CN105159071A - Method for estimating economic performance of industrial model prediction control system in iterative learning strategy - Google Patents

Abstract

The invention discloses a method for estimating the economic performance of an industrial model prediction control system in an iterative learning strategy. The economic performance and system state of first iteration are estimated by calculating process data and the weight parameter, whether the economic performance of the system is converged during the first iteration is determined according to the system state, and if the economic performance is converged, iteration is stopped, and a control scheme of the first iteration is used for industrial control; and if the economic performance is not converged, the weight parameter of iteration of the next time is obtained by estimating the system state of the first iteration and utilizing the ILC learning rate, the input increment and the output increment which enable optimal economic performance of the iteration of the next time are solved, process data of the iteration of the next time is calculated via the process data of the first iteration, and the iteration of the next time is implemented. A linear programming problem is posed via the data obtained online, the linear programming problem can be used to provide net setting points for model prediction control, and thus, the economic performance of a controller is improved.

Description

The economic performance appraisal procedure of industrial model predictive control system under a kind of iterative learning strategy

Technical field

The present invention relates to the economic performance assessment technology field of the process industry such as oil, chemical industry model predictive control system, be specifically related to the economic performance appraisal procedure meter of industrial model predictive control system under a kind of iterative learning strategy.

Background technology

Actual industrial process is all carry out continuous print production.At present, industry member faces again various challenge, the fluctuation of the prices of raw and semifnished materials, the impact of environmental factor, fierceness of market competition etc.These factors result in larger investment risk.In order to ensure the economic benefit continued, increasing enterprise considers to use Advanced process control (APC) to improve the quality of products, and reduces energy consumption, ensures production security.In the past in the middle of 10 years, Model Predictive Control widespread use in the industrial production, especially petrochemical industry.Model Predictive Control can be improved the quality of products by improving input and output variance relation to a certain extent, and this business efficiency for enterprise is extremely important.

At present, be all adopt model predictive controller in most of Advanced process control practical application, reason is that Model Predictive Control effectively can process constraint and can be applied in multi-input multi-output system.A major issue of the PREDICTIVE CONTROL that designs a model controller is the weight parameter determining controller.In addition, suitable Model Predictive Control set point is selected also very important, selects more suitable set point can improve the economic performance of control system.

In order to fill up the blank being realized the control method that can reach between Optimum Economic performance by the setting of adjustment model PREDICTIVE CONTROL weight parameter, need design optimal weights parameter regulation means.Adjustment weight parameter can change input and output variance relation, but is in general difficult to the analytic relationship obtaining input and output variance and weight parameter, because usually controller weighting parameter cannot be adjusted to each value.

The setting parameter that (ILC) carrys out adjustment model PREDICTIVE CONTROL is controlled by iterative learning.Iterative Learning Control Algorithm is used in mechanical arm the earliest and controls and on robot controlling, iterative learning controls to carry out repetitive learning by a deviation for before constantly utilizing batch of control.Iterative learning controls the control out being done mechanical arm the earliest by the Uchiyama of Japan.Then in 1984, Arimoto etc. describe this method in English.Iterative learning controls to start to develop fast from then on.The research that in great majority process control in the past, relevant iterative learning controls is all concentrate on batch process.Iterative learning control also with and Model Predictive Control combine and be applied in batch process.According to the Survey of Iterative Learning Control article of Wang etc., iterative learning controls to be divided into direct iteration study control and Indirect iteration study control.The input whether being directly used in real system according to the output being iterative learning control of classification.

But existing iterative learning controls to be Off-line control, cannot disturb, need matching Input output Relationship curve to model prediction, and calculated amount is large, causes realtime control not high.

Summary of the invention

For the deficiencies in the prior art, the economic performance appraisal procedure of industrial model predictive control system under a kind of iterative learning strategy that the present invention proposes.

Under iterative learning strategy, an economic performance appraisal procedure for industrial model predictive control system, comprises the steps:

(1) initialization iterations i is 0, collects process data and the weight parameter of current generation from MPC, calculates input variance during i-th iteration respectively output variance input-mean with output average estimate the economic performance of i-th iteration

(2) the input increment of economic performance optimum and output increment when utilizing the stepping type of described MPC economic performance function to solve to make the i-th+1 time iteration according to the result of calculation of step (1);

(3) the LaGrange parameter vector detecting input and output according to the result of step (1) and (2) determines system state η during current iteration _i, η _i=0 or ± 1;

(4) according to system state η _iwhen judging current iteration, whether systematic economy performance restrains:

If convergence, then think current iteration time economic performance optimum, and utilize the control program of current iteration to carry out Industry Control;

Otherwise, proceed as follows:

(4-1) according to system state η during current iteration _iutilize weight parameter during ILC learning rate acquisition the i-th+1 time iteration;

(4-2) according to the input-mean calculated in step (1) with output average and the input increment that obtains of step (2) and output increment process data when calculating the i-th+1 time iteration;

(4-3) upgrading current iteration number of times is i+1, and returns step (1) based on the result of step (4-1) and (4-2).

Process data in described step (1) refers to the input and output at setting-up time point place, and when industrial model predictive control system exists multiple input variable, corresponding input setting value should be a multi-C vector.Now, variance is inputted input-mean should be all multi-C vector, the input variable of corresponding one of the component of multi-C vector in each dimension.In like manner, when there is multiple output variable in industrial model predictive control system, output variance with output average also should be multi-C vector, and the corresponding output variable of its component in each dimension.

According to system state η during following formula determination current iteration in described step (3) _i:

${\mathrm{\η}}_i=\left\{\begin{array}{cc}1,& B_u>0and{B}_y=0\\ -1,& B_u=0and{B}_y>0\\ 0,& B_u>0and{B}_y>0\\ 0,& B_u=0and{B}_y=0\end{array}\right.,$

Wherein, β _nfor the n-th component of the importation in described LaGrange parameter vector, β _pfor p component of the output in described LaGrange parameter vector.

β _nand β _pdetermined by following formula:

$u_n^{\min }+z_{{\mathrm{\α}}_n/2}({\mathrm{\σ}}_{u_{i,n}^s}+{\mathrm{\Δ\σ}}_{u_{i,n}})-u_{i,n}^s\≤{\mathrm{\Δu}}_{i,n}^s\≤u_n^{\max }-z_{{\mathrm{\α}}_n/2}({\mathrm{\σ}}_{u_{i,n}^s}+{\mathrm{\Δ\σ}}_{u_{i,n}})-u_{i,n}^s$

$y_p^{\min }+z_{{\mathrm{\α}}_p/2}({\mathrm{\σ}}_{y_{i,p}}+{\mathrm{\Δ\σ}}_{y_{i,p}})-y_{i,p}^s\≤\underset{n=1}{\overset{N}{\Σ}}{k}_{np}{\mathrm{\Δu}}_{i,n}^s\≤y_p^{\max }-z_{{\mathrm{\α}}_p/2}({\mathrm{\σ}}_{y_{i,p}}+{\mathrm{\Δ\σ}}_{y_{i,p}})-y_{i,p}^s,$

Wherein, be respectively lower limit and the upper limit of input increment restriction, with the n-th component of variance and p component of output variance is inputted, n=1 when being respectively i-th iteration ..., N, N are the number of described industrial model predictive control system input, p=1,2 ..., P, P are the number that described industrial model predictive control system exports,

with input n-th component of increment of variance and p component of the increment of output variance when being respectively i-th iteration, the result according to described step (1) and step (2) calculates,

with when being respectively i-th iteration, the n-th input-mean and p export average,

the increment of the n-th input-mean when being i-th iteration,

with be respectively lower limit and the upper limit of output increment constraint,

K _npfor the steady-state gain of described industrial model predictive control system,

with be respectively the confidence interval coefficient of not arranging to retrain of input and output.

Described step (2) is specific as follows:

$\begin{array}{c}\underset{{\mathrm{\Δy}}_i^s,{\mathrm{\Δu}}_i^s}{\max }{J}_{i+1}\left(y_i^s+{\mathrm{\Δy}}_i^s,u_i^s+{\mathrm{\Δu}}_i^s,{\mathrm{\σ}}_{y_i^s},{\mathrm{\σ}}_{u_i^s}\right)=\\ J_i\left(y_i^s,u_i^s,{\mathrm{\σ}}_{y_i^s},{\mathrm{\σ}}_{u_i^s}\right)+\underset{{\mathrm{\Δy}}_i^s,{\mathrm{\Δu}}_i^s}{\max }\[\underset{p=1}{\overset{P}{\Σ}}{C}_y^{\left(p\right)}{\mathrm{\Δy}}_{i,p}^s-\underset{n=1}{\overset{N}{\Σ}}{C}_u^{\left(n\right)}{\mathrm{\Δu}}_{i,n}^s\]\end{array}$

$\begin{array}{cc}s.t.& {\mathrm{\Δy}}_{i,p}^s=\underset{n=1}{\overset{N}{\Σ}}{k}_{np}{\mathrm{\Δu}}_{i,n}^s\end{array}$

$u_n^{\min }+z_{{\mathrm{\α}}_n/2}{\mathrm{\σ}}_{u_{i,n}^s}-u_{i,n}^s\≤{\mathrm{\Δu}}_{i,n}^s\≤u_n^{\max }-z_{{\mathrm{\α}}_n/2}{\mathrm{\σ}}_{u_{i,n}^s}-u_{i,n}^s$

$y_p^{\min }+z_{{\mathrm{\α}}_p/2}{\mathrm{\σ}}_{y_{i,p}^s}-y_{i,p}^s\≤\underset{n=1}{\overset{N}{\Σ}}{k}_{np}{\mathrm{\Δu}}_{i,n}^s\≤y_p^{\max }-z_{{\mathrm{\α}}_p/2}{\mathrm{\σ}}_{y_{i,p}^s}-y_{i,p}^s$

$0\≤\left|{\mathrm{\Δu}}_{i,n}^s\right|\≤{\mathrm{\Δu}}_i^{\max }$

$0\≤\left|{\mathrm{\Δy}}_{i,p}^s\right|\≤{\mathrm{\Δy}}_p^{\max }$

Wherein, for the optimized coefficients of the n-th input of described industrial model predictive control system,

for the n-th optimized coefficients exported of described industrial model predictive control system, be respectively input increment and the output increment of economic performance optimum when making the i-th+1 time iteration.

The lower limit of input increment restriction and the upper limit the lower limit of output increment constraint and the upper limit the steady-state gain k of industrial model predictive control system _np, do not arrange retrain confidence interval coefficient and the optimized coefficients of each input and output presets according to embody rule situation.

When judging current iteration in described step (4) by the following method, whether systematic economy performance restrains:

If system state η during current iteration _i=0 or systematic error when being less than preset value, then think current iteration time systematic economy performance convergence;

Otherwise, think current iteration time systematic economy performance do not restrain.

In the present invention, systematic error is: || b-Au||, wherein:

B is default output increment matrix, the steady state gain matrix of system when A is current iteration, the input Increment Matrix of system when u is current iteration.

As preferably, the span of described preset value is 10 ^-6~ 10 ^-5.

Described step (4-1) comprises the steps:

If η _i=η _i-1=η _i-2=...=η ₀, then weight parameter λ when obtaining the i-th+1 time iteration according to following formulae discovery _i+1:

λ _i+1＝λ _i+l(η _i,λ _i)η _i，

Wherein, λ _iit is weight parameter during i-th iteration; L (η _i, λ _i) change step of weight parameter when being i-th iteration:

$l({\mathrm{\&eta;}}_i,{\mathrm{\&lambda;}}_i)=\left\{\begin{array}{cccc}{l}_1{\mathrm{\&lambda;}}_i,& l_1>1& if& {\mathrm{\&eta;}}_i=1\\ l_2{\mathrm{\&lambda;}}_i,& 0<l_2<1& if& {\mathrm{\&eta;}}_i=-1\end{array}\right.;$

Otherwise, according to weight parameter λ during following method calculating the i-th+1 time iteration _i+1:

λ _i+1＝0.5(λ _i-1+λ _i),ifη _i≠η _i-1

。

λ _i+1＝0.5(λ _i+λ _τ),ifη _i＝η _i-1＝...＝η _τ+1≠η _r

Wherein, η _iunder representing corresponding four kinds of situations, system state is expressed, l (η _i) be depend on system state η _ilearning rate.L (η _i) mainly consider following two kinds of situations: (1) B _u> 0 and B _y=0, B _u=0 and B _y> 0.Because when economic performance objective function increases continuously, the increase that weight parameter should be dull or reduction; That is, now η _i+1=η _i=η _i-1=η _i-2=....According to the learning rate of current system state, can judge whether to need by increasing λ, or reduce λ, or allow it constantly increase economic performance objective function.

If economic performance objective function experienced by a reduction and a situation about increasing two continuous print (λ), i.e. η _i-1≠ η _i, now optimum λ mono-fixes on λ _iand λ _i-1interval in the middle of.If η _i=η _i-1=η _i-2=...=η _τ+1≠ η _τwherein system state occurs and only after phase points τ, changes symbol, now weight parameter λ should meet λ _τ> λ ^*> λ _i> λ _i-1> λ _i-2> ... > λ _τ+1or λ _τ< λ ^*< λ _i< λ _i-1< λ _i-2< ... < λ _τ+1.At this time show optimum λ ^*at interval λ _τand λ _ibetween, the next stage should from 0.5 (λ _τ+ λ _i) start to consider interval reduction.

Compared with prior art, the present invention has following effect:

A () constructs linear programming problem by the data obtained online, therefore can avoid the nonlinear programming problem solving LQG Performance Evaluation; Only utilize linear programming problem just can provide a new set point (namely inputting setting value) to Model Predictive Control, carry out the economic performance of lifting controller with this.

B design that () controls based on the iterative learning of sensitivity analysis can on-line tuning Model Predictive Control weight parameter, thus promotes economic performance further.

C the economic performance of () system can constantly promote in each data collection phase until arrive optimum economic performance, and the Optimum Economic performance that can reach surely, and need to obtain LQG curve by homing method unlike the assessment of LQG economic performance, and optimum point cannot reach.

Embodiment

Describe the present invention below in conjunction with specific embodiment.

The industrial model predictive control system of (rectification column) detachment process of the present embodiment comprises two input: return velocity u ₁, steam flow u ₂; Two outputs: product of distillation y ₁, bottom product y ₂, also have a disturbance: feed quantity d simultaneously.System model (i.e. industrial model predictive control system) can be expressed as follows:

$\left[\begin{array}{c}{y}_1\\ y_2\end{array}\right]=\left[\begin{array}{cc}\frac{2.56e^{-s}}{16.7s+1}& \frac{-5.76e^{-3s}}{21s+1}\\ \frac{1.32e^{-7s}}{10.9s+1}& \frac{-5.82e^{-3s}}{14.4s+1}\end{array}\right]\left[\begin{array}{c}{u}_1\\ u_2\end{array}\right]+\left[\begin{array}{c}\frac{3.8e^{-8s}}{14.9s+1}\\ \frac{4.9e^{-7s}}{13.2s+1}\end{array}\right]d$

Wherein, d is the white noise of zero mean unit variance.

Note product of distillation y ₁input-mean be note bottom product y ₂input-mean be for with confidence level be 99.7%.The up-and-down boundary value of input and output is as follows:

$u_1^{\max }=3,u_1^{\min }=-3,u_2^{\max }=5,u_2^{\min }=-5$

$y_1^{\max }=15,y_1^{\min }=-15,y_2^{\max }=15,y_2^{\min }=-15,$

Wherein, the subscript of variable: max represents coboundary value, and min represents lower boundary value.

The model predictive controller initial situation of the present embodiment is as follows:

The optimized coefficients matrix Q=diag [101] of the output of industrial model predictive control system (i.e. model predictive controller), namely $C_y^{\left(p\right)}(p=1)=C_{y_1}=10,C_{y_2}^{\left(p\right)}(p=1,...,P)=1.$

The optimized coefficients matrix Q=diag [101] of the input of industrial model predictive control system (i.e. model predictive controller), namely $C_u^{\left(n\right)}(n=1)=C_{u_1}=3,C_u^{\left(n\right)}(n=2)=C_{u_2}=3.$

The prediction time domain of MPC (Model Predictive Control) controller selects M=20.

The input constraint of model predictive controller is selected :-1≤Δ u ₁≤ 1 and-5≤Δ u ₂≤ 5.

The initial setting exported clicks and is selected as initial weight parameter is set to λ ₀=0.01.

The underlying model predictive controller (i.e. industrial model predictive control system) of this rectification column, its economic performance appraisal procedure is as follows:

Step 1: carry out to controller the setting that initialization completes initial parameter according to model predictive controller initial situation, comprises weight matrix and constraint condition (i.e. the up-and-down boundary values of input and output).

In the present embodiment, the process data of current generation comprises the input and output of current generation 0 ~ 3000 sampling time point forward.

Step 2: operation controller (model predictive controller, i.e. MPC), initialization iterations i is 0, collects process data and the weight parameter of current generation from MPC, calculates input variance during i-th iteration respectively output variance input-mean with output average estimate the economic performance of i-th iteration $J_i=J_i(y_i^s,u_i^s,{\mathrm{\&sigma;}}_{y_i^s},{\mathrm{\&sigma;}}_{u_i^s});$

In the present embodiment, the process data of current generation comprises the input and output of current generation 0 ~ 3000 sampling time point forward.

Step 3: the input increment of economic performance optimum and output increment when utilizing the stepping type of MPC economic performance function to solve to make the i-th+1 time iteration according to the result of calculation of step 2; Obtain especially by solving following function:

$\begin{array}{c}\underset{{\mathrm{\&Delta;y}}_i^s,{\mathrm{\&Delta;u}}_i^s}{\max }{J}_{i+1}\left(y_i^s+{\mathrm{\&Delta;y}}_i^s,u_i^s+{\mathrm{\&Delta;u}}_i^s,{\mathrm{\&sigma;}}_{y_i^s},{\mathrm{\&sigma;}}_{u_i^s}\right)=\\ J_i\left(y_i^s,u_i^s,{\mathrm{\&sigma;}}_{y_i^s},{\mathrm{\&sigma;}}_{u_i^s}\right)+\underset{{\mathrm{\&Delta;y}}_i^s,{\mathrm{\&Delta;u}}_i^s}{\max }\&lsqb;\underset{p=1}{\overset{P}{\&Sigma;}}{C}_y^{\left(p\right)}{\mathrm{\&Delta;y}}_{i,p}^s-\underset{n=1}{\overset{N}{\&Sigma;}}{C}_u^{\left(n\right)}{\mathrm{\&Delta;u}}_{i,n}^s\&rsqb;\end{array}$

$\begin{array}{cc}s.t.& {\mathrm{\&Delta;y}}_{i,p}^s=\underset{n=1}{\overset{N}{\&Sigma;}}{k}_{np}{\mathrm{\&Delta;u}}_{i,n}^s\end{array}$

$u_n^{\min }+z_{{\mathrm{\&alpha;}}_n/2}{\mathrm{\&sigma;}}_{u_{i,n}^s}-u_{i,n}^s\&le;{\mathrm{\&Delta;u}}_{i,n}^s\&le;u_n^{\max }-z_{{\mathrm{\&alpha;}}_n/2}{\mathrm{\&sigma;}}_{u_{i,n}^s}-u_{i,n}^s$

$y_p^{\min }+z_{{\mathrm{\&alpha;}}_p/2}{\mathrm{\&sigma;}}_{y_{i,p}^s}-y_{i,p}^s\&le;\underset{n=1}{\overset{N}{\&Sigma;}}{k}_{np}{\mathrm{\&Delta;u}}_{i,n}^s\&le;y_p^{\max }-z_{{\mathrm{\&alpha;}}_p/2}{\mathrm{\&sigma;}}_{y_{i,p}^s}-y_{i,p}^s$

$0\&le;\left|{\mathrm{\&Delta;u}}_{i,n}^s\right|\&le;{\mathrm{\&Delta;u}}_i^{\max }$

$0\&le;\left|{\mathrm{\&Delta;y}}_{i,p}^s\right|\&le;{\mathrm{\&Delta;y}}_p^{\max }$

Wherein, for the optimized coefficients of the n-th input of described industrial model predictive control system, need during use to preset

for the n-th optimized coefficients exported of described industrial model predictive control system, need during use to preset

be respectively input increment and the output increment of economic performance optimum when making the i-th+1 time iteration.

Step 4: the LaGrange parameter vector detecting input and output according to the result of step (2) and (3) determines system state η during current iteration _i, η _i=0 or ± 1;

According to system state η during following formula determination current iteration _i:

${\mathrm{\&eta;}}_i=\left\{\begin{array}{cc}1,& B_u>0and{B}_y=0\\ -1,& B_u=0and{B}_y>0\\ 0,& B_u>0and{B}_y>0\\ 0,& B_u=0and{B}_y=0\end{array}\right.,$

Wherein, β _nfor the n-th component of the importation in LaGrange parameter vector, β _pfor p component of the output in described LaGrange parameter vector, determine according to following formula:

$u_n^{\min }+z_{{\mathrm{\&alpha;}}_n/2}({\mathrm{\&sigma;}}_{u_{i,n}^s}+{\mathrm{\&Delta;\&sigma;}}_{u_{i,n}})-u_{i,n}^s\&le;{\mathrm{\&Delta;u}}_{i,n}^s\&le;u_n^{\max }-z_{{\mathrm{\&alpha;}}_n/2}({\mathrm{\&sigma;}}_{u_{i,n}^s}+{\mathrm{\&Delta;\&sigma;}}_{u_{i,n}})-u_{i,n}^s$

$y_p^{\min }+z_{{\mathrm{\&alpha;}}_p/2}({\mathrm{\&sigma;}}_{y_{i,p}}+{\mathrm{\&Delta;\&sigma;}}_{y_{i,p}})-y_{i,p}^s\&le;\underset{n=1}{\overset{N}{\&Sigma;}}{k}_{np}{\mathrm{\&Delta;u}}_{i,n}^s\&le;y_p^{\max }-z_{{\mathrm{\&alpha;}}_p/2}({\mathrm{\&sigma;}}_{y_{i,p}}+{\mathrm{\&Delta;\&sigma;}}_{y_{i,p}})-y_{i,p}^s,$

Wherein, be respectively lower limit and the upper limit of input increment restriction, with the n-th component of variance and p component of output variance is inputted, n=1 when being respectively i-th iteration ..., N, N are the number of described industrial model predictive control system input, p=1,2 ..., P, P are the number that described industrial model predictive control system exports,

with input n-th component of increment of variance and p component of the increment of output variance when being respectively i-th iteration, the result according to described step (1) and step (2) calculates,

with when being respectively i-th iteration, the n-th input-mean and p export average,

the increment of the n-th input-mean when being i-th iteration,

with be respectively lower limit and the upper limit of output increment constraint,

K _npfor the steady-state gain of described industrial model predictive control system,

with be respectively the confidence interval coefficient of not arranging to retrain of input and output.

Step 5: according to system state η _iwhen judging current iteration, whether systematic economy performance restrains:

If convergence, then think current iteration time economic performance optimum, and utilize the control program of current iteration to carry out Industry Control;

Otherwise, proceed as follows:

Step 5-1: according to system state η during current iteration _iutilize weight parameter during ILC learning rate acquisition the i-th+1 time iteration:

If η _i=η _i-1=η _i-2=...=η ₀, then weight parameter λ when obtaining the i-th+1 time iteration according to following formulae discovery _i+1:

λ _i+1＝λ _i+l(η _i,λ _i)η _i，

Wherein, λ _iit is weight parameter during i-th iteration; L (η _i, λ _i) change step of weight parameter when being i-th iteration:

$l({\mathrm{\&eta;}}_i,{\mathrm{\&lambda;}}_i)=\left\{\begin{array}{cccc}{l}_1{\mathrm{\&lambda;}}_i,& l_1>1& if& {\mathrm{\&eta;}}_i=1\\ l_2{\mathrm{\&lambda;}}_i,& 0<l_2<1& if& {\mathrm{\&eta;}}_i=-1\end{array}\right.;$

Otherwise, according to weight parameter λ during following method calculating the i-th+1 time iteration _i+1:

λ _i+1＝0.5(λ _i-1+λ _i),ifη _i≠η _i-1

λ _i+1＝0.5(λ _i+λ _τ),ifη _i＝η _i-1＝…＝η _τ+1≠η _τ。

Step 5-2: according to the input-mean calculated in step 2 with output average and the input increment that obtains of step 3 and output increment process data when calculating the i-th+1 time iteration;

Step 5-3: upgrading current iteration number of times is i+1, and returns step 2 based on the result of step 5-1 and step 5-2.

When judging current iteration in the present embodiment by the following method, whether systematic economy performance restrains:

If system state η during current iteration _i=0 or systematic error be less than preset value (in the present embodiment, the span of preset value be 10 ^-6) time, then think current iteration time systematic economy performance convergence;

Otherwise, think current iteration time systematic economy performance do not restrain.

In the present embodiment, systematic error is defined as: || b-Au||, wherein:

B is default output increment matrix, the steady state gain matrix of system when A is current iteration, the input Increment Matrix of system when u is current iteration.

By above method, the online raising of the economic performance to system can be realized.

Above-described embodiment has been described in detail technical scheme of the present invention and beneficial effect; be understood that and the foregoing is only most preferred embodiment of the present invention; be not limited to the present invention; all make in spirit of the present invention any amendment, supplement and equivalent to replace, all should be included within protection scope of the present invention.

Claims (8)

1. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy, is characterized in that, comprise the steps:

(1) initialization iterations i is 0, collects process data and the weight parameter of current generation from MPC, calculates input variance during i-th iteration respectively output variance input-mean with output average estimate the economic performance of i-th iteration

(2) the input increment of economic performance optimum and output increment when utilizing the stepping type of described MPC economic performance function to solve to make the i-th+1 time iteration according to the result of calculation of step (1);

(3) the LaGrange parameter vector detecting input and output according to the result of step (1) and (2) determines system state η during current iteration _i, η _i=0 or ± 1;

(4) according to system state η _iwhen judging current iteration, whether systematic economy performance restrains:

If convergence, then think current iteration time economic performance optimum, and utilize the control program of current iteration to carry out Industry Control;

Otherwise, proceed as follows:

(4-1) according to system state η during current iteration _iutilize weight parameter during ILC learning rate acquisition the i-th+1 time iteration;

(4-2) according to the input-mean calculated in step (1) with output average and the input increment that obtains of step (2) and output increment process data when calculating the i-th+1 time iteration;

(4-3) upgrading current iteration number of times is i+1, and returns step (1) based on the result of step (4-1) and (4-2).

2. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 1, is characterized in that, according to system state η during following formula determination current iteration in described step (3) _i:

${\mathrm{\&eta;}}_i=\left\{\begin{array}{cc}1,& B_u>0and{B}_y=0\\ -1,& B_u=0and{B}_y>0\\ 0,& B_u>0and{B}_y>0\\ 0,& B_u=0and{B}_y=0\end{array}\right.,$

Wherein, $B_u=\underset{n=1,...,2N}{max}{\mathrm{\&beta;}}_n,B_y=\underset{p=2N+1,...,2N+2P}{max}{\mathrm{\&beta;}}_p,$ β _nfor the n-th component of the importation in described LaGrange parameter vector, β _pfor p component of the output in described LaGrange parameter vector.

3. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 2, is characterized in that, β _nand β _pdetermined by following formula:

$u_n^{\min }+z_{{\mathrm{\&alpha;}}_n/2}({\mathrm{\&sigma;}}_{u_{i,n}^s}+{\mathrm{\&Delta;\&sigma;}}_{u_{i,n}})-u_{i,n}^s\&le;{\mathrm{\&Delta;u}}_{i,n}^s\&le;u_n^{\max }-z_{{\mathrm{\&alpha;}}_n/2}({\mathrm{\&sigma;}}_{u_{i,n}^s}+{\mathrm{\&Delta;\&sigma;}}_{u_{i,n}})-u_{i,n}^s$

$y_p^{\min }+z_{{\mathrm{\&alpha;}}_p/2}({\mathrm{\&sigma;}}_{y_{i,p}}+{\mathrm{\&Delta;\&sigma;}}_{y_{i,p}})-y_{i,p}^s\&le;\underset{n=1}{\overset{N}{\&Sigma;}}{k}_{np}{\mathrm{\&Delta;u}}_{i,n}^s\&le;y_p^{\max }-z_{{\mathrm{\&alpha;}}_p/2}({\mathrm{\&sigma;}}_{y_{i,p}}+{\mathrm{\&Delta;\&sigma;}}_{y_{i,p}})-y_{i,p}^s,$

Wherein, be respectively lower limit and the upper limit of input increment restriction,

with the n-th component of variance and p component of output variance is inputted, n=1 when being respectively i-th iteration ..., N, N are the number of described industrial model predictive control system input, p=1,2 ..., P, P are the number that described industrial model predictive control system exports,

with input n-th component of increment of variance and p component of the increment of output variance when being respectively i-th iteration, the result according to described step (1) and step (2) calculates,

with when being respectively i-th iteration, the n-th input-mean and p export average,

the increment of the n-th input-mean when being i-th iteration,

with be respectively lower limit and the upper limit of output increment constraint,

K _npfor the steady-state gain of described industrial model predictive control system,

with be respectively the confidence interval coefficient of not arranging to retrain of input and output.

4. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 3, it is characterized in that, described step (2) is specific as follows:

$\begin{array}{c}\underset{{\mathrm{\&Delta;y}}_i^s,{\mathrm{\&Delta;u}}_i^s}{\max }{J}_{i+1}\left(y_i^s+{\mathrm{\&Delta;y}}_i^s,u_i^s+{\mathrm{\&Delta;u}}_i^s,{\mathrm{\&sigma;}}_{y_i^s},{\mathrm{\&sigma;}}_{u_i^s}\right)=\\ J_i\left(y_i^s,u_i^s,{\mathrm{\&sigma;}}_{y_i^s},{\mathrm{\&sigma;}}_{u_i^s}\right)+\underset{{\mathrm{\&Delta;y}}_i^s,{\mathrm{\&Delta;u}}_i^s}{\max }\&lsqb;\underset{p=1}{\overset{P}{\&Sigma;}}{C}_y^{\left(p\right)}{\mathrm{\&Delta;y}}_{i,p}^s-\underset{n=1}{\overset{N}{\&Sigma;}}{C}_u^{\left(n\right)}{\mathrm{\&Delta;u}}_{i,n}^s\&rsqb;\end{array}$

$\begin{array}{cc}s.t.& {\mathrm{\&Delta;y}}_{i,p}^s=\underset{n=1}{\overset{N}{\&Sigma;}}{k}_{np}{\mathrm{\&Delta;u}}_{i,n}^s\end{array}$

$u_n^{\min }+z_{{\mathrm{\&alpha;}}_n/2}{\mathrm{\&sigma;}}_{u_{i,n}^s}-u_{i,n}^s\&le;{\mathrm{\&Delta;u}}_{i,n}^s\&le;u_n^{\max }-z_{{\mathrm{\&alpha;}}_n/2}{\mathrm{\&sigma;}}_{u_{i,n}^s}-u_{i,n}^s$

$y_p^{\min }+z_{{\mathrm{\&alpha;}}_p/2}{\mathrm{\&sigma;}}_{y_{i,p}^s}-y_{i,p}^s\&le;\underset{n=1}{\overset{N}{\&Sigma;}}{k}_{np}{\mathrm{\&Delta;u}}_{i,n}^s\&le;y_p^{\max }-z_{{\mathrm{\&alpha;}}_p/2}{\mathrm{\&sigma;}}_{y_{i,p}^s}-y_{i,p}^s$

$0\&le;\left|{\mathrm{\&Delta;u}}_{i,n}^s\right|\&le;{\mathrm{\&Delta;u}}_i^{\max }$

$0\&le;\left|{\mathrm{\&Delta;y}}_{i,p}^s\right|\&le;{\mathrm{\&Delta;y}}_p^{\max }$

Wherein, for the optimized coefficients of the n-th input of described industrial model predictive control system,

for the n-th optimized coefficients exported of described industrial model predictive control system,

be respectively input increment and the output increment of economic performance optimum when making the i-th+1 time iteration.

5. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 4, is characterized in that, when judging current iteration in described step (4) by the following method, whether systematic economy performance restrains:

If system state η during current iteration _i=0 or systematic error when being less than preset value, then think current iteration time systematic economy performance convergence;

Otherwise, think current iteration time systematic economy performance do not restrain.

6. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 5, it is characterized in that, systematic error is: || b-Au||, wherein:

B is default output increment matrix, the steady state gain matrix of system when A is current iteration, the input Increment Matrix of system when u is current iteration.

7. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 5, it is characterized in that, the span of described preset value is 10 ^-6~ 10 ^-5.

8. the economic performance appraisal procedure of industrial model predictive control system under iterative learning strategy as claimed in claim 7, it is characterized in that, described step (4-1) comprises the steps:

If η _i=η _i-1=η _i-2=...=η ₀, then weight parameter λ when obtaining the i-th+1 time iteration according to following formulae discovery _i+1:

λ _i+1＝λ _i+l(η _i,λ _i)η _i，

Wherein, λ _iit is weight parameter during i-th iteration; L (η _i, λ _i) change step of weight parameter when being i-th iteration:

$l({\mathrm{\&eta;}}_i,{\mathrm{\&lambda;}}_i)=\left\{\begin{array}{cccc}{l}_1{\mathrm{\&lambda;}}_i,& l_1>1& if& {\mathrm{\&eta;}}_i=1\\ l_2{\mathrm{\&lambda;}}_i,& 0<l_2<1& if& {\mathrm{\&eta;}}_i=-1\end{array}\right.;$

Otherwise, according to weight parameter λ during following method calculating the i-th+1 time iteration _i+1:

λ _i+1＝0.5(λ _i-1+λ _i),ifη _i≠η _i-1

λ _i+1＝0.5(λ _i+λ _τ),ifη _i＝η _i-1＝…＝η _τ+1≠η _τ。