic mixing optimized Algorithm of Skew Tent mapping, may further comprise the steps:

Starting condition: establishing a class optimization problem is:

minf(x _i)，x _imin≤x _i≤x _imax，

i＝1，2，…，n.

Wherein, f (x _i) be any optimization aim function;

1) chooses n the initial value x that fine difference is arranged in (0,1) interval _i ⁰(i=1,2 ..., n),, make current optimum Chaos Variable as the initial value of Chaos Variable

Step-up error scope ε＞0; The maximum iteration time M of chaos optimization coarse search and pattern search is set ₁, M ₂And binary search number of times M ₃

2) with x _i ⁰As the initial point of chaos iteration, utilize Skew Tent mapping to generate chaos sequence, and foundation and optimization range [x _Imin, x _Imax] between corresponding relation, iteration is carried out M ₁The optimizing of inferior coarse search chaos, if

Then obtain current optimum point

And optimum solution

3) carry out M ₂Inferior pattern search fine searching fast seeking, if

Then

4) check stopping criterion for iteration is if twice Function Optimization value is poor Or t＞M ₃(t=1,2 ..., M ₃), then algorithm finishes, and obtains x _i ^*And f ^*Otherwise make x _i ^*Add microvariations,

Economy β ∈ (0,1), t=t+1 changes step 2 over to).

Images