CN107045287A

CN107045287A - Coordinated Control Systems and control method based on Prediction and Control Technology

Info

Publication number: CN107045287A
Application number: CN201710328757.7A
Authority: CN
Inventors: 吕剑虹; 陈昊洋; 于吉
Original assignee: Southeast University
Current assignee: Southeast University
Priority date: 2017-05-11
Filing date: 2017-05-11
Publication date: 2017-08-15

Abstract

The invention discloses a kind of Coordinated Control Systems based on Prediction and Control Technology and control method, including backfeed loop and calorific value correction loop, the master controller of the backfeed loop uses generalized predictive controller, and the calorific value corrective loop uses RBF neural controller；The generalized predictive controller includes being used to calculate each polynomial Diophantine equation computing module on backward shift operator z, and calculating matrix G solves module and solves module for the controlled quentity controlled variable for solving the controlled quentity controlled variable in predictive control algorithm.The present invention effectively increases the regulation performance of unit load, reduces the fluctuation of key parameter, substantially increases the operation stability of unit.

Description

Coordinated Control Systems and control method based on Prediction and Control Technology

Technical field

The present invention relates to Thermal power engneering and automatically controlling, more particularly to a kind of unit based on Prediction and Control Technology Coordinated control system and control method.

Background technology

Traditional coordination system typically uses the regulation scheme of " load instruction feedforward "+" PID/feedback ", due to steam coal amount pair The inertia time of main steam pressure influence is long, and conventional PID control is often difficult to control effectively to such Large Time Delay Process； And for control of steam temperature, due to during unit load increase and decrease, unreasonable, the control of steam temperature inferior quality of coal-water ratio control. Variable load rate is only capable of being set to 3~6MW/min, and AGC examination speed and precision are significantly affected；Significantly varying duty When, main vapour pressure deviation is maximum up to 1.4MPa, and has obvious regulation vibration, and can have more than 1.0MPa for a long time Control deviation；Control performance is very poor, operations staff frequently solve manually intervene in the case of, Stream temperature degree often up to 608 DEG C with On.And due to unit Fuel- Water Rate control and main-stream control harmony it is poor, Stream temperature degree often fall to for a long time 590 DEG C with Under, it is minimum up to 575 DEG C.

Because the inertia time of boiler is long, and the inertia time of steam turbine is very short, and in unit running process, boiler does not catch up with Energy imbalance caused by steam turbine is to cause the basic reason that unit load regulation performance is poor, key parameter fluctuation is big, and Because coal often changes in real process, easily occurs the fluctuation of main vapour pressure and Stream temperature degree.

Therefore, there is the fluctuation of the key parameters such as Load Regulation ability undesirable, vapour pressure steam temperature greatly in unit and system can not be very Adapt to the practical problem such as coal type change well, traditional coordinated control system can not be well controlled effect.

The content of the invention

Goal of the invention：To solve the deficiencies in the prior art, the present invention provides one kind by using PREDICTIVE CONTROL and nerve net Network is controlled, and can effectively shift to an earlier date the adjustment of boiler heat load, it is ensured that boiler can be with the well-coordinated control system of steam turbine and control Method processed.

Technical scheme：A kind of Coordinated Control Systems based on Prediction and Control Technology, including backfeed loop and calorific value are rectified Positive loop, the master controller of the backfeed loop uses generalized predictive controller, to control and adjust controlled device in advance；It is described Calorific value corrective loop uses RBF neural controller, and with the calorific value correction coefficient at look-ahead current time, quilt is controlled in advance Control object.

Further, the generalized predictive controller includes：Lost for calculating each on the polynomial of backward shift operator z Kind figure equation computing module, calculating matrix G solves module and the controlled quentity controlled variable for solving the controlled quentity controlled variable in predictive control algorithm Solve module.

A kind of control method using the control system, comprises the following steps：

(1) using the calculating of the generalized predictive controller amount of being controlled

(11) each multinomial on backward shift operator z is calculated,

(12) evaluator solution matrix G,

(13) controlled quentity controlled variable in predictive control algorithm is solved；

(2) coal calorific value is corrected using RBF neural network

Coal of the input to subsequent time is used as using the coal calorific value correction coefficient of current time and the first two sampling instant Calorific value correction coefficient is planted to be predicted；Including：

(21) cluster centre is calculated,

(22) weights are calculated.

Further, the output of process discrete differential equation of predictive controller is in the step (11)：

A(z^-1) y (t)=B (z^-1)u(t-1)+C(z^-1) w (t)/Δ,

Wherein, A (z^-1)、B(z^-1)、C(z^-1) all it is multinomial on backward shift operator z：

U (t), y (t) represent the input and output of controlled device；Δ=1-z^-1Represent difference operator；W (t) represents to disturb at random It is dynamic；Coefficient in each multinomial is constant, is drawn by experiment；n_a、n_b、n_cMost high-order term number of times in each multinomial is represented respectively.

Further, polynomial solving matrix G in the step (12) is calculated as follows：

For the optimal value predicted, using following Diophantine equation to G_j(z^-1) solved：

1=E_j(z^-1)A(z^-1)Δ+z^-jF_j(z^-1)

E_j(z^-1)B(z^-1)=G_j(z^-1)+z^-jH_j(z^-1),

Wherein, j=1,2,3 ..., N₁, N₁For maximum predicted time domain, and

E_j(z^-1)=e₀+e₁z^-1+…+e_j-1z^-j+1

G_j(z^-1)=g₀+g₁z^-1+…+g_j-1z^-j+1

Wherein, e₀,e₁..., e_j-1Represent the multinomial E solved by first Diophantine equation_j(z^-1) in each coefficient,Represent the multinomial F solved by first Diophantine equation_j(z^-1) in each coefficient, g₀,g₁,…,g_j-1Represent Multinomial G_j(z^-1) in every coefficient,Representative polynomial H_j(z^-1) in every coefficient.

When j takes 1,2,3 respectively ..., N₁When, by E_j(z^-1)B(z^-1) substitute into second Diophantine Equation Solution draw it is multinomial Formula G_j(z^-1) inCoefficientObtained coefficient is constituted into matrix, rectangular is drawn Formula G：

Wherein N₁For maximum predicted time domain, N_uFor control time domain.

Further, in the step (13), A (z are solved by Diophantine equation^-1) and bring difference equation into, and by E_j(z^-1)B(z^-1) it is substituted for G_j(z^-1)+z^-jH_j(z^-1), and omit (z^-1), and by formula：Δ u (t)=(1-z^-1) t in u (t) changes t into + j can be obtained：

Y (t+j)=G_jΔu(t+j-1)+F_jy(t)+H_jΔu(t-1)+E_jW (t+j),

When j takes 1,2,3 respectively ..., N₁When, equation can be write as to the rectangular of matrix form, i.e. generalized predictive control Formula is：

Y=Gu+Fy (t)+H Δs u (t-1)+E,

Wherein：

Δ u (t-1)=(1-z-¹)u(t-1)

y^T=[y (t+1) ..., y (t+N₁)]

u^T=[Δ u (t) ..., Δ u (t+N_u-1)]

Wherein, y is defined_r ^TThe setting value exported for controlled device：

y_r ^T=[y_r(t+1),…,y_r(t+N₁)],

Definition J is performance index function：

J=(y-y_r)^T(y-y_r)+λu^TU,

Wherein, λ is control weighting constant；

It is 0 that it is made to the performance index function derivation, and when can obtain J and taking minimum value, u is：

U=(G^TG+λI)^-1G^T[y_r-Fy(t)-HΔu(t-1)]。

Further, the step (21) includes：

(a) algorithm initialization：H different initial cluster centers are selected, and make k=1, are randomly selected from experimental data H different samples, which are inputted as primary data center, is designated as c_i(k), i=1,2 ..., h；

(b) distance for the cluster centre that all sample inputs are selected with each is calculated | | X_j-c_i(k) | |, wherein, i=1, 2 ..., h, j=1,2 ..., N；

(c) to input X_j, classified by minimal distance principle：Work as | | X_j-c_i(k) | | when obtaining minimum value, i=1, 2 ..., h, X_jIt is classified as the i-th class；

(d) all kinds of new cluster centres are recalculated：

Wherein, N_iFor the number of the sample included in ith cluster domain；X is all inputs for belonging to the i-th class；

If (e) c_i(k+1)≠c_i(k) step (b), is gone to, otherwise cluster terminates；

(f) extension constant, extension constant σ=κ d of hidden node are determined according to the distance between each cluster centre_i,

Wherein, d_iFor the distance between i-th data center and other his nearest data centers, i.e., κ is overlap coefficient.

Further, in the step (22), according to the data center of the hidden node of determination and extension constant, output is drawn Weight vector w=[w₁,w₂,…,w_h] hidden layer output battle array H be：

Wherein,E is natural constant, i=1,2 ..., N j=1,2 ..., h；

Then neural network model is finally output as：

R=Hw,

Weighted vector is calculated using the method for least square：

W=(H^TH)^-1H^Tr。

Beneficial effect：Compared with prior art, a kind of unit cooperative control based on Prediction and Control Technology proposed by the present invention System and control method processed, the inside Stream temperature degree forecast model of predictive controller use by past Stream temperature degree influenceed it is linear Function model, can effectively advancement by PREDICTIVE CONTROL so that boiler can get caught up in the energy requirement of steam turbine, The regulation performance of unit load is improved, the fluctuation of key parameter is effectively reduced, the operation stability of unit is substantially increased And security；Calorific value correction coefficient model based on RBF neural, which is used, has received the non-linear of calorific value correction coefficient influence Function model, can adapt to the change of coal, and the trimming process for making up traditional BTU correction calorific values lags far behind the change of coal The problem of, by storing the coal calorific value correction coefficient of each sampling instant of early stage, and it is used as the training sample of RBF neural Data, set up the nonlinear neural network model of coal calorific value correction coefficient, then with the model to coal calorific value correction coefficient Future value carry out recursion and estimate, make up the delayed of BTU trimming processes by the pre-set time estimated, improve the stabilization of unit Property.

Brief description of the drawings

Fig. 1 is the systematic schematic diagram of the embodiment of the present invention；

Fig. 2 is the coal calorific value calibration model figure based on RBF neural.

Embodiment

With reference to specific embodiment, the technical scheme that the present invention is furture elucidated.

As shown in figure 1, a kind of Coordinated Control Systems based on Prediction and Control Technology disclosed in the embodiment of the present invention, bag Backfeed loop and calorific value correction loop are included, the master controller of the backfeed loop uses generalized predictive controller, the calorific value school Positive loop uses RBF neural controller.

Wherein, generalized predictive controller includes：Diophantine equation computing module, for calculating each on backward shift operator z Multinomial；Calculating matrix G solves module, the polynomial solving matrix G drawn by Diophantine equation computing module；Controlled quentity controlled variable Module is solved, for solving the controlled quentity controlled variable in predictive control algorithm.The inside Stream temperature degree forecast model of generalized predictive controller Using the linear function model on last time Stream temperature degree.

Neural network controller includes coal calorific value correction coefficient model and the correction of coal calorific value based on RBF neural The recursion prediction module of coefficient.

Master controller controls the deviation between main vapour pressure and setting value using generalized predictive controller, and the feedforward of coal amount passes through Main vapour pressure setting value and load instruction instruct and fed water to coal amount to instruct and adjust in advance, and calorific value corrective loop is by storing calorific value The calorific value correction coefficient of correction coefficient prediction future time instance is to coal type change frequently situation advancement.

(11) each multinomial on backward shift operator z is calculated

The output of process (i.e. main vapour pressure) of predictive controller can be expressed as with difference equation：

A(z^-1) y (t)=B (z^-1)u(t-1)+C(z^-1) w (t)/Δ,

Wherein, A (z^-1)、B(z^-1)、C(z^-1) all it is multinomial on backward shift operator z, formula is as follows：

U (t), y (t) represent the input and output of controlled device；Δ=1-z^-1Represent difference operator；W (t) represents to disturb at random It is dynamic；Coefficient is constant in each multinomial, is all drawn by experiment；n_a、n_b、n_cMost high-order term number of times in each multinomial is represented respectively.

(12) evaluator solution matrix G

1=E_j(z^-1)A(z^-1)Δ+z^-jF_j(z^-1)

E_j(z^-1)B(z^-1)=G_j(z^-1)+z^-jH_j(z^-1),

Wherein, j=1,2,3 ..., N₁, N₁For maximum predicted time domain, and

E_j(z^-1)=e₀+e₁z^-1+…+e_j-1z^-j+1

G_j(z^-1)=g₀+g₁z^-1+…+g_j-1z^-j+1

Wherein, N₁For maximum predicted time domain, N_uFor control time domain.

(13) controlled quentity controlled variable in predictive control algorithm is solved

A (z are solved by Diophantine equation^-1) and bring difference equation into, and by E_j(z^-1)B(z^-1) it is substituted for G_j(z^-1)+z-jH_j (z^-1), and omit (z^-1), separately by formula Δ u (t)=(1-z^-1) t in u (t) changes t+j into and can obtain：

Y (t+j)=G_jΔu(t+j-1)+F_jy(t)+H_jΔu(t-1)+E_jW (t+j),

Y=Gu+Fy (t)+H Δs u (t-1)+E,

Wherein：

Δ u (t-1)=(1-z^-1)u(t-1)

y^T=[y (t+1) ..., y (t+N₁)]

u^T=[Δ u (t) ..., Δ u (t+N_u-1)]

Wherein, y is defined_r ^TThe setting value exported for controlled device：

y_r ^T=[y_r(t+1),…,y_r(t+N₁)],

Definition J is performance index function：

J=(y-y_r)^T(y-y_r)+λu^TU,

Wherein, λ is control weighting constant；

U=(G^TG+λI)-¹G^T[y_r-Fy(t)-HΔu(t-1)]。

(2) coal calorific value is corrected using RBF neural network

Coal calorific value calibration model based on RBF neural as shown in Figure 2, is sampled using current time and the first two The coal calorific value correction coefficient at moment is predicted as input to the coal calorific value correction coefficient of subsequent time；Input and become in figure Measure x₁,x₂,…,x_nAs coal calorific value correction coefficient, w₁,w₂,…,w_nFor function f (X, c_k) weights, f in the present embodiment (X,c_k) be expressed from the next：

Wherein, | | | | distance is represented, σ is extension constant, and X is sample input vector, c_kFor center of a sample.

Sample input vector X transposition is X^T, and X^TIncluding current time calorific value correction coefficient r (k), k-1 moment calorific values school Positive coefficient r (k-1), the coal calorific value correction coefficient r (k-2) at k-2 moment, is shown below：

X^T=[r (k-2), r (k-1), r (k)],

Using the coal calorific value correction coefficient at k+1 moment as output, to the coal calorific value school at subsequent time (i.e. k+1 moment) Positive coefficient is predicted, i.e.,：

R=r (k+1),

The control method in the calorific value correction coefficient loop based on RBF neural includes：

(21) cluster centre is calculated

(a) algorithm initialization：H different initial cluster centers are selected, and make k=1.Can be random from experimental data H different samples inputs are chosen as primary data center and are designated as c_i(k), i=1,2 ..., h；

(b) distance for the cluster centre that all sample inputs are selected with each is calculated | | X_j-c_i(k) | |, wherein i=1, 2 ..., h, j=1,2 ..., N；

(d) all kinds of new cluster centres are recalculated：

Wherein, N_iFor the number of the sample included in ith cluster domain, X is all inputs for belonging to the i-th class；

If (e) c_i(k+1)≠c_i(k) step (b), is gone to, otherwise cluster terminates；

(f) extension constant, extension constant σ=κ d of hidden node are determined according to the distance between each cluster centre_i, wherein d_i For the distance between i-th data center and other his nearest data centers, i.e.,κ is overlapping Coefficient.

(22) weights are calculated.

Because the data center of hidden node and extension constant are determined, output weight vector w=[w₁,w₂,…,w_h] hidden layer Exporting battle array H is：

Wherein,E is natural constant, i=1,2 ..., N j=1,2 ..., h；

The then output of neural network model finally, i.e. calorific value correction coefficient is：

R=Hw,

Weighted vector is calculated using the method for least square：

W=(H^TH)^-1H^Tr。

Control principle：PREDICTIVE CONTROL can be adjusted according to the controlled deviation in the future time section predicted, if with PREDICTIVE CONTROL realizes the heat load adjustment of boiler, then " can effectively shift to an earlier date " adjustment of boiler heat load, it is ensured that boiler can be with Steam turbine is fully coordinated.

The calorific value corrective loop uses current time and the first two sampling instant using the neural network model based on RBF Coal calorific value correction coefficient the coal calorific value correction coefficient of subsequent time is predicted as input.Although coal calorific value Change has certain randomness and complexity, but the change of coal calorific value or a continuous process, always can be with one again Miscellaneous non-linear continuous function is described, cognitive highly complex nonlinear change process, often can be by nerve Network technology sets up the nonlinear model of its change procedure, by the analysis to model and estimates, and grasps its changing rule.This hair It is bright traditional coal BTU correction on the basis of, by storing the coal calorific value correction coefficient of each sampling instant of early stage, and conduct The training sample data of RBF neural, set up the nonlinear neural network model of coal calorific value correction coefficient, the model are used afterwards Future value progress recursion to coal calorific value correction coefficient is estimated, and BTU trimming processes are made up by the pre-set time estimated It is delayed.

Embodiment：

Below by taking coordinated control system of certain electricity generating corporation, Ltd 600MW supercritical units using the present invention as an example, in detail Illustrate the present invention.

(1) with control performance during 10MW/min speed varying duties

550MW is dropped to from 650MW with 10MW/min speed, rises back 650MW again after stable, one has been carried out Secondary forward and reverse amplitude 50MW load change test.

The overall performance evaluation of optimization system varying duty under 10MW/min speed：

1. spatial load forecasting：Actual load is in strict accordance with setting Changing load-acceleration change, and dynamic process is steady, dead-beat, mistake Tune amount very little；Actual speed rate, response time, dynamic control deviation, steady state controling precision are satisfied by requiring.

2. main vapour pressure is controlled：Same trend change is kept with sliding pressure setting value, dynamic process is steady, dead-beat and mistake Adjust, maximum dynamic error is only 0.3~0.4MPa.

3. Stream temperature degree is controlled：Control is very steady, and the average maximum dynamic error during varying duty is only 2~3 DEG C.

(2) the AGC runnabilities of Optimal Control System

1. spatial load forecasting：Speed is set to 9MW/min, and actual load changes in strict accordance with setting Changing load-acceleration, dynamic Process is steady, dead-beat, overtravel very little；Actual speed rate, response time, dynamic control deviation, steady state controling precision are equal Meet and require.

2. main vapour pressure is controlled：In steady-state process, almost indifference is controlled；In significantly AGC varying duties and open, stop the mill stage, Also only there is 0.3~0.5MPa deviation in a short time, the actual main vapour pressure of unit is set with sliding pressure in most times Value remains that same trend changes, and dynamic process is steady, dead-beat and toning.

3. Stream temperature degree is controlled：In steady-state process, Temperature Deviation in the range of ± 1.0 DEG C, controls almost indifference all the time, Steam temperature is very steady.In significantly varying duty and open, stop mill, steam temperature also only has 5~8 DEG C of decline in the short time, and by optimization The intelligent Fuel- Water Rate control strategy of system, it is not necessary to which operations staff intervenes manually quickly to be pulled back near setting value.Fixed When value is set to 600.8 DEG C, under various disturbance operating modes, steam temperature maximum can be controlled within 603 DEG C, illustrate optimization system pair In suppressing, overtemperature is highly effective, and unit further can improve steam temperature setting value in operation from now on, to obtain more preferable operation Economy.

As can be seen here, after Optimal Control System input, the AGC runnabilities of unit, the Control platform of main steam temperature are equal Obtain obviously improving.

Claims

1. a kind of Coordinated Control Systems based on Prediction and Control Technology, it is characterised in that：Rectified including backfeed loop and calorific value Positive loop, the master controller of the backfeed loop uses generalized predictive controller, to control and adjust controlled device in advance；It is described Calorific value corrective loop uses RBF neural controller, and with the calorific value correction coefficient at look-ahead current time, quilt is controlled in advance Control object.

2. a kind of Coordinated Control Systems based on Prediction and Control Technology according to claim 1, it is characterised in that institute Stating generalized predictive controller includes：For calculating each polynomial Diophantine equation computing module on backward shift operator z, meter Matrix G is calculated to solve module and solve module for the controlled quentity controlled variable for solving the controlled quentity controlled variable in predictive control algorithm.

3. the control method of any one of a kind of use claim 1 to 2 control system, it is characterised in that including following step Suddenly：

(11) each multinomial on backward shift operator z is calculated,

(12) evaluator solution matrix G,

(2) coal calorific value is corrected using RBF neural network

Coal heat of the input to subsequent time is used as using the coal calorific value correction coefficient of current time and the first two sampling instant Value correction coefficient is predicted；Including：

(21) cluster centre is calculated,

(22) weights are calculated.

4. a kind of control method according to claim 3, it is characterised in that：The mistake of predictive controller in the step (11) Journey exports discrete differential equation：

A(z^-1) y (t)=B (z^-1)u(t-1)+C(z^-1) w (t)/Δ,

Wherein, u (t), y (t) represent the input and output of controlled device；Δ=1-z^-1Represent difference operator；W (t) represents random Disturbance；Coefficient in each multinomial is constant, is drawn by experiment；n_a、n_b、n_cMost high-order term number of times in each multinomial is represented respectively.

5. a kind of control method according to claim 4, it is characterised in that calculate as follows in the step (12) Polynomial solving matrix G：

1=E_j(z^-1)A(z^-1)Δ+z^-jF_j(z^-1)

E_j(z^-1)B(z^-1)=G_j(z^-1)+z^-jH_j(z^-1),

Wherein, j=1,2,3 ..., N₁, N₁For maximum predicted time domain, and

E_j(z^-1)=e₀+e₁z^-1+…+e_j-1z^-j+1

G_j(z^-1)=g₀+g₁z^-1+…+g_j-1z^-j+1

When j takes 1,2,3 respectively ..., N₁When, by E_j(z^-1)B(z^-1) substitute into second Diophantine Equation Solution draw multinomial G_j (z^-1) inCoefficientObtained coefficient is constituted into matrix, matrix form G is drawn：

Wherein N₁For maximum predicted time domain, N_uFor control time domain.

6. a kind of control method according to claim 5, it is characterised in that in the step (13), by Diophantine equation Solve A (z^-1) and bring difference equation into, and by E_j(z^-1)B(z^-1) it is substituted for G_j(z^-1)+z^-jH_j(z^-1), and omit (z^-1), and will Formula：Δ u (t)=(1-z^-1) t in u (t) changes t+j into and can obtain：

Y (t+j)=G_jΔu(t+j-1)+F_jy(t)+H_jΔu(t-1)+E_jW (t+j),

When j takes 1,2,3 respectively ..., N₁When, the matrix form of equation can be write as matrix form, i.e. generalized predictive control be：

Y=Gu+Fy (t)+H Δs u (t-1)+E,

Wherein：

Δ u (t-1)=(1-z^-1)u(t-1)

y^T=[y (t+1) ..., y (t+N₁)]

u^T=[Δ u (t) ..., Δ u (t+N_u-1)]

Wherein, y is defined_r ^TThe setting value exported for controlled device：

y_r ^T=[y_r(t+1),…,y_r(t+N₁)],

Definition J is performance index function：

J=(y-y_r)^T(y-y_r)+λu^TU,

Wherein, λ is control weighting constant；

U=(G^TG+λI)-¹G^T[y_r-Fy(t)-HΔu(t-1)]。

7. a kind of control method according to claim 3, it is characterised in that the step (21) includes：

(a) algorithm initialization：H different initial cluster centers are selected, and make k=1, h are randomly selected from experimental data Different samples inputs as primary data center and is designated as c_i(k), i=1,2 ..., h；

(b) distance for the cluster centre that all sample inputs are selected with each is calculated | | X_j-c_i(k) | |, wherein i=1,2 ..., h, J=1,2 ..., N；

(c) to input X_j, classified by minimal distance principle：Work as | | X_j-c_i(k) | | when obtaining minimum value, i=1,2 ..., H, X_jIt is classified as the i-th class；

(d) all kinds of new cluster centres are recalculated：

<mrow> <msub> <mi>c</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>+</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <msub> <mi>N</mi> <mi>i</mi> </msub> </mfrac> <mi>&Sigma;</mi> <mi>X</mi> <mo>,</mo> </mrow>

If (e) c_i(k+1)≠c_i(k) step (b), is gone to, otherwise cluster terminates；

Wherein, d_iFor the distance between i-th data center and other his nearest data centers, i.e.,κ For overlap coefficient.

8. a kind of control method according to claim 7, it is characterised in that in the step (22), according to the hidden of determination The data center of node and extension constant, draw output weight vector w=[w₁,w₂,…,w_h] hidden layer output battle array H be：

Wherein,E is natural constant, i=1,2 ..., N j=1,2 ..., h；Then neutral net Model is finally output as：

R=Hw,

Weighted vector is calculated using the method for least square：

W=(H^TH)^-1H^Tr。