CN106209474A - A kind of network control system tracking and controlling method based on predictive compensation - Google Patents

Abstract

The invention discloses a kind of network control system tracking and controlling method based on predictive compensation, neural network forecast controller is by predicting that generator and time delay compensator two parts form, and the controlling increment sequence of k moment neural network forecast generator output is:The controlling increment of time delay compensator t is:The present invention shortens prediction step to introduce the least forecast error for cost, improves the predictive controller compensation effect to network delay, compares traditional neural network forecast and controls to be more easy to realize.

Description

A kind of network control system tracking and controlling method based on predictive compensation

Technical field

The present invention relates to network control system, particularly relate to network control system based on predictive compensation and follow the tracks of control Method processed.

Background technology

The closed-loop control system formed by communication network is referred to as network control system (networked control Systems, be abbreviated NCSs), NCSs have convenient for installation and maintenance, motility is high and is prone to the advantages such as reconstruct.Control based on network system In system, the parts such as controlled device, sensor, controller and executor connect and compose closed loop system by real-time network.Drawing of network Enter, on the one hand overcome the remote distributed in space, provide basis for realizing remotely monitor and control.On the other hand, network topology Inevitable time delay and packet loss during structure, Internet resources are limited, bag transmission communicates, all propose to choose to traditional control method War.At present, Control Methods of Networked Control Systems is broadly divided into two classes: network is directly used the parameter confirmed in advance by a class method Describing and introduce system model, be analyzed by control theory with time delay in traditional theory, this kind of theory has time-delay system to manage Opinion, stochastic control theory, Optimal Control Theory, switched system are theoretical.This kind of method simplifies network shadow in the controls Ring, can use for reference existing theoretical result, but non-Active Compensation time delay, it is impossible to systematic function is substantially improved.Another kind of by network and System Independent modeling collaborative design, with the systematic function of satisfied requirement, is the focus of research at present.

Neural network forecast controls to be a kind of collaborative design method, compensates network delay by the way of prediction.For time discrete Linearly time-invariant system state can not survey situation, existing method design controller end observer, provides according to feedback channel Delay information estimate the State Viewpoint measured value before some steps, and by the system state space model some steps of recursion current and future Status predication value, is realized closed loop control by fixed gain feedback of status, can promote systematic function, but need designated lane, difficult real Existing, cost height；Another kind of improvement neural network forecast control method, realizes prediction closed-loop control system, nothing in conjunction with optimal control thought Need designated lane, but underuse controller end Given information, add prediction step.

Summary of the invention

For above-mentioned problems of the prior art, the invention provides a kind of control based on network based on predictive compensation System tracking and controlling method.For the tracking control problem being widely present in engineering reality, equal at forward path and feedback channel Design closed loop controller in the case of there is random delay, fully use Given information effective compensation time delay.

The technical solution adopted in the present invention is: a kind of network control system tracing control side based on predictive compensation Method, neural network forecast controller, by predicting that generator and time delay compensator two parts form, comprises the following steps:

1) network control system model is set up

Constant multiple-input and multiple-output controlled system during one class discrete-time linear, its augmented system state space description is such as Under:

$\left\{\begin{array}{c}\stackrel{\‾}{x}(t+1)=\stackrel{\‾}{A}\stackrel{\‾}{x}\left(t\right)+\stackrel{\‾}{B}\mathrm{\Δ}u\left(t\right)\\ y\left(t\right)=\stackrel{\‾}{C}\stackrel{\‾}{x}\left(t\right)\end{array}\right.$

Wherein,Δ u (t)=u (t)-u (t-1)；x(t)∈Rⁿ、u(t)∈R^mWith y (t) ∈ R^lRespectively State vector, control input and control output, w (t) ∈ R for controlled device^qFor deterministic perturbation, A ∈ R^n×n,B∈R^n×m,C ∈R^l×nFor coefficient matrix, it is fully controllable that system meets (A, B), and (A, C) is the most considerable.System output y (t) follows the tracks of external reference Input y₀T () realizes asymptotic tracking, i.e. tracking error e (t) meets:

The coefficient matrix of augmented system is respectively as follows:

$\stackrel{\‾}{A}=\left(\begin{array}{cc}A& B\\ 0& I\end{array}\right),\stackrel{\‾}{B}=\left(\begin{array}{c}B\\ I\end{array}\right),\stackrel{\‾}{C}=\left(\begin{array}{cc}C& 0\end{array}\right)$

System mode is not directly surveyed, and sets up state observer:

$\left\{\begin{array}{c}\stackrel{\‾}{x}(t+1|t)=\stackrel{\‾}{A}\stackrel{\‾}{x}(t|t-1)+\stackrel{\‾}{B}\mathrm{\Δ}u\left(t\right)+\stackrel{\‾}{L}\[y\left(t\right)-y\left(t\right|t-1)\]\\ y\left(t\right|t-1)=\stackrel{\‾}{C}x(t|t-1)\end{array}\right.$

Wherein, x (t+1 | t) it is the current time State Viewpoint measured value to subsequent time, x (t | t-1) and y (t | t-1) respectively For previous moment to the State Viewpoint measured value of current time and output observation, L ∈ R^n*lFor observer gain.Augmentation observer increases Benefit

2) prediction generator

The controlling increment sequence of k moment neural network forecast generator output is:

$\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}\left(k\right|k-{\mathrm{\τ}}_{sc,k})=-K_{{\mathrm{\τ}}_{sc,k}}\·\[{\mathrm{\Γ}}_{{\mathrm{\τ}}_{sc,k}}\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}+1|k-{\mathrm{\τ}}_{sc,k})-Y_0\left(k\right)\]$

Wherein, It is 1 × (N_u+τ_sc,k-1) block matrix, from τ_sc,k+1 BOB(beginning of block) is unit matrix I ∈ R^m×m；It is τ that k time data wraps in the real-time time delay in feedback channel_sc,k, N_u-1 for optimizing interval Length；For N_pThe block matrix of × 1,For lower triangular matrix, meet j-i≤τ_sc,k-1 AndY₀K () is that system follows desired value sequence: Y₀(k)=[y₀(k+1)…y₀(k+N_p)]^T；R, Q are power Weight coefficient matrix；System output and control input are respectively y (k-τ_sc,k) and u (k-τ_sc,k),For the k moment The State Viewpoint measured value of state observer latest update:

3) time delay compensator

The controlling increment of t is:

Wherein,It is 1 × (N_u-1) block matrix, τ^* _ca,tBlock is unit matrix I ∈ R^m×m,For t Up-to-date controlling increment sequence, τ in moment compensator^* _sc,tAnd τ^* _ca,tIt is respectively the up-to-date optimal control increasing that time delay compensator is chosen The time delay that amount sequence pair is answered.

Networking tracing control closed loop system is: x (t+1)=Ax (t)+B [u (t-1)+Δ u (t)].

Compared with prior art, beneficial effects of the present invention: under the background of tracing control, to the network in control system Inducement is analyzed, and provides network delay and packet loss solution in actual applications.Meanwhile, pre-at legacy network On the basis of surveying control method, shorten prediction step with the least forecast error of introducing for cost, improve predictive controller pair The compensation effect of network delay, compares traditional neural network forecast and controls to be more easy to realize.

Accompanying drawing explanation

Accompanying drawing 1 is the integrated stand composition of neural network forecast control system.

Accompanying drawing 2 is the fundamental diagram of networking Predictive Control System.

Accompanying drawing 3 is the feedback channel network random delay τ of front 50 sampling instants_scCurve.

Accompanying drawing 4 is the forward path network random delay τ of front 50 sampling instants_caCurve.

Accompanying drawing 5 is that the angular velocity of the servo-control system that predictive controller is constituted follows curve.

Detailed description of the invention

Below in conjunction with the accompanying drawings the detailed description of the invention of the present invention is described further.

Real network, owing to topological structure, Internet resources are limited and the feature such as bag transmission, produces three kinds of shapes to control system The impact of formula: time delay, packet loss and data incorrect order.Wherein the impact of time delay is relatively big, and other two kinds can be by data transmission mechanism etc. Effect is converted to time delay.Considering network inducement delay, neural network forecast control system integral frame is as shown in Figure 1.

Neural network forecast controller, by predicting that generator and time delay compensator two parts form (shown in accompanying drawing 1 dotted line), passes The neural network forecast control method of system needs designated lane transmission of control signals, is finally constituted closed loop by fixed gain feedback of status mode System.The inventive method, without designated lane, shortens prediction step, finally realizes with the feedback of status form of fixed structure variable element Closed loop control.

A kind of network control system tracking and controlling method based on predictive compensation, comprises the following steps:

Step 1: set up network control system model

Constant multiple-input and multiple-output controlled system during one class discrete-time linear, its state space description is as follows:

$\left\{\begin{array}{c}x(t+1)=Ax\left(t\right)+Bu\left(t\right)+B_{w}w\left(t\right)\\ y\left(t\right)=Cx\left(t\right)+D_{w}w\left(t\right)\end{array}\right.---\left(1\right)$

Wherein x (t) ∈ Rⁿ、u(t)∈R^mWith y (t) ∈ R^lIt is respectively the state vector of controlled device, controls input and control Output, w (t) ∈ R^qFor deterministic perturbation, A ∈ R^n×n,B∈R^n×m,B_w∈R^n×q,C∈R^l×n,D_w∈R^l×qFor sytem matrix, it is It is fully controllable that system meets (A, B), and (C, D) is the most considerable.

System output y (t) is made to follow the tracks of external reference input y₀T () realizes asymptotic tracking, and w (t) ≡ 0, i.e. tracking error e T () meets:

$\underset{t\→\∝}{\lim }e\left(t\right)=\underset{t\→\∝}{\lim }\[y_0\left(t\right)-y\left(t\right)\]=0---\left(2\right)$

System mode is not directly surveyed, and sets up state observer observer state, and state observer is described as:

$\left\{\begin{array}{c}x(t+1|t)=Ax\left(t\right|t-1)+Bu\left(t\right)+L\[y\left(t\right)-y\left(t\right|t-1)\]\\ y\left(t\right|t-1)=Cx\left(t\right|t-1)\end{array}\right.---\left(3\right)$

Wherein: x (t+1 | t) it is the current time State Viewpoint measured value to subsequent time, x (t | t-1) and y (t | t-1) respectively For previous moment to the State Viewpoint measured value of current time and output observation, L ∈ R^n*lFor observer gain.

Step 2: prediction generator design

Consider that each control module of closed loop system meets time synchronized, and except necessary status information, each packet with Time labelling record delivery time；All there is random delay τ in feedback channel and forward path_scAnd τ_ca, delay, τ_scAnd τ_caFull respectively FootThe upper boundN_pFor forecast interval length, the upper boundN_u-1 for optimizing siding-to-siding block length, N_u≥N_p；Control output and the acquisition controlling to input all are existed delayed when calculating controlled quentity controlled variable by tracking control unit, it was predicted that generator It is τ that k time data wraps in the real-time time delay in feedback channel_sc,k, corresponding control output and control input are denoted as_y(k-τ_sc,k) and u (k-τ_sc,k)。

The object function choosing tracking control system is

J=[Y₀(k)-Y(k|k-τ_sc,k)]^TQ[Y₀(k)-Y(k|k-τ_sc,k)]+ΔU^T(k|k-τ_sc,k)RΔU(k|k-τ_sc,k) (4)

Wherein, Y (k | k-τ_sc,k) be output sequence: Y (k | k-τ_sc,k)=[y (k+1 | k-τ_sc,k)…y(k+N_p|k-τ_sc,k) ]^T, Δ U (k | k-τ_sc,k) it is controlling increment sequence: Δ U (k | k-τ_sc,k)=[Δ u (k-τ_sc,k+1|k-τ_sc,k)……Δu(k+ N_u-1|k-τ_sc,k)]^T, Y₀K () is that system follows desired value sequence: Y₀(k)=[y₀(k+1)…y₀(k+N_p)^T], R, Q are weight system Matrix number.

Being minimised as target with object function (4), original system augmented conversion is as follows:

$\left\{\begin{array}{c}\stackrel{\‾}{x}(t+1)=\stackrel{\‾}{A}\stackrel{\‾}{x}\left(t\right)+\stackrel{\‾}{B}\mathrm{\Δ}u\left(t\right)\\ y\left(t\right)=\stackrel{\‾}{C}\stackrel{\‾}{x}\left(t\right)\end{array}\right.---\left(5\right)$

Wherein,Δ u (t)=u (t)-u (t-1). the sytem matrix of augmented system is respectively as follows:

$\stackrel{\‾}{A}=\left(\begin{array}{cc}A& B\\ 0& I\end{array}\right),\stackrel{\‾}{B}=\left(\begin{array}{c}B\\ I\end{array}\right),\stackrel{\‾}{C}=\left(\begin{array}{cc}C& 0\end{array}\right)$

It is described as follows after the change of observer augmentation:

$\left\{\begin{array}{c}\stackrel{\‾}{x}(t+1|t)=\stackrel{\‾}{A}\stackrel{\‾}{x}\left(t\right|t-1)+\stackrel{\‾}{B}\mathrm{\Δu}\left(t\right)+\stackrel{\‾}{L}[y\left(t\right)-y\left(t\right|t-1)]\\ y\left(t\right|t-1)=\stackrel{\‾}{C}x\left(t\right|t-1)\end{array}\right.---\left(6\right)$

Wherein, observer gainAnd

Based on above augmented system, and output sequence Y (k | k-τ_sc,k) it is:

$Y\left(k\right|k-{\mathrm{\τ}}_{sc,k})=\[y(k+1|k-{\mathrm{\τ}}_{sc,k}),...,y(k+N_p|k-{\mathrm{\τ}}_{sc,k})\]=\stackrel{\‾}{C}\·\[\stackrel{\‾}{x}(k+1|k-{\mathrm{\τ}}_{sc,k}),...,\stackrel{\‾}{x}(k+N_p|k-{\mathrm{\τ}}_{sc,k})\]---\left(7\right)$

The state estimation value expression in following some moment is:

$\stackrel{\‾}{x}(k+{\mathrm{\τ}}_{ca,k}|k-{\mathrm{\τ}}_{sc,k})={\stackrel{\‾}{A}}^{{\mathrm{\τ}}_{sc,k}+{\mathrm{\τ}}_{ca,k}-1}(\stackrel{\‾}{A}-\stackrel{\‾}{L}\stackrel{\‾}{C})\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}|k-{\mathrm{\τ}}_{sc,k}-1)+\underset{i=1}{\overset{{\mathrm{\τ}}_{sc,k}+{\mathrm{\τ}}_{ca,k}}{\Σ}}\[{\stackrel{\‾}{A}}^{{\mathrm{\τ}}_{sc,k}+{\mathrm{\τ}}_{ca,k}-i}\stackrel{\‾}{B}\mathrm{\Δ}u(k-{\mathrm{\τ}}_{sc,k}+i-1)\]+{\stackrel{\‾}{A}}^{{\mathrm{\τ}}_{sc,k}+{\mathrm{\τ}}_{ca,k}-1}\stackrel{\‾}{L}y(k-{\mathrm{\τ}}_{sc,k})---\left(8\right)$

Formula (8) is substituted into formula (7):

$Y\left(k\right|k-{\mathrm{\τ}}_{sc,k})={\mathrm{\Γ}}_{{\mathrm{\τ}}_{sc,k}}\·\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}+1|k-{\mathrm{\τ}}_{sc,k})+{\mathrm{\Θ}}_{{\mathrm{\τ}}_{{}_{sc,k}}}\·\mathrm{\Δ}U\left(k\right|k-{\mathrm{\τ}}_{sc,k})---\left(9\right)$

WhereinFor N_pThe block matrix of × 1,For lower triangular matrix, meet j-i ≤τ_sc,k-1 and State Viewpoint measured value for k moment state observer latest update.

Can be calculated by following expression:

$\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}+1|k-{\mathrm{\τ}}_{sc,k})=(\stackrel{\‾}{A}-\stackrel{\‾}{L}\stackrel{\‾}{C})\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}|k-{\mathrm{\τ}}_{sc,k}-1)+\stackrel{\‾}{B}\mathrm{\Δ}u(k-{\mathrm{\τ}}_{sc,k})+\stackrel{\‾}{L}y(k-{\mathrm{\τ}}_{sc,k})---\left(10\right)$

The object function that formula (9) substitution formula (4) describes is sought optimal value, order:

${\mathrm{\ΔU}}^{*}\left(k\right|k-{\mathrm{\τ}}_{sc,k})=-{({{\mathrm{\Θ}}^T}_{{\mathrm{\τ}}_{sc,k}}{\mathrm{Q\Θ}}_{{\mathrm{\τ}}_{sc,k}}+R)}^{-1}{{\mathrm{\Θ}}^T}_{{\mathrm{\τ}}_{sc,k}}{\mathrm{Q\Γ}}_{{\mathrm{\τ}}_{sc,k}}\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}+1|k-{\mathrm{\τ}}_{sc,k})+{({{\mathrm{\Θ}}^T}_{{\mathrm{\τ}}_{sc,k}}{\mathrm{Q\Θ}}_{{\mathrm{\τ}}_{sc,k}}+R)}^{-1}{{\mathrm{\Θ}}^T}_{{\mathrm{\τ}}_{sc,k}}{\mathrm{QY}}_0\left(k\right)---\left(11\right)$

By Δ U^*(k|k-τ_sc,kIn), the optimal control increment sequence of future time instance is sent to network compensator by forward path Constituting closed loop, the controlling increment sequence of neural network forecast generator output is:

$\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}\left(k\right|k-{\mathrm{\τ}}_{sc,k})=M_{{\mathrm{\τ}}_{sc,k}}{\mathrm{\ΔU}}^{*}\left(k\right|k-{\mathrm{\τ}}_{sc,k})---\left(12\right)$

Wherein M_τsc,kIt is 1 × (N_u+τ_sc,k-1) block matrix, from τ_sc,k+ 1 BOB(beginning of block) is unit matrix I ∈ R^m×m。

Formula (12) substitutes into formula (11) and obtains:

$\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}\left(k\right|k-{\mathrm{\τ}}_{sc,k})=-K_{{\mathrm{\τ}}_{sc,k}}\·\[{\mathrm{\Γ}}_{{\mathrm{\τ}}_{sc,k}}\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}+1|k-{\mathrm{\τ}}_{sc,k})-Y_0\left(k\right)\]---\left(13\right)$

Wherein,

Step 3: the design of time delay compensator

Based on above predictive control algorithm, the random delay in feedback channel is compensated by neural network forecast generator, and Send controlling increment sequence to network delay compensator, realize closed loop control with the State Feedback Approach of fixed structure variable element.Net Network time delay compensator chooses current up-to-date controlling increment sequence, the controlling increment of current time is sent to executor with compensate before Random delay in passage.

The real-time time delay receiving packet of time delay compensator t time delay compensator end buffer area is denoted as τ respectivelyⁱ _sc,tWith τⁱ _ca,t| i=1,2 ..., then corresponding loopback delay τⁱ _t=τⁱ _sc,t+τⁱ _ca,t.If the up-to-date optimal control that time delay compensator is chosen The time delay that increment sequence is corresponding is respectively τ^* _sc,tAnd τ^* _ca,t, then:

(τ^* _sc,t,τ^* _ca,t)=min (τⁱ _sc,t+τⁱ _ca,t)=min (τⁱ _t) i=1,2 ... (14)

T, in compensator, up-to-date controlling increment sequence isThen the controlling increment of current time is:

$\mathrm{\Δ}u\left(t\right)=S_{{{\mathrm{\τ}}^{*}}_{ca,t}}\·\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}(t-{{\mathrm{\τ}}^{*}}_{ca.t}|t-{{\mathrm{\τ}}^{*}}_t)=-S_{{{\mathrm{\τ}}^{*}}_{ca,t}}{K}_{{{\mathrm{\τ}}^{*}}_{sc,t}}\[{\mathrm{\Γ}}_{{{\mathrm{\τ}}^{*}}_{sc,t}}\stackrel{\‾}{x}(t-{{\mathrm{\τ}}^{*}}_t+1|k-{{\mathrm{\τ}}^{*}}_t)-Y_0(t-{{\mathrm{\τ}}^{*}}_{ca,t})\]---\left(15\right)$

Wherein,It is 1 × (N_u-1) block matrix, τ^* _ca,tBlock is unit matrix I ∈ R^m×m。

Formula (15) shows, the relevant parameter of controlling incrementWithIt is τ^* _sc,tAnd τ^* _ca,tFunction, can wait Valency is that the variable element State Feedback Approach of fixed structure constitutes closed loop system.

Networking tracing control closed loop system describes such as formula (16), corresponding closed-loop control system operation principle such as accompanying drawing 2 Shown in.

$\left\{\begin{array}{c}x(t+1)=Ax\left(t\right)+B\[u(t-1)+\mathrm{\Δ}u\left(t\right)\]\\ \mathrm{\Δ}u\left(t\right)=S_{{{\mathrm{\τ}}^{*}}_{ca,t}}\·\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}(t-{{\mathrm{\τ}}^{*}}_{ca,t}|t-{{\mathrm{\τ}}^{*}}_t)\\ \mathrm{\Δ}{\stackrel{\‾}{U}}^{*}(t-{{\mathrm{\τ}}^{*}}_{ca,t}|t-{{\mathrm{\τ}}^{*}}_t)=-K_{{{\mathrm{\τ}}^{*}}_{sc,t}}\[{\mathrm{\Γ}}_{{{\mathrm{\τ}}^{*}}_{sc,t}}\stackrel{\‾}{x}(t-{{\mathrm{\τ}}^{*}}_t+1|t-{{\mathrm{\τ}}^{*}}_t)-Y_0(t-{{\mathrm{\τ}}^{*}}_{ca,t})\]\end{array}\right.---\left(16\right)$

Embodiment:

Use a kind of based on predictive compensation the network control system tracking and controlling method that the present invention proposes, for explanation originally The effectiveness of literary composition method, it is considered to the design of the Networked controller of DC servo motor control system.Carry out with 0.04s for the cycle Sampling, the discrete time model of servo-control system can be described by following formula:

$G\left(z^{-1}\right)=\frac{A\left(z^{-1}\right)}{B\left(z^{-1}\right)}=\frac{3.5629z^{-2}+2.7739z^{-3}+1.0121z^{-4}}{1-1.2998z^{-1}+0.4343z^{-2}-0.1343z^{-3}}$

This servo control system follow the angular velocity that object is motor, the sytem matrix of corresponding states spatial model As follows:

$A=\left[\begin{array}{ccc}1.2998& -0.4343& 0.1343\\ 1& 0& 0\\ 0& 1& 0\end{array}\right],B=\left[\begin{array}{c}1\\ 0\\ 0\end{array}\right],C={\left[\begin{array}{c}3.5629\\ 2.7739\\ 1.0121\end{array}\right]}^T$

Concrete methods of realizing is as follows:

By state Observer Design principle, withCarry out POLE PLACEMENT USING for target, take L=[0.1334 -0.0400 0.0541]^T。

Matlab emulates, takes reference angle speedI=1,2 ..., N_p. system initial value is provided that

Controlled device: x (0)=[0.1 0.1 0.1]^T, u (0)=0；

State observer: x (0 | 0)=[0.1 0.1 0.1]^T。

When forward path and feedback channel random delay meet respectivelyTime delayTime, use side of the present invention Method design controller, takes N_p=N_u=8.In simulation process, the network random delay practical situation such as accompanying drawing of front 50 sampling instants 3 and accompanying drawing 4 shown in.

The servo control system output valve of the inventive method design follows situation as shown in Figure 5, as shown in Figure 5, The servosystem of present invention design can successfully track reference angle speed near the k=17 moment.The present invention introduce the least pre- Survey under error condition, shorten prediction step so that networking Predictive Control System can tolerate bigger time delay.

It is above presently preferred embodiments of the present invention, not the present invention is made any pro forma restriction, every foundation The technical spirit of the present invention, to any simple modification made for any of the above embodiments, equivalent variations and modification, belongs to inventive technique In the range of scheme.

Claims (1)

1. a network control system tracking and controlling method based on predictive compensation, it is characterised in that neural network forecast controller It is made up of prediction generator and time delay compensator two parts, specifically includes following steps:

Constant multiple-input and multiple-output controlled system during (1) one class discrete-time linear, its augmented system state space description is:

$\left\{\begin{array}{c}\stackrel{\‾}{x}(t+1)=\stackrel{\‾}{A}\stackrel{\‾}{x}\left(t\right)+\stackrel{\‾}{B}\mathrm{\Δ}u\left(t\right)\\ y\left(t\right)=\stackrel{\‾}{C}\stackrel{\‾}{x}\left(t\right)\end{array}\right.$

Wherein,Δ u (t)=u (t)-u (t-1)；x(t)∈Rⁿ、u(t)∈R^mWith y (t) ∈ R^lIt is respectively quilt The state vector of control object, control input and control output, A ∈ R^n×n,B∈R^n×m,C∈R^l×nFor coefficient matrix；System meets (A, B) is fully controllable, and (A, C) is the most considerable；System output y (t) follows the tracks of external reference input y₀T () realizes asymptotic tracking, i.e. Tracking error e (t) meets:State is not directly Survey, set up state observer:

$\left\{\begin{array}{c}\stackrel{\‾}{x}(t+1|t)=\stackrel{\‾}{A}\stackrel{\‾}{x}\left(t\right|t-1)+\stackrel{\‾}{B}\mathrm{\Δ}u\left(t\right)+\stackrel{\‾}{L}\[y\left(t\right)-y\left(t\right|t-1)\]\\ y\left(t\right|t-1)=\stackrel{\‾}{C}x\left(t\right|t-1)\end{array}\right.$

Wherein, x (t+1 | t) is the current time State Viewpoint measured value to subsequent time, before x (t | t-1) and y (t | t-1) is respectively One moment is to the State Viewpoint measured value of current time and output observation, L ∈ R^n*lFor observer gain；Augmentation observer gain

(2) the controlling increment sequence of k moment neural network forecast generator output is:

$\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}\left(k\right|k-{\mathrm{\τ}}_{sc,k})=-K_{{\mathrm{\τ}}_{sck}}\·\[{\mathrm{\Γ}}_{{\mathrm{\τ}}_{sc,k}}\stackrel{\‾}{x}(k-{\mathrm{\τ}}_{sc,k}+1|k-{\mathrm{\τ}}_{sc,k})-Y_0\left(k\right)\]$

Wherein, It is 1 × (N_u+τ_sc,k-1) block matrix, from τ_sc,kOpen for+1 piece Begin as unit matrix I ∈ R^m×m；It is τ that k time data wraps in the real-time time delay in feedback channel_sc,k, N_pFor forecast interval length, N_u-1 for optimizing siding-to-siding block length；For N_pThe block matrix of × 1, For lower triangular matrix, expire Foot j-i≤τ_sc,k-1 andY₀K () is that system follows desired value sequence: Y₀(k)=[y₀(k+1) … y₀ (k+N_p)]^T；R, Q are weight coefficient matrix；System output and control input are respectively y (k-τ_sc,k) and u (k-τ_sc,k),State Viewpoint measured value for k moment state observer latest update:

(3) controlling increment of time delay compensator t is:

$\mathrm{\Δ}u\left(t\right)=S_{{{\mathrm{\τ}}^{*}}_{ca,t}}\·\mathrm{\Δ}{\stackrel{\‾}{U}}^{*}(t-{{\mathrm{\τ}}^{*}}_{ca.t}|t-{{\mathrm{\τ}}^{*}}_t)=-S_{{{\mathrm{\τ}}^{*}}_{ca,t}}{K}_{{{\mathrm{\τ}}^{*}}_{sc,t}}\[{\mathrm{\Γ}}_{{{\mathrm{\τ}}^{*}}_{sc,t}}\stackrel{\‾}{x}(t-{{\mathrm{\τ}}^{*}}_t+1|k-{{\mathrm{\τ}}^{*}}_t)-Y_0(t-{{\mathrm{\τ}}^{*}}_{ca,t})\]$

Wherein,It is 1 × (N_u-1) block matrix, theBlock is unit matrix I ∈ R^m×m,For_tMoment mends Repay up-to-date controlling increment sequence, τ in device^* _sc,tAnd τ^* _ca,tIt is respectively the up-to-date optimal control increment sequence that time delay compensator is chosen Corresponding time delay；Networking tracing control closed loop system is: x (t+1)=Ax (t)+B [u (t-1)+Δ u (t)].