CN102353119B - Control method of VAV (variable air volume) air-conditioning system - Google Patents

Abstract

The invention provides a control scheme of a VAV (variable air volume) air-conditioning system. A neural network predictive control method is used for a tail end VAV-BOX and an air-conditioning unit, thus the hysteresis characteristic of a VAV system can be overcome, the control accuracy is improved, the resonance phenomena of an actuating mechanism can be greatly reduced, the energy-saving effect is improved by above 13%, and the control parameter tuning problem in the project is solved. A pressure independent cascade stage predictive control method is used for the tail end VAV-BOX, thus the control accuracy can also be improved. The air-conditioning unit is provided with four control loops and can automatically select a static pressure control or total air volume control policy with adjustable setting static pressure by using an all-condition integrated control technology, thus the predictive control on a fan can be realized; the primary air volume of all the tail end VAV-BOXes can be detected and the air supply temperature can be adjusted according to the operation condition, thus the problem of too low temperature in partial air-conditioning area in the lowest fresh air operation is solved; the cascade stage predictive control method is adopted for a fresh air ratio control loop, thus the accurate control on the fresh air ratio can be realized and the energy-saving level can be further improved.

Description

A kind of VAV Control Method of Vav Air Conditioning System

Technical field

The present invention relates to a kind of VAV Control Method of Vav Air Conditioning System, belong to civil buildings VAV air quantity variable air conditioner control technology field.

Background technology

VAV air quantity variable air conditioner control system mainly is based on the independent design in a plurality of loops at present, and all be to regulate with the PID control method basically, its robot control system(RCS) needs the commissioning engineer according to the on-the-spot pid parameter of setting of the experience of self, does not possess self-learning capability.Because the time constant of each robot control system(RCS) is different, very easily causes the robot control system(RCS) adaptivity poor, make temperature fluctuation greatly or " resonance " phenomenon of generation actuator.Simultaneously, owing to have certain coupled relation between a plurality of loops in the VAV air conditioning system with variable, even the debugging of single loop and operation are all out of question, when the co-ordination of all loops, whole system also is difficult for realizing stable control, is prone to " resonance " phenomenon of system.

In addition, the setting of ratio, integration, differential parameter has a great impact the performance of control system in the PID control method, because air conditioning system with variable forms complexity, equipment is numerous, various application scenario is also different to the requirement of parameter, need the commissioning engineer to set according to the experience scene of self, debugging has brought very large difficulty to engineering site.Simultaneously, owing to exist summer, winter and conditioning in Transition Season three kinds of different working conditions, the VAV air-conditioning system often needs could substantially satisfy more than 1 year client's control performance requirement at least in practical engineering application.

There is at present minority Design of Variable Air Volume System company to adopt fuzzy PID control method, in order to can in control procedure, automatically adjust pid parameter according to environmental change, improves the adaptive ability of control system.But, because the PID control method can't realize optimum control in essence, namely can't make certain performance indications reach optimum, therefore do not realize the purpose of further saving energy consumption.

In sum, at present the VAV air conditioning system with variable exists that control performance is relatively poor, debugging work load large and the energy-saving effect problem of good aspect not, and these problems affect are to the application of VAV air conditioning system with variable.

Summary of the invention

The present invention is directed to the existing this difficult point of multiple-input and multiple-output nonlinear system aspect control that has the large time delay characteristic of VAV air quantity variable air conditioner, the advantage of comprehensive neutral net, optimum control and PREDICTIVE CONTROL has proposed a kind of intelligence control method.The comprehensive Hamilton-Jacobi-Bellman of the method (HJB) and Eular-Lagrange (EL) optimized algorithm, utilize predicted roll to optimize the thought Training Multilayer Feedforward Neural Networks, become the optimization feedback solution of multiple-input and multiple-output nonlinear system when then it being found the solution as the optimizing feedback control device, can and take the Optimal Control Problem that solves nonlinear system in the moderate situation of memory block capacity in amount of calculation, utilize simultaneously the multi-step prediction rolling optimization to overcome various uncertainties and the complicated impact that changes.

A kind of VAV Control Method of Vav Air Conditioning System comprises following step:

The first step: utilize BP neural network air conditioning area temperature neural network prediction model, terminal air-valve forecast model, air-conditioning unit main air duct duct static pressure forecast model, main air duct air quantity forecast model, wind pushing temperature forecast model, new wind air-valve forecast model and Air Quality Forecast model

1) determines air conditioning area temperature prediction model, terminal air-valve forecast model, air-conditioning unit main air duct duct static pressure forecast model, main air duct air quantity forecast model, wind pushing temperature forecast model, new wind air-valve forecast model and Air Quality Forecast model structure

The input signal of air conditioning area temperature prediction model is outdoor intensity of solar radiation, outdoor temperature, CO2 concentration, indoor temperature, air quantity and terminal valve area, is output as next constantly indoor temperature;

The input signal of terminal air-valve forecast model is terminal valve area and duct static pressure (air-conditioning unit main air duct place), is output as next constantly terminal air quantity;

The input signal of main air duct duct static pressure forecast model is rotation speed of fan, duct static pressure, return air CO2 concentration, outdoor temperature and intensity of solar radiation, is output as next constantly duct static pressure;

The input signal of main air duct air quantity forecast model is rotation speed of fan and duct static pressure, is output as next constantly main air duct air quantity;

The input signal of wind pushing temperature forecast model is wind pushing temperature and water valve aperture, is output as next constantly wind pushing temperature;

The input signal of new wind air-valve forecast model is new wind valve area and duct static pressure (fresh air pipeline place), is output as next constantly resh air requirement;

The input signal of Air Quality Forecast model is new wind valve area and CO2 concentration, is output as next constantly CO2 concentration;

2) collecting sample data;

3) sample data is by formula carried out normalization in (1), (2):

$x_i=\frac{x_{\mathrm{di}}-x_{d\min }}{x_{d\max }-x_{d\min }}---\left(1\right)$

$y_{\mathrm{tl}}=\frac{y_{\mathrm{dl}}-y_{d\min }}{y_{d\max }-y_{d\min }}---\left(2\right)$

Wherein, x _iInput value for neutral net after the normalization; x _DiBe former input value; x _DminBe the minimum of a value in the former input value; x _DmaxBe the maximum in the former input value; y _TlDesired value for neutral net after the normalization; y _DlRepresent former desired value; y _DminRepresent the minimum of a value in the former desired value; y _DmaxBe the maximum in the former desired value;

4) each neural network prediction model is carried out off-line training.

Second step: VAV air conditioning terminal tandem PREDICTIVE CONTROL

1) determines terminal outer shroud predictive control function

Terminal outer shroud predictive control function is: $J_o\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_c-1}{\mathrm{\&Sigma;}}}{L}_o\left(T_o\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_c-1}{\mathrm{\&Sigma;}}}{\left(T_o\right[k]-T_{o,\mathrm{set}}[k\left]\right)}^2$

M wherein _cBe prediction time domain, t ₁The initial time in the prediction time domain, T _o[k] is the air conditioning area temperature in k sampling period, T _{O, set}[k] is the air conditioning area desired temperature in k sampling period, L _oIt is terminal outer shroud object function of k sampling period;

2) determine terminal interior ring predictive control function

The ring predictive control function is in terminal: $J_i\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_i-1}{\mathrm{\&Sigma;}}}{L}_i\left(V_i\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_i-1}{\mathrm{\&Sigma;}}}{\left(V_i\right[k]-V_{i,\mathrm{set}}[k\left]\right)}^2$

V wherein _i[k] is the air quantity in k sampling period, V _{I, set}[k] is the air quantity setting value in k sampling period; M _iBe prediction time domain, L _iIt is ring object function in k the sampling period end;

3) determine terminal interior ring neural network prediction controller and terminal outer shroud neural network prediction controller architecture

As input, terminal valve area is output to ring neural network prediction controller with air quantity setting value, duct static pressure (air-conditioning unit main air duct place) and pipeline air quantity in terminal;

The input parameter of terminal outer shroud neural network prediction controller comprises outdoor temperature, intensity of solar radiation, indoor temperature and air conditioning area desired temperature; Output parameter is the air quantity setting value;

4) ring neural network prediction controller in terminal and terminal outer shroud neural network prediction controller are carried out the online optimizing training.

The 3rd step: VAV air-conditioning unit PREDICTIVE CONTROL

1) determines static pressure control loop, VAV air-conditioning unit total blast volume control loop, VAV air-conditioning unit wind pushing temperature control loop and the VAV air-conditioning unit new wind ratio control loop predictive control function of VAV air-conditioning unit adjustable settings static pressure

The static pressure control loop predictive control function is:

$J_s\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_s-1}{\mathrm{\&Sigma;}}}{L}_s\left(P_s\right[k],U_{\mathrm{fan}}[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_s-1}{\mathrm{\&Sigma;}}}\{{\left(P_s\right[k]-P_{s,\mathrm{set}}[k\left]\right)}^2+U_{\mathrm{fan}}^2[k\left]\right\}$

P wherein _s[k] is the duct static pressure in k sampling period, P _{S, set}[k] is that the air conditioning area in k sampling period is set static pressure, M _sBe prediction time domain, U _Fan[k] is k sampling period blower voltage controlled quentity controlled variable, L _sBe k sampling period static pressure control loop object function;

Total blast volume control loop predictive control function is:

$J_f\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_f-1}{\mathrm{\&Sigma;}}}{L}_f\left(V_{\mathrm{fan}}\right[k],U_{\mathrm{fan}}[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_f-1}{\mathrm{\&Sigma;}}}\{{\left(V_{\mathrm{fan}}\right[k]-V_{\mathrm{fan}.\mathrm{set}}[k\left]\right)}^2+U_{\mathrm{fan}}^2[k\left]\right\}$

V wherein _Fan[k] is the pipeline air quantity in k sampling period, M _fBe prediction time domain, V _Fan.set[k] is each terminal prediction air quantity sum in k sampling period, L _fBe k sampling period total blast volume control loop object function;

Wind pushing temperature control loop predictive control function is:

$J_{\mathrm{st}}\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{st}}\left(T_{\mathrm{st}}\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}{\left(T_{\mathrm{st}}\right[k]-T_{\mathrm{st},\mathrm{set}}[k\left]\right)}^2$

T wherein _St[k] is the temperature in k sampling period, T _{St, set}[k] is the design temperature in k sampling period, M _StBe prediction time domain, L _StBe k sampling period wind pushing temperature control loop object function;

New wind ratio control loop outer shroud predictive control function is:

$J_q\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_q-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{no}}\left(Q\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_q-1}{\mathrm{\&Sigma;}}}{\left(Q\right[k]-Q_{\mathrm{set}}[k\left]\right)}^2$

Q[k wherein] be the air quality in k sampling period, Q _Set[k] is the air quality setting value in k sampling period, M _qBe prediction time domain, L _NoBe k sampling period new wind ratio control loop outer shroud object function;

The ring predictive control function is in the new wind ratio control loop:

$J_{\mathrm{ni}}\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{ni}}-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{ni}}\left(V_{\mathrm{ni}}\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{ni}}-1}{\mathrm{\&Sigma;}}}{\left(V_{\mathrm{ni}}\right[k]-S[k\left]\right)}^2$

V wherein _i[k] is the prediction air-valve discharge quantity of fan in k sampling period, S[k] be the setting air quantity in k sampling period, M _iBe prediction time domain, L _NiIt is ring object function in k the sampling period new wind ratio control loop;

2) determine static pressure control loop, VAV air-conditioning unit total blast volume control loop, VAV air-conditioning unit wind pushing temperature control loop and the VAV air-conditioning unit new wind ratio control loop neural network prediction controller architecture of VAV air-conditioning unit adjustable settings static pressure

The input signal of static pressure neural network prediction controller is output as rotation speed of fan for setting duct static pressure and duct static pressure;

Total blast volume neural network prediction controller input signal is pipeline air quantity and total blast volume, is output as rotation speed of fan;

The input signal of wind pushing temperature neural network prediction controller is output as the water valve aperture for setting wind pushing temperature and wind pushing temperature;

New wind ratio outer shroud neural network prediction controller is input as sets CO2 concentration and return air CO2 concentration, is output as the setting resh air requirement; Ring neural network prediction controller will be set resh air requirement, duct static pressure (fresh air pipeline place) and pipeline air quantity as input in the new wind ratio, and new air valve aperture is output;

3) realize static pressure neural network prediction controller online optimizing in the static pressure control loop;

4) realize total blast volume neural network prediction controller online optimizing in the total blast volume control loop;

5) realize wind pushing temperature neural network prediction controller online optimizing in the wind pushing temperature control loop;

6) realize encircling and outer shroud neural network prediction controller online optimizing in the new wind ratio in the new wind ratio control loop.

The invention has the advantages that:

(1) the present invention utilizes the neural network forecast model, and model can be revised online, and disturbance is had good adaptive ability; Simultaneously, by the Neural Network Self-learning function, can solve the engineering middle controller on-site parameters difficult problem of adjusting;

(2) the present invention adopts the PREDICTIVE CONTROL scheme, can solve the larger problem of actual environment temperature fluctuation that air-conditioning system produces owing to hysteresis characteristic, and can realize optimization control, can further reduce the energy consumption of VAV air-conditioning system;

(3) because the control method that the present invention proposes considers comfort index and energy consumption index is optimized PREDICTIVE CONTROL as optimality criterion, compare with the PID control method, in the suitable situation of selected performance indications parameter, can make VAV air quantity variable air conditioner blower fan system energy efficient more than 13%;

(4) the pressure independent type that the present invention is directed to the VAV analysis of VAV terminals is controlled process characteristic, adopts the tandem forecast Control Algorithm based on neutral net, has improved control accuracy;

(5) the present invention is according to the features of winter, summer and conditioning in Transition Season, and system selects static pressure control strategy or the total blast volume control strategy of adjustable settings static pressure automatically, realizes the accurate control to the blower fan total blast volume;

(6) the present invention has adopted new wind pushing temperature control strategy, by detecting all terminal VAV-BOX primary air flows, operating condition according to VAV-BOX is adjusted wind pushing temperature, has solved because minimum new wind moves, and has caused the excessively low problem of part air conditioning area temperature;

(7) new wind ratio control loop of the present invention adopts the tandem forecast Control Algorithm, realizes the accurate control to new wind ratio, thereby has further improved the energy-saving effect of system;

(8) the neural network prediction system optimizing control that adopts of the present invention can the non-linear and probabilistic impact of resolution system, and the algorithm real-time is good, and it is few to take the memory block, is easy to Project Realization.When realizing on the controller at the scene, it is 1.1MB that whole program takies space, field controller memory block, process committed memory size is 1.3MB only during operation, is applicable to build at present the networked control technologys such as the extensive fieldbus that adopts in building automatic control field, EPA.

Description of drawings

Fig. 1 is air conditioning area temperature prediction model structure figure

Fig. 2 is terminal air-valve forecast model structure chart

Fig. 3 is main air duct duct static pressure forecast model structure chart

Fig. 4 is main air duct air quantity forecast model structure chart

Fig. 5 is wind pushing temperature forecast model structure chart

Fig. 6 is new wind air-valve forecast model structure chart

Fig. 7 is Air Quality Forecast model structure figure

Fig. 8 is VAV air conditioning terminal schematic diagram

Fig. 9 is terminal tandem predictive control loop figure

Figure 10 is terminal interior ring neural network prediction controller architecture figure

Figure 11 is terminal outer shroud neural network prediction controller architecture figure

Figure 12 is terminal outer shroud and interior ring neural network prediction controller searching process figure

Figure 13 is the static pressure predictive control loop figure of adjustable settings static pressure

Figure 14 is total blast volume predictive control loop figure

Figure 15 is wind pushing temperature predictive control loop figure

Figure 16 is new wind ratio predictive control loop figure

Figure 17 is static pressure neural network prediction controller architecture figure

Figure 18 is total blast volume neural network prediction controller architecture figure

Figure 19 is wind pushing temperature predictive controller structure chart

Figure 20 is new wind ratio outer shroud neural network prediction controller architecture figure

Figure 21 is ring neural network prediction controller architecture figure in the new wind ratio

Figure 22 is static pressure neural network prediction controller searching process figure

Figure 23 is total blast volume neural network prediction controller searching process figure

Figure 24 is wind pushing temperature neural network prediction controller searching process figure

Figure 25 is new wind ratio outer shroud and inside and outside neural network prediction controller searching process figure

The specific embodiment

The present invention is described in further detail below in conjunction with drawings and Examples.

The present invention is a kind of VAV Control Method of Vav Air Conditioning System, specifically comprises following step:

The first step: utilize BP neural network air conditioning terminal regional temperature forecast model, terminal air-valve forecast model, air-conditioning unit main air duct duct static pressure forecast model, main air duct air quantity forecast model, wind pushing temperature forecast model, new wind air-valve forecast model and Air Quality Forecast model

1) determines air conditioning terminal regional temperature forecast model, terminal air-valve forecast model, air-conditioning unit main air duct duct static pressure forecast model, main air duct air quantity forecast model, wind pushing temperature forecast model, new wind air-valve forecast model and Air Quality Forecast model structure

As shown in Figure 1, the input signal of air conditioning area temperature prediction model is outdoor intensity of solar radiation, outdoor temperature, CO ₂Concentration, indoor temperature, air quantity and terminal valve opening are output as next constantly indoor temperature;

As shown in Figure 2, the input signal of terminal air-valve forecast model is terminal valve opening and duct static pressure (air-conditioning unit main air duct place), is output as next constantly terminal air quantity;

As shown in Figure 3, the input signal of main air duct duct static pressure forecast model is VAV air-conditioning unit rotation speed of fan, duct static pressure, return air CO2 concentration, outdoor temperature and intensity of solar radiation, is output as next constantly duct static pressure;

As shown in Figure 4, the input signal of main air duct air quantity forecast model is rotation speed of fan and duct static pressure, is output as next constantly main air duct air quantity;

As shown in Figure 5, the input signal of wind pushing temperature forecast model is wind pushing temperature and water valve aperture, is output as next constantly wind pushing temperature;

As shown in Figure 6, the input signal of new wind air-valve forecast model is new air valve aperture and duct static pressure (fresh air pipeline place), is output as next constantly resh air requirement;

As shown in Figure 7, the input signal of Air Quality Forecast model is new wind air-valve valve opening and CO2 concentration, is output as next constantly CO2 concentration.

2) collecting sample data

The sampling time scope be 8 a.m. to 6 pm, sampling time interval 150 seconds, each forecast model gather about 2000 groups of data; Getting the time interval when setting up forecast model is 5 minutes;

Air conditioning area temperature prediction model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and terminal valve area also is divided into ten grades by 0V to 10V interval 1V, gathers outdoor intensity of solar radiation, outdoor temperature, CO ₂Concentration, indoor temperature and air quantity;

Terminal air-valve forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and terminal valve area also is divided into ten grades by 0V to 10V interval 1V, gathers duct static pressure (air-conditioning unit main air duct place) and air quantity;

Main air duct duct static pressure forecast model: after each terminal debugging was finished, air-blower control amount signal all was divided into ten grades by 0V to 10V interval 1V, gathered VAV air-conditioning unit rotation speed of fan, duct static pressure, return air CO2 concentration, outdoor temperature, intensity of solar radiation;

Main air duct air quantity forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, gathers rotation speed of fan, duct static pressure, main air duct air quantity;

Wind pushing temperature forecast model: the water valve aperture is divided into ten grades by 0V to 10V interval 1V, gathers wind pushing temperature and water valve aperture;

New wind air-valve forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and new wind valve area also is divided into ten grades by 0V to 10V interval 1V, gathers duct static pressure (fresh air pipeline place) and air quantity;

The Air Quality Forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and new wind valve area also is divided into ten grades by 0V to 10V interval 1V, gathers CO2 concentration;

3) sample data is by formula carried out normalization in (1), (2):

$x_i=\frac{x_{\mathrm{di}}-x_{d\min }}{x_{d\max }-x_{d\min }}---\left(1\right)$

$y_{\mathrm{tl}}=\frac{y_{\mathrm{dl}}-y_{d\min }}{y_{d\max }-y_{d\min }}---\left(2\right)$

X wherein _iBe the input value of neutral net after the normalization, x _DiBe former input value, x _DminBe the minimum of a value in the former input value, x _DmaxBe the maximum in the former input value; y _TlBe the desired value of neutral net after the normalization, y _DlRepresent former desired value, y _DminRepresent the minimum of a value in the former desired value, y _DmaxBe the maximum in the former desired value;

4) by table 1 parameter neutral net is carried out off-line training:

Table 1 Neural Network Training Parameter table

The neutral net type	Single hidden layer forward direction BP network
		The input layer number	A *
The output layer nodes	1
		The number of hidden nodes	B *
The neuron excitation function	Hidden layer ' tansig ', output layer ' purelin '
		Learning function	′learngdm′

Performance function	' msereg ' (weighted mean square is poor)
		The network training function	' trainbr ' (Bayesian-regularization method)
Power (threshold) value initialization method	' initnw ' (Nguyen-Widrow method)
		The maximum training time	2000Epochs
Target error	0

A ^*: air conditioning area temperature prediction model value 6, terminal air-valve forecast model, wind pushing temperature forecast model, main air duct air quantity forecast model, Air Quality Forecast model and new wind air-valve forecast model value 2, main air duct duct static pressure forecast model value 5;

B ^*：

Second step: VAV air conditioning terminal tandem PREDICTIVE CONTROL

The present invention adopts the tandem forecast Control Algorithm for the VAV air conditioning terminal.Fig. 8 shows VAV air conditioning terminal schematic diagram, and air-valve, pressure sensor and air velocity transducer are installed in the air delivery duct.Temperature sensor and CO ₂Concentration sensor is separately positioned in air conditioning area and the return airway; At outdoor mounting temperature sensor and the intensity of solar radiation sensor of also needing.Terminal tandem control loop as shown in Figure 9, cascade control system inner and outer ring controller is all selected the neural network prediction controller, according to the air conditioning area temperature that gathers, input outer shroud neural network prediction controller, calculate the setting air quantity, ring neural network prediction controller is adjusted terminal valve area in the recycling.Predictive controller not only can improve the trace performance of air quantity setting value, eliminates simultaneously static pressure to the impact of air quantity, has improved control accuracy.

1) determines terminal outer shroud predictive control function

Terminal outer shroud predictive control function is: $J_o\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_c-1}{\mathrm{\&Sigma;}}}{L}_o\left(T_o\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_c-1}{\mathrm{\&Sigma;}}}{\left(T_o\right[k]-T_{o,\mathrm{set}}[k\left]\right)}^2$

M wherein _cBe prediction time domain, t ₁The initial time in the prediction time domain, T _o[k] is the air conditioning area temperature in k sampling period, T _{O, set}[k] is the air conditioning area desired temperature in k sampling period, L _oIt is terminal outer shroud object function of k sampling period;

2) determine terminal interior ring predictive control function

The ring predictive control function is in terminal: $J_i\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_i-1}{\mathrm{\&Sigma;}}}{L}_i\left(V_i\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_i-1}{\mathrm{\&Sigma;}}}{\left(V_i\right[k]-V_{i,\mathrm{set}}[k\left]\right)}^2$

V wherein _i[k] is the air quantity in k sampling period, V _{I, set}[k] is the air quantity setting value in k sampling period; M _iBe prediction time domain, L _iIt is ring object function in k the sampling period end;

3) determine terminal interior ring neural network prediction controller and terminal outer shroud neural network prediction controller architecture

Encircle the neural network prediction controller architecture as shown in figure 10 in terminal, air quantity setting value, duct static pressure (air-conditioning unit main air duct place) are input with the pipeline air quantity, and valve opening is output.The number of hidden nodes of controller neutral net is 5;

The structure of terminal outer shroud neural network prediction controller as shown in figure 11, input parameter comprises outdoor temperature, intensity of solar radiation, indoor temperature and desired temperature, output parameter is the air quantity setting value.The number of hidden nodes of controller neutral net is 8;

4) ring neural network prediction controller in terminal and terminal outer shroud neural network prediction controller are carried out the online optimizing training

Initialize terminal in ring be connected with outer shroud neural network prediction controller that each connects weights, assignment is less random number in [1,1] scope, then computing controller is exported.Prediction time domain M _iAnd M _oBe respectively 3 and 6, predetermined period was got 5 minutes;

Searching process as shown in figure 12, x[k wherein] be k constantly the relevant state variables parameter of outer shroud controlled device (being air conditioning area), i.e. indoor temperature; Be constantly forecast model output of k+1, i.e. k+1 Indoo r prediction temperature constantly; x ^*It is desired temperature; S[k] be constantly air quantity setting value of k; Y[k] be constantly interior ring relevant state variables parameter, i.e. air quantity of k; C[k] be the k terminal valve area of the moment that terminal interior ring neural network prediction controller calculates; U[k] be that the terminal valve area that i.e. optimizing obtains was exported in k control constantly after the optimizing of terminal interior ring neural network prediction controller finished;

The k+1 that exports for terminal air-valve forecast model predicts air quantity constantly.With x[t ₁], x ^*With-1 act on terminal outer shroud neural network prediction controller, obtain set amount S[t ₁], with S[t ₁], Y[t ₁] and-1 act on terminal in ring neural network prediction controller, obtain C[t ₁].Then with C[t ₁] act on terminal air-valve forecast model, obtain predicting air quantity

Keep terminal interior ring neural network prediction controller weights constant, will

S[t ₁] and-1 act on terminal in ring neural network prediction controller, obtain C ' [t ₁+ 1].With C ' [t ₁+ 1] the terminal air-valve forecast model of input obtains

Will

S[t ₁] and-1 act on terminal in ring neural network prediction controller obtain C ' [t ₁+ 2]; With C ' [t ₁+ 2] the terminal air-valve forecast model of input obtains

The data that calculate are preserved.Make λ _i[k+M _i]=0, the λ from backward front respectively calculating formula (3) _i[k] and q _i[k]:

$q_i\left[k\right]=\frac{\&PartialD;f_v{\left(k\right)}^T}{\&PartialD;C\left[k\right]}{\mathrm{\&lambda;}}_i[k+1]+\frac{\&PartialD;L_i{\left(k\right)}^T}{\&PartialD;C\left[k\right]}$

${\mathrm{\&lambda;}}_i\left[k\right]=\frac{\&PartialD;f_v{\left(k\right)}^T}{\&PartialD;y\left[k\right]}{\mathrm{\&lambda;}}_i[k+1]+\frac{\&PartialD;L_i{\left(k\right)}^T}{\&PartialD;y\left[k\right]}+\frac{\&PartialD;g_i{(k;W_i)}^T}{\&PartialD;y\left[k\right]}{q}_i\left[k\right]---\left(3\right)$

F in the formula _v(k) the terminal air-valve forecast model of setting up before the representative; g _i(k; W _i) be terminal interior ring neural network prediction controller equation.According to the q that calculates _i[k], through type (4) and formula (5) are revised the weights of terminal interior ring neural network prediction controller:

$\mathrm{\&Delta;}{W}_i=-{\mathrm{\&mu;}}_i\frac{\&PartialD;g_i{(k,W_i)}^T}{\&PartialD;W_i}{q}_i\left[k\right]---\left(4\right)$

W _i＝W _i+ΔW _i (5)

W wherein _iThe weights battle array of terminal interior ring neural network prediction controller, μ _iThe right value update rate, μ _iSelect 0.05.Constantly revise the weights of terminal interior ring neural network prediction controller, until Δ W _i＜0.001;

With S[t ₁], Y[t ₁] and-1 act on ring neural network prediction controller in terminal after the optimizing, obtain u ' [t ₁].With u[t] act on terminal air-valve forecast model, obtain

Afterwards with x[t ₁],

Act on the air conditioning area forecast model, obtain

Keep terminal outer shroud neural network prediction controller weights constant, will

x ^*With-1 act on terminal outer shroud neural network prediction controller, obtain set amount s ' [t ₁+ 1]; Carry out again interior ring optimizing, with s ' [t ₁+ 1], With-1 act on ring neural network prediction controller in terminal after the optimizing, obtain u ' [t ₁+ 1].Again with u ' [t ₁+ 1] acts on terminal air-valve forecast model, obtain

Afterwards with x[t ₁+ 1],

Act on the air conditioning area forecast model, obtain

Will

x ^*With-1 act on terminal outer shroud neural network prediction controller, obtain set amount s ' [t ₁+ 2]; Carry out again interior ring optimizing, with s ' [t ₁+ 2], With-1 act on ring neural network prediction controller in terminal after the optimizing, obtain u ' [t ₁+ 2].Again with u ' [t ₁+ 2] act on terminal air-valve forecast model, obtain

Afterwards, will

Act on the air conditioning area forecast model, obtain

And the rest may be inferred, utilizes controller neutral net and object forecast model to extrapolate following u ' [t ₁+ i] and

Value, i=3 wherein ..., 6, and the data that calculate are preserved.Make λ _o[k+M _o]=0, the λ from backward front respectively calculating formula (6) _o[k] and q _o[k]:

$q_o\left[k\right]=\frac{\&PartialD;f_z{\left(k\right)}^T}{\&PartialD;u\left[k\right]}{\mathrm{\&lambda;}}_o[k+1]+\frac{\&PartialD;L_o{\left(k\right)}^T}{\&PartialD;u\left[k\right]}$

${\mathrm{\&lambda;}}_o\left[k\right]=\frac{\&PartialD;f_z{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_o[k+1]+\frac{\&PartialD;L_o{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{\&PartialD;g_o{(k;W_o)}^T}{\&PartialD;x\left[k\right]}{q}_o\left[k\right]---\left(6\right)$

F in the formula _z(k) the air conditioning area forecast model of setting up before the representative; g _o(k; W _o) be terminal outer shroud neural network prediction controller equation.According to the q that calculates _o[k] is for k=t ₁+ M _o-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (7) and formula (8) are revised the weights of terminal outer shroud neural network prediction controller:

$\mathrm{\&Delta;}{W}_o=-{\mathrm{\&mu;}}_o\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_o-1}{\mathrm{\&Sigma;}}}\frac{\&PartialD;g_o{(k,W_o)}^T}{\&PartialD;W_o}{q}_o\left[k\right]---\left(7\right)$

W _o＝W _o+ΔW _o (8)

W wherein _oThe weights battle array of terminal outer shroud neural network prediction controller, μ _oThe right value update rate, μ _oSelect 0.05.Constantly revise the weights of terminal outer shroud neural network prediction controller, until Δ W _o＜0.001; After the outer shroud optimizing finishes, with x[t ₁] and x ^*Input terminal outer shroud neural network prediction controller, again with the S[t that obtains ₁] and Y[t ₁] the terminal interior ring neural network prediction controller of input, will export u[t ₁] directly act on terminal air-valve;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

The 3rd step: VAV air-conditioning unit PREDICTIVE CONTROL

The present invention finishes the control of air-conditioning unit by 4 control loops, is respectively static pressure control loop (Figure 13), VAV air-conditioning unit total blast volume control loop (Figure 14), VAV air-conditioning unit wind pushing temperature control loop (Figure 15) and the VAV air-conditioning unit new wind ratio control loop (Figure 16) of VAV air-conditioning unit adjustable settings static pressure.

The present invention is according to different operating modes, realizes control to air-conditioning unit air quantity by two kinds of control strategies: when the static pressure of system static pressure monitoring point descends when reaching setting value, select the static pressure predictive control strategy (Figure 13) of adjustable settings static pressure; When the static pressure of system static pressure monitoring point is higher than setting value, then move total blast volume predictive control strategy (Figure 14).

The present invention is different from conventional VAV air-conditioning system and all adopts the control strategy of deciding wind pushing temperature, variable air rate, but adopt different control strategies according to different working conditions, when moving owing to minimum resh air requirement operating mode to solve conditioning in Transition Season part air conditioning area, cause the excessively low problem of temperature.Concrete control program is: control system detects all terminal VAV-BOX primary air flows, when a certain VAV-BOX primary air flow of appearance is lower than nominal air delivery 30%, reduces the water valve aperture, improves 0.5 ℃ of wind pushing temperature; When a certain terminal VAV-BOX primary air flow greater than 70% the time, increase the water valve aperture, reduce by 0.5 ℃ of wind pushing temperature (Figure 15).

The conventional control strategy of VAV air amount is the aperture of the new wind air-valve of control and return air air-valve, and for the recuperation of heat unit is set on the roof, adopt the vertical system that concentrates new wind, because the dynamic and static pressure of vertical VMC relation, the resh air requirement adjustment of adjacent floor can affect the quantity delivered of the new wind of this layer.The present invention adopts the tandem PREDICTIVE CONTROL of pressure independent type in the new wind ratio control loop, according to the return air CO2 concentration that gathers, input outer shroud neural network prediction controller calculates the setting resh air requirement, and ring neural network prediction controller is adjusted new wind valve area in the recycling.Outer shroud is to carry out PREDICTIVE CONTROL according to air conditioning area CO2 concentration, and interior ring is the PREDICTIVE CONTROL of resh air requirement being carried out the pressure independent type, thereby reach the accurate control to resh air requirement, under the prerequisite that satisfies the air conditioning area environmental quality, realize maximum energy-saving control (Figure 16).

1) determines VAV air-conditioning unit static pressure control loop, VAV air-conditioning unit total blast volume control loop, VAV air-conditioning unit wind pushing temperature control loop and VAV air-conditioning unit new wind ratio control loop predictive control function

The static pressure control loop predictive control function is:

$J_s\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_s-1}{\mathrm{\&Sigma;}}}{L}_s\left(P_s\right[k],U_{\mathrm{fan}}[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_s-1}{\mathrm{\&Sigma;}}}\{{\left(P_s\right[k]-P_{s,\mathrm{set}}[k\left]\right)}^2+U_{\mathrm{fan}}^2[k\left]\right\}$

P wherein _s[k] is the duct static pressure in k sampling period, P _{S, set}[k] is that the air conditioning area in k sampling period is set static pressure, M _sBe prediction time domain, U _Fan[k] is k sampling period blower voltage controlled quentity controlled variable, L _sBe k sampling period static pressure control loop object function;

Total blast volume control loop predictive control function is:

$J_f\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_f-1}{\mathrm{\&Sigma;}}}{L}_f\left(V_{\mathrm{fan}}\right[k],U_{\mathrm{fan}}[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_f-1}{\mathrm{\&Sigma;}}}\{{\left(V_{\mathrm{fan}}\right[k]-V_{\mathrm{fan}.\mathrm{set}}[k\left]\right)}^2+U_{\mathrm{fan}}^2[k\left]\right\}$

V wherein _Fan[k] is the pipeline air quantity in k sampling period, M _fBe prediction time domain, V _Fan.set[k] is each terminal prediction air quantity sum in k sampling period, L _fBe k sampling period total blast volume control loop object function;

Wind pushing temperature control loop predictive control function is:

$J_{\mathrm{st}}\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{st}}\left(T_{\mathrm{st}}\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}{\left(T_{\mathrm{st}}\right[k]-T_{\mathrm{st},\mathrm{set}}[k\left]\right)}^2$

T wherein _St[k] is the temperature in k sampling period, T _{St, set}[k] is the design temperature in k sampling period, M _StBe prediction time domain, L _StBe k sampling period wind pushing temperature control loop object function;

New wind ratio control loop outer shroud predictive control function is:

$J_q\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_q-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{no}}\left(Q\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_q-1}{\mathrm{\&Sigma;}}}{\left(Q\right[k]-Q_{\mathrm{set}}[k\left]\right)}^2$

Q[k wherein] be the air quality in k sampling period, Q _Set[k] is the air quality setting value in k sampling period, M _qBe prediction time domain, L _NoBe k sampling period new wind ratio control loop outer shroud object function;

The ring predictive control function is in the new wind ratio control loop:

$J_{\mathrm{ni}}\left[k\right]=\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{ni}}-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{ni}}\left(V_{\mathrm{ni}}\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{ni}}-1}{\mathrm{\&Sigma;}}}{\left(V_{\mathrm{ni}}\right[k]-S[k\left]\right)}^2$

V wherein _i[k] is the prediction air-valve discharge quantity of fan in k sampling period, S[k] be the setting air quantity in k sampling period, M _iBe prediction time domain, L _NiIt is ring object function in k the sampling period new wind ratio control loop;

2) determine VAV air-conditioning unit static pressure control loop, VAV air-conditioning unit total blast volume control loop, VAV air-conditioning unit wind pushing temperature control loop and VAV air-conditioning unit new wind ratio control loop neural network prediction controller architecture

The input signal of static pressure neural network prediction controller is output as rotation speed of fan (control signal of motor frequency conversion device is 0～10V voltage) for setting duct static pressure and duct static pressure; Hidden node is 3 (Figure 17);

Total blast volume neural network prediction controller input signal is pipeline air quantity and total blast volume, is output as rotation speed of fan (control signal of motor frequency conversion device is 0～10V voltage); Hidden node is 3 (Figure 18);

The input signal of wind pushing temperature neural network prediction controller is output as water valve aperture (control signal of electric valve executing mechanism is 0～10V voltage) for setting wind pushing temperature and wind pushing temperature; Hidden node is 3 (Figure 19);

New wind ratio outer shroud neural network prediction controller input signal is output as the setting resh air requirement for setting CO2 concentration and return air CO2 concentration; Hidden node is 4 (Figure 20); Ring neural network prediction controller will be set resh air requirement, duct static pressure (fresh air pipeline place) and pipeline air quantity as input in the new wind ratio, and new air valve aperture is output (control signal of electric valve executing mechanism is 0～10V voltage); Hidden node is 5 (Figure 21);

3) static pressure neural network prediction controller online optimizing in the static pressure predictive control loop of realization adjustable settings static pressure

Initialize each connection weights of static pressure neural network prediction controller, assignment is less random number in [1,1] scope, then computing controller output.Prediction time domain M _sIt was 2 steps, for the right value update rate μ of static pressure neural network prediction controller _sSelect 0.05, predetermined period was got 5 minutes.

Searching process as shown in figure 22, x[k wherein] parameter be duct static pressure, u _Fan[k] is the air-blower control amount, x ^*To set duct static pressure.With x[t ₁], x ^*With-1 act on static pressure neural network prediction controller, controlled amount u _Fan[t ₁], then with u _Fan[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u _Fan[t ₁], x[t ₁] and-1 act on main air duct duct static pressure forecast model, obtain

Keep static pressure neural network prediction controller weights constant, with x[t ₁+ 1], x ^*With-1 act on controller, obtain

Will

X[t ₁+ 1] and-1 act on main air duct duct static pressure forecast model, obtains The data that calculate are preserved.Make λ _s[k+M _s]=0, the λ from backward front respectively calculating formula (9) _s[k] and q _s[k]:

$q_s\left[k\right]=\frac{\&PartialD;f_s{\left(k\right)}^T}{\&PartialD;u_{\mathrm{fan}}\left[k\right]}{\mathrm{\&lambda;}}_s[k+1]+\frac{\&PartialD;L_s{\left(k\right)}^T}{\&PartialD;u_{\mathrm{fan}}\left[k\right]}$

${\mathrm{\&lambda;}}_s\left[k\right]=\frac{\&PartialD;f_s{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_s[k+1]+\frac{\&PartialD;L_s{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{\&PartialD;g_s{(k;W_s)}^T}{\&PartialD;x\left[k\right]}{q}_s\left[k\right]---\left(9\right)$

F in the formula _s(k) the main air duct duct static pressure forecast model of setting up before the representative, g _s(k; W _s) be static pressure neural network prediction controller equation.According to the q that calculates _s[k] is for k=t ₁+ M _s-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (10) and formula (11) are revised the weights of static pressure neural network prediction controller:

$\mathrm{\&Delta;}{W}_s=-{\mathrm{\&mu;}}_s\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_s-1}{\mathrm{\&Sigma;}}}\frac{\&PartialD;g_s{(k,W_s)}^T}{\&PartialD;W_s}{q}_s\left[k\right]---\left(10\right)$

W _s＝W _s+ΔW _s (11)

W wherein _sThe weights battle array of static pressure neural network prediction controller, μ _sThe right value update rate, μ _sSelect 0.05.Constantly revise the weights of static pressure neural network prediction controller, until Δ W _s＜0.001;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

4) realize total blast volume neural network prediction controller online optimizing in the total blast volume control loop

Initialize each connection weights of total blast volume neural network prediction controller, assignment is less random number in [1,1] scope, then computing controller output.Prediction time domain M _fIt was 2 steps, for the right value update rate μ of total blast volume neural network prediction controller _fSelect 0.05, predetermined period was got 5 minutes.

Searching process as shown in figure 23, x[k wherein] parameter be the main air duct air quantity, u _Fan[k] is the air-blower control amount, x ^*It is each regional air quantity setting value sum.With x[t ₁], x ^*With-1 act on total blast volume neural network prediction controller, controlled amount u _Fan[t ₁], then with u _Fan[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u _Fan[t ₁], x[t ₁] and-1 act on main air duct air quantity forecast model, obtain Keep total blast volume neural network prediction controller weights constant, with x[t ₁+ 1], x ^*With-1 act on controller, obtain

Will

X[t ₁+ 1] and-1 act on main air duct air quantity forecast model, obtains

The data that calculate are preserved.Make λ _f[k+M _f]=0, the λ from backward front respectively calculating formula (12) _f[k] and q _f[k]:

$q_f\left[k\right]=\frac{\&PartialD;f_f{\left(k\right)}^T}{\&PartialD;u_{\mathrm{fan}}\left[k\right]}{\mathrm{\&lambda;}}_f[k+1]+\frac{\&PartialD;L_f{\left(k\right)}^T}{\&PartialD;u_{\mathrm{fan}}\left[k\right]}$

${\mathrm{\&lambda;}}_f\left[k\right]=\frac{\&PartialD;f_f{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_f[k+1]+\frac{\&PartialD;L_f{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{\&PartialD;g_f{(k;W_f)}^T}{\&PartialD;x\left[k\right]}{q}_f\left[k\right]---\left(12\right)$

F in the formula _f(k) the main air duct air quantity forecast model of setting up before the representative, g _f(k; W _f) be total blast volume neural network prediction controller equation.According to the q that calculates _f[k] is for k=t ₁+ M _f-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (13) and formula (14) are revised the weights of total blast volume neural network prediction controller:

$\mathrm{\&Delta;}{W}_f=-{\mathrm{\&mu;}}_f\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_f-1}{\mathrm{\&Sigma;}}}\frac{\&PartialD;g_f{(k,W_f)}^T}{\&PartialD;W_f}{q}_f\left[k\right]---\left(13\right)$

W _f＝W _f+ΔW _f (14)

W wherein _fThe weights battle array of total blast volume neural network prediction controller, μ _fThe right value update rate, μ _fSelect 0.05.Constantly revise the weights of total blast volume neural network prediction controller, until Δ W _f＜0.001;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

5) realize wind pushing temperature neural network prediction controller online optimizing in the wind pushing temperature control loop

Initialize each connection weights of wind pushing temperature neural network prediction controller, assignment is less random number in [1,1] scope, then computing controller output.Prediction time domain M _StIt was 2 steps, for the right value update rate μ of wind pushing temperature neural network prediction controller _StSelect 0.05, predetermined period was got 5 minutes.

Searching process as shown in figure 24, x[k wherein] parameter be wind pushing temperature, u[k] be the water valve aperture, x ^*To set wind pushing temperature.With x[t ₁], x ^*With-1 act on wind pushing temperature neural network prediction controller, controlled amount u[t ₁], then with u[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u[t ₁], x[t ₁] and-1 act on the wind pushing temperature forecast model, obtain Keep wind pushing temperature neural network prediction controller weights constant, with x[t ₁+ 1], x ^*With-1 act on controller, obtain u ' [t ₁+ 1], with u ' [t ₁+ 1], x[t ₁+ 1] and-1 act on the wind pushing temperature forecast model, obtains The data that calculate are preserved.Make λ _St[k+M _St]=0, the λ from backward front respectively calculating formula (15) _St[k] and q _St[k]:

$q_{\mathrm{st}}\left[k\right]=\frac{\&PartialD;f_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;u\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{st}}[k+1]+\frac{\&PartialD;L_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;u\left[k\right]}$

${\mathrm{\&lambda;}}_{\mathrm{st}}\left[k\right]=\frac{\&PartialD;f_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{st}}[k+1]+\frac{\&PartialD;L_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{\&PartialD;g_{\mathrm{st}}{(k;W_{\mathrm{st}})}^T}{\&PartialD;x\left[k\right]}{q}_{\mathrm{st}}\left[k\right]---\left(15\right)$

F in the formula _St(k) the wind pushing temperature forecast model of setting up before the representative, g _St(k; W _St) be wind pushing temperature neural network prediction controller equation.According to the q that calculates _St[k] is for k=t ₁+ M _St-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (16) and formula (17) are revised the weights of wind pushing temperature neural network prediction controller:

$\mathrm{\&Delta;}{W}_{\mathrm{st}}=-{\mathrm{\&mu;}}_{\mathrm{st}}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}\frac{\&PartialD;g_{\mathrm{st}}{(k,W_{\mathrm{st}})}^T}{\&PartialD;W_{\mathrm{st}}}{q}_{\mathrm{st}}\left[k\right]---\left(16\right)$

W _st＝W _st+ΔW _st (17)

W wherein _StThe weights battle array of wind pushing temperature neural network prediction controller, μ _StThe right value update rate, μ _StSelect 0.05.Constantly revise the weights of wind pushing temperature neural network prediction controller, until Δ W _St＜0.001;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

6) realize neural network prediction controller online optimizing in the new wind ratio control loop

Initialize in the new wind ratio ring and be connected each connection weights with outer shroud neural network prediction controller, assignment is less random number in [1,1] scope, then computing controller output is if output valve is all in [0.7,1] scope, then use this group controller initial weight, otherwise random initializtion again.Prediction time domain M _NiAnd M _qBe respectively 3 and 6, for the right value update rate μ selection 0.05 of ring and outer shroud neural network prediction controller in the new wind ratio, predetermined period was got 5 minutes;

Searching process as shown in figure 25, x[k wherein] be constantly CO2 concentration of k;

Be constantly forecast model output of k, i.e. k+1 return air CO2 concentration constantly;

Be t ₁Return air CO2 concentration constantly; x ^*It is the CO2 concentration set point; S[k] constantly set resh air requirement for k; Y[k] be that k encircles the relevant state variables parameter constantly, comprise pipeline air quantity and duct static pressure; O[k] the constantly new wind valve area of the k that calculates for interior ring neural network prediction controller; U[k] be k control output constantly after interior ring neural network prediction controller optimizing finishes, i.e. the new wind valve area that optimizing obtains;

The k+1 that exports for new wind air-valve forecast model predicts air quantity constantly.With x[t ₁], x ^*With-1 act on new wind ratio outer shroud neural network prediction controller, obtain set amount S[t ₁], with S[t ₁], Y[t ₁] and-1 act on ring neural network prediction controller in the new wind ratio, obtain O[t ₁].Then with O[t ₁] act on respectively new wind air-valve forecast model, obtain predicting air quantity

It is constant to keep encircling neural network prediction controller weights in the new wind ratio, will

S[t ₁] and-1 act on ring neural network prediction controller in the new wind ratio, obtain O ' [t ₁+ 1].With O ' [t ₁+ 1] the new wind air-valve forecast model of input obtains

Will

S[t ₁] and-1 act on that ring neural network prediction controller obtains O ' [2] in the new wind ratio.With the new wind air-valve forecast model of O ' [2] input, obtain The data that calculate are preserved.Make λ _Ni[k+2]=0, the λ from backward front respectively calculating formula (18) _Ni[k] and q _Ni[k]:

$q_{\mathrm{ni}}\left[k\right]=\frac{\&PartialD;f_{\mathrm{nv}}{\left(k\right)}^T}{\&PartialD;O\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{ni}}[k+1]+\frac{\&PartialD;L_{\mathrm{ni}}{\left(k\right)}^T}{\&PartialD;O\left[k\right]}$

${\mathrm{\&lambda;}}_{\mathrm{ni}}\left[k\right]=\frac{\&PartialD;f_{\mathrm{nv}}{\left(k\right)}^T}{\&PartialD;y\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{ni}}[k+1]+\frac{\&PartialD;L_{\mathrm{ni}}{\left(k\right)}^T}{\&PartialD;y\left[k\right]}+\frac{\&PartialD;g_{\mathrm{ni}}{(k;W_{\mathrm{ni}})}^T}{\&PartialD;y\left[k\right]}{q}_{\mathrm{ni}}\left[k\right]---\left(18\right)$

F in the formula _Nv(k) the new wind air-valve forecast model of setting up before the representative; g _Ni(k; W _Ni) be ring neural network prediction controller equation in the new wind ratio; According to the q that calculates _Ni[k], through type (19) and formula (20) are revised the weights of ring neural network prediction controller in the new wind ratio:

$\mathrm{\&Delta;}{W}_{\mathrm{ni}}=-{\mathrm{\&mu;}}_{\mathrm{ni}}\frac{\&PartialD;g_{\mathrm{ni}}{(k,W_{\mathrm{ni}})}^T}{\&PartialD;W_{\mathrm{ni}}}{q}_{\mathrm{ni}}\left[k\right]---\left(19\right)$

W _ni＝W _ni+ΔW _ni (20)

W wherein _NiThe weights battle array of ring neural network prediction controller in the new wind ratio, μ _NiThe right value update rate, μ _NiSelect 0.05.Constantly revise the weights of ring neural network prediction controller in the new wind ratio, until Δ W _Ni＜0.001;

With S[t ₁], Y[t ₁] and-1 new wind ratio that acts on after the optimizing in ring neural network prediction controller, obtain u[t ₁].Then with u[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u[t ₁] act on new wind air-valve forecast model, obtain

Afterwards with x[t ₁],

Act on the air conditioning area forecast model, obtain

Keep new wind ratio outer shroud neural network prediction controller weights constant, with x[t ₁+ 1], x ^*With-1 act on new wind ratio outer shroud neural network prediction controller, obtain set amount s ' [t ₁+ 1]; Carry out again interior ring optimizing, with s ' [t ₁+ 1],

With ring neural network prediction controller in-1 new wind ratio that acts on after the optimizing, obtain u ' [t ₁+ 1].Again with u ' [t ₁+ 1] acts on new wind air-valve forecast model, obtain Afterwards, with x[t ₁+ 1],

Act on the air conditioning area forecast model, obtain Will

x ^*With-1 act on new wind ratio outer shroud neural network prediction controller, obtain set amount s ' [t ₁+ 2]; Carry out again interior ring optimizing, with s ' [t ₁+ 2],

With ring neural network prediction controller in-1 new wind ratio that acts on after the optimizing, obtain u ' [t ₁+ 2].Again with u ' [t ₁+ 2] act on new wind air-valve forecast model, obtain

Afterwards, will

Act on the air conditioning area forecast model, obtain

And the rest may be inferred, utilizes controller neutral net and object forecast model to extrapolate following u ' [t ₁+ i] and

Value, i=3 wherein ..., 6, and the data that calculate are preserved.Make λ _No[k+M _q]=0, the λ from backward front respectively calculating formula (21) _No[k] and q _No[k]:

$q_{\mathrm{no}}\left[k\right]=\frac{\&PartialD;f_n{\left(k\right)}^T}{\&PartialD;u\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{no}}[k+1]+\frac{\&PartialD;L_{\mathrm{no}}{\left(k\right)}^T}{\&PartialD;u\left[k\right]}$

${\mathrm{\&lambda;}}_{\mathrm{no}}\left[k\right]=\frac{\&PartialD;f_n{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{no}}[k+1]+\frac{\&PartialD;L_{\mathrm{no}}{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{\&PartialD;g_{\mathrm{no}}{(k;W_{\mathrm{no}})}^T}{\&PartialD;x\left[k\right]}{q}_{\mathrm{no}}\left[k\right]---\left(21\right)$

F in the formula _n(k) the Air Quality Forecast model of setting up before the representative; g _No(k; W _No) be new wind ratio outer shroud neural network prediction controller equation.According to the q that calculates _No[k] is for k=t ₁+ M _q-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (22) and formula (23) are revised the weights of new wind ratio outer shroud neural network prediction controller:

$\mathrm{\&Delta;}{W}_{\mathrm{no}}=-{\mathrm{\&mu;}}_{\mathrm{no}}\underset{k=t_1}{\overset{{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HDA0000082058780000023.png,HDA0000082058780000033.png"\; state="\{\{state\}\}">t</figure-callout>}}_1+M_q-1}{\mathrm{\&Sigma;}}}\frac{\&PartialD;g_{\mathrm{no}}{(k,W_{\mathrm{no}})}^T}{\&PartialD;W_{\mathrm{no}}}{q}_{\mathrm{no}}\left[k\right]---\left(22\right)$

W _no＝W _no+ΔW _no (23)

W wherein _NoThe weights battle array of new wind ratio outer shroud neural network prediction controller, μ _NoThe right value update rate, μ _NoSelect 0.05.Constantly revise the weights of new wind ratio outer shroud neural network prediction controller, until Δ W _No＜0.001; After the outer shroud optimizing finishes, with x[t ₁] and x ^*Input new wind ratio outer shroud neural network prediction controller is again with the S[t that obtains ₁] and Y[t ₁] the interior ring of input new wind ratio neural network prediction controller, will export u[t ₁] directly act on new wind air-valve;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

Claims (8)

1. a VAV Control Method of Vav Air Conditioning System is characterized in that, comprises following step:

The first step: utilize BP neural network air conditioning area temperature prediction model, terminal air-valve forecast model, air-conditioning unit main air duct duct static pressure forecast model, main air duct air quantity forecast model, wind pushing temperature forecast model, new wind air-valve forecast model and Air Quality Forecast model

1) determines air conditioning area temperature prediction model, terminal air-valve forecast model, air-conditioning unit main air duct duct static pressure forecast model, main air duct air quantity forecast model, wind pushing temperature forecast model, new wind air-valve forecast model and Air Quality Forecast model structure

The input signal of air conditioning area temperature prediction model is outdoor intensity of solar radiation, outdoor temperature, CO2 concentration, indoor temperature, air quantity and terminal valve opening, is output as next constantly indoor temperature;

The input signal of terminal air-valve forecast model is terminal valve area and air-conditioning unit main air duct duct static pressure, is output as next constantly terminal air quantity;

The input signal of main air duct duct static pressure forecast model is VAV air-conditioning unit rotation speed of fan, duct static pressure, return air CO2 concentration, outdoor temperature and intensity of solar radiation, is output as next constantly duct static pressure;

The input signal of main air duct air quantity forecast model is rotation speed of fan and duct static pressure, is output as next constantly main air duct air quantity;

The input signal of wind pushing temperature forecast model is wind pushing temperature and water valve aperture, is output as next constantly wind pushing temperature;

The input signal of new wind air-valve forecast model is new wind valve area and fresh air pipeline duct static pressure, is output as next constantly resh air requirement;

The input signal of Air Quality Forecast model is new wind valve area and CO2 concentration, is output as next constantly CO2 concentration;

2) collecting sample data;

3) sample data is by formula carried out normalization in (1), (2):

$x_i=\frac{x_{\mathrm{di}}-x_{d\min }}{x_{d\max }-x_{d\min }}---\left(1\right)$

$y_{\mathrm{tl}}=\frac{y_{\mathrm{dl}}-y_{d\min }}{y_{d\max }-y_{d\min }}---\left(2\right)$

X wherein _iBe the input value of neutral net after the normalization, x _DiBe former input value, x _DminBe the minimum of a value in the former input value, x _DmaxBe the maximum in the former input value; y _TlBe the desired value of neutral net after the normalization, y _DlRepresent former desired value; y _DminRepresent the minimum of a value in the former desired value; y _DmaxBe the maximum in the former desired value;

4) neutral net is carried out off-line training;

Second step: VAV air conditioning terminal tandem PREDICTIVE CONTROL

1) determines terminal outer shroud predictive control function

Terminal outer shroud predictive control function is: $J_o\left[k\right]=\underset{k=t_1}{\overset{t_1+M_c-1}{\mathrm{\&Sigma;}}}{L}_o\left(T_o\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_1+M_c-1}{\mathrm{\&Sigma;}}}{\left(T_o\right[k]-T_{o,\mathrm{set}}[k\left]\right)}^2$

M wherein _cBe prediction time domain, t ₁The initial time in the prediction time domain, T _o[k] is the air conditioning area temperature in k sampling period, T _{O, set}[k] is the air conditioning area desired temperature in k sampling period, L _oIt is terminal outer shroud object function of k sampling period;

2) determine terminal interior ring predictive control function

The ring predictive control function is in terminal: $J_i\left[k\right]=\underset{k=t_1}{\overset{t_1+M_i-1}{\mathrm{\&Sigma;}}}{L}_i\left(V_i\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_1+M_i-1}{\mathrm{\&Sigma;}}}{\left(V_i\right[k]-V_{i,\mathrm{set}}[k\left]\right)}^2$

V wherein _i[k] is the air quantity in k sampling period, V _{I, set}[k] is the air quantity setting value in k sampling period; M _iBe prediction time domain, L _iIt is ring object function in k the sampling period end;

3) determine terminal interior ring neural network prediction controller and terminal outer shroud neural network prediction controller architecture

Terminal tandem control loop inner and outer ring controller is all selected the neural network prediction controller, and as input, valve opening is output to terminal interior ring neural network prediction controller with air quantity setting value, air-conditioning unit main air duct duct static pressure and pipeline air quantity; The input parameter of terminal outer shroud neural network prediction controller comprises outdoor temperature, intensity of solar radiation, indoor temperature and air conditioning area desired temperature; Output parameter is the air quantity setting value;

4) ring neural network prediction controller in terminal and terminal outer shroud neural network prediction controller are carried out the online optimizing training, according to the air conditioning area temperature that gathers, input outer shroud neural network prediction controller, calculate the setting air quantity, ring neural network prediction controller is adjusted terminal valve area in the recycling;

The 3rd step: VAV air-conditioning unit PREDICTIVE CONTROL

1) determines static pressure control loop, VAV air-conditioning unit total blast volume control loop, VAV air-conditioning unit wind pushing temperature control loop and the VAV air-conditioning unit new wind ratio tandem control loop predictive control function of VAV air-conditioning unit adjustable settings static pressure

The static pressure control loop predictive control function is:

$J_s\left[k\right]=\underset{k=t_1}{\overset{t_1+M_s-1}{\mathrm{\&Sigma;}}}{L}_s\left(P_s\right[k],U_{\mathrm{fan}}[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_1+M_s-1}{\mathrm{\&Sigma;}}}\{{\left(P_s\right[k]-P_{s,\mathrm{set}}[k\left]\right)}^2+U_{\mathrm{fan}}^2[k\left]\right\}$

P wherein _s[k] is the duct static pressure in k sampling period, P _{S, set}[k] is that the air conditioning area in k sampling period is set static pressure, M _sBe prediction time domain, U _Fan[k] is k sampling period blower voltage controlled quentity controlled variable, L _sBe k sampling period static pressure control loop object function;

Total blast volume control loop predictive control function is:

$J_f\left[k\right]=\underset{k=t_1}{\overset{t_1+M_f-1}{\mathrm{\&Sigma;}}}{L}_f\left(V_{\mathrm{fan}}\right[k],U_{\mathrm{fan}}[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_1+M_f-1}{\mathrm{\&Sigma;}}}\{{\left(V_{\mathrm{fan}}\right[k]-V_{\mathrm{fan}.\mathrm{set}}[k\left]\right)}^2+U_{\mathrm{fan}}^2[k\left]\right\}$

V wherein _Fan[k] is the pipeline air quantity in k sampling period, M _fBe prediction time domain, V _Fan.set[k] is each terminal prediction air quantity sum in k sampling period, L _fBe k sampling period total blast volume control loop object function;

Wind pushing temperature control loop predictive control function is:

$J_{\mathrm{st}}\left[k\right]=\underset{k=t_1}{\overset{t_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{st}}\left(T_{\mathrm{st}}\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}{\left(T_{\mathrm{st}}\right[k]-T_{\mathrm{st},\mathrm{set}}[k\left]\right)}^2$

T wherein _St[k] is the temperature in k sampling period, T _{St, set}[k] is the design temperature in k sampling period, M _StBe prediction time domain, L _StBe k sampling period wind pushing temperature control loop object function;

New wind ratio control loop outer shroud predictive control function is:

$J_q\left[k\right]=\underset{k=t_1}{\overset{t_1+M_q-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{no}}\left(Q\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_1+M_q-1}{\mathrm{\&Sigma;}}}{\left(Q\right[k]-Q_{\mathrm{set}}[k\left]\right)}^2$

Q[k wherein] be the air quality in k sampling period, Q _Set[k] is the air quality setting value in k sampling period, M _qBe prediction time domain, L _NoBe k sampling period new wind ratio control loop outer shroud object function;

The ring predictive control function is in the new wind ratio control loop:

$J_{\mathrm{ni}}\left[k\right]=\underset{k=t_1}{\overset{t_i+M_{\mathrm{ni}}-1}{\mathrm{\&Sigma;}}}{L}_{\mathrm{ni}}\left(V_{\mathrm{ni}}\right[k],k)=\frac{1}{2}\underset{k=t_1}{\overset{t_i+M_{\mathrm{ni}}-1}{\mathrm{\&Sigma;}}}{\left(V_{\mathrm{ni}}\right[k]-S[k\left]\right)}^2$

V wherein _i[k] is the prediction air-valve discharge quantity of fan in k sampling period, S[k] be the setting air quantity in k sampling period, M _iBe prediction time domain, L _NiIt is ring object function in k the sampling period new wind ratio control loop;

2) determine static pressure control loop, VAV air-conditioning unit total blast volume control loop, VAV air-conditioning unit wind pushing temperature control loop and the VAV air-conditioning unit new wind ratio tandem control loop neural network prediction controller architecture of VAV air-conditioning unit adjustable settings static pressure

The input signal of static pressure neural network prediction controller is output as rotation speed of fan for setting duct static pressure and duct static pressure;

Total blast volume neural network prediction controller input signal is pipeline air quantity and total blast volume, is output as rotation speed of fan;

The input signal of wind pushing temperature neural network prediction controller is output as the water valve aperture for setting wind pushing temperature and wind pushing temperature;

New wind ratio tandem control loop inner and outer ring controller is all selected the neural network prediction controller, and new wind ratio outer shroud neural network prediction controller is input as sets CO2 concentration and return air CO2 concentration, is output as the setting resh air requirement; Ring neural network prediction controller will be set air quantity, fresh air pipeline static pressure and pipeline air quantity as input in the new wind ratio, and new air valve aperture is output;

3) when the static pressure decline of system static pressure monitoring point reaches setting value, select the static pressure predictive control strategy of adjustable settings static pressure, realize static pressure neural network prediction controller online optimizing in the static pressure control loop;

4) when the static pressure of system static pressure monitoring point is higher than setting value, operation total blast volume predictive control strategy is realized total blast volume neural network prediction controller online optimizing in the total blast volume control loop;

5) realize wind pushing temperature neural network prediction controller online optimizing in the wind pushing temperature control loop;

6) realize encircling and outer shroud neural network prediction controller online optimizing in the new wind ratio in the new wind ratio tandem control loop, according to the return air CO2 concentration that gathers, input outer shroud neural network prediction controller, calculate the setting resh air requirement, ring neural network prediction controller is adjusted new wind valve area in the recycling.

2. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1 is characterized in that, the first step 2) the collecting sample data are specially:

The sampling time scope be 8 a.m. to 6 pm, sampling time interval 150 seconds, each forecast model gather about 2000 groups of data; Getting the time interval when setting up forecast model is 5 minutes;

Air conditioning area temperature prediction model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, simultaneously, terminal valve area also is divided into ten grades by 0V to 10V interval 1V, gathers outdoor intensity of solar radiation, outdoor temperature, CO2 concentration, indoor temperature and air quantity;

Terminal air-valve forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and simultaneously, terminal valve area also is divided into ten grades by 0V to 10V interval 1V, gathers air-conditioning unit main air duct duct static pressure and air quantity;

Main air duct duct static pressure forecast model: after each terminal debugging was finished, air-blower control amount signal all was divided into ten grades by 0V to 10V interval 1V, gathered VAV air-conditioning unit rotation speed of fan, duct static pressure, return air CO2 concentration, outdoor temperature, intensity of solar radiation;

Main air duct air quantity forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, gathers rotation speed of fan, duct static pressure, main air duct air quantity;

Wind pushing temperature forecast model: the water valve aperture is divided into ten grades by 0V to 10V interval 1V, gathers wind pushing temperature and water valve aperture;

New wind air-valve forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and simultaneously, new wind valve area also is divided into ten grades by 0V to 10V interval 1V, gathers fresh air pipeline duct static pressure and air quantity;

The Air Quality Forecast model: air-blower control amount signal all is divided into ten grades by 0V to 10V interval 1V, and is same, and new wind valve area also is divided into ten grades by 0V to 10V interval 1V, gathers CO2 concentration.

3. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1 is characterized in that, the first step 4) carry out off-line training according to table 1 pair neutral net:

Table 1 Neural Network Training Parameter table

The neutral net type Single hidden layer forward direction BP network The input layer number A* The output layer nodes 1 The number of hidden nodes B*

The neuron excitation function Hidden layer ' tansig', output layer ' purelin' Learning function 'learngdm' Performance function 'msereg' The network training function 'trainbr' Power (threshold) value initialization method 'initnw' The maximum training time 2000Epochs Target error 0

A*: air conditioning area temperature prediction model value 6, terminal air-valve forecast model, Air Quality Forecast model, wind pushing temperature forecast model, main air duct air quantity forecast model and new wind air-valve forecast model value 2, main air duct duct static pressure forecast model value 5;

B*：

4. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1, it is characterized in that, second step 4) be specially: initialize terminal in ring be connected each connection weights with the outer shroud nerve network controller, assignment is [1,1] less random number in the scope, then computing controller output; Prediction time domain M _iAnd M _oBe respectively 3 and 6, predetermined period was got 5 minutes;

Searching process is: establish x[k] be constantly the relevant state variables parameter of outer shroud controlled device (being air conditioning area), i.e. indoor temperature of k;

Be constantly forecast model output of k+1, i.e. k+1 Indoo r prediction temperature constantly; X* is desired temperature; S[k] be constantly air quantity setting value of k; Y[k] be constantly interior ring relevant state variables parameter, i.e. air quantity of k; C[k] be the k terminal air-valve valve opening of the moment that terminal interior ring neural network prediction controller calculates; U[k] be that the terminal air-valve valve opening that i.e. optimizing obtains was exported in k control constantly after the optimizing of terminal interior ring neural network prediction controller finished;

The k+1 that exports for valve end air-valve forecast model predicts air quantity constantly, with x[t ₁], x* and-1 acts on terminal outer shroud neural network prediction controller, obtains set amount S[t ₁], with S[t ₁], Y[t ₁] and-1 act on terminal in ring neural network prediction controller, obtain C[t ₁], then with C[t ₁] act on terminal air-valve forecast model, obtain predicting air quantity Keep terminal interior ring neural network prediction controller weights constant, will

S[t ₁] and-1 act on terminal in ring neural network prediction controller, obtain C ' [t ₁+ 1], with C ' [t ₁+ 1] the terminal air-valve forecast model of input obtains

Will

S[t ₁] and-1 act on terminal in ring neural network prediction controller, obtain C ' [t ₁+ 2]; With C ' [t ₁+ 2] the terminal air-valve forecast model of input obtains

The data that calculate are preserved, make λ _i[k+M _i]=0, the λ from backward front respectively calculating formula (3) _i[k] and q _i[k]:

$q_i\left[k\right]=\frac{{\&PartialD;f}_v{\left(k\right)}^T}{\&PartialD;C\left[k\right]}{\mathrm{\&lambda;}}_i[k+1]+\frac{{\&PartialD;L}_i{\left(K\right)}^T}{\&PartialD;C\left[k\right]}$ （3）

${\mathrm{\&lambda;}}_i\left[k\right]=\frac{{\&PartialD;f}_v{\left(k\right)}^T}{\&PartialD;y\left[k\right]}{\mathrm{\&lambda;}}_i[k+1]+\frac{{\&PartialD;L}_i{\left(k\right)}^T}{\&PartialD;y\left[k\right]}+\frac{{\&PartialD;g}_i{(k;W_i)}^T}{\&PartialD;y\left[k\right]}{q}_i\left[k\right]$

F in the formula _v(k) the terminal air-valve forecast model of setting up before the representative; g _i(k; W _i) be terminal interior ring neural network prediction controller equation; According to the q that calculates _i[k], through type (4) and formula (5) are revised the weights of terminal interior ring neural network prediction controller:

${\mathrm{\&Delta;W}}_i=-{\mathrm{\&mu;}}_i\frac{{\&PartialD;g}_i{(k,W_i)}^T}{{\&PartialD;W}_i}{q}_i\left[k\right]---\left(4\right)$

W _i＝W _i+ΔW _i (5)

W wherein _iThe weights battle array of terminal interior ring neural network prediction controller, μ _iThe right value update rate, μ _iSelect 0.05, constantly revise the weights of terminal interior ring neural network prediction controller, until Δ W _i＜0.001;

With S[t ₁], Y[t ₁] and-1 act on ring neural network prediction controller in terminal after the optimizing, obtain u ' [t ₁], with u'[t ₁] act on terminal air-valve forecast model, obtain

Afterwards with x[t ₁],

Act on the air conditioning area forecast model, obtain

Keep terminal outer shroud neural network prediction controller weights constant, will

X* and-1 acts on terminal outer shroud neural network prediction controller, obtains set amount s ' [t ₁+ 1]; Carry out again interior ring optimizing, with s ' [t ₁+ 1],

With-1 act on ring neural network prediction controller in terminal after the optimizing, obtain u ' [t ₁+ 1], again with u ' [t ₁+ 1] acts on terminal air-valve forecast model, obtain

Afterwards with x[t ₁+ 1] x[1], Act on the air conditioning area forecast model, obtain

Will

X* and-1 acts on terminal outer shroud neural network prediction controller, obtains set amount s ' [t ₁+ 2]; Carry out again interior ring optimizing, with s ' [t ₁+ 2],

With-1 act on ring neural network prediction controller in terminal after the optimizing, obtain u ' [t ₁+ 2], again with u ' [t ₁+ 2] act on terminal air-valve forecast model, obtain Afterwards, will

Act on the air conditioning area forecast model, obtain

And the rest may be inferred, utilizes controller neutral net and object forecast model to extrapolate following u ' [t ₁+ i] and Value, i=3 wherein ..., 6, and the data that calculate are preserved, λ made _o[k+M _o]=0, the λ from backward front respectively calculating formula (6) _o[k] and q _o[k]: $q_o\left[k\right]=\frac{{\&PartialD;f}_z{\left(k\right)}^T}{\&PartialD;u\left[k\right]}{\mathrm{\&lambda;}}_o[k+1]+\frac{{\&PartialD;L}_o{\left(k\right)}^T}{\&PartialD;u\left[k\right]}$ （6）

${\mathrm{\&lambda;}}_0\left[k\right]=\frac{{\&PartialD;f}_z{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_o[k+1]+\frac{{\&PartialD;L}_o{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{{\&PartialD;g}_o{(k;W_o)}^T}{\&PartialD;x\left[k\right]}{q}_o\left[k\right]$

F in the formula _z(k) the air conditioning area forecast model of setting up before the representative, g _o(k; W _o) be terminal outer shroud neural network prediction controller equation, according to the q that calculates _o[k] is for k=t ₁+ M _o-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (7) and formula (8) are revised the weights of terminal outer shroud neural network prediction controller:

${\mathrm{\&Delta;W}}_o=-{\mathrm{\&mu;}}_o\underset{k=t_1}{\overset{t_1+M_o-1}{\mathrm{\&Sigma;}}}\frac{{\&PartialD;g}_o{(k,W_o)}^T}{{\&PartialD;W}_o}{q}_o\left[k\right]---\left(7\right)$

W _o＝W _o+ΔW _o (8)

W wherein _oThe weights battle array of terminal outer shroud neural network prediction controller, μ _oThe right value update rate, μ _oSelect 0.05, constantly revise the weights of terminal outer shroud neural network prediction controller, until Δ W _o＜0.001; After the outer shroud optimizing finishes, with x[t ₁] and x* input terminal outer shroud neural network prediction controller, again with the S[t that obtains ₁] and Y[t ₁] the terminal interior ring neural network prediction controller of input, will export u[t ₁] directly act on terminal air-valve;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

5. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1, it is characterized in that, the 3rd step 3) be specially: each that initializes static pressure neural network prediction controller connects weights, and assignment is [1,1] less random number in the scope, then computing controller output; Prediction time domain M _sIt was 2 steps, for the right value update rate μ of static pressure neural network prediction controller _sSelect 0.05, predetermined period was got 5 minutes;

Searching process: establish x[k] parameter be duct static pressure, u _Fan[k] is the air-blower control amount, and x* sets duct static pressure, with x[t ₁], x* and-1 acts on static pressure neural network prediction controller, controlled amount u _Fan[t ₁], then with u _Fan[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u _Fan[t ₁], x[t ₁] and-1 act on main air duct duct static pressure forecast model, obtain

Keep static pressure neural network prediction controller weights constant, with x[t ₁+ 1], x* and-1 acts on controller, obtains u ' _Fan[t ₁+ 1], with u ' _Fan[t ₁+ 1], x[t ₁+ 1] and-1 act on main air duct duct static pressure forecast model, obtains The data that calculate are preserved, make λ _s[k+M _s]=0, the λ from backward front respectively calculating formula (9) _s[k] and q _s[k]: $q_s\left[k\right]=\frac{{\&PartialD;f}_s{\left(k\right)}^T}{{\&PartialD;u}_{\mathrm{fan}}\left[k\right]}{\mathrm{\&lambda;}}_s[k+1]+\frac{{\&PartialD;L}_s{\left(K\right)}^T}{{\&PartialD;u}_{\mathrm{fan}}\left[k\right]}$ （9）

${\mathrm{\&lambda;}}_s\left[k\right]=\frac{{\&PartialD;f}_s{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_s[k+1]+\frac{{\&PartialD;L}_s{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{{\&PartialD;g}_s{(k;W_s)}^T}{\&PartialD;x\left[k\right]}{q}_s\left[k\right]$

F in its Chinese style _s(k) the main air duct duct static pressure forecast model of setting up before the representative, g _s(k; W _s) be static pressure neural network prediction controller equation, according to the q that calculates _s[k] is for k=t ₁+ M _s-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (10) and formula (11) are revised the weights of static pressure neural network prediction controller:

${\mathrm{\&Delta;W}}_s=-{\mathrm{\&mu;}}_s\underset{k=t_1}{\overset{t_1+M_s-1}{\mathrm{\&Sigma;}}}\frac{{\&PartialD;g}_s{(k,W_s)}^T}{{\&PartialD;W}_s}{q}_s\left[k\right]---\left(10\right)$

W _s＝W _s+ΔW _s (11)

W wherein _sThe weights battle array of static pressure neural network prediction controller, μ _sThe right value update rate, μ _sSelect 0.05, constantly revise the weights of static pressure neural network prediction controller, until Δ W _s＜0.001;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

6. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1, it is characterized in that, the 3rd step 4) be specially: each that initializes total blast volume neural network prediction controller connects weights, and assignment is [1,1] less random number in the scope, then computing controller output; Prediction time domain M _fIt was 2 steps, for the right value update rate μ of total blast volume neural network prediction controller _fSelect 0.05, predetermined period was got 5 minutes;

Searching process: establish x[k] parameter be the main air duct air quantity, u _Fan[k] is the air-blower control amount, and x* is each regional air quantity setting value sum, with x[t ₁], x* and-1 acts on total blast volume neural network prediction controller, controlled amount u _Fan[t ₁], then with u _Fan[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u _Fan[t ₁], x[t ₁] and-1 act on main air duct air quantity forecast model, obtain

Keep total blast volume neural network prediction controller weights constant, with x[t ₁+ 1], x* and-1 acts on controller, obtains u ' _Fan[t ₁+ 1], with u ' _Fan[t ₁+ 1], x[t ₁+ 1] and-1 act on main air duct air quantity forecast model, obtains

The data that calculate are preserved, make λ _f[k+M _f]=0, the λ from backward front respectively calculating formula (12) _f[k] and q _f[k]:

$q_f\left[k\right]=\frac{{\&PartialD;f}_f{\left(k\right)}^T}{{\&PartialD;u}_{\mathrm{fan}}\left[k\right]}{\mathrm{\&lambda;}}_f[k+1]+\frac{{\&PartialD;L}_f{\left(K\right)}^T}{{\&PartialD;u}_{\mathrm{fan}}\left[k\right]}$ （12）

${\mathrm{\&lambda;}}_f\left[k\right]=\frac{{\&PartialD;f}_f{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_f[k+1]+\frac{{\&PartialD;L}_f{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{{\&PartialD;g}_f{(k;W_f)}^T}{\&PartialD;x\left[k\right]}{q}_f\left[k\right]$

F wherein _f(k) the main air duct air quantity forecast model of setting up before the representative, g _f(k; W _f) be total blast volume neural network prediction controller equation, according to the q that calculates _f[k] is for k=t ₁+ M _f-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (13) and formula (14) are revised the weights of total blast volume neural network prediction controller:

${\mathrm{\&Delta;W}}_f=-{\mathrm{\&mu;}}_f\underset{k=t_1}{\overset{t_1+M_f-1}{\mathrm{\&Sigma;}}}\frac{{\&PartialD;g}_f{(k,W_f)}^T}{{\&PartialD;W}_f}{q}_f\left[k\right]---\left(13\right)$

W _f＝W _f+ΔW _f (14)

W wherein _fThe weights battle array of total blast volume neural network prediction controller, μ _fThe right value update rate, μ _fSelect 0.05, constantly revise the weights of total blast volume neural network prediction controller, until Δ W _f＜0.001;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

7. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1, it is characterized in that, the 3rd step 5) be specially: each that initializes wind pushing temperature neural network prediction controller connects weights, and assignment is [1,1] less random number in the scope, then computing controller output; Prediction time domain M _StIt was 2 steps, for the right value update rate μ of wind pushing temperature neural network prediction controller _StSelect 0.05, predetermined period was got 5 minutes;

Searching process: establish x[k] be wind pushing temperature, u[k] be the water valve aperture, x* sets wind pushing temperature, control system detects all terminal VAV-BOX primary air flows, when a certain VAV-BOX primary air flow of appearance is lower than nominal air delivery 30%, reduce the water valve aperture, will set wind pushing temperature x* and improve 0.5 ℃; When a certain terminal VAV-BOX primary air flow greater than 70% the time, increase the water valve aperture, will set 0.5 ℃ of wind pushing temperature x* reduction;

With x[t ₁], x* and-1 acts on wind pushing temperature neural network prediction controller, controlled amount u[t ₁], then with u[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u[t ₁], x[t ₁] and-1 act on the wind pushing temperature forecast model, obtain

Keep wind pushing temperature neural network prediction controller weights constant, with x[t ₁+ 1], x* and-1 acts on controller, obtains u'[t ₁+ 1], with u'[t ₁+ 1], x[t ₁+ 1] and-1 act on the wind pushing temperature forecast model, obtains

The data that calculate are preserved, make λ _St[k+M _St]=0, the λ from backward front respectively calculating formula (15) _St[k] and q _St[k]:

$q_{\mathrm{st}}\left[k\right]=\frac{{\&PartialD;f}_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;u\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{st}}[k+1]+\frac{{\&PartialD;L}_{\mathrm{st}}{\left(K\right)}^T}{\&PartialD;u\left[k\right]}$ （15）

${\mathrm{\&lambda;}}_{\mathrm{st}}\left[k\right]=\frac{{\&PartialD;f}_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{st}}[k+1]+\frac{{\&PartialD;L}_{\mathrm{st}}{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{{\&PartialD;g}_{\mathrm{st}}{(k;W_{\mathrm{st}})}^T}{\&PartialD;x\left[k\right]}{q}_{\mathrm{st}}\left[k\right]$

F wherein _St(k) the wind pushing temperature forecast model of setting up before the representative, g _St(k; W _St) be wind pushing temperature neural network prediction controller equation, according to the q that calculates _St[k] is for k=t ₁+ M _St-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (16) and formula (17) are revised the weights of wind pushing temperature neural network prediction controller:

${\mathrm{\&Delta;W}}_{\mathrm{st}}=-{\mathrm{\&mu;}}_{\mathrm{st}}\underset{k=t_1}{\overset{t_1+M_{\mathrm{st}}-1}{\mathrm{\&Sigma;}}}\frac{{\&PartialD;g}_{\mathrm{st}}{(k,W_{\mathrm{st}})}^T}{{\&PartialD;W}_{\mathrm{st}}}{q}_{\mathrm{st}}\left[k\right]---\left(16\right)$

W _st＝W _st+ΔW _st (17)

W wherein _StThe weights battle array of wind pushing temperature neural network prediction controller, μ _StThe right value update rate, μ _StSelect 0.05, constantly revise the weights of wind pushing temperature neural network prediction controller, until Δ W _St＜0.001;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

8. a kind of VAV Control Method of Vav Air Conditioning System according to claim 1, it is characterized in that, the 3rd step 6) be specially and initialize in the new wind ratio ring and be connected each connection weights with outer shroud neural network prediction controller, assignment is less random number in [1,1] scope, then computing controller output, if output valve is all [0.7,1] in the scope, then use this group controller initial weight, otherwise random initializtion again; Prediction time domain M _NiAnd M _qBe respectively 3 and 6, for the right value update rate μ selection 0.05 of ring and outer shroud neural network prediction controller in the new wind ratio, predetermined period was got 5 minutes;

Searching process: establish x[k] be constantly CO2 concentration of k;

Be constantly forecast model output of k, i.e. k+1 return air CO2 concentration constantly;

Be t ₁Return air CO2 concentration constantly; X* is the CO2 concentration set point; S[k] constantly set resh air requirement for k; Y[k] be that k encircles the relevant state variables parameter constantly, comprise pipeline air quantity and duct static pressure; O[k] the constantly new wind valve area of the k that calculates for interior ring neural network prediction controller; U[k] be k control output constantly after interior ring neural network prediction controller optimizing finishes, i.e. the new wind valve area that optimizing obtains;

The k+1 that exports for new wind air-valve forecast model predicts air quantity constantly, with x[t ₁], x* and-1 acts on new wind ratio outer shroud neural network prediction controller, obtains set amount S[t ₁], with S[t ₁], Y[t ₁] and-1 act on ring neural network prediction controller in the new wind ratio, obtain O[t ₁], then with O[t ₁] act on respectively new wind air-valve forecast model, obtain predicting air quantity It is constant to keep encircling neural network prediction controller weights in the new wind ratio, will

S[t ₁] and-1 act on ring neural network prediction controller in the new wind ratio, obtain O ' [t ₁+ 1], with O ' [t ₁+ 1] the new wind air-valve forecast model of input obtains

Will S[t ₁] and-1 act on that ring neural network prediction controller obtains O ' [2] in the new wind ratio, with the new wind air-valve forecast model of O ' [2] input, obtain The data that calculate are preserved, make λ _Ni[k+2]=0, the λ from backward front respectively calculating formula (18) _Ni[k] and q _Ni[k]:

$q_{\mathrm{ni}}\left[k\right]=\frac{{\&PartialD;f}_{\mathrm{nv}}{\left(k\right)}^T}{\&PartialD;O\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{ni}}[k+1]+\frac{{\&PartialD;L}_{\mathrm{ni}}{\left(k\right)}^T}{\&PartialD;O\left[k\right]}$ （18）

${\mathrm{\&lambda;}}_{\mathrm{ni}}\left[k\right]=\frac{{\&PartialD;f}_{\mathrm{nv}}{\left(k\right)}^T}{\&PartialD;y\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{ni}}[k+1]+\frac{{\&PartialD;L}_{\mathrm{ni}}{\left(k\right)}^T}{\&PartialD;y\left[k\right]}+\frac{{\&PartialD;g}_{\mathrm{ni}}{(k;W_{\mathrm{ni}})}^T}{\&PartialD;v\left[k\right]}{q}_{\mathrm{ni}}\left[k\right]$

F in the formula _Nv(k) the new wind air-valve forecast model of setting up before the representative; g _Ni(k; W _Ni) be ring neural network prediction controller equation in the new wind ratio; According to the q that calculates _Ni[k], through type (19) and formula (20) are revised the weights of ring neural network prediction controller in the new wind ratio:

${\mathrm{\&Delta;W}}_{\mathrm{ni}}=-{\mathrm{\&mu;}}_{\mathrm{ni}}\frac{{\&PartialD;g}_{\mathrm{ni}}{(k,W_{\mathrm{ni}})}^T}{{\&PartialD;W}_{\mathrm{ni}}}{q}_{\mathrm{ni}}\left[k\right]---\left(19\right)$

W _ni＝W _ni+ΔW _ni (20)

W wherein _NiThe weights battle array of ring neural network prediction controller in the new wind ratio, μ _NiThe right value update rate, μ _NiSelect 0.05, constantly revise the weights of ring neural network prediction controller in the new wind ratio, until Δ W _Ni＜0.001;

With S[t ₁], Y[t ₁] and-1 new wind ratio that acts on after the optimizing in ring neural network prediction controller, obtain u[t ₁], then with u[t ₁] act on controlled device, obtain x[t ₁+ 1], again with u[t ₁] act on new wind air-valve forecast model, obtain Afterwards with x[t ₁],

Act on the air conditioning area forecast model, obtain Keep new wind ratio outer shroud neural network prediction controller weights constant, with x[t ₁+ 1], x* and-1 acts on new wind ratio outer shroud neural network prediction controller, obtains set amount s ' [t ₁+ 1]; Carry out again interior ring optimizing, with s ' [t ₁+ 1],

With ring neural network prediction controller in-1 new wind ratio that acts on after the optimizing, obtain u ' [t ₁+ 1], again with u ' [t ₁+ 1] acts on new wind air-valve forecast model, obtain

Afterwards, with x[t ₁+ 1], Act on the air conditioning area forecast model, obtain

Will

X* and-1 acts on new wind ratio outer shroud neural network prediction controller, obtains set amount s ' [t ₁+ 2]; Carry out again interior ring optimizing, with s ' [t ₁+ 2],

With ring neural network prediction controller in-1 new wind ratio that acts on after the optimizing, obtain u ' [t ₁+ 2], again with u ' [t ₁+ 2] act on new wind air-valve forecast model, obtain

Afterwards, will

Act on the air conditioning area forecast model, obtain

And the rest may be inferred, utilizes controller neutral net and object forecast model to extrapolate following u ' [t ₁+ i] and

Value, i=3 wherein ..., 6, and the data that calculate are preserved, λ made _No[k+M _q]=0, the λ from backward front respectively calculating formula (21) _No[k] and q _No[k]:

$q_{\mathrm{no}}\left[k\right]=\frac{{\&PartialD;f}_n{\left(k\right)}^T}{\&PartialD;u\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{no}}[k+1]+\frac{{\&PartialD;L}_{\mathrm{no}}{\left(k\right)}^T}{\&PartialD;u\left[k\right]}$ （21）

${\mathrm{\&lambda;}}_{\mathrm{no}}\left[k\right]=\frac{{\&PartialD;f}_n{\left(k\right)}^T}{\&PartialD;x\left[k\right]}{\mathrm{\&lambda;}}_{\mathrm{no}}[k+1]+\frac{{\&PartialD;L}_{\mathrm{no}}{\left(k\right)}^T}{\&PartialD;x\left[k\right]}+\frac{{\&PartialD;g}_{\mathrm{no}}{(k;W_{\mathrm{no}})}^T}{\&PartialD;x\left[k\right]}{q}_{\mathrm{no}}\left[k\right]$

F in the formula _n(k) the Air Quality Forecast model of setting up before the representative, g _No(k; W _No) be new wind ratio outer shroud neural network prediction controller equation, according to the q that calculates _No[k] is for k=t ₁+ M _q-1 ..., t ₁+ 2, t ₁+ 1, t ₁, through type (22) and formula (23) are revised the weights of new wind ratio outer shroud neural network prediction controller:

${\mathrm{\&Delta;W}}_{\mathrm{no}}=-{\mathrm{\&mu;}}_{\mathrm{no}}\underset{k=t_1}{\overset{t_1+M_q-1}{\mathrm{\&Sigma;}}}\frac{{\&PartialD;g}_{\mathrm{no}}{(k,W_{\mathrm{no}})}^T}{{\&PartialD;W}_{\mathrm{no}}}{q}_{\mathrm{no}}\left[k\right]---\left(22\right)$

W _no＝W _no+ΔW _no (23)

W wherein _NoThe weights battle array of new wind ratio outer shroud neural network prediction controller, μ _NoThe right value update rate, μ _NoSelect 0.05, constantly revise the weights of new wind ratio outer shroud neural network prediction controller, until Δ W _No＜0.001; After the outer shroud optimizing finishes, with x[t ₁] and x* input new wind ratio outer shroud neural network prediction controller, again with the S[t that obtains ₁] and Y[t ₁] the interior ring of input new wind ratio neural network prediction controller, will export u[t ₁] directly act on new wind air-valve;

The next sampling period then repeats aforesaid operations, and the value of each moment controlled quentity controlled variable after calculating respectively is until control procedure finishes.

Images