CN102902203A - Time series prediction and intelligent control combined online parameter adjustment method and system - Google Patents

Abstract

The invention discloses a time series prediction-based wavelet neural network online PID adjustment method and a system using the same. The method specifically comprises the steps of initiating the parameters, computing the control parameters and rectifying the online adjustment parameters, computing the control amount, computing or acquiring the system output and computing the prediction result; the system specifically comprises a control decision device, an online adjuster, a control executer, a controlled object, an online predictor, a control perturbation source and a prediction perturbation source; the control decision device is used for realizing the parameter initiation; the online adjuster is used for computing the control parameters and rectifying the online adjustment algorithm parameters; the control executer is used for computing the control amount; the online predictor is used for computing the prediction result; and the control decision device is also used for judging whether the algorithm is finished. According to the method, the wavelet neural network and the classic control method are combined to solve the problem of dependence on the parameter configuration work before the system operates in the control field, so that the control system has the effects of prediction, learning, online parameter optimization and self-adaptation.

Description

The parameter on-line tuning method and system that time series forecasting is combined with Based Intelligent Control

The skill technical field

The present invention proposes a kind of time series forecasting the parameter on-line tuning method of being combined with Based Intelligent Control and the system that has adopted the method, can be applicable to the fields such as control and Based Intelligent Control, artificial intelligence and aid decision making.

Background technology

Up to the present, the feature of the control method of control scientific research is: the control method of (1) control system is made of control algolithm and control parameter; (2) control algolithm is described with the rule (model, formula) of determining; (3) there are two kinds of forms in the configuration of control parameter and setting method: (3.1) control parameter is carried out initial configuration, is tentatively adjusted and finally adjust before the control system commencement of commercial operation, the control parameter immobilizes such as proportional integral derivative control method, fuzzy control method etc. after the commencement of commercial operation; (3.2) the control parameter is carried out initialization and is tentatively adjusted before the control system commencement of commercial operation, and control system self is adjusted to the control parameter according to the rule of determining and target after the commencement of commercial operation, for example neural network control method; (4) foundation of control parameter adjustment can be divided into again two kinds of situations: (4.1) can adjust methods such as proportional integral derivative control, fuzzy control, ANN (Artificial Neural Network) Control according to history, the currency of each state of control system; (4.2) can also adjust according to the predicted value of state, for example PREDICTIVE CONTROL.

Feature (1) is determined by Mathematical Modeling Methods and Computer Organization Principles.At present Mathematical Modeling Methods is the effective method of describing objective law, and computing machine is to realize the most effective instrument of control.Therefore be subjected to the restriction of above-mentioned two basic subjects development based on control algolithm and the development of the control method of control parameter model.Breakthrough for feature (1) need to break through above-mentioned two Research foundations.

Feature (2) is subjected to Artificial Intelligence Science development constraint to a great extent, because the artificial intelligence main method is mainly in the field of the methods such as inference system, Neurocomputing Science at present, so the pattern of control algolithm is approximate with intelligent algorithm to a great extent.Be breakthrough to artificial intelligence approach for the breakthrough key of feature (2).

This research improves mainly for feature (3) and feature (4), and final wavelet neural network, time series forecasting technology are carried out combination with classical control method, have created a kind of new algorithm.At first the present Research of the classic algorithm that relates to of feature (3) and feature (4) is introduced:

Proportional integral derivative control is the widest control method of classic applications, the core concept of the method is that the embodiment that will control is divided into three kinds of situations, namely execute continuously control, assault and execute control and delay to execute three kinds of modes of control, thereby solve three category features of control system controlled device, namely reach a certain target need to continue stressed, need to be subjected to energetically when starting or when needing fast response, the stressed excessive system that causes is unstable.The advantage of the method is simple in structure, and it is convenient to implement.The problem of the method is lack of wisdom, can't tackle online the variation that control system occurs, and when for example controlled device changed, control system self can't be adjusted the control parameter, thereby tackles this variation; The method self can't targetedly be attempted the optimal control parameter intelligently according to environmental change in addition.

Fuzzy control method can be in advance with for the control parameter configuration of different situations in control system, when situation changes, the control parameter can change the control parameter according to predefined rule, solved the problem of control system Adaptive change, but three class features that can't as classical proportional integral derivative control, tackle controlled device; Same fuzzy control method self also can't targetedly be attempted the optimal control parameter intelligently according to environmental change.

The characteristics of neural network control method are targetedly to attempt intelligently the optimal control parameter according to environmental change, and the process of adjustment parameter can be nonlinear.This method has had certain intelligent, and its reason one is because the algorithm basis of the method is the biology mathematical model of human brain.The problem of the method is three class features that can't tackle controlled device as classical proportional integral derivative control.

The characteristics of forecast Control Algorithm be the control parameter adjustment according to increased control implemented after result's prediction, therefore control has possessed advanced; But the problem of the method is except tackling three class features of controlled device as classical proportional integral derivative control, also exist the method to the Dependence Problem of plant model, if forecast model is inaccurate, it is inaccurate then can to cause predicting the outcome, Forecasting Methodology is based on the mathematical model of controlled device, and the accuracy that predicts the outcome and reliability lack guarantee;

Summary of the invention

The problem that this present invention mainly solves:

The present invention proposes a kind of time series forecasting the parameter on-line tuning method of being combined with Based Intelligent Control and the system that has adopted the method, can be applicable to the fields such as control and Based Intelligent Control, artificial intelligence and aid decision making.The method has solved the Dependence Problem of control field to the front parameter configuration work of system's operation with wavelet neural network and classical control method combination; Output adopts the vector time series method to process to system, so that control method has had prediction effect, and this prediction has the statistical test assurance; This technical scheme combines wavelet neural network and forecasting techniques simultaneously with classical control method, so that control system has had comparatively senior artificial intelligence, have prediction, study, parameter on-line optimization, adaptive effect.

Innovative point of the present invention is:

(1) the employing wavelet neural network has been realized the control parameter on-line optimization to classical control method;

(2) adopted the vector time series method to realize PREDICTIVE CONTROL to control system;

(3) adopted simultaneously the realization of wavelet neural network and vector time series method with the control parameter on-line optimization of prediction effect;

(4) said method that proposes for the present invention has designed the utility model hardware system that adopts the method; Write complete computer simulation program, can verify every effect of the method and control result's stability; This program and subroutine thereof can directly apply in the utility model hardware system that has adopted the method and go;

The technical scheme that the present invention takes:

To achieve these goals, address the above problem, the present invention has taked following technical scheme:

1. the time series forecasting parameter on-line tuning method of being combined with Based Intelligent Control is characterized in that: comprise following 5 steps, namely step 1.0 is to step 1.5:

Step 1.0 parameter initialization:

This step is that each variable that relates in the computation process is given respectively initial value; Comprise following 12 sub-steps, namely step 1.0.1 is to step 1.0.12:

1.0.1 the current execution number of times of set algorithm t, execution total degree T:

When algorithm moves the situation of limited number of time continuously: establish current execution number of times initial value t=1; The value principle of carrying out and adjust total degree T is: operation 0.8T the rear control result of total degree of assurance system (1300) can be stabilized near the control target (1000); Wherein can stable implication be: after the execution number of times reaches 0.8T, control result (1300) y _System(t) with control target (1000) r _In(t) error between is in ± 3%, referring to formula 1:

$\left|\frac{y_{\mathrm{system}}\left(t\right)-r_{\mathrm{in}}\left(t\right)}{r_{\mathrm{in}}\left(t\right)}\right|<3\%$ Formula 1

The method to set up of carrying out and adjust total degree T is:

Carry out total degree T=5000 1.0.1.1 establish;

1.0.1.2 operational system, and carry out T this method Overall Steps, adjust according to operation result again;

If 1.0.1.3 the operation result display algorithm is carried out after 0.8T time, control result (1300) can't be stable at control target (1000), then T is increased 10%, and jumps to step 1.0.1.2;

If 1.0.1.4 the operation result display algorithm namely reaches before the 0.5T control result (1300) y when carrying out number of times after control result (1300) being converged on control target (1000) for a long time _System(t) with control target (1000) r _In(t) error between then reduces 10% with T in ± 3%, and jumps to step 1.0.1.2;

The result of above-mentioned steps is: operation 0.8T the rear control result of total degree of system (1300) can be stabilized near the control target (1000), and the so far work that arranges of T is finished;

When algorithm moves situation about never stopping continuously: establish current execution number of times initial value t=1; If carry out total degree T=0, this algorithm constantly circulates when T=0, and algorithm can not stop;

1.0.2 set each control target (1000) value r constantly _In(t), r _In(t) can be a constant r _In, also can be the function take t as independent variable;

Start constantly T of prediction 1.0.3 set _Predict, T _PredictBe positive integer, according to engineering experience, its span is referenced as: 5%T≤T _Predict≤ 10%T;

1.0.4 set the number K of the historical data of online prediction algorithm (401) needs _Predict, Kpredict is positive integer, its value determines that by the test of hypothesis statistical indicator of on-line prediction algorithm according to engineering experience, its span is referenced as: 10≤K _Predict≤ 5%T;

1.0.5 the setting prediction step is L _Predict, L _PredictBe positive integer, according to engineering experience, its span is referenced as: 1≤L _Predict≤ 10%T;

1.0.6 setup control as a result initial value is zero, i.e. y _System(t)=0;

1.0.7 setup control prediction of result value is zero, i.e. y _Predict(t)=0;

Initial value is zero 1.0.8 setting predicts the outcome, i.e. y _Out(t)=0;

1.0.9 set each layer of wavelet neural network neuronal quantity M, Q, L, span generally is M≤10, M≤Q≤5M, in the positive integer interval in L≤M interval;

Wherein M represents the input layer quantity of wavelet neural network, and namely wavelet neural network is made of M input layer; Q represents the hidden layer neuron quantity of wavelet neural network, and namely wavelet neural network is made of Q hidden layer neuron; L represents the output layer neuronal quantity of wavelet neural network, and namely wavelet neural network is made of L output layer neuron;

1.0.10 set input layer, hidden layer, each the neuronic input variable of output layer in the computation process t=1 time $\begin{array}{cc}{\mathrm{net}}_i^{\left(1\right)}\left(t\right)& (i=\mathrm{1,2},...,M),\end{array}$ $\begin{array}{cc}{\mathrm{net}}_j^{\left(2\right)}\left(t\right)& (j=\mathrm{1,2},...,Q),\end{array}$ $\begin{array}{cc}{\mathrm{net}}_k^{\left(3\right)}\left(t\right)& (k=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,L),\end{array}$ The connection weight value matrix of input layer and hidden layer

Each neuronic connection weight value matrix of hidden layer output layer

The activation functions change of scale parameter matrix a of hidden layer neuron _j(t) and b _j(t) each element in is the at random decimal in 0 to 1 open interval;

Ground floor---the input layer of the upper right footmark of above-mentioned variable " (1) " expression neural network, hidden layer represents with " (2) ", output layer represents with " (3) ", bottom right footmark " i " expression input layer sequence number, " j " expression hidden layer neuron sequence number, " k " expression output layer neuron sequence number;

1.0.11 set learning rate η and the inertial coefficient α of neural network: wherein the span of η is the decimal in 0.01 to 0.7 closed interval; The span of α is the decimal in the 0.01-0.2 closed interval;

1.0.12 the initial value of setup control amount u (t) and margin of error error (t) and time sequential value are zero: u (t)=0, u (t-1)=0, u (t-2)=0, error (t)=0, error (t-1)=0, error (t-2)=0;

Step 1.1 is calculated the control parameter K _cAnd correction on-line tuning algorithm parameter w:

This step is according to control target (1000) r _In(t) and (1400) y that predicts the outcome _Out(t-1), adopt on-line tuning algorithm (101), calculate control parameter (1100) K _cAnd revise on-line tuning algorithm parameter w (t+1) (t); On-line tuning algorithm in the parameter on-line tuning method that on-line tuning algorithm (101) employing time series forecasting is combined with Based Intelligent Control perhaps adopts control parameter on-line tuning method non-time-based sequence prediction or non-neural net method on-line tuning control parameter;

Wherein adopt the on-line tuning algorithm steps in the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control as follows,

1.1.1 calculate the control parameter K _c:

1.1.1.1 calculate the input of input layer

The input of input layer

Calculating referring to formula 46:

$\left\{\begin{array}{c}{\mathrm{net}}_1^{\left(1\right)}\left(t\right)=r_{\mathrm{in}}\left(t\right)\\ {\mathrm{net}}_2^{\left(1\right)}\left(t\right)=y_{\mathrm{out}}(t-1)\\ {\mathrm{net}}_3^{\left(1\right)}\left(t\right)=\mathrm{error}\left(t\right)\end{array}\right.$ Formula 2

Variable wherein

3 neuronic inputs of expression input layer; 3 neuronic inputs of input layer are respectively (1400) y that predicts the outcome _Out(t-1), control target (1000) r _In(t), departure error (t), the calculating of error (t) is referring to formula 47:

Error (t)=y _Out(t-1)-r _In(t) formula 3

1.1.1.2 calculate the output of input layer

The output of neural network input layer

Calculating referring to formula 48:

$O_i^{\left(1\right)}\left(t\right)={\mathrm{net}}_i^{\left(1\right)}\left(t\right),$ (i=1,2,3) formula 4

Wherein, variable is used respectively in input layer, hidden layer, each neuronic output of output layer

$(i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,M),$ $\begin{array}{cc}{O}_j^{\left(2\right)}\left(t\right)& (j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,Q),\end{array}$ $\begin{array}{cc}{O}_k^{\left(3\right)}\left(t\right)& (k=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,L)\end{array}$ Expression;

1.1.1.3 calculate the input of hidden layer

The input of neural network hidden layer

Calculating referring to formula 49:

${\mathrm{net}}_j^{\left(2\right)}\left(t\right)=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)\&CenterDot;O_i^{\left(1\right)}\left(t\right),$ (i=1,2 ..., Q) formula 5

Wherein

Weighting weights between expression i neuron of input layer and j neuron of hidden layer;

1.1.1.4 calculate the output of hidden layer

The output of neural network hidden layer Calculating referring to formula 50:

$O_j^{\left(2\right)}\left(t\right)={\mathrm{\&psi;}}_{a,b}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right),$ (j=1,2 ..., Q) formula 6

ψ wherein _{A, b}(t) be wavelet function, i.e. the activation functions of hidden layer neuron; a _j(t), b _j(t) be wavelet function change of scale parameter, i.e. the activation functions change of scale parameter of hidden layer neuron; Get the wavelet function ψ that satisfies the framework condition _{A, b}The computing method of (t), seeing are referring to formula 51:

${\mathrm{\&psi;}}_{a,b}\left(t\right)=\cos \left(1.75t\right)\&CenterDot;e^{-\frac{t^2}{2}}$ Formula 7

ψ _{A, b}The computing method of a mediation number (t) are referring to formula 52:

${\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(t\right)=-1.75\&CenterDot;\sin \left(1.75t\right)\&CenterDot;e^{-\frac{t^2}{2}}-t\&CenterDot;\cos \left(1.75t\right)\&CenterDot;e^{-\frac{t^2}{2}}$ Formula 8

Wherein sin (x), cos (x) represent respectively sine function and the cosine function value of x;

1.1.1.5 calculate the input of output layer

The input of neural network output layer

Calculating referring to formula 53:

${\mathrm{net}}_k^{\left(3\right)}\left(t\right)=\underset{j=1}{\overset{Q}{\mathrm{\&Sigma;}}}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;O_j^{\left(2\right)}\left(t\right),$ (k=1,2 ..., L) formula 9

Wherein

Weighting weights between expression j neuron of hidden layer and k neuron of output layer;

1.1.1.6 calculate the output of output layer

The output of neural network output layer

Calculating referring to formula 54:

$O_k^{\left(3\right)}\left(t\right)=g\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right),$ (k=1,2 ..., L) formula 10

Wherein g (x) gets non-negative Sigmoid function for the neuronic activation functions of output layer, and the calculating of g (x) is referring to formula 55:

$g\left(x\right)==\frac{e^x}{e^x+e^{-x}}$ Formula 11

The calculating of its first order derivative of g (x) is referring to formula 56:

$g^{\&prime;}\left(x\right)==\frac{2}{{(e^x+e^{-x})}^2}$ Formula 12

1.1.1.7 calculate control parameter (1100) K _p(t), K _i(t), K _d(t):

Control parameter (1100) K _p(t), K _i(t), K _d(t) calculating is referring to formula 57:

$\left\{\begin{array}{c}{K}_p\left(t\right)=O_1^{\left(3\right)}\left(t\right)\\ K_i\left(t\right)=O_2^{\left(3\right)}\left(t\right)\\ K_d\left(t\right)=O_3^{\left(3\right)}\left(t\right)\end{array}\right.$ Formula 13

Wherein, K _p(t), K _i(t), K _d(t) be control parameter (1100) K that the neuronic input of output layer corresponds to respectively classical proportional integral derivative control method _c(t);

1.1.1.8 control parameter K _c(t) calculate end;

1.1.2 revise on-line tuning algorithm parameter w (t):

On-line tuning algorithm parameter w (t) in the on-line tuning algorithm in the online PID setting method of wavelet neural network comprises following four parameters: the weighting weights between i neuron of input layer and j neuron of hidden layer

Weighting weights between j neuron of hidden layer and k neuron of output layer

Wavelet function change of scale parameter a _j(t), b _j(t); Namely

a _j(t), b _j(t) };

The step 1.1.2.1 error of calculation and performance index:

The calculating of error e rror (t) is referring to formula 58:

Error (t)=y _Out(t-1)-r _In(t) formula 14

The calculating of performance index E (t) is referring to formula 59:

$E\left(t\right)=\frac{1}{2}{\mathrm{error}}^2\left(t\right)=\frac{1}{2}{(r_{\mathrm{in}}\left(t\right)-y_{\mathrm{out}}(t-1))}^2$ Formula 15

Step 1.1.2.2 adopts gradient descent method to the weights coefficient between neural network hidden layer and the output layer

Adjust:

Weights coefficient between hidden layer and the output layer

Correction

Calculating referring to formula 60:

$\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)}$ Formula 16

In the formula 60, Computing formula referring to formula 61:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\left(t\right)\mathrm{jk}}^{\left(3\right)}\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)}$ Formula 17

In the formula 61,

Calculating referring to formula 62:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;y_{\mathrm{out}}\left(t\right)}=\mathrm{error}\left(t\right)\&CenterDot;\frac{\&PartialD;\mathrm{error}\left(t\right)}{\&PartialD;y_{\mathrm{out}}\left(t\right)}=\mathrm{error}\left(t\right)$ Formula 18

In the formula 61, because

The unknown is with approximate sign function Replace, Calculating referring to formula 63;

$\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;\mathrm{\&Delta;u}\left(t\right)}=\mathrm{sgn}\left(\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;\mathrm{\&Delta;u}\left(t\right)}\right)$ Formula 19

The sign function value of sgn (x) expression x wherein, if i.e. x＞o then sgn (x)=1, if x=o then sgn (x) if=0 x＜0 then sgn (x)=-1;

In the formula 61,

Calculating referring to formula 64:

$\left\{\begin{array}{c}\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_1^{\left(3\right)}\left(t\right)}=\mathrm{error}\left(t\right)-\mathrm{error}(t-1)\\ \frac{\&PartialD;u\left(t\right)}{\&PartialD;O_2^{\left(3\right)}\left(t\right)}=\mathrm{error}\left(t\right)\\ \frac{\&PartialD;u\left(t\right)}{\&PartialD;O_3^{\left(3\right)}\left(t\right)}=\mathrm{error}\left(t\right)-2\&CenterDot;\mathrm{error}(t-1)+\mathrm{error}(t-2)\end{array}\right.$ Formula 20

In the formula 61,

Calculating referring to formula 65:

$\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}=g^{\&prime;}\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right)$ Formula 21

In the formula 61,

Calculating referring to formula 66:

$\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)}=O_j^{\left(2\right)}\left(t\right)$ Formula 22

In sum,

Computing formula can be expressed as formula 67:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)}=\mathrm{error}\left(t\right)\&CenterDot;\mathrm{sgn}\left(\frac{\&PartialD;y\left(t\right)}{\&PartialD;\mathrm{\&Delta;u}\left(t\right)}\right)\&CenterDot;\frac{\&PartialD;\mathrm{\&Delta;u}\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;g^{\&prime;}\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right)\&CenterDot;O_j^{\left(2\right)}\left(t\right)$ Formula 23

In sum,

Computing formula can be expressed as formula 68:

$\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}(t-1)\&PlusMinus;\mathrm{\&eta;}\&CenterDot;{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;O_j^{\left(2\right)}\left(t\right)$ Formula 24

Wherein,

Calculating referring to formula 69:

${\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}$ Formula 25

$=\mathrm{error}\left(t\right)\&CenterDot;\mathrm{sgn}\left(\frac{\&PartialD;y\left(t\right)}{\&PartialD;u\left(t\right)}\right)\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;g^{\&prime;}\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right)(k=\mathrm{1,2},3)$

1.1.2.3 adopt gradient descent method to the weights coefficient between neural network input layer and the hidden layer

Adjust:

Calculate the weights coefficient between input layer and the hidden layer

Modified value

See formula 70:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(2\right)}(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}$ Formula 26

In the formula 70, Calculating referring to formula 71:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;O_j^{\left(2\right)}\left(t\right)}$

$\&CenterDot;\frac{\&PartialD;O_j^{\left(2\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}$ Formula 27

$={\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;O_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_j^{\left(2\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}$

In the formula 71, Calculating referring to formula 72:

$\frac{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}=w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)$ Formula 28

In the formula 71,

Calculating referring to formula 73:

$\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;\mathrm{net}}_j^{\left(2\right)}\left(t\right)}={\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;\frac{1}{a_j\left(t\right)}$ Formula 29

In the formula 71, Calculating referring to formula 74:

$\frac{{\&PartialD;\mathrm{net}}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}=O_i^{\left(1\right)}\left(t\right)$ Formula 30

In sum, Computing formula can be expressed as formula 75:

$\frac{\&PartialD;E\left(t\right)}{{\&PartialD;w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;{\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;\frac{1}{a_j\left(t\right)}\&CenterDot;O_i^{\left(1\right)}\left(t\right)$ Formula 31

In the formula 75

In formula 69, provide;

1.1.2.3 calculate the activation functions change of scale parameter a of hidden layer neuron _j(t) modified value Δ a _j(t):

The activation functions change of scale parameter a of hidden layer neuron _j(t) modified value Δ a _j(t) computing formula can be expressed as formula 76:

${\mathrm{\&Delta;a}}_j\left(t\right)=\mathrm{\&alpha;}\&CenterDot;{\mathrm{\&Delta;a}}_j(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{{\&PartialD;a}_j\left(t\right)}$ Formula 32

In the formula 76,

Calculating referring to formula 77:

$\frac{\&PartialD;E\left(t\right)}{{\&PartialD;a}_j\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{{\&PartialD;O}_k^{\left(3\right)}}\&CenterDot;\frac{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;a}_j\left(t\right)}$ Formula 33

In the formula 77, $\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}$ Calculating referring to formula 69:

In the formula 77,

Calculating referring to formula 72:

In the formula 77,

Calculating referring to formula 78:

$\frac{\&PartialD;O_j^{\left(2\right)}}{\&PartialD;a_j\left(t\right)}={\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;(-\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j^2\left(t\right)})$ Formula 34

In sum,

Computing formula can be expressed as formula 79:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;a_j\left(t\right)}=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;{\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;(-\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j^2\left(t\right)})$ Formula 35

1.1.2.4 calculate the activation functions change of scale parameter b of hidden layer neuron _j(t) modified value Δ b _j(t):

The activation functions change of scale parameter b of hidden layer neuron _j(t) modified value Δ b _j(t) computing formula can be expressed as formula 80:

$\mathrm{\&Delta;}{b}_j\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{b}_j(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{\&PartialD;b_j\left(t\right)}$ Formula 36

In the formula 80,

Calculating referring to formula 81:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;b_j\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;b}_j\left(t\right)}$ Formula 37

In the formula 80, $\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}$ Calculating referring to formula 69:

In the formula 80,

Calculating referring to formula 72:

In the formula 80,

Calculating referring to formula 82:

$\frac{\&PartialD;O_j^{\left(2\right)}\left(t\right)}{\&PartialD;b_j\left(t\right)}={\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^p\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)(-\frac{1}{a_j\left(t\right)})$ Formula 38

In sum,

Computing formula can be expressed as formula 83:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;b_j\left(t\right)}=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;{\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)(-\frac{1}{a_j\left(t\right)})$ Formula 39

Arrive w (t+1) 1.1.2.5 revise on-line tuning algorithm parameter w (t):

The on-line tuning algorithm parameter that current the t time calculating uses is:

Aj (t), b _j(t) }; Revised result be calculate for the t+1 time the on-line tuning algorithm parameter w (t+1) that will use= $\{w_{\mathrm{ij}}^{\left(2\right)}(t+1),$ $w_{\mathrm{jk}}^{\left(3\right)}(t+1),$ a _j(t+1)，b _j(t+1)}

Weighting weights between j neuron of hidden layer and k neuron of output layer

Corrected Calculation referring to formula 84:

$w_{\mathrm{jk}}^{\left(3\right)}(t+1)=w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)+\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)$ Formula 40

Weighting weights between i neuron of input layer and j neuron of hidden layer

Corrected Calculation referring to formula 85:

$w_{\mathrm{ij}}^{\left(2\right)}(t+1)=w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)+\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)$ Formula 41

The activation functions change of scale parameter a of hidden layer neuron _j(t) corrected Calculation is referring to formula 86:

a _j(t+1)=a _j(t)+Δ a _j(t) formula 42

The activation functions change of scale parameter b of hidden layer neuron _j(tt) corrected Calculation is referring to formula 87:

b _j(t+1)=b _j(t)+Δ b _j(t) formula 43

Step 1.2 is calculated controlled quentity controlled variable u (t):

This step adopts control algolithm (201), calculates controlled quentity controlled variable (1200) u (t);

Step 1.3 gathers or computing system output y _Sjstem(t):

This step calculates control result (1300) y _System(t) or directly collect control result (1300) y _System(t);

Step 1.4 is calculated the y that predicts the outcome _Out(t):

This step adopts on-line prediction algorithm (401), calculates (1400) y that predicts the outcome _Out(t);

Whether step 1.5 evaluation algorithm finishes:

Algorithm is unclosed situation still: if T=0, if perhaps T ≠ 0 and t＜T, then t=t+1 jumps to step 1.1, has so far finished the t time calculating;

The situation that algorithm finishes: if T ≠ 0 and t 〉=T have so far finished whole T time and calculated;

2. the step 1.2 of the time series forecasting according to claim 1 parameter on-line tuning method of being combined with Based Intelligent Control is calculated controlled quentity controlled variable, it is characterized in that:

The input of algorithm is control parameter (1100) K _c(t), controlled quentity controlled variable (1200) u (t) and error e rror (t) time sequential value;

The output of algorithm is controlled quentity controlled variable (1200) u (t);

Control algolithm (201) can employing, positional PID control calculation, increment type PID control algolithm, FUZZY ALGORITHMS FOR CONTROL, expert's control algolithm;

When adopting the increment type PID control algolithm, the calculating of u (t) is referring to formula 44:

U (t)=u (t-1)+K _p(t) (error (t)-error (t-1)+K _i(t) error (t) formula 44

+K _d(t)·(error(t)-2error(t-1)+error(t-2))

3. the step 1.3 of the time series forecasting according to claim 1 parameter on-line tuning method of being combined with Based Intelligent Control gathers or computing system output, it is characterized in that:

When this system was the situation of Computer Simulation application, controlled device (300) adopted the control system controlled device mathematical model of determining, the input of this mathematical model is controlled quentity controlled variable u (t), and the result who draws by Computer Simulation is as a result y of control _System(t); Above-mentioned Computer Simulation namely by controlled device being set up continuous transfer function model, and then adopts z conversion to carry out discretize to this continuous transport function, obtains according to u (t) and y _System(t) time sequential value calculates y _System(t) computing formula; The z conversion realizes by c2d () function and tfdata () function in Matlab; The specific implementation process sees also embodiment part in the patent specification;

When this system is the situation of true control system: therefore control as a result y _System(t) do not obtain by Computer Simulation, but obtain by true control system output is sampled;

4. the step 1.4 of the time series forecasting according to claim 1 parameter on-line tuning method of being combined with Based Intelligent Control is calculated and is predicted the outcome, and it is characterized in that:

Judge at first whether current t arrives to start to predict it is to start constantly T of prediction constantly _Predict:

If current t is constantly less than starting constantly T of prediction _Predict, then do not start prediction, control result (1300) y _System(t) directly serve as (1400) y that predicts the outcome _Our(t), i.e. y _Out(t)=y _System(t);

If current t is constantly more than or equal to starting constantly T of prediction _Predict, then start prediction, gather control disturbance amount (1500) r of control disturbance source (500) generation of current time _Oef(t) and prediction disturbance quantity (1600) r that produces of prediction disturbing source (600) _Orf(t);

The input of on-line prediction algorithm (401) is respectively: the time series K of control parameter (1100) _c(t), K _c(t-1), K _c(t-2) ..., K _c(t-K _Predict), time series u (t), the u (t-1) of controlled quentity controlled variable (1200), u (t-2) ..., u (t-K _Predict), control the result (1300) time series y _System(t), y _System(t-1), y _System(t-2) ..., y _System(t-K _Predict), the time series r of control disturbance amount (1500) _Oef(t), r _Oef(t-1), r _Oef(t-2) ..., r _Oef(t-K _Predict), the prediction disturbance quantity (1600) time series r _Orf(t), r _Otf(t-1), r _Orf(t-2) ..., r _Orf(t-K _Predict);

Wherein, control disturbance source (500) be except controlled quentity controlled variable (1200) to other devices in the controling environment of controlled device generation effect; Prediction disturbing source (600) be except control parameter (1100), controlled quentity controlled variable (1200), control the result (1300) to on-line prediction device (400) exert an influence control environment in other devices; Control disturbance amount (1500) is to react on controlled device (300) by what control disturbance source (500) produced with controlled quentity controlled variable (1200), and the input that its output is exerted an influence, in ideal conditions, and control disturbance amount (1500) r _Oef(t) can for constant 0, namely ignore; Wherein, prediction disturbance quantity (1600) be produced by prediction disturbing source (600) with control parameter (1100), controlled quentity controlled variable (1200), control result (1300) reacts on on-line prediction device (400), and the input that the output of on-line prediction algorithm (401) is exerted an influence, in the ideal case, namely can ignore in the situation of the disturbing influence of other devices in the control system internal and external environment, control disturbance source (500) and prediction disturbing source (600) can be ignored, namely there are not control disturbance source (500) and prediction disturbing source (600), at this moment control disturbance amount (1500) r _Oef(t) and prediction disturbance quantity (1600) r _Orf(t) be constant 0;

The output of on-line prediction algorithm (401) is control prediction of result value (1401) y _Predict(t);

Current t is constantly more than or equal to starting constantly T of prediction _PredictIn this case, will control prediction of result value (1401) y _Predict(t) as predicting the outcome (1400) y _Our(t), i.e. y _Out(t)=y _Predict(t);

In sum, (1400) y predicts the outcome _Our(t) value is seen formula 45:

$y_{\mathrm{out}}\left(t\right)=\left\{\begin{array}{ccc}{y}_{\mathrm{system}}\left(t\right)& ,& t<T_{\mathrm{predict}}\\ y_{\mathrm{predict}}\left(t\right)& ,& t\&GreaterEqual;T_{\mathrm{predict}}\end{array}\right.$ Formula 45

5. on-line prediction algorithm according to claim 4 (401) is characterized in that:

It is VARMA method, neural net prediction method, linear regression Forecasting Methodology, non-linear regression Forecasting Methodology, curvilinear function approximating method that on-line prediction algorithm (401) can adopt classical vector time series Forecasting Methodology, also can adopt satisfied input is that output is the prediction algorithm of control prediction of result value (1401) by time series control parameter (1100), time series controlled quentity controlled variable (1200), time series control result (1300), time series control disturbance amount (1500), time series forecasting disturbance quantity (1600);

When adopting the VARMA algorithm, the span of nAR and nMA parameter is the positive integer in 5 to 10 closed intervals; The VARMA prediction algorithm is realized by vgxset (), vgxvarx (), three functions of vgxpred () in Matlab; The specific implementation process sees also embodiment part in the patent specification;

6. adopted the system of the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control, it is characterized in that: by the 〇 device that---------------on-line prediction device (400), the 5th device---control disturbance source (500) and the 6th device---predict that (600) seven groups of devices of disturbing source form for controlled device (300), the 4th device for control actuator (200), the 3rd device for on-line tuning device (100), the second device for control decision device (0), first device;

1.1 respectively organize the inner formation of device and input, output interface situation:

1.1.0 the 〇 device--control decision device (0) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC programmable logic controller (PLC) or computing machine or server or portable terminal;

When the 〇 device---when control decision device (0) adopted computing machine (1), its output interface can be Ethernet interface, by netting twine and the network switching equipment and first device---the input interface of on-line tuning device (100) is connected;

1.1.1 first device---on-line tuning device (100) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC programmable logic controller (PLC) or computing machine or server;

When first device---when on-line tuning device (100) adopts computing machine (100), two input interface can be Ethernet interface, by netting twine and the network switching equipment and 〇 device---control decision device (0), the 4th device---output interface of on-line prediction device (400) is connected; Two output interface can be Ethernet interface, and the input interface of on-line prediction device (400) is connected---to control actuator (200), the 4th device---by netting twine and the network switching equipment and the second device; Above-mentioned Ethernet interface can multiplexing same Ethernet interface;

1.1.2 the second device---control actuator (200) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC programmable logic controller (PLC) or computing machine or server or PID controller cooperation frequency converter or driver;

When the second device---when control actuator (200) adopts PLC programmable logic controller (PLC) (201) and frequency converter (202), its input interface can be the Ethernet interface of PLC programmable logic controller (PLC) (201), by netting twine and the network switching equipment and first device---and the output interface of on-line tuning device (100) is connected; Its output interface can be that (202) three in frequency converter exchanges output interface, installs with the 3rd by cable---the input interface of controlled device (300) is connected;

The second device---the annexation of control actuator (200) interior arrangement is as follows: between PLC programmable logic controller (PLC) (201) and the frequency converter (202), can be connected by RS485 interface or Ethernet interface or industrial bus interface;

The second device---control actuator (200) is as retransmission unit, realization is from the 6th device, and------connection of on-line prediction device (400) is meticulous with the course of work to be: PLC programmable logic controller (PLC) (201) can receive from the 6th device as input interface by sensor interface---predict the information of disturbing source (600), and install to the 4th by Ethernet interface---, and on-line prediction device (400) is transmitted to predict that disturbing source (600) installs to the 4th;

1.1.3 the 3rd installs---controlled device (300) can be motor or temperature control device or voltage-controlled devices or electromagnetic field or production system or economic system;

When the 3rd device---when controlled device (300) adopts motor (301), load fan (302) and during speed measuring coder (303), one of them input interface can be that three of motor (301) exchange input interface, by cable and the second device---three of control actuator (200) exchange output interface and are connected; Wherein another input is the 5th device that load fan (302) is subject to---the disturbance of the air of the rapid flow of cooling fan (501) output of control disturbance source (500); The input interface of on-line prediction device (400) is connected the 3rd device---RS232 that controlled device (300) output interface can be by speed measuring coder (303) or Ethernet or industrial bus interface and the 4th install---;

The 3rd installs---and the annexation of controlled device (300) interior arrangement is as follows: motor (301) moment of torsion output main shaft is connected by the moment of torsion entering spindle of gearing with load fan (302), motor (301) band dynamic load fan (302) rotates, thereby consists of one group of output and the relation of inputting; The moment of torsion entering spindle of load fan (302) is fixed with again the code-disc of speed measuring coder (303) simultaneously, and fan (302) rotation drives code-disc and rotates, thereby consists of one group of output and the relation of inputting;

Can produce heat in motor (301) course of work, affect the serviceability of motor, and then the result of impact prediction; Therefore can produce heat in motor (301) course of work and can be used as the 4th device---one of input of on-line prediction device (400); This heat dissipation problem solves by cooling fan (501), cooling fan (501) also produces load fan (302) when solving heat dissipation problem and disturbs, and therefore the 5th installs---and the output of control disturbance source (500) also is one of input of load fan (302);

1.1.4 the 4th installs---on-line prediction device (400) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC or computing machine or server;

When the 4th device---when on-line prediction device (400) adopts computing machine (401), two input interface can be Ethernet interface, and the Ethernet output interface of the PLC programmable logic controller (PLC) (201) of---on-line tuning device (100), second device---control actuator (200) is connected by netting twine and the network switching equipment and first device; Be connected by RS232 or Ethernet or industrial bus interface between the speed measuring coder (303) of its another input interface and controlled device (300);

1.1.5 the 5th device---control disturbance source (500) can be the devices that affects the device of Air Flow or affect the device of humidity, the device that affects temperature, barrier, generation interference;

When the 5th device---when control disturbance source (500) adopt cooling fan (501), its fan rotates the moving air that produces motor (301) is lowered the temperature, also can exert an influence to load fan (302) simultaneously, namely as one of input of load fan (302);

1.1.6 the 6th device---prediction disturbing source (600) can be the device that affects the device of Air Flow or affect the device of humidity, the device that affects temperature, barrier, generation interference.

When thinking the 6th device---when one of prediction disturbing source (600) was the thermal value of motor (301), the input of temperature sensor (601) was the temperature of motor (301); The three-wire system connection is adopted in the output of temperature sensor (601) and the second device---the sensor input interface of the PLC programmable logic controller (PLC) (201) of control actuator (200) is connected;

1.2 respectively organize annexation and signal transitive relation between device:

Control decision device (0) is output as control target (1000) r of system _InThis control target (1000) is connected with the input of on-line tuning device (100), becomes one of input of on-line tuning device (100); Two of the input of on-line tuning device (100) is (1400) y that predict the outcome _OurOn-line controller calculates control parameter (1100) K according to control target (1000) and predict the outcome (1400) by on-line tuning algorithm (101) _c, this control parameter (1100) is the input of control actuator (200), is again one of input of on-line prediction device (400); Control actuator (200) calculates controlled quentity controlled variable (1200) u according to the control parameter (1100) of its input by control algolithm, and this controlled quentity controlled variable (1200) is one of input of controlled device (300), be again on-line prediction device (400) input two; Two of the input of controlled device (300) is control disturbance amount (1500) r that control disturbance source (500) produces _Oef, controlled device (300) produces control result (1300) y under the acting in conjunction of controlled quentity controlled variable (1200) and these two inputs of control disturbance amount (1500) _System, this control result (1300) be on-line prediction device (400) input three; Four of the input of on-line prediction device (400) is prediction disturbance quantity (1600) r that produced by prediction disturbing source (600) _Orf, on-line prediction device (400) is according to control parameter (1100), controlled quentity controlled variable (1200), control result (1300), prediction disturbance quantity (1600) r _OrfThe time series history value, calculate control prediction of result value (1401) y by on-line prediction algorithm (401) _PredictWith (1400) y that predicts the outcome _Out, this (1400) input as on-line tuning device (100) that predicts the outcome;

1.3 the course of work after system starts is as follows:

After system starts each installed according to 1.3.0 process 0 to 1.3.5 process 5 these six process operations:

1.3.0 process 0:

〇 device---control decision device (0) is realized step 1.0 parameter initialization of " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control ";

1.3.1 process 1:

First device---on-line tuning device (100) is realized the step 1.1 calculating control parameter K of " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " _cAnd correction on-line tuning algorithm parameter w;

1.3.2 process 2:

The second device---control actuator (200) is realized " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.2 calculating controlled quentity controlled variable u (t);

1.3.3 process 3:

The 3rd device---controlled device (300) realizes that " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.3 gathers or computing system output y _System(t);

1.3.4 process 4:

The 4th device---on-line prediction device (400) realization " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.4 is calculated the y that predicts the outcome _Out(t);

1.3.5 process 5:

〇 device---control decision device (0) realizes according to parameters, state and the output of current system whether " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.5 evaluation algorithm finishes.

Advantage compared with prior art of the present invention:

The parameter on-line tuning method and system advantage compared with prior art that the time series forecasting that the present invention proposes is combined with Based Intelligent Control is as follows: (1) is compared with the proportional integral derivative control method of classics, the method has the on-line tuning ability, the control parameter can be realized tentatively arranging before control system is moved and control system can be started, system is in operational process, can be according to controlled device and the characteristics that control environment, Automatic Optimal control parameter, realize the more control target of optimization, the control target not only can be control output, also can be the self-defining performance function value of formulating according to control system each several part state value; (2) compare with fuzzy control method, the method has the ability of intelligence learning, need not to provide clear and definite control parameter rule before system's operation, controller can be according to situation and the characteristics of system's operation, on-line study produces the control tuning method, and will be imposed in the computation process of control parameter; (3) compare with neural network control method, the learning functionality of the method has the characteristic that can not be absorbed in local minimum point; (4) compare with PREDICTIVE CONTROL, this algorithm can be broken away from the dependence to the plant model degree of accuracy; (5) compare with existing on-line tuning PID method, the method has more senior intelligence learning ability, need not to provide clear and definite Optimization about control parameter rule before system's operation, controller can be according to situation and the characteristics of system's operation, on-line study produces the Optimization about control parameter rule, and will be imposed in the computation process of control parameter;

Description of drawings

Accompanying drawing 1 has adopted the structural drawing of the system of the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control

The parameter on-line tuning method flow diagram that accompanying drawing 2 time series forecastings are combined with Based Intelligent Control

The device annexation figure of accompanying drawing 3 systems

Embodiment

Provide one embodiment of the present of invention below in conjunction with accompanying drawing 1 to accompanying drawing 3: the disturbance rejection Motor Rotating Speed Control System.The control target of this system is to wish before system's operation, only needs rough configuration control parameter to get final product start-up system.In system's operational process, according to actual working environment and disturbed condition, the online control parameter of adjusting, the load that motor is dragged reaches as early as possible and is stable at rotating speed of target.

(1) lectotype selection:

The model of the computing machine that adopts in the present embodiment as shown in Figure 1, (1), computing machine (101) and computing machine (401) is IBM x3650; The model that the PLC programmable logic controller (PLC) (201) of control actuator (200) adopts is Siemens S7-400, and the model that frequency converter (202) adopts is Siemens MM430; Motor (301) adopts Siemens 1LG0090-4AA21; Load fan (302) and cooling fan (501) all adopt Siemens fan of frequency converter W2D210-EA10-11; Temperature sensor (601) adopts the Pt100 temperature sensor;

(2) equipment connection:

Use netting twine the Ethernet network interface of computing machine (1), computing machine (101) and computing machine (401) and PLC programmable logic controller (PLC) (201) to be connected to each network interface of same switch;

Be connected by the RS485 interface between PLC programmable logic controller (PLC) (201) and the frequency converter (202); Three of frequency converter (202) exchange output interfaces, and three by cable and motor (301) exchange input interface and are connected;

Motor (301) moment of torsion output main shaft is connected by the moment of torsion entering spindle of gearing with load fan (302), and motor (301) band dynamic load fan (302) rotates; The moment of torsion entering spindle of load fan (302) is fixed with again the code-disc of speed measuring coder (303) simultaneously, and fan (302) rotates and drives the code-disc rotation;

The thermal value of motor (301) can exert an influence to motor (301) rotating speed, and temperature sensor (601) is deployed in the upper collecting temperature that is used for of motor (301); The output of temperature sensor (601) adopts the three-wire system connection to be connected with the sensor input interface of PLC programmable logic controller (PLC) (201); PLC programmable logic controller (PLC) (201) receives information from temperature sensor (601) by sensor interface as input interface, and transmits to the computing machine (401) of on-line prediction device (400) by Ethernet interface;

Cooling fan (501) rotates the moving air that produces lowers the temperature to motor (301), also can exert an influence to load fan (302) simultaneously;

Control decision device (0) is used for parameter on-line tuning method step 1.0 parameter initialization that the realization time series forecasting is combined with Based Intelligent Control; On-line tuning device (100) has adopted on-line tuning algorithm (101), is used for implementation method step 1.1 and calculates the control parameter K _cAnd correction on-line tuning algorithm parameter w; Control actuator (200) has adopted control algolithm (201), is used for implementation method step 1.2 and calculates controlled quentity controlled variable u (t); Controlled device (300) is reached the control target for generation of control system output, is used for the collection of implementation method step 1.3 or computing system output y _System(t); On-line prediction device (400) has adopted on-line prediction algorithm (401), is used for performing step 1.4 and calculates the Yout (t) that predicts the outcome;

(3) course of work after system starts is as follows:

After system starts each installed according to 1.3.0 process 0 to 1.3.5 process 5 these six process operations:

1.3.0 process 0:

〇 device---control decision device (0) is realized step 1.0 parameter initialization of " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control ":

Above-mentioned steps 1.0 is given respectively initial value for each variable that relates in the computation process; Comprise following 12 sub-steps, namely step 1.0.1 is to step 1.0.12:

The current execution number of times of step 1.0.1 set algorithm t, execution total degree T:

Present embodiment belongs to algorithm and moves continuously situation about never stopping: establish current execution number of times initial value t=1; If carry out total degree T=0, this algorithm constantly circulates when T=0, can not stop;

Step 1.0.2 sets each control target (1000) value r constantly _In(t)=1; Control target implication in this embodiment is the rotating speed of the motor drag object that obtains of expectation, 1000 rev/mins of 1 expressions;

Step 1.0.3 sets and starts constantly T of prediction _Predict=51;

Step 1.0.4 sets the number K of the historical data of online prediction algorithm (401) needs _Predict=50;

It is L that step 1.0.5 sets prediction step _Predict=10;

Step 1.0.6 setup control as a result initial value is zero, i.e. y _System(t)=0; y _System(t) represent that in this embodiment motor drags the in real time true rotating speed of load;

Step 1.0.7 setup control prediction of result value is zero, i.e. y _Predict(t)=0; y _Predict(t) represent that in this embodiment motor drags the predicted value of load speed

It is zero that step 1.0.8 sets the initial value that predicts the outcome, i.e. y _Out(t)=0, y _Out(t) the on-line tuning device that is used for of expression predictive controller output drags load speed as the motor of inputting;

Step 1.0.9 sets each layer of wavelet neural network neuronal quantity M=3, Q=10, L=3;

Wherein M represents the input layer quantity of wavelet neural network, and namely wavelet neural network is made of M input layer; Q represents the hidden layer neuron quantity of wavelet neural network, and namely wavelet neural network is made of Q hidden layer neuron; L represents the output layer neuronal quantity of wavelet neural network, and namely wavelet neural network is made of L output layer neuron;

When wavelet neural network combines with the PID control method, can be with (1400) y that predicts the outcome _Out(t-1), control target (1000) r _In(t), departure error (t) is as the input of neural network, this moment M=3; Can also with constant 1 also as the input one of, this moment M=4, M=5,6 ... situation by that analogy; Can be with control parameter (1100) K of PID control method _c(t) be scale parameter K _p(t), integral parameter K _i(t), differential parameter K _d(t) as the output of neural network, this moment L=3; Can also with constant 1 also as output one of, this moment L=4, L=5,6 ... situation by that analogy; Hidden layer node is counted certain value that Q can be in being made as 8-12, establishes Q=10 herein;

Step 1.0.10 sets input layer, hidden layer, each the neuronic input variable of output layer in the computation process t=1 time $\begin{array}{cc}{\mathrm{net}}_i^{\left(1\right)}\left(t\right)& (i=\mathrm{1,2}...,M)\end{array}$ $\begin{array}{cc}{\mathrm{net}}_j^{\left(2\right)}\left(t\right)& (\mathrm{1,2}...,M)\end{array}$ $\begin{array}{cc}{\mathrm{net}}_k^{\left(3\right)}\left(t\right)& (k=\mathrm{1,2},...,L),\end{array}$ The connection weight value matrix of input layer and hidden layer

Each neuronic connection weight value matrix of hidden layer output layer

The activation functions change of scale parameter matrix a of hidden layer neuron _j(t) and b _j(t) each element in is the at random decimal in 0 to 1 open interval;

Ground floor---the input layer of the upper right footmark of above-mentioned variable " (1) " expression neural network, hidden layer represents with " (2) ", output layer represents with " (3) ", bottom right footmark " i " expression input layer sequence number, " j " expression hidden layer neuron sequence number, " k " expression output layer neuron sequence number;

The learning rate η of step 1.0.11 setting neural network=0.2; The inertial coefficient α of setting neural network=0.05;

The initial value of step 1.0.12 setup control amount u (t) and margin of error error (t) and time sequential value are zero: u (t)=0, u (t-1)=0, u (t-2)=0, error (t)=0, error (t-1)=0, error (t-2)=0;

1.3.1 process 1:

First device---on-line tuning device (100) is realized the step 1.1 calculating control parameter K of " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " _cAnd correction on-line tuning algorithm parameter w:

Above-mentioned steps 1.1 is according to control target (1000) r _In(t) and (1400) y that predicts the outcome _Our(t-1), adopt on-line tuning algorithm (101), calculate control parameter (1100) K _cAnd revise on-line tuning algorithm parameter w (t+1) (t); On-line tuning algorithm in the parameter on-line tuning method that the employing time series forecasting is combined with Based Intelligent Control in the present embodiment:

Step 1.1.1 calculates control parameter K c:

Step 1.1.1.1 calculates the input of input layer

The input of input layer

Calculating referring to formula 46:

$\left\{\begin{array}{c}{\mathrm{net}}_1^{\left(1\right)}\left(t\right)=r_{\mathrm{in}}\left(t\right)\\ {\mathrm{net}}_2^{\left(1\right)}\left(t\right)=y_{\mathrm{out}}(t-1)\\ {\mathrm{net}}_3^{\left(1\right)}\left(t\right)=\mathrm{error}\left(t\right)\end{array}\right.$ Formula 46

Variable wherein 3 neuronic inputs of expression input layer; 3 neuronic inputs of input layer are respectively (1400) y that predicts the outcome _Our(t-1), control target (1000) r _In(t), departure error (t), the calculating of error (t) is referring to formula 47:

Error (t)=y _Out(t-1)-r _In(t) formula 47

Step 1.1.1.2 calculates the output of input layer

The output of neural network input layer

Calculating referring to formula 48:

$O_i^{\left(1\right)}\left(t\right)={\mathrm{net}}_i^{\left(1\right)}\left(t\right),$ (i=1,2,3) formula 48

Wherein, variable is used respectively in input layer, hidden layer, each neuronic output of output layer

$(i=\mathrm{1,2},...,M),$ $\begin{array}{cc}{O}_j^{\left(2\right)}\left(t\right)& (j=\mathrm{1,2},...,Q),\end{array}$ $\begin{array}{cc}{O}_k^{\left(3\right)}\left(t\right)& (k=\mathrm{1,2},...,L)\end{array}$ Expression;

Step 1.1.1.3 calculates the input of hidden layer

The input of neural network hidden layer

Calculating referring to formula 49:

${\mathrm{net}}_j^{\left(2\right)}\left(t\right)=\underset{i-1}{\overset{M}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)\&CenterDot;O_i^{\left(1\right)}\left(t\right),$ (j=1,2 ..., Q) formula 49

Wherein

Weighting weights between expression i neuron of input layer and j neuron of hidden layer;

Step 1.1.1.4 calculates the output of hidden layer

The output of neural network hidden layer

Calculating referring to formula 50:

$O_j^{\left(2\right)}\left(t\right)={\mathrm{\&psi;}}_{a,b}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}-b_j\left(t\right)}{a_j\left(t\right)}\right),$ (j=1,2 ..., Q) formula 50

ψ wherein _{A, b}(t) be wavelet function, i.e. the activation functions of hidden layer neuron; a _j(t), b _j(t) be wavelet function change of scale parameter, i.e. the activation functions change of scale parameter of hidden layer neuron; Get the wavelet function ψ that satisfies the framework condition _{A,, b}The computing method of (t), seeing are referring to formula 51:

${\mathrm{\&psi;}}_{a,b}\left(t\right)=\cos \left(1.75t\right)\&CenterDot;e^{-\frac{t^2}{2}}$ Formula 51

ψ _{A,, b}The computing method of a mediation number (t) are referring to formula 52:

${\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(t\right)=-1.75\&CenterDot;\sin \left(1.75t\right)\&CenterDot;e^{-\frac{t^2}{2}}-t\&CenterDot;\cos \left(1.75t\right)\&CenterDot;e^{-\frac{t^2}{2}}$ Formula 52

Wherein sin (x), cos (x) represent respectively sine function and the cosine function value of x;

Step 1.1.1.5 calculates the input of output layer

The input of neural network output layer

Calculating referring to formula 53:

${\mathrm{net}}_k^{\left(3\right)}\left(t\right)=\underset{j=1}{\overset{Q}{\mathrm{\&Sigma;}}}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;O_j^{\left(2\right)}\left(t\right),$ (k=1,2 ..., L) formula 53

Wherein

Weighting weights between expression j neuron of hidden layer and k neuron of output layer;

Step 1.1.1.6 calculates the output of output layer

The output of neural network output layer

Calculating referring to formula 54:

$O_k^{\left(3\right)}\left(t\right)=g\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right),$ (k=1,2 ..., L) formula 54

Wherein g (x) gets non-negative Sigmoid function for the neuronic activation functions of output layer, and the calculating of g (x) is referring to formula 55:

$g\left(x\right)==\frac{e^x}{e^x+e^{-x}}$ Formula 55

The calculating of its first order derivative of g (x) is referring to formula 56:

$g^{\&prime;}\left(x\right)==\frac{2}{{(e^x+e^{-x})}^2}$ Formula 56

Step 1.1.1.7 calculates control parameter (1100) K _p(t), K _i(t), K _d(t):

Control parameter (1100) K _p(t), K _i(t), K _d(t) calculating is referring to formula 57:

$\left\{\begin{array}{c}{K}_p\left(t\right)=O_1^{\left(3\right)}\left(t\right)\\ K_i\left(t\right)=O_2^{\left(3\right)}\left(t\right)\\ K_d\left(t\right)=O_3^{\left(3\right)}\left(t\right)\end{array}\right.$ Formula 57

Wherein, K _p(t), K _i(t), K _d(t) be control parameter (1100) K that the neuronic input of output layer corresponds to respectively classical proportional integral derivative control method _c(t);

Step 1.1.1.8 controls parameter K _c(t) calculate end;

Step 1.1.2 revises on-line tuning algorithm parameter w (t):

On-line tuning algorithm parameter w (t) in the on-line tuning algorithm in the online PID setting method of wavelet neural network comprises following four parameters: the weighting weights between i neuron of input layer and j neuron of hidden layer

Weighting weights between j neuron of hidden layer and k neuron of output layer

Wavelet function change of scale parameter a _j(t), b _j(t); Namely

a _j(t), b _j(t) };

The step 1.1.2.1 error of calculation and performance index:

The calculating of error e rror (t) is referring to formula 58:

Error (t)=r _In(t)-y _Our(t-1) formula 58

The calculating of performance index E (t) is referring to formula 59:

$E\left(t\right)=\frac{1}{2}{\mathrm{error}}^2\left(t\right)=\frac{1}{2}{(r_{\mathrm{in}}\left(t\right)-y_{\mathrm{out}}(t-1))}^2$ Formula 59

Step 1.1.2.2 adopts gradient descent method to the weights coefficient between neural network hidden layer and the output layer

Adjust:

Weights coefficient between hidden layer and the output layer Correction

Calculating referring to formula 60:

$\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)}$ Formula 60

In the formula 60,

Computing formula referring to formula 61:

Formula 61

In the formula 61,

Calculating referring to formula 62:

Formula 62

In the formula 61, because

The unknown is with approximate sign function

Replace, Calculating referring to formula 63;

Formula 63

The sign function value of sgn (x) expression x wherein, if i.e. x＞0 then sgn (x)=1, if x=0 then sgn (x) if=0 x＜0 then sgn (x)=-1;

In the formula 61, Calculating referring to formula 64:

Formula 64

In the formula 61, Calculating referring to formula 65:

Formula 65

In the formula 61, Calculating referring to formula 66:

Formula 66

In sum, Computing formula can be expressed as formula 67:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)}=\mathrm{error}\left(t\right)\&CenterDot;\mathrm{sgn}\left(\frac{\&PartialD;y\left(t\right)}{\&PartialD;\mathrm{\&Delta;u}\left(t\right)}\right)\&CenterDot;\frac{\&PartialD;\mathrm{\&Delta;u}\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;g^{\&prime;}\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right)\&CenterDot;O_j^{\left(2\right)}\left(t\right)$ Formula 67

In sum,

Computing formula can be expressed as formula 68:

$\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{w}_{\mathrm{jk}}^{\left(3\right)}(t-1)\&PlusMinus;\mathrm{\&eta;}\&CenterDot;{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;O_j^{\left(2\right)}\left(t\right)$ Formula 68

Wherein, Calculating referring to formula 69:

${\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}$ Formula 69

$=\mathrm{error}\left(t\right)\&CenterDot;\mathrm{sgn}\left(\frac{\&PartialD;y\left(t\right)}{\&PartialD;u\left(t\right)}\right)\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;g^{\&prime;}\left({\mathrm{net}}_k^{\left(3\right)}\left(t\right)\right)$ (k＝1，2，3)

Step 1.1.2.3 adopts gradient descent method to the weights coefficient between neural network input layer and the hidden layer

Adjust:

Calculate the weights coefficient between input layer and the hidden layer

Modified value See formula 70:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{w}_{\mathrm{ij}}^{\left(2\right)}(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}$ Formula 70

In the formula 70,

Calculating referring to formula 71:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{\&PartialD;O_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_k^{\left(3\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;O_j^{\left(2\right)}\left(t\right)}$

$\&CenterDot;\frac{\&PartialD;O_j^{\left(2\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}$ Formula 71

$={\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;O_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;O_j^{\left(2\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}$

In the formula 71,

Calculating referring to formula 72:

$\frac{\&PartialD;{\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{\&PartialD;O_j^{\left(2\right)}\left(t\right)}=w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)$ Formula 72

In the formula 71,

Calculating referring to formula 73:

$\frac{\&PartialD;O_j^{\left(2\right)}\left(t\right)}{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}={\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;\frac{1}{a_j\left(t\right)}$ Formula 73

In the formula 71,

Calculating referring to formula 74:

$\frac{\&PartialD;{\mathrm{net}}_j^{\left(2\right)}\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}=O_i^{\left(1\right)}\left(t\right)$ Formula 74

In sum,

Computing formula can be expressed as formula 75:

$\frac{\&PartialD;E\left(t\right)}{\&PartialD;w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)}=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;{\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;\frac{1}{a_j\left(t\right)}\&CenterDot;O_i^{\left(1\right)}\left(t\right)$ Formula 75

In the formula 75

In formula 69, provide;

Step 1.1.2.3 calculates the activation functions change of scale parameter a of hidden layer neuron _j(t) modified value Δ a _j(t):

The activation functions change of scale parameter a of hidden layer neuron _j(t) modified value Δ a _j(t) computing formula can be expressed as formula 76:

$\mathrm{\&Delta;}{a}_j\left(t\right)=\mathrm{\&alpha;}\&CenterDot;\mathrm{\&Delta;}{a}_j(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{\&PartialD;a_j\left(t\right)}$ Formula 76

In the formula 76,

Calculating referring to formula 77:

$\frac{\&PartialD;E\left(t\right)}{{\&PartialD;a}_j\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;a}_j\left(t\right)}$ Formula 77

In the formula 77, $\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}$ Calculating referring to formula 69:

In the formula 77,

Calculating referring to formula 72:

In the formula 77,

Calculating referring to formula 78:

$\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;a}_j\left(t\right)}={\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;(-\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j^2\left(t\right)})$ Formula 78

In sum,

Computing formula can be expressed as formula 79:

$\frac{\&PartialD;E\left(t\right)}{{\&PartialD;a}_j\left(t\right)}=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)\&CenterDot;{\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)\&CenterDot;\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j^2\left(t\right)}\right)$ Formula 79

Step 1.1.2.4 calculates the activation functions change of scale parameter b of hidden layer neuron _j(t) modified value Δ b _j(t):

The activation functions change of scale parameter b of hidden layer neuron _j(t) modified value Δ b _j(t) computing formula can be expressed as formula 80:

${\mathrm{\&Delta;b}}_j\left(t\right)=\mathrm{\&alpha;}\&CenterDot;{\mathrm{\&Delta;b}}_j(t-1)-\mathrm{\&eta;}\&CenterDot;\frac{\&PartialD;E\left(t\right)}{{\&PartialD;b}_j\left(t\right)}$ Formula 80

In the formula 80,

Calculating referring to formula 81:

$\frac{\&PartialD;E\left(t\right)}{{\&PartialD;b}_j\left(t\right)}=\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;b}_j\left(t\right)}$ Formula 81

In the formula 80, $\frac{\&PartialD;E\left(t\right)}{\&PartialD;\mathrm{yout}\left(t\right)}\&CenterDot;\frac{\&PartialD;\mathrm{yout}\left(t\right)}{\&PartialD;u\left(t\right)}\&CenterDot;\frac{\&PartialD;u\left(t\right)}{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}\&CenterDot;\frac{{\&PartialD;O}_k^{\left(3\right)}\left(t\right)}{{\&PartialD;\mathrm{net}}_k^{\left(3\right)}\left(t\right)}$ Calculating referring to formula 69:

In the formula 80,

Calculating referring to formula 72:

In the formula 80,

Calculating referring to formula 82:

$\frac{{\&PartialD;O}_j^{\left(2\right)}\left(t\right)}{{\&PartialD;b}_j\left(t\right)}={\mathrm{\&psi;}}_{a,b}^{\&prime;}\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)(-\frac{1}{a_j\left(t\right)})$ Formula 82

In sum,

Computing formula can be expressed as formula 83:

$\frac{\&PartialD;E\left(t\right)}{{\&PartialD;b}_j\left(t\right)}=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_k^{\left(3\right)}\left(t\right)\&CenterDot;w_{\mathrm{jk}}^{\left(3\right)}\left(t\right){\mathrm{\&psi;}}_{a,b}^{\&prime;}\&CenterDot;\left(\frac{{\mathrm{net}}_j^{\left(2\right)}\left(t\right)-b_j\left(t\right)}{a_j\left(t\right)}\right)(-\frac{1}{a_j\left(t\right)})$ Formula 83

Step 1.1.2.5 revises on-line tuning algorithm parameter w (t) and arrives w (t+1):

The on-line tuning algorithm parameter that current the t time calculating uses is:

a _j(t), b _j(t) }; Revised result be calculate for the t+1 time the on-line tuning algorithm parameter w (t+1) that will use= $\{w_{\mathrm{ij}}^{\left(2\right)}(t+1),$ $w_{\mathrm{jk}}^{\left(3\right)}(t+1),$ a _j(t+1)，b _j(t+1)}

Weighting weights between j neuron of hidden layer and k neuron of output layer

Corrected Calculation referring to formula 84:

$w_{\mathrm{jk}}^{\left(3\right)}(t+1)=w_{\mathrm{jk}}^{\left(3\right)}\left(t\right)+{\mathrm{\&Delta;w}}_{\mathrm{jk}}^{\left(3\right)}\left(t\right)$ Formula 84

Weighting weights between i neuron of input layer and j neuron of hidden layer

Corrected Calculation referring to formula 85:

$w_{\mathrm{ij}}^{\left(2\right)}(t+1)=w_{\mathrm{ij}}^{\left(2\right)}\left(t\right)+{\mathrm{\&Delta;w}}_{\mathrm{ij}}^{\left(2\right)}\left(t\right)$ Formula 85

The activation functions change of scale parameter a of hidden layer neuron _j(t) corrected Calculation is referring to formula 86:

a _j(t+1)=a _j(t)+Δ a _j(t) formula 86

The activation functions change of scale parameter b of hidden layer neuron _j(t) corrected Calculation is referring to formula 87:

b _j(t+1)=b _j(t)+Δ b _j(t) formula 87

1.3.2 process 2:

The second device---control actuator (200) is realized " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.2 calculating controlled quentity controlled variable u (t);

Above-mentioned steps 1.2 adopts control algolithm (201), calculates controlled quentity controlled variable (1200) u (t); The input of algorithm is control parameter (1100) K _c(t), controlled quentity controlled variable (1200) u (t) and error e rror (t) time sequential value; The output of algorithm is controlled quentity controlled variable (1200) u (t);

Control algolithm (201) adopts the increment type PID control algolithm, and when adopting the increment type PID control algolithm, the calculating of u (t) is referring to formula 88:

U (t)=u (t-1)+K _p(t) (error (t)-error (t-1))+K _i(t) error (t) formula 88

+K _d(t)·(error(t)-2error(t-1)+error(t-2))

1.3.3 process 3:

The 3rd device---controlled device (300) realizes that " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.3 gathers or computing system output y _System(t):

Above-mentioned steps 1.3 controlled result (1300) y _System(t), because this embodiment is true motor drive control system, therefore control as a result y _System(t) do not obtain by Computer Simulation, but obtain by the load speed measurement code-disc sampling that motor is dragged;

In addition, for calculating control result (1300) y in the computer emulation method of mentioning in the claim 1 _System(t), provide an embodiment herein:

When the mathematical model of controlled device is described by formula 89

$\left(s\right)=\frac{523500}{s^3+{87.35s}^2+10470s}$ Formula 89

This moment y _System(t) calculating is referring to formula 90:

y _system(t)＝-den(2)·y _system(t-1)-den(3)·y _system(t-2)-den(4)

Y _System(t-3)+num (2) y _System(t-1)+num (3) formula 90

·y _system(t-2)+num(4)·y _system(t-3)

Wherein conversion obtains through z by formula 89 for den (2), den (3), den (4), num (2), num (3), num (4); Above-mentioned z conversion realizes by c2d () function and tfdata () function in Matlab, realizes that the source program of this algorithm is as follows:

sys=tf(523500,[1,87.35,10470,0]);

dsys=c2d(sys,ts,'tustin');

[num,den]=tfdata(dsys,'v');

1.3.4 process 4:

The 4th device---on-line prediction device (400) realization " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.4 is calculated the Yout (t) that predicts the outcome:

Above-mentioned steps 1.4 adopts on-line prediction algorithm (401), calculates (1400) yout (t) that predicts the outcome;

Can to adopt the vector time series Forecasting Methodology be the VARMA method to on-line prediction algorithm (401) in the present embodiment; When adopting the VARMA algorithm, the span of nAR and nMA parameter is the positive integer in 5 to 10 closed intervals; The VARMA prediction algorithm is realized by vgxset (), vgxvarx (), three functions of vgxpred () in Matlab, below provides the example that a method is implemented:

VARMA algorithmic function source program is as follows:

functionypredict=myarma5in2ar2ma(Y,PredictLength)

Ylog=diff(log(Y));

S=cei1(0.1*size(Ylog,1));

Ypre1=Ylog(1:S,:);

Yest1=Ylog((S+1):end,:);

VAR2MA2=vgxset('n',5,'nAR',2,'nMA',2,'Series',{'yout','u','kp','ki','kd'});

[EstSpec1,EstStdErrors1]=vgxvarx(VAR2MA2,Yest1,[],Ypre1,'CovarType','Diagonal','IgnoreMA','yes');

[ypredict,FYCov1]=vgxpred(EstSpec1,PredictLength,[],Yest1);

ypredict=[log(Y(end,:));ypredict];

ypredict=exp(cumsum(ypredict));

Judge at first whether current t arrives to start to predict it is to start constantly T of prediction constantly _Predict:

If current t is constantly less than starting constantly T of prediction _Predict, then do not start prediction, control result (1300) y _System(t) directly serve as (1400) y that predicts the outcome _Our(t), i.e. y _Our(t)=y _System(t);

If current t is constantly more than or equal to starting constantly T of prediction _Predict, then start prediction, adopt the on-line prediction algorithm to calculate.In the present embodiment, ignore the impact of control disturbance source (500) and prediction disturbing source (600), namely think control disturbance amount (1500) r _Oef(t) and prediction disturbance quantity (1600) r _Orf(t) be constant 0;

The input of on-line prediction algorithm (401) is respectively: the time series K of control parameter (1100) _c(t), K _c(t-1), K _c(t-2) ..., K _c(t-K _Predict), time series u (t), the u (t-1) of controlled quentity controlled variable (1200), u (t-2) ..., u (t-K _Predict), control the result (1300) time series y _System(t), y _System(t-1), y _System(t-2) ..., y _System(t-K _Predict);

To control prediction of result value (1401) y _Predict(t) as predicting the outcome (1400) y _Out(t), i.e. y _Out(t)=y _Predict(t);

In sum, (1400) y predicts the outcome _Out(t) value is seen formula 91:

$y_{\mathrm{out}}\left(t\right)=\left\{\begin{array}{ccc}{y}_{\mathrm{system}}\left(t\right)& ,& t<T_{\mathrm{predict}}\\ y_{\mathrm{predict}}\left(t\right)& ,& t\&GreaterEqual;T_{\mathrm{predict}}\end{array}\right.$ Formula 91

1.3.5 process 5:

〇 device---control decision device (0) realizes according to parameters, state and the output of current system whether " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.5 evaluation algorithm finishes:

Above-mentioned steps 1.5 has two case branch to consist of:

Algorithm is unclosed situation still: if T=0, if perhaps T ≠ 0 and t＜T, then t=t+1 jumps to step 1.1, namely reenters 1.3.1 process 1 to 1.3.5 process 5, has so far finished the t time calculating;

The situation that algorithm finishes: if T ≠ 0 and t 〉=T have so far finished whole T time and calculated.

Claims (6)

1. the time series forecasting parameter on-line tuning method of being combined with Based Intelligent Control is characterized in that: comprise following 5 steps, namely step 1.0 is to step 1.5:

Step 1.0 parameter initialization:

This step is that each variable that relates in the computation process is given respectively initial value; Comprise following 12 sub-steps, namely step 1.0.1 is to step 1.0.12:

1.0.1 the current execution number of times of set algorithm t, execution total degree T:

When algorithm moves the situation of limited number of time continuously: establish current execution number of times initial value t=1; The value principle of carrying out and adjust total degree T is: operation 0.8T the rear control result of total degree of assurance system (1300) can be stabilized near the control target (1000); Wherein can stable implication be: after the execution number of times reaches 0.8T, control result (1300) y _System(t) with control target (1000) r _In(t) error between is in ± 3%, referring to formula 1:

Formula 1

The method to set up of carrying out and adjust total degree T is:

Carry out total degree T=5000 1.0.1.1 establish;

1.0.1.2 operational system, and carry out T this method Overall Steps, adjust according to operation result again;

If 1.0.1.3 the operation result display algorithm is carried out after 0.8T time, control result (1300) can't be stable at control target (1000), then T is increased 10%, and jumps to step 1.0.1.2;

If 1.0.1.4 the operation result display algorithm namely reaches before the 0.5T control result (1300) y when carrying out number of times after control result (1300) being converged on control target (1000) for a long time _System(t) with control target (1000) r _In(t) error between then reduces 10% with T in ± 3%, and jumps to step 1.0.1.2;

The result of above-mentioned steps is: operation 0.8T the rear control result of total degree of system (1300) can be stabilized near the control target (1000), and the so far work that arranges of T is finished;

When algorithm moves situation about never stopping continuously: establish current execution number of times initial value t=1; If carry out total degree T=0, this algorithm constantly circulates when T=0, and algorithm can not stop;

1.0.2 set each control target (1000) value r constantly _In(t), r _In(t) can be a constant r _In, also can be the function take t as independent variable;

Start constantly T of prediction 1.0.3 set _Pteaict, T _PredictBe positive integer, according to engineering experience, its span is referenced as: 5%T≤T _Predict≤ 10%T;

1.0.4 set the number K of the historical data of online prediction algorithm (401) needs _Predict, K _PredictBe positive integer, its value determines that by the test of hypothesis statistical indicator of on-line prediction algorithm according to engineering experience, its span is referenced as: 10≤K _Preaict≤ 5%T;

1.0.5 the setting prediction step is L _Predict, L _PredictBe positive integer, according to engineering experience, its span is referenced as: 1≤L _Predict≤ 10%T;

1.0.6 setup control as a result initial value is zero, i.e. y _System(t)=0;

1.0.7 setup control prediction of result value is zero, i.e. y _Predict(t)=0;

Initial value is zero 1.0.8 setting predicts the outcome, i.e. y _Our(t)=0;

1.0.9 set each layer of wavelet neural network neuronal quantity M, Q, L, span generally is M≤10, M≤Q≤5M, in the positive integer interval in L≤M interval;

Wherein M represents the input layer quantity of wavelet neural network, and namely wavelet neural network is made of M input layer; Q represents the hidden layer neuron quantity of wavelet neural network, and namely wavelet neural network is made of Q hidden layer neuron; L represents the output layer neuronal quantity of wavelet neural network, and namely wavelet neural network is made of L output layer neuron;

1.0.10 set input layer, hidden layer, each the neuronic input variable of output layer in the computation process t=1 time

The connection weight value matrix of input layer and hidden layer

Each neuronic connection weight value matrix of hidden layer output layer The activation functions change of scale parameter matrix a of hidden layer neuron _j(t) and b _j(t) each element in is the at random decimal in 0 to 1 open interval;

Ground floor---the input layer of the upper right footmark of above-mentioned variable " (1) " expression neural network, hidden layer represents with " (2) ", output layer represents with " (3) ", bottom right footmark " i " expression input layer sequence number, " j " expression hidden layer neuron sequence number, " k " expression output layer neuron sequence number;

1.0.11 set learning rate η and the inertial coefficient α of neural network: wherein the span of η is the decimal in 0.01 to 0.7 closed interval; The span of α is the decimal in the 0.01-0.2 closed interval;

1.0.12 the initial value of setup control amount u (t) and margin of error error (t) and time sequential value are zero: u (t)=0, u (t-1)=0, u (t-2)=0, error (t)=0, error (t-1)=0, error (t-2)=0;

Step 1.1 is calculated the control parameter K _cAnd correction on-line tuning algorithm parameter w:

This step is according to control target (1000) r _In(t) and (1400) y that predicts the outcome _Our(t-1), adopt on-line tuning algorithm (101), calculate control parameter (1100) K _cAnd revise on-line tuning algorithm parameter w (t+1) (t); On-line tuning algorithm in the parameter on-line tuning method that on-line tuning algorithm (101) employing time series forecasting is combined with Based Intelligent Control perhaps adopts control parameter on-line tuning method non-time-based sequence prediction or non-neural net method on-line tuning control parameter;

Wherein adopt the on-line tuning algorithm steps in the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control as follows,

1.1.1 calculate the control parameter K _c:

1.1.1.1 calculate the input of input layer

The input of input layer

Calculating referring to formula 2:

Formula 2

Variable wherein

3 neuronic inputs of expression input layer; 3 neuronic inputs of input layer are respectively (1400) y that predicts the outcome _Out(t-1), control target (1000) r _In(t), departure error (t), the calculating of error (t) is referring to formula 3:

Error (t)=y _Out(t-1)-r _In(t) formula 3

1.1.1.2 calculate the output of input layer

The output of neural network input layer Calculating referring to formula 4:

(i=1,2,3) formula 4

Wherein, variable is used respectively in input layer, hidden layer, each neuronic output of output layer

Expression;

1.1.1.3 calculate the input of hidden layer

The input of neural network hidden layer

Calculating referring to formula 5:

(j=1,2 ..., Q) formula 5

Wherein

Weighting weights between expression i neuron of input layer and j neuron of hidden layer;

1.1.1.4 calculate the output of hidden layer

The output of neural network hidden layer Calculating referring to formula 6:

(j=1,2 ..., Q) formula 6

ψ wherein _{A, b}(t) be wavelet function, i.e. the activation functions of hidden layer neuron; a _j(t), b _j(t) be wavelet function change of scale parameter, i.e. the activation functions change of scale parameter of hidden layer neuron; Get the wavelet function ψ that satisfies the framework condition _{A, b}The computing method of (t), seeing are referring to formula 7:

Formula 7

ψ _{A, b}The computing method of a mediation number (t) are referring to formula 8:

Formula 8

S wherein _In(x), cos (x) represents respectively sine function and the cosine function value of x;

1.1.1.5 calculate the input of output layer

The input of neural network output layer Calculating referring to formula 9:

(k=1,2 ..., L) formula 9

Wherein Weighting weights between expression j neuron of hidden layer and k neuron of output layer;

1.1.1.6 calculate the output of output layer

The output of neural network output layer

Calculating referring to formula 10:

(k=1,2 ..., L) formula 10

Wherein g (x) gets non-negative Sigmoid function for the neuronic activation functions of output layer, and the calculating of g (x) is referring to formula 11:

Formula 11

The calculating of its first order derivative of g (x) is referring to formula 12:

Formula 12

1.1.1.7 calculate control parameter (1100) K _p(t), K _i(t), K _d(t):

Control parameter (1100) K _p(t), K _i(t), K _d(t) calculating is referring to formula 13:

Formula 13

Wherein, K _p(t), K _i(t), K _d(t) be control parameter (1100) K that the neuronic input of output layer corresponds to respectively classical proportional integral derivative control method _c(t);

1.1.1.8 control parameter K _c(t) calculate end;

1.1.2 revise on-line tuning algorithm parameter w (t):

On-line tuning algorithm parameter w (t) in the on-line tuning algorithm in the online PID setting method of wavelet neural network comprises following four parameters: the weighting weights between i neuron of input layer and j neuron of hidden layer

Weighting weights between j neuron of hidden layer and k neuron of output layer

Wavelet function change of scale parameter aj (t), b _j(t); Namely

a _j(t), b _j(t) };

The step 1.1.2.1 error of calculation and performance index:

The calculating of error e rror (t) is referring to formula 14:

Error (t)=y _Out(t-1)-r _In(t) formula 14

The calculating of performance index E (t) is referring to formula 15:

Formula 15

Step 1.1.2.2 adopts gradient descent method to the weights coefficient between neural network hidden layer and the output layer

Adjust:

Weights coefficient between hidden layer and the output layer

Correction

Calculating referring to formula 16:

Formula 16

In the formula 16,

Computing formula referring to formula 17:

Formula 17

In the formula 17,

Calculating referring to formula 18:

Formula 18

In the formula 17, because The unknown is with approximate sign function

Replace,

Calculating referring to formula 19;

Formula 19

The sign function value of sgn (x) expression x wherein, if i.e. x＞0 then sgn (x)=1, if x=0 then sgn (x) if=o x＜0 then sgn (x)=-1;

In the formula 17,

Calculating referring to formula 20:

Formula 20

In the formula 17,

Calculating referring to formula 21:

Formula 21

In the formula 17,

Calculating referring to formula 22:

Formula 22

In sum, Computing formula can be expressed as formula 23:

Formula 23

In sum,

Computing formula can be expressed as formula 24:

Formula 24

Wherein, Calculating referring to formula 25:

Formula 25

(k＝1，2，3)

1.1.2.3 adopt gradient descent method to the weights coefficient between neural network input layer and the hidden layer

Adjust:

Calculate the weights coefficient between input layer and the hidden layer

Modified value See formula 26:

Formula 26

In the formula 26,

Calculating referring to formula 27:

Formula 27

In the formula 27,

Calculating referring to formula 28:

Formula 28

In the formula 27,

Calculating referring to formula 29:

Formula 29

In the formula 27, Calculating referring to formula 30:

Formula 30

In sum,

Computing formula can be expressed as formula 31:

Formula 31

In the formula 31

In formula 25, provide;

1.1.2.3 calculate the activation functions change of scale parameter a of hidden layer neuron _j(t) modified value Δ a _j(t):

The activation functions change of scale parameter a of hidden layer neuron _j(t) modified value Δ a _j(t) computing formula can be expressed as formula 32:

Formula 32

In the formula 32,

Calculating referring to formula 33:

Formula 33

In the formula 33,

Calculating referring to formula 25:

In the formula 33,

Calculating referring to formula 28:

In the formula 33,

Calculating referring to formula 34:

Formula 34

In sum,

Computing formula can be expressed as formula 35:

Formula 35

1.1.2.4 calculate the activation functions change of scale parameter b of hidden layer neuron _j(t) modified value Δ b _j(t):

The activation functions change of scale parameter b of hidden layer neuron _j(t) modified value Δ b _j(t) computing formula can be expressed as formula 36:

Formula 36

In the formula 36, Calculating referring to formula 37:

Formula 37

In the formula 36,

Calculating referring to formula 25:

In the formula 36,

Calculating referring to formula 28:

In the formula 36,

Calculating referring to formula 38:

Formula 38

In sum,

Computing formula can be expressed as formula 39:

Formula 39

Arrive w (t+1) 1.1.2.5 revise on-line tuning algorithm parameter w (t):

The on-line tuning algorithm parameter that current the t time calculating uses is:

Aj (t), b _j(t) }; Revised result be calculate for the t+1 time the on-line tuning algorithm parameter w (t++1) that will use=

a _j(t+1), b _j(t+1) }

Weighting weights between j neuron of hidden layer and k neuron of output layer

Corrected Calculation referring to formula 40:

Formula 40

Weighting weights between i neuron of input layer and j neuron of hidden layer Corrected Calculation referring to formula 41:

Formula 41

The activation functions change of scale parameter a of hidden layer neuron _j(t) corrected Calculation is referring to formula 42:

a _j(t+1)=a _j(t)+Δ a _j(t) formula 42

The activation functions change of scale parameter b of hidden layer neuron _j(t) corrected Calculation is referring to formula 43:

b _j(t+1)=b _j(t)+Δ b _j(t) formula 43

Step 1.2 is calculated controlled quentity controlled variable u (t):

This step adopts control algolithm (201), calculates controlled quentity controlled variable (1200) u (t);

Step 1.3 gathers or computing system output y _System(t):

This step calculates control result (1300) y _System(t) or directly collect control result (1300) y _System(t);

Step 1.4 is calculated the y that predicts the outcome _Out(t):

This step adopts on-line prediction algorithm (401), calculates (1400) y that predicts the outcome _Our(t);

Whether step 1.5 evaluation algorithm finishes:

Algorithm is unclosed situation still: if T=0, if perhaps T ≠ 0 and t＜T, then t=t+1 jumps to step 1.1, has so far finished the t time calculating;

The situation that algorithm finishes: if T ≠ 0 and t 〉=T have so far finished whole T time and calculated.

2. the step 1.2 of the time series forecasting according to claim 1 parameter on-line tuning method of being combined with Based Intelligent Control is calculated controlled quentity controlled variable, it is characterized in that:

The input of algorithm is control parameter (1100) K _c(t), controlled quentity controlled variable (1200) u (t) and error e rror (t) time sequential value;

The output of algorithm is controlled quentity controlled variable (1200) u (t);

Control algolithm (201) can employing, positional PID control calculation, increment type PID control algolithm, FUZZY ALGORITHMS FOR CONTROL, expert's control algolithm;

When adopting the increment type PID control algolithm, the calculating of u (t) is referring to formula 44:

U (t)=u (t-1)+K _p(t) (error (t)-error (t-1))+K _i(t) error (t) formula 44

+K _d(t)·(error(t)-2error(t-1)+error(t-2))。

3. the step 1.3 of the time series forecasting according to claim 1 parameter on-line tuning method of being combined with Based Intelligent Control gathers or computing system output, it is characterized in that:

When this system was the situation of Computer Simulation application, controlled device (300) adopted the control system controlled device mathematical model of determining, the input of this mathematical model is controlled quentity controlled variable u (t), and the result who draws by Computer Simulation is as a result y of control _System(t); Above-mentioned Computer Simulation namely by controlled device being set up continuous transfer function model, and then adopts z conversion to carry out discretize to this continuous transport function, obtains according to u (t) and y _System(t) time sequential value calculates y _System(t) computing formula; The z conversion realizes by c2d () function and tfdata () function in Matlab; The specific implementation process sees also embodiment part in the patent specification;

When this system is the situation of true control system: therefore control as a result y _System(t) do not obtain by Computer Simulation, but obtain by true control system output is sampled.

4. the step 1.4 of the time series forecasting according to claim 1 parameter on-line tuning method of being combined with Based Intelligent Control is calculated and is predicted the outcome, and it is characterized in that:

Judge at first whether current t arrives to start to predict it is to start constantly T of prediction constantly _Predict:

If current t is constantly less than starting constantly T of prediction _Predict, then do not start prediction, control result (1300) y _System(t) directly serve as (1400) y that predicts the outcome _Out(t), i.e. y _Out(t)=y _System(t);

If current t is constantly more than or equal to starting constantly T of prediction _Predict, then start prediction, gather control disturbance amount (1500) r of control disturbance source (500) generation of current time _Oef(t) and prediction disturbance quantity (1600) r that produces of prediction disturbing source (600) _Orf(t);

The input of on-line prediction algorithm (401) is respectively: the time series K of control parameter (1100) _c(t), K _c(t-1), K _c(t-2) ..., K _c(t-K _Predict), time series u (t), the u (t-1) of controlled quentity controlled variable (1200), u (t-2) ..., u (t-K _Predict), control the result (1300) time series y _System(t), y _System(t-1), y _System(t-2) ..., y _System(t-K _Predict), the time series r of control disturbance amount (1500) _Oef(t), r _Oef(t-1), r _Oef(t-2) ..., r _Oef(t-K _Predict), the prediction disturbance quantity (1600) time series r _Orf(t), r _Otf(t-1), r _Orf(t-2) ..., r _Orf(t-K _Predict);

Wherein, control disturbance source (500) be except controlled quentity controlled variable (1200) to other devices in the controling environment of controlled device generation effect; Prediction disturbing source (600) be except control parameter (1100), controlled quentity controlled variable (1200), control the result (1300) to on-line prediction device (400) exert an influence control environment in other devices; Control disturbance amount (1500) is to react on controlled device (300) by what control disturbance source (500) produced with controlled quentity controlled variable (1200), and the input that its output is exerted an influence, in ideal conditions, and control disturbance amount (1500) r _Oef(t) can for constant 0, namely ignore; Wherein, prediction disturbance quantity (1600) be produced by prediction disturbing source (600) with control parameter (1100), controlled quentity controlled variable (1200), control result (1300) reacts on on-line prediction device (400), and the input that the output of on-line prediction algorithm (401) is exerted an influence, in the ideal case, namely can ignore in the situation of the disturbing influence of other devices in the control system internal and external environment, control disturbance source (500) and prediction disturbing source (600) can be ignored, namely there are not control disturbance source (500) and prediction disturbing source (600), at this moment control disturbance amount (1500) r _Oef(t) and prediction disturbance quantity (1600) r _Orf(t) be constant 0;

The output of on-line prediction algorithm (401) is control prediction of result value (1401) y _Predict(t);

Current t is constantly more than or equal to starting constantly T of prediction _PredictIn this case, will control prediction of result value (1401) y _Predict(t) as predicting the outcome (1400) y _Out(t), i.e. y _Our(t)=y _Predict(t);

In sum, (1400) y predicts the outcome _Out(t) value is seen formula 45:

Formula 45.

5. on-line prediction algorithm according to claim 4 (401) is characterized in that:

It is VARMA method, neural net prediction method, linear regression Forecasting Methodology, non-linear regression Forecasting Methodology, curvilinear function approximating method that on-line prediction algorithm (401) can adopt classical vector time series Forecasting Methodology, also can adopt satisfied input is that output is the prediction algorithm of control prediction of result value (1401) by time series control parameter (1100), time series controlled quentity controlled variable (1200), time series control result (1300), time series control disturbance amount (1500), time series forecasting disturbance quantity (1600);

When adopting the VARMA algorithm, the span of nAR and nMA parameter is the positive integer in 5 to 10 closed intervals; The VARMA prediction algorithm is realized by vgxset (), vgxvarx (), three functions of vgxpred () in Matlab; The specific implementation process sees also embodiment part 5 in the patent specification

6. adopted the system of the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control, it is characterized in that: by the 〇 device that---------------on-line prediction device (400), the 5th device---control disturbance source (500) and the 6th device---predict that (600) seven groups of devices of disturbing source form for controlled device (300), the 4th device for control actuator (200), the 3rd device for on-line tuning device (100), the second device for control decision device (0), first device;

1.1 respectively organize the inner formation of device and input, output interface situation:

1.1.0 〇 device---control decision device (0) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC programmable logic controller (PLC) or computing machine or server or portable terminal;

When the 〇 device---when control decision device (0) adopted computing machine (1), its output interface can be Ethernet interface, by netting twine and the network switching equipment and first device---the input interface of on-line tuning device (100) is connected;

1.1.1 first device---on-line tuning device (100) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC programmable logic controller (PLC) or computing machine or server;

When first device---when on-line tuning device (100) adopts computing machine (100), two input interface can be Ethernet interface, by netting twine and the network switching equipment and 〇 device---control decision device (0), the 4th device---output interface of on-line prediction device (400) is connected; Two output interface can be Ethernet interface, and the input interface of on-line prediction device (400) is connected---to control actuator (200), the 4th device---by netting twine and the network switching equipment and the second device; Above-mentioned Ethernet interface can multiplexing same Ethernet interface;

1.1.2 the second device---control actuator (200) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC programmable logic controller (PLC) or computing machine or server or PID controller cooperation frequency converter or driver;

When the second device---when control actuator (200) adopts PLC programmable logic controller (PLC) (201) and frequency converter (202), its input interface can be the Ethernet interface of PLC programmable logic controller (PLC) (201), by netting twine and the network switching equipment and first device---and the output interface of on-line tuning device (100) is connected; Its output interface can be that (202) three in frequency converter exchanges output interface, installs with the 3rd by cable---the input interface of controlled device (300) is connected;

The second device---the annexation of control actuator (200) interior arrangement is as follows: between PLC programmable logic controller (PLC) (201) and the frequency converter (202), can be connected by RS485 interface or Ethernet interface or industrial bus interface;

The second device---control actuator (200) is as retransmission unit, realization is from the 6th device, and------connection of on-line prediction device (400) is meticulous with the course of work to be: PLC programmable logic controller (PLC) (201) can receive from the 6th device as input interface by sensor interface---predict the information of disturbing source (600), and install to the 4th by Ethernet interface---, and on-line prediction device (400) is transmitted to predict that disturbing source (600) installs to the 4th;

1.1.3 the 3rd installs---controlled device (300) can be motor or temperature control device or voltage-controlled devices or electromagnetic field or production system or economic system;

When the 3rd device---when controlled device (300) adopts motor (301), load fan (302) and during speed measuring coder (303), one of them input interface can be that three of motor (301) exchange input interface, by cable and the second device---three of control actuator (200) exchange output interface and are connected; Wherein another input is the 5th device that load fan (302) is subject to---the disturbance of the air of the rapid flow of cooling fan (501) output of control disturbance source (500); The input interface of on-line prediction device (400) is connected the 3rd device---RS232 that controlled device (300) output interface can be by speed measuring coder (303) or Ethernet or industrial bus interface and the 4th install---;

The 3rd installs---and the annexation of controlled device (300) interior arrangement is as follows: motor (301) moment of torsion output main shaft is connected by the moment of torsion entering spindle of gearing with load fan (302), motor (301) band dynamic load fan (302) rotates, thereby consists of one group of output and the relation of inputting; The moment of torsion entering spindle of load fan (302) is fixed with again the code-disc of speed measuring coder (303) simultaneously, and fan (302) rotation drives code-disc and rotates, thereby consists of one group of output and the relation of inputting;

Can produce heat in motor (301) course of work, affect the serviceability of motor, and then the result of impact prediction; Therefore can produce heat in motor (301) course of work and can be used as the 4th device---one of input of on-line prediction device (400); This heat dissipation problem solves by cooling fan (501), cooling fan (501) also produces load fan (302) when solving heat dissipation problem and disturbs, and therefore the 5th installs---and the output of control disturbance source (500) also is one of input of load fan (302);

1.1.4 the 4th installs---on-line prediction device (400) can be embedded device or Single Chip Microcomputer (SCM) system or industrial computer or PLC or computing machine or server;

When the 4th device---when on-line prediction device (400) adopts computing machine (401), two input interface can be Ethernet interface, and the Ethernet output interface of the PLC programmable logic controller (PLC) (201) of---on-line tuning device (100), second device---control actuator (200) is connected by netting twine and the network switching equipment and first device; Be connected by RS232 or Ethernet or industrial bus interface between the speed measuring coder (303) of its another input interface and controlled device (300);

1.1.5 the 5th device---control disturbance source (500) can be the devices that affects the device of Air Flow or affect the device of humidity, the device that affects temperature, barrier, generation interference;

When the 5th device---when control disturbance source (500) adopt cooling fan (501), its fan rotates the moving air that produces motor (301) is lowered the temperature, also can exert an influence to load fan (302) simultaneously, namely as one of input of load fan (302);

1.1.6 the 6th device---prediction disturbing source (600) can be the device that affects the device of Air Flow or affect the device of humidity, the device that affects temperature, barrier, generation interference.

When thinking the 6th device---when one of prediction disturbing source (600) was the thermal value of motor (301), the input of temperature sensor (601) was the temperature of motor (301); The three-wire system connection is adopted in the output of temperature sensor (601) and the second device---the sensor input interface of the PLC programmable logic controller (PLC) (201) of control actuator (200) is connected;

1.2 respectively organize annexation and signal transitive relation between device:

Control decision device (0) is output as control target (1000) r of system _InThis control target (1000) is connected with the input of on-line tuning device (100), becomes one of input of on-line tuning device (100); Two of the input of on-line tuning device (100) is (1400) y that predict the outcome _OutOn-line controller calculates control parameter (1100) K according to control target (1000) and predict the outcome (1400) by on-line tuning algorithm (101) _c, this control parameter (1100) is the input of control actuator (200), is again one of input of on-line prediction device (400); Control actuator (200) calculates controlled quentity controlled variable (1200) u according to the control parameter (1100) of its input by control algolithm, and this controlled quentity controlled variable (1200) is one of input of controlled device (300), be again on-line prediction device (400) input two; Two of the input of controlled device (300) is control disturbance amount (1500) r that control disturbance source (500) produces _Oef, controlled device (300) produces control result (1300) y under the acting in conjunction of controlled quentity controlled variable (1200) and these two inputs of control disturbance amount (1500) _System, this control result (1300) be on-line prediction device (400) input three; Four of the input of on-line prediction device (400) is prediction disturbance quantity (1600) r that produced by prediction disturbing source (600) _Orf, on-line prediction device (400) is according to control parameter (1100), controlled quentity controlled variable (1200), control result (1300), prediction disturbance quantity (1600) r _OrfThe time series history value, calculate control prediction of result value (1401) y by on-line prediction algorithm (401) _PredictWith (1400) y that predicts the outcome _Out, this (1400) input as on-line tuning device (100) that predicts the outcome;

1.3 the course of work after system starts is as follows:

After system starts each installed according to 1.3.0 process 0 to 1.3.5 process 5 these six process operations:

1.3.0 process 0:

〇 device---control decision device (0) is realized step 1.0 parameter initialization of " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control ";

1.3.1 process 1:

First device---on-line tuning device (100) is realized the step 1.1 calculating control parameter K of " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " _cAnd correction on-line tuning algorithm parameter w;

1.3.2 process 2:

The second device---control actuator (200) is realized " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.2 calculating controlled quentity controlled variable u (t);

1.3.3 process 3:

The 3rd device---controlled device (300) realizes that " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.3 gathers or computing system output y _System(t);

1.3.4 process 4:

The 4th device---on-line prediction device (400) realization " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.4 is calculated the y that predicts the outcome _Out(t);

1.3.5 process 5:

〇 device---control decision device (0) realizes according to parameters, state and the output of current system whether " the parameter on-line tuning method that time series forecasting is combined with Based Intelligent Control " step 1.5 evaluation algorithm finishes.

Images