You are on page 1of 7

TELKOMNIKA, Vol. 11, No. 9, September 2013, pp.

5030~5036
e-ISSN: 2087-278X

5030

The Predictive Control Method of VAV Air Conditioning
System
Jiejia Li*, Xiaofeng Li, Yang Cao

School of Information and Control Engineering, Shenyang Jianzhu University, 110168, Shenyang,
Liaoning, China
*Corresponding author, email: 814686683@qq.com

Abstract
Aiming at the characteristics which variable air volume air conditioning system is multi-variable,
nonlinear and uncertain system, normal fuzzy neural network is hard to meet the requirements which
dynamic control of multi-variable. In this paper, we put forward a recursive neural network predictive
control strategy based on wavelet neural network model. Through recursive wavelet neural network
predictor on line established controlled object’s mathematical model, and using Elman neural network
controller on line corrected information we get, thus to improve control effect. The simulation results show
that recursive wavelet neural network predictive control has stronger robustness and adaptive ability, high
control precision, better and reliable control effect and other advantages.
Keywords: variable air volume aie conditioning (VAV), recursive wavelet neural network (RWNN), elman
neural network, predictive control.
Copyright © 2013 Universitas Ahmad Dahlan. All rights reserved.

1. Introduction
Variable air volume (VAV) air conditioning system is a full air system of multi-variable
and nonlinear. In VAV air conditioning system, when the air conditioning load changes or indoor
air parameters’ set value changes, it will adjust air flow into the room automatically, and adjust
the air conditioning parameters to set value, to meet comfort and energy-saving requirements
[1]. Therefore, this paper puts forward a kind of recursive neural network predictive control
method based on T-S fuzzy model, to establish reference model online, to complete the
identification of system’s dynamic characteristics, and using neural network control method to
improve control effect.

2. The Working Principle of VAV Air Conditioning System

Figure 1. Typical VAV System Working Principle Diagram

Received March 14, 2013; Revised May 31, 2013; Accepted June 14, 2013

u (k ) is control output.Among them: r(k) is temperature setting . Structure Diagram of Control System 3. y * (k ) is predictive temperature. Application of Recursive Neural Network based on Wavelet Neural Network Model in Predictive Control of VAV System The VAV air conditioning control system structure is shown in Figure 2. In this network structure. e2 is the error of output and predictive value. and predict object’s future output. This paper will complete the design of predictor and controller. The controllers are composed of recursive wavelet neural network predictor and RBF neural network controller [3]. we need good control strategy to meet the needs of people [2]. In this paper mainly completes temperature control of system. So. Recursive Neural Network based on Wavelet Neural Network Model Predictor Design y * ( k  1) Wj H j ( k  1) Vkj ( k  1) Oi (k ) Wij Wkj x1 ( k ) xn (k ) x11 ( k  1) xkj ( k  1) x mm ( k  1) Figure 3. it regulate new air volume. Single Output Recursive Wavelet Neural Network In this paper the design of predictor adopts recursive wavelet neural network. the input of hidden layer is composed of The Predictive Control Method of VAV Air Conditioning System (Jiejia LI) . Its working principle is: According to control requirements. and then according to the future output to adjust control strategy online. put air volume into VAV box. The structure diagram is shown in Figure 3. e1 r (k ) e(k ) y  (k ) u (k ) Z1 y (k ) e2 Figure 2. 3. and then adjust VAV box to complete control function. The typical VAV system is shown in Figure 1. Structure Diagram of Multi-input. time-varying and uncertainty system. The predictor uses the characteristics which recursive neural network based on wavelet neural network model can effective control non-linear. e1 is the difference of control output. in order to achieve temperature control requirements. y (k ) is output temperature. e(k) is temperature deviation between room temperature and temperature setting.1. to improve the comfort and energy saving effect.e-ISSN: 2087-278X TELKOMNIKA  5031 VAV air conditioning system is mainly composed of VAV air conditioning units and VAV Box. to achieve optimal control of air conditioning [4].

so input of associated layer is expressed as x kj ( k  1)   H j ( k ) . time-varying. for VAV air conditioning system which has dynamic. we can get: j (k 1)  j (k) j (k) [j (k) j (k 1)] (4) ij (k 1)  ij (k) ij (k) [ij (k) ij (k 1)] (5) kj(k 1) kj(k) kj(k) [kj(k) kj(k 1)] (6)  j (k 1)   j (k)  j (k) [ j (k)  j (k 1)] (7) Among them: 6 m m E E y*(k1) j   *  [y(k1)y*(k1)](ijOi (k)kjVkj(k1)j ) j y (k1) j k1 j1 i1 (8) m n m E E y*(k 1) ij     *  [y(k 1)  y*(k 1)]j'(ijOi (k) kjVkj(k 1) j )Oi (k) ij y (k 1) ij j1 i1 k1 (9) TELKOMNIKA Vol. i  1. x2 (k) . at k  1 moment.Hidden layer and associated layer both have m neurons.  j is the weight between hidden layer and output layer. [5] so recursive network structure has good dynamic characteristics. this paper takes n=3. The input layer of network has n neurons.  is feedback gain [5].. 11. uncertain system. x3 (k ) . Oi (k) is input layer’s output.  ij is the weight between input layer andhidden layer..So. y ( k  1) and y * ( k  1) represent actual output and reference model output of predictor. nonlinear. At k moment. recursive wavelet neural network’s output is: m n m j 1 i 1 k 1 y* (k  1)    j ( ij Oi (k )   kjVkj (k  1)   j ) (1)  j is wavelet function’s translation coefficient.  kj is the weight between associated layer and hidden layer.n . No.2.  () is wavelet function. actual output delay y(k 1) of controller object and the error y(k)y*(k) of object’s actual output and predicted output. output of hidden layer is H j (k  1) . Take wavelet function:  ( x )  (1  x ) e 2  x2 2 (2) Define the error performance index function of this network: E  1 [ y ( k  1 )  y * ( k  1 )] 2 2 (3) Among them. The learning of parameters takes adapt gradient descent method. it has good control effect. September 2013: 5030 – 5036 . 9. neurons of input layer are x1(k). on behalf of output u(k) of neural network controller.5032 e-ISSN: 2087-278X  output of input layer and output of associated layer. Associated layer’s output is Vkj ( k  1) .

e-ISSN: 2087-278X TELKOMNIKA  5033 m n m E E y*(k1) kj   *  [y(k1)y*(k 1)]j'(ijOi (k)kjVkj(k 1)j )Vkj(k 1) kj y (k 1) kj j1 i1 k1 j   m n m E E y*(k 1)  *   [y(k 1)  y*(k 1)]j'(ijOi (k) kjVkj(k 1) j ) j y (k 1) j j1 i1 k1  is leaning rate. The Predictive Control Method of VAV Air Conditioning System (Jiejia LI) .1) .  (10) (11) both in the range of (0. and it can be thought as a step time-delay matrix. In addition to the input layer. y (k )  w( 3) x(k )   w( 2) w(1)     u (k  1) Figure 4. and  . The nonlinear state space expression described by the modified Elman neural network is: (13) xk   f W 1 xc k   W 2u k  1   xc k   xk  1  xc k  1   y k   g W 3 xk  For this paper.  is inertial parameter. 3. but Bp neural network has the drawbacks of the slow convergence speed and easily getting into local minimum. and (14) (15) x2 is error variation e . The output value is control value u (k ) [9]. hidden layer and output layer. It can simulate the high order system [8]. Elman network has also a unique structural unit which is used for memory of hidden-layer output at previous time.2. Elman Neural Network Controller Design At present. there are two input values. Structure of Modified Elman Neural Network x k   x k  1  x k  1 cl cl l Where xcl  k is the output of structural unit and (12) xl  k  is the output of hidden layer. And the basic Elman network can be improved by adding a self-feedback connection of a fixed-gain   [7]. Elman neural network could not only have fast learning speed but also not easily get into local minimum [6]. Bp neural network is widely used in the VAV air conditioning control field. and α is a self-feedback gain factor. As shown in Figure 4. x1 is error value e .

 i0   rd ( k  1)  y i ( k  1)  g i' ( ) .. 2. 2. 2. y ( k  1) is system’s actual output. Hence.. n. l  1.. n (22) (23) xc  k  depends on il1 ... Where define y  k 1 and r  k 1 as actual output and expected output respectively. 11.. m (21) i 1 Then.... E P   i0 x j  k  ij3 i  1. n  ij3  f j' · u q  k  1  (19) (20) Let. r m E P x j  k  0 3 1 =     i  ij   jl i 1   1jl Because j  1. Hence: y  k  1 E P    ri ( k  1)  y i ( k  1)  i 3    rd  k  1   y i  k  1   g i' () x j  k  3   ij   ij Where ij3 (17) is the weight linking hidden neurons and output neurons. E p x j  k  E P   2   jq  x j  k    2jq    m i 1 0 i j  1..... q  1. (18) Then. 2. 2. 9. . No.. r (k  1) is desired output.. n.. . x j  k  r    n 1  f j    ji x ci  k     2ji u i  k  1    1     jl   i 1 i 1  n  x  k     f j' ·  x cl  k     1ji ci 1    jl  i 1  n x  k  1    f j' ·  xl  k  1    1ji i 1   jl  i 1  1 jl  TELKOMNIKA Vol. September 2013: 5030 – 5036 (24) .. 2 .... Let.5034 e-ISSN: 2087-278X  Define the performance index function: 1 [ r ( k  1)  y ( k  1)] 2 2 Ep  (16) In formula.  jh =    i0 ij3  f j' · . m . EP h =   j uq  k  1  2jq j  1...

n.e-ISSN: 2087-2 278X TELK KOMNIKA Also o.. 2.. l  1.  i0   y d i  k   y i  k   g i' · (30) (  jhh    i0 ij3  f j' · (31) ( m i 1 According to the above. x j  k    1jl  f j' · x cl  k  f j' ·x cl k   f j' ·x l k  1    f j' ·x cl k  1  (32) ( (33) ( Then.... m. n. 2. 2. 2..... S In order to verity the e reasonable eness of the e control sch heme...... In th his paper. m its tra ansfer function is: The Predictive P Con ntrol Method d of VAV Air Conditioning C g System (Jie ejia LI) . Simulation and a Conclus sion 4... m x j  k  i 1   1jl   1jl     i0 ij3  j  1. a the dependence of o xl  k  1 and  1jl caan be ignoredd. control object is second--order lag model. n (27) ( (28) ( (29) ( Wherre. w radial basis functiion’s center and a radial ba asis function’’s width. m we get g iterative algorithm o of output’s weight.. 2... Hence.. x j  k  1jl ca an be ignored d in the modified Elman neural n ' ' = f j · xcl  k  = xl  k  1 f j ·   1jl  5035 (26) ( Then the e learning allgorithm will degenerate e into the fo ollowing stan ndard BP lea arning algorrithm: ij3    i0 x j  k  i  1... 2.   ij   E p (25) (   ij But the dependence of o x l  k  1  and netwo ork [10]. 2 n q  1.....  2jq    jhuq  k  1 j  1. Hence. We m made a simu ulation experriment. (34) ( According g to gradien nt descent method. r j  1.

recurrsive wavele et neural ne etwork predictive control has small overshoot. 112(1): 145--161. [7] Song S Xiaoyu,Zheng Yan,Chang Chun nguang. systems. Functtional equivalence between radial basiss functions a and fuzzy inte erence E Transon Neu ural Networks. ASHRAE Trans saction (S0001-2505). Model Pred dictive Contro ol of Supply Air A Temperatu ure and Outsiide Air In ntake Rate of a VAV Air-Handling Unit. 62(12): 2760-2768. 30(2): 82-85. IEEE L Zhanming. F A neural ne etwork approa ach for on-line fault detection of nitrogen n sensors in alterna ated active sludge s treatm ment plants. 11. 1993. [10] Caccavale C F. Iam marino. 33(4)): 878-881. 1999. Intellige ent anti-skid brake controlle er using a neu ural network. 2012.The simulation compparison diagram of recuursive wavelet neural nnetwork preddictive contrrol and comm mon neural network contrrol is shown in i Figure 5. Energy Analysis for f Air C Conditioning System Using Fuzzy F Logic Controller. A Va ariable Precision Rough Se et Based Model for E Emergency Sccheduling. Min nicomputer min iniature system m. [4] Som S Akumarr. and itts adjusting time is abou ut 800s . 2004. 9(1): 139 9~150. Chandrasekh har J. Ap pplication of Particle P Swarm m Algorithm with w Inertia We eight PID Con ntrol in V Variable Air Vo olume System. 2006 6. Computation C o Lyapunov Exponents of of o Dynamic Systems S with Spherical Surface S S Symmetry. [5] LIU L Jian gong.Electroma agnetic Speed d Regulation Technology of Shearer. Hishamud ddin Jamalud ddin. S Pierri F. k  11. [9] Chen C Ning. 9.  10 . (7): 611-621. 20 010. 2004. T2  6. No. and its oovershoot is large. Fig gure 5.S Syeriff. 40(6): 858-864. References [1] S Shui yuan. Journal of Sh henyang Jianz zhu Universityy. set desirred temperatture is ( T 1 s  1 )( T 2 s  1 ) 24 C  . c 2011. An n adaptive learning algorithm m for T-S fuzzzy model base ed RBF [3] Li n neural networkk. [2] Jang JSR. 1(1). But recursive r wa avelet neural network pre edictive control has quicck response speed and small oversshoot. 2011. World Jourrnal of E Engineering. W Water Scienc ce and T Technology. [6] Henry H Nasutio on. 2 KOMNIKA Vol.5036 e--ISSN: 2087-278X  G (s)  kke   s . Control C E Ebgineering Prractice. Wang W Jun. Take T1  14. Info ormation and control. M Ma asi. Ro onald A Prerzz. 4(1): 15 56-158. 27(3):: 608-612. The erefore. high h control precision. Journal of La an zhou Unive ersity of Techn nology. Kan ng Aihong. [8] Ma M Shaohua. recu ursive wavelet neural ne etwork predictive contro ol has better dynamic performance p e. Septemberr 2013: 5030 0 – 5036 TELK . Comp parison Diagram of Recursive Wavele et Neural Ne etwork Predicctive Control and Common Neural N Network Control Figure 5 shows that: When the input tempe erature chan nges from 19  C to 25  C . So. Jamaluddin Mohd. strong s robusstness and good control performance p e. P Digiulio. C TEL LKOMNIKA. Su un CT. Long Suo. 2011. short adjus sting time. the eural networkk control is at a about 1250 adjussting time of common ne 0s .