CN102069095A - Statistical learning-based method for predicting and controlling finish rolling temperature in fine rolling - Google Patents

Abstract

The invention discloses a statistical learning-based method for predicting and controlling finish rolling temperature in fine rolling, and relates to the field of temperature control. The conventional fine rolling temperature control has typical hysteresis and is difficult to get the ideal control effect. The invention provides the statistical learning-based method for predicting and controlling the finish rolling temperature in fine rolling, in each control cycle, the threading speed and the water amount of an appointed frame are used as independent variables, the finish rolling temperature is used as a dependent variable, and a statistical prediction model is established through fine rolling process data for performing real-time prediction on the finish rolling temperature; and then the receding horizon optimization algorithm is adopted for giving out a finite horizon receding optimization control decision based on the given finish rolling target temperature and the prediction model under the premise that the predicted value of the prediction mode is consistent with the changing trend of the actual value. The method can give out the on-line real-time temperature prediction according to real-time dynamic situation of the fine rolling, make a real-time optimization control decision for the finish rolling temperature according to the temperature prediction and further meet the requirement of real-time property.

Description

A kind of prediction of finish rolling finishing temperature and control method based on statistical learning

Technical field

The present invention relates to the hot continuous rolling production control method, be specifically related to a kind of prediction of finish rolling finishing temperature and control method based on statistical learning.

Background technology

In hot-strip was produced, the control accuracy of mm finishing mill unit finishing temperature directly had influence on the structure property of final products.The control of fine-rolling strip steel total length finishing temperature is the important subject during hot rolling is produced always, also is one of difficult point.Existing hot-rolled strip production line, generally be to wear tape speed and temperature acceleration by design in advance, and shower water between frame is carried out PID control, realize control to finishing temperature, but final rolling temperature control has typical hysteresis quality, is difficult to obtain ideal control effect with classical PID control method.

For this reason, people attempt improving, and explore new method.Carry out the prediction and the control of finish rolling finishing temperature as the adopting process model.Because in the finish rolling production process, the factor that influences the finish rolling finishing temperature is numerous, as cooling water inflow, mill speed and the acceleration etc. between the keeping warm mode of delay table, high pressure descaling pattern, each frame.The adopting process model carries out temperature forecast, consider very many factors, comprises various process variables and various thermal conduction study physics coefficients etc., and this brings very big uncertainty for the accuracy of process modeling.Simultaneously, it is dynamic in real time that process modeling also is difficult to tracing process.

Summary of the invention

The purpose of this invention is to provide a kind of prediction of finish rolling finishing temperature and control method based on statistical learning, by process data is carried out real-time sampling and statistical learning, set up the real-time estimate model of finish rolling finishing temperature, and the finish rolling finishing temperature is carried out real-time optimization control, solve the operating mode of can't monitoring in real time that prior art exists and set up final rolling temperature control model, and the not high problem of control accuracy that causes by control hysteresis.

For achieving the above object, the present invention takes following technical scheme:

A kind of prediction of finish rolling finishing temperature and control method based on statistical learning, form by following steps:

(1) input/output variable and the order thereof of setting sampling period, forecast model, wherein the input variable of forecast model is worn tape speed for specifying the frame water yield and finish to gauge frame, and the output variable of forecast model is a finishing temperature;

(2) in each fixing sampling period, carry out following steps:

The input/output variable of a, collection forecast model, and, determine model structure according to the forecast model variable order that step (1) is set, construct corresponding inputoutput data regression vector matrix and modeling training dataset;

B, the inputoutput data regression vector matrix that step (a) is obtained carry out denoising;

C, based on the modeling training dataset that step (a) obtains, set up the non-linear input and output forecast model of finishing temperature with statistical method;

D, check forecast model are to the consistent degree of the predicted value and the actual value variation tendency of finishing temperature, if basically identical, then with nonlinear model in the present operating point linearisation, obtain linear model; Otherwise, do not carry out subsequent control and calculate, directly return step (a);

E, the linear model that obtains based on step (d), selecting to specify the frame water yield is control variables, and with reference to physical device and process conditions, definition finite time-domain rolling optimization problem;

The optimization problem that obtains in f, the solution procedure (e) obtains corresponding control decision vector, first element of this vector after amplitude limiting processing as the control decision in current sampling period;

(3) repeating step (a)---step (f) to control task finishes, and makes up final rolling temperature prediction and control system;

In the described step (a),, all can gather new the appointment frame water yield, finish to gauge frame and wear tape speed and finishing temperature, and new data are joined the modeling training dataset in each sampling period.

Described step (b) comprises the steps:

According to finish rolling equipment and process conditions, determine variable finishing temperature, the bound of specifying the frame water yield and finish to gauge frame to wear tape speed and rate of change thereof;

Bound with variate-value and variable rate is carried out denoising to all real time datas.

Statistical method in the described step (c) is for setting up statistical models, and the kernel function in the statistical models is the RBF kernel function, and concrete nonlinear model form is:

$y\left(x\right)=\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_k\exp \{-{\left|\right|x-x_k\left|\right|}_2^2/{\mathrm{\σ}}^2\}+b$

Wherein, x _kBe support vector, α _kBe corresponding support vector x _kWeight coefficient, x is current regression data vector, y (x) is and the corresponding output of x, i.e. finishing temperature predicted value.

Described step (d) comprises the steps:

Contrast finishing temperature actual value sequence T _a({ T _a(t), T _a(t-1), T _a(t-2), L}) and model prediction value sequence T _p({ T _p(t-1), T _p(t-2), L}), if between finishing temperature actual value sequence and the model prediction value sequence continuous some sampling periods of the AC compounent of deviation be no more than the AC compounent upper limit of appointment, then assert the predicted value and the actual value basically identical of current time finishing temperature.

If forecast model then is located at the present operating point linearisation with non-linear mould to the predicted value and the actual value basically identical of finishing temperature, gained linear model form is as follows:

$T\left(t\right)=b_{1}q(t-1)+...+b_{n_q}q(t-n_q)-a_{1}T(t-1)-...-a_{n_T}(t-n_T)+$

$p_0+{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">c</figure-callout>}}_1*v(t-1)+...+c_{n_v}*v(t-n_v)$

If forecast model is inconsistent to the predicted value and the actual value of finishing temperature, then do not carry out follow-up control and calculate, directly return step (a).

Described step (e) comprises the steps:

With reference to physical device and process conditions, definition optimization aim function comprises temperature control deviation penalty term and controlled quentity controlled variable increment penalty term in the function; Selected optimization time domain and control time domain, thus be defined as follows rolling Optimization of Time Domain problem, and concrete formula is:

$\min J\left(t\right)=\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_i{[T_r(t+i)-\hat{T}(t+i)]}^2+\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j{\left[\mathrm{\&Delta;q}(t+j-1)\right]}^2$

Wherein, P is the prediction time domain, and M is the control time domain, T _rBe with reference to given, u _i(1≤i≤P) and λ _j(1≤j≤M) is the weight coefficient of corresponding penalty term.

Described step (f) comprises the steps:

By matrix computations, obtain following optimum control increment sequence [Δ q (t), Δ q (t+1), L, Δ q (t+M-1)];

With reference to physical device and process conditions, determine the bound constraint and the rate of change constraint of the water yield.If Δ q (t) transfinites, it is carried out amplitude limit;

Obtaining the control corresponding sequence of decisions by the control increment sequence is

[q(t-1)+Δq(t)，q(t-1)+Δq(t)+Δq(t+1)，L，q(t-1)+Δq(t)+L+Δq(t+M-1)]

Wherein first element q (t) of sequence=q (t-1)+Δ q (t) if q (t) transfinites, also will carry out amplitude limit, and final gained q (t) is exactly the control decision of current time.

Adopt the inventive method, have the following advantages:

1, Chang Yong nonlinear prediction model comprises neutral net and fuzzy model etc., and its learning algorithm all is based on the empiric risk minimization principle.All there is " over-fitting " problem in learning algorithm based on this principle.The present invention is based on statistical learning and structural risk minimization principle, set up the statistical models of finishing temperature, can take into account the empiric risk and the popularization ability of learning algorithm.This model has robustness and sparse property, and need not the line solver quadratic programming problem, can satisfy the real-time requirement of rapid system line modeling fully.

2, select for use among the present invention based on RBF as model and function, can institute's established model can fit the nonlinear characteristic of final rolling temperature system effectively.

3, this method can dynamically provide online real time temperature prediction in real time according to finish rolling, and survey is carried out the real-time optimization control decision to finishing temperature according to temperature, and then by control module control finishing temperature, thereby realize the control of finish rolling finishing temperature.

Description of drawings

Fig. 1 is a high-level schematic functional block diagram of the present invention

Fig. 2 is the FB(flow block) of finish rolling finishing temperature prediction of the present invention and control

The specific embodiment

Below in conjunction with drawings and Examples the present invention is described in detail.

Embodiment

As shown in Figure 2, a kind of prediction of finish rolling finishing temperature and control method based on statistical learning, form by following steps:

(1) input/output variable and the order thereof of setting sampling period, forecast model, wherein the input variable of forecast model is worn tape speed for specifying the frame water yield and finish to gauge frame, and the output variable of forecast model is a finishing temperature.According to finish rolling production technology and appointed condition, selecting the sampling period is about 1 second.Determine after the sampling period, can select frame water yield q, finish to gauge frame to wear the order of tape speed v and finishing temperature T accordingly.The variable order elects 3 as.With the frame of regulating water amount serves as to specify frame, and other frame water yield then keeps constant.

(2) in each fixing sampling period, carry out following steps:

The input/output variable of a, collection forecast model, and, determine model structure according to the forecast model variable order that step (1) is set, construct corresponding inputoutput data regression vector matrix and modeling training dataset; Current t input and output regression vector constantly is:

x(t)＝[q(t-1)，L，q(t-n _q)，v(t-1)，L，v(t-n _v)，T(t-1)，L，T(t-n _T)].?(1)

Wherein, n _q, n _vAnd n _TIt is respectively the order that frame water yield q, finish to gauge frame are worn tape speed v and finishing temperature T.A regression vector and corresponding finishing temperature constitute a training data, and for example: { x (t), T (t) }, it is remembered work { x as the sample in the training sample set _k, T _k.

Begun to current time t by the rolling initial moment, all regression vectors are formed the regression vector matrix

And the regression vector matrix is formed the modeling training dataset with corresponding finishing temperature

Wherein N is the number of data centralization training sample.

B, the inputoutput data regression vector matrix that step (a) is obtained carry out denoising, according to finish rolling equipment and process conditions, determine finishing temperature, specify the frame water yield and finish to gauge frame to wear the bound of variate-value such as tape speed and rate of change thereof.Bound with variate-value and variable rate is tested to all real time datas, removes all training samples that variate-value constituted that transfinite.

C, the modeling training dataset that obtains based on step (a) are set up the non-linear input and output forecast model of finishing temperature with statistical method, and concrete modeling method is:

At first, set up following matrix equation based on the statistical learning principle

$\left[\begin{array}{cc}0& {\stackrel{\&RightArrow;}{1}}^{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">T</figure-callout>}}\\ \stackrel{\&RightArrow;}{1}& \mathrm{\&Omega;}+{\mathrm{\&gamma;}}^{-1}I\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\&alpha;}\end{array}\right]=\left[\begin{array}{c}0\\ T\end{array}\right]---\left(2\right)$

Wherein, T=[T ₁, T ₂..., T _N] ^T

α=[α ₁..., α _N] ^TΩ is a square formation, and the capable element of its k row l is

K () is a kernel function.

Can try to achieve α and b by (2) formula, thereby obtain following nonlinear model:

$y\left(x\right)=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_k\exp \{-{\left|\right|x-x_k\left|\right|}_2^2/{\mathrm{\&sigma;}}^2\}+b---\left(3\right)$

Wherein, x _kBe support vector, α _kBe corresponding support vector x _kWeight coefficient, b is the model correction factor, x is current regression data vector (form is seen formula 1), T (x) is and the corresponding prediction of x output, i.e. finishing temperature predicted value.

D, check forecast model are to the consistent degree of the predicted value and the actual value variation tendency of finishing temperature, if basically identical, then with nonlinear model in the present operating point linearisation, obtain linear model; Otherwise, do not carry out subsequent control and calculate, directly return step (a); Specifically in the following manner:

The actual value sequence T of contrast finishing temperature reality _a(T _a(t), T _a(t-1), T _a(t-2), L}) and the model prediction value sequence $\hat{T}\left(\right\{\hat{T}(t-1),\hat{T}(t-2),L\left\}\right):$

Calculate the sliding average of prediction deviation

Judge the AC compounent of the prediction deviation of current time

Whether set up, if continuous m ₂Err (i)≤Err_Ac is all arranged, ((m in the individual sampling period ₂-1)≤and i≤t), then assert the predicted value and the actual value variation tendency basically identical of current time finishing temperature;

Wherein, Err_Ac is the AC compounent upper limit of the predicted value deviation of appointment.m ₁Be the time cycle of calculation deviation slip average, m2 is a round of visits of judging that predicted value is whether stable.In actual finish rolling is produced, debug the value of setting that can get these three parameters by working control, generally specify m ₁=15, m ₂=8, Err_Ac=10 ℃.

If forecast model is consistent with the actual value variation tendency to the predicted value of finishing temperature, then with nonlinear model in the present operating point linearisation.If current sampling instant is t, then current sampling point is x (t), conveniently makes x for expressing ₀=x (t).With formula (4) at an x ₀Linearisation type in place's has

$T\left(x\right)$

$=T\left(x\right){|}_{x=x_0}+\frac{\&PartialD;T}{\&PartialD;x\left(1\right)}{|}_{x=x_0}[x\left(1\right)-x_0\left(1\right)]+L+\frac{\&PartialD;T}{\&PartialD;x(n_q+n_v+n_T)}{|}_{x=x_0}[x(n_q+n_v+n_T)-x_0(n_q+n_v+n_T)]$

$=T\left(x\right){|}_{x=x_0}-\frac{\&PartialD;T}{\&PartialD;x\left(1\right)}{|}_{x=x_0}{x}_0\left(1\right)-...-\frac{\&PartialD;T}{\&PartialD;x(n_q+n_v+n_T)}{|}_{x=x_0}{x}_0(n_q+n_v+n_T)$

$+\frac{\&PartialD;T}{\&PartialD;x\left(1\right)}{|}_{x=x_0}x\left(1\right)+...+\frac{\&PartialD;T}{\&PartialD;x(n_q+n_v+n_T)}{|}_{x=x_0}x(n_q+n_v+n_T)$

$=p_0+{\mathrm{<figure-callout\; id="1"\; label="b"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">b</figure-callout>}}_1*x\left(1\right)+...+b_{n_q}*x\left(n_q\right)+{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">c</figure-callout>}}_1*x(n_q+1)+...+c_{n_v}*x(n_q+n_v)$

$-a_1*x(n_q+n_v+1)-...-a_{n_T}*x(n_q+n_v+n_T)---\left(4\right)$

Wherein, $\frac{\&PartialD;T}{\&PartialD;x\left(i\right)}=-\frac{2}{{\mathrm{\&sigma;}}^2}\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\left\{{\mathrm{\&alpha;}}_k\exp (-{\left|\right|x-x_k\left|\right|}_2^2/{\mathrm{\&sigma;}}^2)\right[x\left(i\right)-x_k\left(i\right)\left]\right\}(1\&le;i\&le;n_q+n_v+n_T),$

$b=\frac{\&PartialD;T}{\&PartialD;x\left(i\right)}{|}_{x=x_0}(1\&le;i\&le;n_q),$

$c_i=\frac{\&PartialD;T}{\&PartialD;x(n_q+i)}{|}_{x=x_0}(1\&le;i\&le;n_v),$

$a_i=\frac{\&PartialD;T}{\&PartialD;x(n_q{+n}_v+i)}{|}_{x=x_0}(1\&le;i\&le;n_T),$

$p_0=T\left(x\right){|}_{x=x_0}-\frac{\&PartialD;T}{\&PartialD;x\left(1\right)}{|}_{x=x_0}{x}_0\left(1\right)-...-\frac{\&PartialD;T}{\&PartialD;x(n_q+n_v+n_T)}{|}_{x=x_0}{x}_0(n_q+n_v+n_T)$

According to the arrangement of variable in the regression vector of formula (1), the model after the linearisation is as can be known

$T\left(t\right)=a_{1}T(t-1)+...+a_{n_T}T(t-n_T)=b_{1}q(t-1)+...+b_{n_q}q(t-n_q)$

$+c_1-v(t-1)+...+c_{n_v}*v(t-n_v)+p_0---\left(5\right)$

Promptly

A(z ^-1)T(t)＝B(z ^-1)q(t-1)+C(z ^-1)v(t-1)+p ₀ (6)

Wherein, $A\left(z^{-1}\right)=1+a_1{z}^{-1}+...+a_{n_T}{z}^{-n_T}$

$B\left(z^{-1}\right)={\mathrm{<figure-callout\; id="1"\; label="b"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">b</figure-callout>}}_1+b_2{z}^{-1}+...+b_{n_q}{z}^{-n_q+1},$

$C\left(z^{-1}\right)={\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">c</figure-callout>}}_1+c_2{z}^{-1}+...+c_{n_v}{z}^{-{\mathrm{<figure-callout\; id="1"\; label="n\; v\; +"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">n</figure-callout>}}_v+1}.$

E, the linear model that obtains based on step (d), selecting to specify the frame water yield is control variables, and with reference to physical device and process conditions, definition finite time-domain rolling optimization problem; The finish to gauge frame is worn tape speed and is preestablished, and belongs to known dynamic variable quantity, and therefore selecting to specify the frame water yield is control variables.

By formula (6), have

D(z ^-1)T(t)＝B(z ^-1)Δq(t-1)+C(z ^-1)Δv(t-1) (7)

Wherein, Δ=1-z ^-1, p ₀Irrelevant with the time, so Δ gp ₀=0.

The predictive equation of finishing temperature T is as follows as can be known by formula (8):

$C_D\hat{T}=C_B\mathrm{\&Delta;q}+C_C\mathrm{\&Delta;v}+H_B{\mathrm{\&Delta;q}}_p+H_D{T}_p+H_B{\mathrm{\&Delta;q}}_p---\left(8\right)$

Wherein,

Be following predicted temperature constantly,

Be historical temperature information, the rest may be inferred for other symbol.

Know C by (9) formula _DReversible, and have

$\hat{T}=C_D^{-1}{C}_B\mathrm{\&Delta;q}+C_D^{-1}{C}_C\mathrm{\&Delta;c}+C_D^{-1}{H}_B{\mathrm{\&Delta;q}}_p+C_D^{-1}{H}_D{T}_p+C_D^{-1}{H}_B{\mathrm{\&Delta;q}}_p=C_D^{-1}{C}_B\mathrm{\&Delta;q}+Q---\left(9\right)$

Wherein,

It is the current time Given information.

With reference to physical device and process conditions, definition optimization aim function comprises temperature control deviation penalty term and controlled quentity controlled variable increment penalty term in the function; Selected optimization time domain and control time domain, thus be defined as follows rolling Optimization of Time Domain problem:

$\min J\left(t\right)=\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_t{[T_r(t+i)-\hat{T}(t+i)]}^2+\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j{\left[\mathrm{\&Delta;q}(t+j-1)\right]}^2---\left(10\right)$

Wherein, P is the prediction time domain, and M is the control time domain, T _rBe with reference to given, u _i(1≤i≤P) and λ _j(1≤j≤M) is the weight coefficient of corresponding penalty term.

The optimization problem that obtains in f, the solution procedure (e) obtains corresponding control decision vector, first element of this vector after amplitude limiting processing as the control decision in current sampling period; Detailed process has for formula (11) is changed into vector form

$\min J={(T_r-\hat{T})}^T\mathrm{\&Phi;}(T_r-\hat{T})+{\mathrm{\&Delta;q}}^T\mathrm{\&Lambda;\&Delta;q}---\left(12\right)$

Wherein, T _r=[T _r(t+1), L, T _r(t+P)] ^T,

Δ q=[Δ q (t), L, Δ q (t+M-1)] ^TΦ is with μ _iBe the diagonal matrix of diagonal element, Λ is with λ _jDiagonal matrix for diagonal element.

Ask the minimum of a value of object function J, have

$\frac{\&PartialD;J}{\&PartialD;\mathrm{\&Delta;q}}=-2\frac{\&PartialD;{\hat{T}}^T}{\&PartialD;\mathrm{\&Delta;q}}\mathrm{\&Phi;}(T_r-\hat{T})+2\mathrm{\&Lambda;\&Delta;q}=0$

In conjunction with (9) formula, have

$\mathrm{\&Delta;q}=[C_B^T{\left(C_D^{-1}\right)}^T\mathrm{\&Phi;}{C}_D^{-1}{C}_B+\mathrm{\&Lambda;}{]}^{-1}{C}_B^T{\left(C_D^{-1}\right)}^T\mathrm{\&Phi;}[T_r-Q]---\left(13\right)$

The control corresponding sequence of decisions is

[q(t-1)+Δq(t)，q(t-1)+Δq(t)+Δq(t+1)，L，q(t-1)+Δq(t)+L+Δq(t+M-1)]

With reference to physical device and process conditions, determine the bound constraint and the rate of change constraint of the water yield,

$\left\{\begin{array}{c}0\&le;q(t+j-1)\&le;q_{\max }\\ \left|\mathrm{\&Delta;q}(t+j-1)\right|<{\mathrm{\&Delta;<figure-callout\; id="1"\; label="q\; max"\; filenames="H2009102727927E0000011.png"\; state="\{\{state\}\}">q</figure-callout>}}_{\max }\end{array}\right.$ 1≤j≤M

If Δ q (t) transfinites, it is carried out amplitude limit:

$\left\{\begin{array}{cc}\mathrm{\&Delta;q}\left(t\right)>{\mathrm{\&Delta;q}}_{\max }& \mathrm{\&Delta;q}\left(t\right)={\mathrm{\&Delta;q}}_{\max }\\ \mathrm{\&Delta;q}\left(t\right)<-{\mathrm{\&Delta;q}}_{\max }& \mathrm{\&Delta;q}\left(t\right)=-{\mathrm{\&Delta;q}}_{\max }\end{array}\right.$

First element q (t) in the sequence=q (t-1)+Δ q (t) if q (t) transfinites, also will carry out amplitude limit.

$\left[\begin{array}{cc}q\left(t\right)<0& q\left(t\right)=0\\ q\left(t\right)>q_{\max }& q\left(t\right)=q_{\max }\end{array}\right]$

Final gained q (t) is exactly the control decision of current time.

(3) repeating step (a)---step (f) to control task finishes, and makes up final rolling temperature prediction and control system;

The operation principle of the inventive method is dynamically to provide online real time temperature prediction in real time according to finish rolling, and survey is carried out real-time rolling optimization control decision to finishing temperature according to temperature, thereby realizes the control of finish rolling finishing temperature.In the real work, as shown in Figure 2, the functional module that realizes the inventive method comprises: the final rolling temperature system, set up the statistical model module, the estimated performance analysis module, the nonlinear model linearization block, rolling Optimization of Time Domain control module, switch module, wherein set up statistical model module performing step (b), (c) and the data acquisition (d) arrangement, denoising and modeling function, and export the model of obtaining to the nonlinear model linearization block respectively, the estimated performance analysis module is arranged on the switch control module of setting up between statistical model module and the nonlinear model linearization block and is used to control the information of statistical model module to the output of nonlinear model linearization block of setting up.The function of estimated performance analysis module and nonlinear model linearization block performing step (e), definition finite time-domain rolling optimization problem, and the result imported rolling Optimization of Time Domain control module, rolling Optimization of Time Domain control module execution in step (f) and function (g), and then control final rolling temperature system regulates the finish rolling finishing temperature.

Claims (7)

1. the finish rolling finishing temperature based on statistical learning is predicted and control method, it is characterized in that being made up of following steps:

(1) input/output variable and the order thereof of setting sampling period, forecast model, wherein the input variable of forecast model is worn tape speed for specifying the frame water yield and finish to gauge frame, and the output variable of forecast model is a finishing temperature;

(2) in each fixing sampling period, carry out following steps:

The input/output variable of a, collection forecast model, and, determine model structure according to the forecast model variable order that step (1) is set, construct corresponding inputoutput data regression vector matrix and modeling training dataset;

B, the inputoutput data regression vector matrix that step (a) is obtained carry out denoising;

C, based on the modeling training dataset that step (a) obtains, set up the non-linear input and output forecast model of finishing temperature with statistical method;

D, check forecast model are to the consistent degree of the predicted value and the actual value variation tendency of finishing temperature, if basically identical, then with nonlinear model in the present operating point linearisation, obtain linear model; Otherwise, do not carry out subsequent control and calculate, directly return step (a);

E, the linear model that obtains based on step (d), selecting to specify the frame water yield is control variables, definition finite time-domain rolling optimization problem;

The optimization problem that obtains in f, the solution procedure (e) obtains corresponding control decision vector, first element of this vector after amplitude limiting processing as the control decision in current sampling period;

(3) repeating step (a)---step (f) to control task finishes, and makes up final rolling temperature prediction and control system.

2. a kind of prediction of finish rolling finishing temperature and control method according to claim 1 based on statistical learning, it is characterized in that in the described step (a), in each sampling period, the capital is gathered the new appointment frame water yield, finish to gauge frame and is worn tape speed and finishing temperature, and new data are joined the modeling training dataset.

3. a kind of prediction of finish rolling finishing temperature and control method based on statistical learning according to claim 1 is characterized in that described step (b) comprises the steps:

According to finish rolling equipment and process conditions, determine variable finishing temperature, the bound of specifying the frame water yield and finish to gauge frame to wear tape speed and rate of change thereof;

Bound with variate-value and variable rate is carried out denoising to all real time datas.

4. a kind of finish rolling finishing temperature based on statistical learning described in claim 1 is predicted and control method, it is characterized in that statistical method in the described step (c) is for setting up statistical models, kernel function in the statistical models is the RBF kernel function, and concrete nonlinear model form is:

$y\left(x\right)=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_k\exp \{-{\left|\right|x-x_k\left|\right|}_2^2/{\mathrm{\&sigma;}}^2\}+b$

Wherein, x _kBe support vector, α _kBe corresponding support vector x _kWeight coefficient, x is current regression data vector, y (x) is and the corresponding output of x, i.e. finishing temperature predicted value.

5. a kind of prediction of finish rolling finishing temperature and control method based on statistical learning according to claim 1 is characterized in that described step (d) comprises the steps:

Contrast finishing temperature actual value sequence T _a({ T _a(t), T _a(t-1), T _a(t-2), L}) and model prediction value sequence T _p({ T _p(t-1), T _p(t-2), L}), if between finishing temperature actual value sequence and the model prediction value sequence continuous some sampling periods of the AC compounent of deviation be no more than the AC compounent upper limit of appointment, then assert the predicted value and the actual value basically identical of current time finishing temperature.

If forecast model then is located at the present operating point linearisation with non-linear mould to the predicted value and the actual value basically identical of finishing temperature, gained linear model form is as follows:

$T\left(t\right)=b_{1}q(t-1)+...+b_{n_q}q(t-n_q)-a_{1}T(t-1)-...-a_{n_T}(t-n_T)+$

$p_0+c_1*v(t-1)+...+c_{n_v}*v(t-n_v)$

If forecast model is inconsistent to the predicted value and the actual value of finishing temperature, then do not carry out follow-up control and calculate, directly return step (a).

6. a kind of prediction of finish rolling finishing temperature and control method based on statistical learning according to claim 1 is characterized in that described step (e) comprises the steps:

With reference to physical device and process conditions, definition optimization aim function comprises temperature control deviation penalty term and controlled quentity controlled variable increment penalty term in the function; Selected optimization time domain and control time domain, thus be defined as follows rolling Optimization of Time Domain problem, and concrete formula is:

$\min J\left(t\right)=\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}{\mathrm{\&mu;}}_i{[T_r(t+i)-\hat{T}(t+i)]}^2+\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j{\left[\mathrm{\&Delta;q}(t+j-1)\right]}^2$

Wherein, P is the prediction time domain, and M is the control time domain, T _rBe with reference to given, u _i(1≤i≤P) and λ _j(1≤j≤M) is the weight coefficient of corresponding penalty term.

7. a kind of prediction of finish rolling finishing temperature and control method based on statistical learning according to claim 1 is characterized in that described step (f) comprises the steps:

By matrix computations, obtain following optimum control increment sequence [Δ q (t), Δ q (t+1), L, Δ q (t+M-1)];

With reference to physical device and process conditions, determine the bound constraint and the rate of change constraint of the water yield.If Δ q (t) transfinites, it is carried out amplitude limit;

Obtaining the control corresponding sequence of decisions by the control increment sequence is

[q(t-1)+Δq(t)，q(t-1)+Δq(t)+Δq(t+1)，L，q(t-1)+Δq(t)+L+Δq(t+M-1)]

Wherein first element q (t) of sequence=q (t-1)+Δ q (t) if q (t) transfinites, also will carry out amplitude limit, and final gained q (t) is the control decision of current time.

Images