CN104833154A - Chilled water loop control method based on fuzzy PID and neural internal model - Google Patents

Abstract

The invention discloses a chilled water loop control method based on a fuzzy PID and a neural internal model. An air conditioner performs real-time control over a compressor through the method after collecting the water outlet temperature of chilled water, wherein the compressor is a controlled object in a chilled water loop. The method includes the steps that step1, a target value of the controlled object is set; step2, a fuzzy PID controller and a neural network estimator together form internal model control; step3, the controlled object is trained through the neural network estimator to be a controlled object model in an ideal state; step4, deviation is adjusted through the fuzzy PID controller, finally, self-adaptation control over the chilled water loop of a variable-air-volume air conditioner system is achieved, and stable output of the controlled object is achieved. The fuzzy PID and the neural internal model control are combined, the defects of nonlinearity and time varying in the chilled water loop can be overcome, and accurate and rapid control is achieved.

Description

Based on the chilled water circuit control method of fuzzy and neural internal mold

Technical field

The present invention relates to a kind of chilled water circuit control method based on fuzzy and neural internal mold, belong to air quantity variable air conditioner technical field.

Background technology

Since the sixties in 20th century, air conditioning system with variable is born in the U.S., because the features such as it is energy-conservation, comfortable are widely used in various Working Life place, but due to the aggravation of energy crisis, makes people more and more recognize energy-conservation necessity.As the attention of air conditioning system with variable nature owing to attracting numerous scholar occupying main status in building energy consumption, the final purpose of air conditioning system with variable is the requirement making the humiture in room reach people, and the Energy Saving Control therefore how realizing air-conditioning system under the prerequisite meeting the requirement of people's comfort level becomes the hot issue of current field of air conditioning.

As everyone knows, wind can be divided in air conditioning system with variable to circulate and water circulation two parts, and the important component part of chilled water circuit just in water circulation, chilled water through evaporator laggard enter surface cooler carry out heat exchange with the air being about to send into room, by controlling the air themperature in wind pushing temperature conditioning chamber, finally reach the requirement meeting people's comfort level.Therefore chilled water seems particularly important at the leaving water temperature after evaporator in air conditioning system with variable, but due to chilled water circuit have non-linear, time the feature such as change, conventional control mode can not play good control effects.

Summary of the invention

This invention proposes a kind of chilled water circuit control method based on fuzzy and neural internal mold, is combined by fuzzy with Neural network internal model control, can overcome in chilled water circuit non-linear, time the feature such as change, realization controls object accurately and rapidly.

Technical scheme of the present invention is as follows:

A kind of chilled water circuit control method based on fuzzy and neural internal mold, after it is characterized in that air-conditioning collects chilled water leaving water temperature, controlled in real time compressor by method, compressor is the controlled device in chilled water circuit, and described method comprises the steps:

(1) the desired value y of controlled device, is set _set;

(2), fuzzy controller and neural network estimator form internal model control jointly;

(3), controlled device becomes plant model ideally through the training of neural network estimator;

(4), when there is mismatch in controlled device or receives external disturbance d, the measured value y of chilled water leaving water temperature and the setting value y of chilled water leaving water temperature _mbetween can produce deviation, i.e. Δ y, this deviation is imported the input of chilled water control loop, again by fuzzy controller for deviation e, adjust, finally realize the Self Adaptive Control of air conditioning system with variable chilled water circuit, realize the stable output of controlled device.

In above-mentioned steps (3), neural network estimator uses BP neutral net, adopts three-layer network, i.e. input layer, hidden layer, output layer; Calculation procedure is:

(21), initialize: be minimum random number by all connection weight value coefficient assignment;

(22) input value of the ideal model of chilled water circuit, is utilized to train;

(23) the actual output of neutral net, is calculated;

(24) the deviation delta y of the desired value of ideal model and the actual output of neutral net, is calculated;

(25) the weight coefficient w of output layer, is adjusted _ki;

(26) the weight coefficient w of hidden layer, is adjusted _ij;

(27), the 23rd step is returned, until error meets the demands.

The beneficial effect that the present invention reaches:

Fuzzy controller combines with internal model control by the present invention, be applied in the chilled water control loop of air conditioning system with variable, take full advantage of both advantages, can overcome non-linear in air-conditioning system, time the feature such as change, control compressor horsepower fast and accurately, maintain wind pushing temperature in setting value, finally make air-conditioned room can meet the comfortableness requirement of people.

Have the following advantages compared to the control mode in traditional chilled water leaving water temperature loop:

1. can overcome non-linear, the time-varying characteristics of air-conditioning system;

2. fuzzy controller and neural network estimator all have certain adaptive ability, maintain the stable of control system when can there is random perturbation in the external world;

3. avoid the parameter tuning problem of traditional PID control;

4. because final controlled device is compressor, so adopt the control algolithm in the present invention to produce larger energy-saving effect.

Accompanying drawing explanation

Fig. 1 is air conditioning system with variable chilled water circuit;

Fig. 2 is the schematic diagram of chilled water circuit control method;

Fig. 3 is the workflow diagram of the inventive method;

Fig. 4 is the structure chart of BP neutral net.

Detailed description of the invention

Below in conjunction with accompanying drawing, the invention will be further described.Following examples only for technical scheme of the present invention is clearly described, and can not limit the scope of the invention with this.

As shown in Figure 1, chilled water circuit in air conditioning system with variable comprises multiple links such as evaporimeter, chilled water pump, surface cooler, main operation principle is after setting chilled water leaving water temperature, the power of compressor can be regulated, maintain chilled water effluent temperature constancy, and then enable wind pushing temperature maintain setting value, ensure that air-conditioning system can be stablized, continuous print runs.In addition because the power of compressor obtains adjustment, also energy-conservation effect is served further.Chilled water circuit control algolithm needs the chilled water leaving water temperature collected, and controls in real time compressor, and controlled device real in visible chilled water circuit is compressor.According to engineering experience, the chilled water leaving water temperature control loop in air-conditioning system uses first order inertial loop with delay link as transfer function usually.

In engineering reality, the load variations of air conditioning system with variable can embody in the change of air output and wind pushing temperature, and chilled water directly contacts with the air in feeding room, therefore the size of chilled water leaving water temperature changing value can represent the size of load variations too, is realized by the change of frequency of compressor to the control of chilled water leaving water temperature in system.

As shown in Figure 2 and Figure 3, a kind of chilled water circuit control method based on fuzzy and neural internal mold, comprises the steps:

(1) the desired value y of controlled device, is set _set;

(2), fuzzy controller and neural network estimator form internal model control jointly;

(3), controlled device becomes plant model ideally through the training of neural network estimator;

(4), when there is mismatch in controlled device or receives external disturbance d, the measured value y of chilled water leaving water temperature and the setting value y of chilled water leaving water temperature _mbetween can produce deviation, i.e. Δ y, this deviation is imported the input of chilled water control loop, again by fuzzy controller for deviation e, adjust, finally realize the Self Adaptive Control of air conditioning system with variable chilled water circuit, realize the stable output of controlled device.

In above-mentioned steps (3), the training of neural network estimator is as follows:

The neural network estimator adopted in this invention uses more common BP neutral net, adopt three-layer network, i.e. input layer, hidden layer, output layer, because BP neutral net is a kind of training algorithm utilizing error back propagation, constantly adjust the connection weights of each layer, finally overcome the feature such as non-linear and to idealize accurately the characteristic of controlled device.

The general principle of BP Learning Algorithm is gradient steepest descent method, uses gradient search technology, and make the error mean square of the real output value of network and desired output difference for minimum, while running medial error back transfer, finishing connects weights.

The structure of BP neutral net as shown in Figure 4.Can see there be M input node in Fig. 4, L output node, the hidden layer of network has q neuron, wherein x ₁, x ₂..., x _mbe respectively the input of BP neutral net, y ₁, y ₂..., y _mfor the output of BP neutral net, y _setfor the target of neutral net exports, e _kfor the output error of neutral net.

At the learning phase of training network, controlled device uses the ideal model of chilled water circuit:

$G\left(s\right)=\frac{{\mathrm{Ke}}^{-\mathrm{\τ}s}}{(T_{1}s+1)}\mathrm{\τ}$

K, T in formula ₁be respectively proportionality coefficient and time constant, τ then represents delay, and the ideal model of the chilled water circuit in air-conditioning system is the one order inertia transfer function with pure delay link, and the current airconditioning control field that is chosen at of this model gains public acceptance.

Suppose after N number of training sample, neutral net can represent this ideal model accurately, and input, the output mode of first getting one of them sample P are trained network, hidden layer i-th neuron being input as under sample p effect:

${\mathrm{net}}_i^p=\underset{j=1}{\overset{M}{\Σ}}{w}_{ij}{o}_j^p-{\mathrm{\θ}}_i=\underset{j=1}{\overset{M}{\Σ}}{w}_{ij}{x}_j^p-{\mathrm{\θ}}_i$

In formula with be respectively the input and output of input node j under sample p effect, the activation primitive due to input layer is 1, and the node for input layer is equal; w _ijthe connection weights between input node j and hidden layer neuron i, wherein i=1,2 ..., q; θ _ifor the threshold value of hidden layer neuron i; M is the quantity of input layer;

Hidden layer i-th neuronic output is:

$o_j^p=g\left({\mathrm{net}}_i^p\right)$

The activation primitive that in formula, g () is hidden layer, selects Sigmoid function, that is:

$g\left(x\right)=\frac{1}{1+\exp (-x)}$

To hidden layer activation primitive can obtain after differentiating:

$g^{\′}\left({\mathrm{net}}_i^p\right)=g\left({\mathrm{net}}_i^p\right)\[1-g\left({\mathrm{net}}_i^p\right)\]=o_i^p(1-o_i^p)$

I-th neuronic output of hidden layer using by the input one of of weight coefficient forward direction to an output layer kth neuron as him, the kth of output layer is individual to be neuronicly always input as:

${\mathrm{net}}_k^p=\underset{i=1}{\overset{q}{\Σ}}{w}_{ki}{o}_i^p-{\mathrm{\θ}}_k$

K=1 in formula, 2 ..., L, w _kifor the connection weights between hidden layer neuron i and output layer neuron k; θ _kit is the threshold value of output layer neuron k; Q is the nodes of hidden layer;

A kth neuronic reality of output layer exports and is:

$o_k^p=g\left({\mathrm{net}}_k^p\right)$

Output layer activation primitive differentiation function be:

$g^{\′}\left({\mathrm{net}}_k^p\right)=g\left({\mathrm{net}}_k^p\right)\[1-g\left({\mathrm{net}}_k^p\right)\]=o_k^p(1-o_k^p)$

If the output valve of the output layer of neutral net and the output valve y of ideal model _kinconsistent, then deviation signal backpropagation returned, the connection weights between many each layers adjust, until output layer neuron obtains desired output valve, and known sample is trained one by one, until complete the training of N number of sample.

The error function produced for the input pattern of arbitrary sample p is:

$J_p=\frac{1}{2}\underset{k=1}{\overset{L}{\Σ}}{(y_k^p-o_k^p)}^2$

Overall error function then for all N number of samples is:

$J=\underset{p=1}{\overset{N}{\Σ}}{J}_p=\frac{1}{2}\underset{p=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{L}{\Σ}}{(y_k^p-o_k^p)}^2---\left(8\right)$

Because error is back transfer, so the adjustment of connection weights is by output layer, neutral net is finally restrained.According to gradient descent method, the neuron of output layer connects weights finishing formula and is:

${\mathrm{\Δw}}_{ki}=-\mathrm{\η}\frac{\∂J_p}{\∂w_{ki}}=-\mathrm{\η}\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}\·\frac{\∂{\mathrm{net}}_k^p}{\∂w_{ki}}=-\mathrm{\η}\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}\·\frac{\∂}{\∂w_{ki}}\left(\underset{i=1}{\overset{q}{\Σ}}{w}_{ki}{o}_i^p-{\mathrm{\θ}}_k\right)=-\mathrm{\η}\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}{o}_i^p$

In formula, η is learning rate, η > 0.

Definition ${\mathrm{\δ}}_k^p=-\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}=-\frac{\∂J_p}{\∂o_k^p}\·\frac{\∂o_k^p}{\∂{\mathrm{net}}_k^p}=(y_k-o_k^p)\·g^{\′}\left({\mathrm{net}}_k^p\right)=(y_k-o_k^p)\·o_k^p(1-o_k^p)$

Therefore the correction formula of the connection weights of any neuron k of output layer is:

${\mathrm{\Δw}}_{ki}={\mathrm{\η\δ}}_k^p{o}_i^p=\mathrm{\η}(y_k-o_k^p)\·o_i^p(1-o_k^p)$

In formula represent the output of output node k when sample p effect; represent the output of hidden layer node i when sample p effect; y _krepresent the target output value of the output node k when sample p inputs.

According to gradient descent method, can learn that the correction formula of hidden layer each neuron connection weights is:

${\mathrm{\Δw}}_{ij}=-\mathrm{\η}\frac{\∂J_p}{\∂w_{ij}}=-\mathrm{\η}\frac{\∂J_p}{\∂{\mathrm{net}}_i^p}\·\frac{\∂{\mathrm{net}}_i^p}{\∂w_{ij}}=-\mathrm{\η}\frac{\∂J_p}{\∂{\mathrm{net}}_i^p}\·\frac{\∂}{\∂w_{ij}}\left(\underset{j=1}{\overset{M}{\Σ}}{w}_{ij}{o}_j^p-{\mathrm{\θ}}_i\right)=-\mathrm{\η}\frac{\∂J_p}{\∂{\mathrm{net}}_i^p}{o}_j^p$

In formula, η is learning rate, η > 0.

Definition ${\mathrm{\δ}}_i^p=-\frac{\∂J_p}{\∂{\mathrm{net}}_i^p}=-\frac{\∂J_p}{\∂o_i^p}\·\frac{\∂o_i^p}{\∂{\mathrm{net}}_i^p}=-\frac{\∂J_p}{\∂o_i^p}\·g^{\′}\left({\mathrm{net}}_i^p\right)=-\frac{\∂J_p}{\∂o_i^p}\·o_i^p(1-o_i^p)$

Change due to a unit output of hidden layer can affect the input of all output units be connected with this unit, namely

$\begin{array}{c}-\frac{\∂J_p}{\∂o_i^p}=-\underset{k=1}{\overset{L}{\Σ}}\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}\·\frac{\∂{\mathrm{net}}_k^p}{\∂o_i^p}=-\underset{k=1}{\overset{L}{\Σ}}\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}\·\frac{\∂}{\∂o_i^p}\left(\underset{i=1}{\overset{q}{\Σ}}{w}_{ki}{o}_i^p-{\mathrm{\θ}}_k\right)\\ =\underset{k=1}{\overset{L}{\Σ}}\left(-\frac{\∂J_p}{\∂{\mathrm{net}}_k^p}\right)\·w_{ki}=\underset{k=1}{\overset{L}{\Σ}}{\mathrm{\δ}}_k^p\·w_{ki}\end{array}$

Then therefore the connection weights correction formula of any neuron i of hidden layer is:

${\mathrm{\Δw}}_{ij}={\mathrm{\η\δ}}_i^p{o}_j^p=(\underset{k=1}{\overset{L}{\Σ}}{\mathrm{\δ}}_k^p\·w_{ki})\·o_i^p(1-o_i^p)o_j^p$

In formula, for the output of hidden layer node i when sample p effect; for the output of input node j when sample p effect, i.e. the input of input node j.

Visible in the learning process of BP neutral net, make error function J connect weights along the direction adjustment declining the fastest, can obtain output layer with the increment formula that any neuron k with i of hidden layer is connected weights under the effect of all samples is:

$w_{ki}(k+1)=w_{ki}\left(k\right)+\mathrm{\η}\underset{p=1}{\overset{N}{\Σ}}{\mathrm{\δ}}_k^p{o}_i^p$

$w_{ij}(k+1)=w_{ij}\left(k\right)+\mathrm{\η}\underset{p=1}{\overset{N}{\Σ}}{\mathrm{\δ}}_i^p{o}_i^p$

The calculation procedure of BP neural network estimator is:

A. initialize: be minimum random number by all connection weight value coefficient assignment;

B. the input value of the ideal model of chilled water circuit is utilized to train;

C. the actual output of neutral net is calculated;

D. the deviation delta y of the desired value of ideal model and the actual output of neutral net is calculated;

E. the weight coefficient w of output layer is adjusted _ki;

F. the weight coefficient w of hidden layer is adjusted _ij;

G. the 3rd step is returned, until error meets the demands.

Fuzzy-adaptation PID control in step (4), compared with regulatory PID control mode, has the advantages such as regulating time is short, steady-state behaviour good, strong robustness, can adapt to the change of air-conditioning system load better.The Fuzzy Self-adaptive PID structure of the present invention's design is documented in " CN201410403382-is based on fuzzy and the variable air rate room temp. control method with predictive control algorithm ", and the present invention does not describe in detail.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the prerequisite not departing from the technology of the present invention principle; can also make some improvement and distortion, these improve and distortion also should be considered as protection scope of the present invention.

Claims (2)

1. the chilled water circuit control method based on fuzzy and neural internal mold, after it is characterized in that air-conditioning collects chilled water leaving water temperature, controlled in real time compressor by method, compressor is the controlled device in chilled water circuit, and described method comprises the steps:

(1) the desired value y of controlled device, is set _set;

(2), fuzzy controller and neural network estimator form internal model control jointly;

(3), controlled device becomes plant model ideally through the training of neural network estimator;

(4), when there is mismatch in controlled device or receives external disturbance d, the measured value y of chilled water leaving water temperature and the setting value y of chilled water leaving water temperature _mbetween can produce deviation, i.e. Δ y, this deviation is imported the input of chilled water control loop, again by fuzzy controller for deviation e, adjust, finally realize the Self Adaptive Control of air conditioning system with variable chilled water circuit, realize the stable output of controlled device.

2. the chilled water circuit control method based on fuzzy and neural internal mold according to claim 1, it is characterized in that in described step (3), neural network estimator uses BP neutral net, adopt three-layer network, i.e. input layer, hidden layer, output layer; Calculation procedure is:

(21), initialize: be minimum random number by all connection weight value coefficient assignment;

(22) input value of the ideal model of chilled water circuit, is utilized to train;

(23) the actual output of neutral net, is calculated;

(24) the deviation delta y of the desired value of ideal model and the actual output of neutral net, is calculated;

(25) the weight coefficient w of output layer, is adjusted _ki;

(26) the weight coefficient w of hidden layer, is adjusted _ij;

(27), the 23rd step is returned, until error meets the demands.