Sunteți pe pagina 1din 5

9th IEEE International Symposium on Applied Computational Intelligence and Informatics • May 15-17, 2014 • Timişoara, Romania

Comparison of Modified Smith Predictor and


PID Controller Tuned by Genetic Algorithms for
Greenhouse Climate Control
Eugen Horatiu Gurban, Gheorghe-Daniel Andreescu, Senior Member, IEEE
Politehnica University of Timisoara, Dept. of Automation and Applied Informatics, Timisoara, Romania
e-mails: eugen.gurban@aut.upt.ro, daniel.andreescu@aut.upt.ro

Abstract—This paper develops a solution for greenhouse


climate control employing a modified Smith predictor. For II. GREENHOUSE CLIMATE MODEL
the well-known greenhouse climate model with dead-time, The highly used greenhouse microclimate model
having high nonlinearities and strong coupling, a feedback- describing the air temperature and humidity evolution was
feedforward linearization & decoupling method employing proposed by Albright et al. [1] :
measured disturbances is used. The equivalent process has
an integrator plus dead time (IPDT) behavior. The un- dTin 1 ⎡
= Si ( t ) − λ Q fog ⎤⎦
delayed process outputs, mandatory required in the dt ρ pV ⎣
C
linearized & decoupled method, are obtained by using a (1.1)

process internal model. A modified Smith predictor is V(t ) UA
employed to control the equivalent process that uses in its − ⎡⎣Tin ( t ) − Tout ( t )⎤⎦ − [ Tin ( t ) − Tout ( t )]
structure the same process internal model. The control V ρ C pV
system performances with two fitted controller types, i.e., dwin 1 1 ⎡ Si ( t ) ⎤
the modified Smith predictor and a PID controller tuned by
dt
= Q fog ( t ) +
V V ⎢α λ − βT win ( t )⎥
genetic algorithms, are analyzed and compared in nominal ⎣ ⎦
• (1.2)
case and with parameter variations for robustness test.
V( t )
− ⎡ win ( t ) − wout ( t )⎤⎦
V ⎣
I. INTRODUCTION where: Tin and win are the greenhouse internal air
The complex, nonlinear, coupled MIMO greenhouse temperature (oC) and absolute humidity (g/m3); Tout and
climate model proposed in [1], [2] is highly used for wout are the outside air temperature (oC) and absolute
development and testing solutions of greenhouse climate humidity (g/m3); V represents the greenhouse space
control for temperature and humidity. volume (m3); α and βT coefficients are used for modeling
One problem, identified in the papers that employ this the plant evapotranspiration rate: α shading and leaf area
model with measured disturbances, is the use of feedback- index, βT thermodynamic constants and other
feedforward linearization and decoupling procedure [3], evapotranspiration influencing factors; λ is the latent heat
[4] to a model with delayed outputs (undelayed outputs of vaporization; ρ is the air density (kg/m3); UA is the heat
being mandatory). Even if in steady state this proves to be transfer coefficient (W/K); Qfog is the humidifier debit;
a viable solution, in transient regime that is no longer the V is the ventilation rate (m3/s); Si is the intercepted solar
case. A solution to provide undelayed system outputs is to radiant energy (W).
use a process internal model [5]. The equivalent The associated state space model has the following
greenhouse climate process with decoupling is reduced to expression:
two decoupled channels for temperature and humidity
UA 1 UA
with integrator plus dead time (IPDT) behavior. x1 ( t ) = − x1 ( t ) + v1 ( t ) + v2 ( t )
In this paper, a control solution based on the Matausek ρ C pV ρ C pV ρ C pV
(2.1)
modified Smith predictor [6] is developed for decoupled 1 1 λ
greenhouse temperature and humidity control. The output + v2 ( t ) u1 ( t ) − x1 ( t ) u1 ( t ) − u2 ( t )
V V ρ C pV
variables without time delay, required by the Smith
predictor structure and also by the feedback-feedforward β α 1
x2 ( t ) = − T x2 ( t ) + v ( t ) + ⎡⎣v3 ( t ) − x2 ( t ) ⎤⎦ u1 ( t )
linearization method, are obtained by using the process V λV 1 V
(2.2)
internal model. For this control structure, the system 1
outputs do not present overshoot and both control + u2 ( t )
V
commands show a smooth profile.
y1 ( t ) = x1 ( t ) (2.3)
Based on simulation results, the paper gives the system
performance comparison between the Matausek modified y2 ( t ) = x2 ( t ) (2.4)
Smith predictor and other structure as PID controller
tuned by genetic algorithms [5], both controllers fitted to where the state variable x1 and x2 are the greenhouse air
greenhouse climate control system for temperature and temperature and absolute humidity; the two command
humidity decoupled channels. The system responses are variables u1 and u2 corresponds to V and Qfog; the three
compared in the case of nominal model, and when disturbances v1,v2 and v3 correspond to Si, Tout and wout.
considering modeling uncertainties for robustness study.

978-1-4799-4694-5/14/$31.00 ©2014 IEEE – 79 –


E. H. Gurban and G.-D. Andreescu • Comparison of Modified Smith Predictor and PID Controller Tuned by Genetic Algorithms…

Because the temperature and humidity sensors are not III. GREENHOUSE CLIMATE CONTROL EMPLOYING
situated in the actuator proximities (propagation delay), MODIFIED SMITH PREDICTOR
greenhouses being processes with distributed parameters,
and also the sensors present delay in measurement, the In this section, a structure for greenhouse climate
dead times d1, d2 are considered for each output (y1, y2). control is proposed by using linearization and decoupling
procedure, internal model and a modified Smith predictor.
The linearization and decoupling method with
measured disturbances proposed in [3] is used in the For processes with dead time, the effect of command
greenhouse climate particular case detailed in [5], [7]. An variables can be observed in the system outputs just after
equivalent system is obtained that has two decoupled this dead time. When the process is characterized by a
channels for temperature and humidity with an integrator significant dead time, the high performances of the control
plus dead time (IPDT) behavior. The linearization and system are difficult to obtain, i.e., fast transient response
decoupling method introduces two new control input (high value for gain crossover frequency) and small
overshoot (satisfactory phase margin) [8]. For
variables: û1 , û2 temperature and humidity rate of change.
compensating the process dead time, a well-known Smith
yi = uˆ i , where i = 1 or 2 (3) predictor control structure [9] (Fig. 2) can be used, where
For achieving this behavior by using undelayed system P(s)e− Ls is the real process and P̂(s)e− L̂s is the modeled
outputs, the actuator commands have the following form: process.
u1( t ) = [U A y1( t ) + λβT y2 ( t ) − ( 1 + α )v1( t )
(4.1)
−U Av2 ( t ) + ρ C pVuˆ 1( t ) + λVuˆ 2 ( t )] / Δ( t )

u2 (t ) = [( y2 ( t ) − v3 ( t )) ⋅ (U A y1( t ) − v1( t )
−U Av2 ( t ) + ρCpVuˆ 1( t )) + ρCp( − y1( t ) + v2 ( t ))( βT y2 ( t ) (4.2)
α
− v1( t ) + Vuˆ 2 ( t )] / Δ( t )
λ
Δ( t ) = ρ C p ( v2 ( t ) − y1( t )) + λ( v3 ( t ) − y2 ( t )) (4.3)
Figure 2. Smith predictor structure
In the real case, the delayed system outputs are
measured, but because in the computation of the command In the case when the delay free part and the dead time
variables u1, u2 the undelayed outputs values are required, of the process and of the process model are equal
a solution for estimating these values was proposed in [5] ˆ and L = Lˆ , the Smith predictor
respectively, i.e., P(s) = P(s)
by using the internal model of the greenhouse climate control structure allows controlling the process without
process (Fig. 1a). The greenhouse climate with decoupling dead time, because the dead time is outside of the control
from Fig. 1a has an equivalent integrator plus dead time loop (Fig. 3).
(IPDT) behavior (Fig. 1b).

Figure 3. Smith predictor equivalent structure

For the greenhouse climate model with decoupling


having an equivalent IPDT behavior, the transfer function
of the temperature and humidity channels have the form:

H(s) = K / s e − Ls , with K = 1, L = 30s (5)


For IPDT processes, the classical Smith predictor
structure is unable to provide a null steady state error
-a- when considering a constant load disturbance [8], [10]. To
eliminate this disadvantage, several modified Smith
predictor structures where proposed [11], [12], [13].
A modified Smith predictor, which adds an additional
feedback loop for compensating the effect of load
disturbances, was proposed by Matausek et al. [6]. The
modified Smith predictor control structure is presented in
Fig. 4, where the C(s) controller is a proportional
controller type (Kr) and M(s) is a filtered PD structure:
-b- K0 (Td s + 1 ) T
Figure 1. Linearization and decoupling using the greenhouse climate M (s) = , with T f = d . (6)
Tf s + 1 10
internal model, b) equivalent IPDT decoupled model

– 80 –
9th IEEE International Symposium on Applied Computational Intelligence and Informatics • May 15-17, 2014 • Timişoara, Romania

Figure 4. Matausek modified Smith predictor structure

The output of the P̂(s) block estimates the undelayed


state variable x̂ . The output error Δ y = y − ˆy has a
delayed action. In steady state, it contains information on
the disturbances d, and thus this disturbance is Figure 5. Greenhouse climate control structure with decoupling using
the Matausek modified Smith predictor
compensated by a feedforward loop using M(s) block that
delivers the disturbance estimation d̂ .
The feedback variable f for the controller Kr contains
two signals: i) the state estimation x̂ with quick action in
transient regime, and ii) the delayed output error Δ y that
is an added compensation in steady state.
The greenhouse climate internal model delivers the
undelayed estimation of output signals, which are used by
the linearization and decoupling block, but also by the
modified Smith predictor.
For computing the parameter values Kr and K0, the
following expressions are used [6]:
π
− φm Figure 6. Greenhouse climate control structure with decoupling using
2 1
K0 = , Kr = (7) PID controllers tuned by genetic algorithms
2 π 2 2 KTr
KL ( 1 − α ) + ( − φm ) α
2
IV. SIMULATION TESTS
where Tr is the time constant of the systems closed loop A. Dusk Transient Regime
transfer function. The derivative time constant (Td) is
chosen to be proportional to the dead time L [8]: For testing the greenhouse climate control system with
modified Smith predictor, a dusk (diurnal to nocturnal)
Td = α L, 0 ≤ α ≤ 1. (8) transient regime scenario is used by employing a 6
minutes ramp set-point decrease for both greenhouse
with the suggested parameters [6]: α = 0.4 and φm = 64D temperature and absolute humidity. The simulation
(phase margin). conditions are presented in Table I.
Taking into account the IPDT behavior for the two TABLE I.
decoupled channels of temperature and humidity, the GREENHOUSE VARIABLES AND PARAMETER VALUES FOR
DIURNAL TO NOCTURNAL TRANSIENT REGIME
following parameter values where obtained:
Tr = 0.4, K0 = 0.024, Td = 12, Tf = 1.2, thus α 0.1249
d1, d2 30 s
0.024( 12s + 1 )
M (s) = . S i, v1 300 W/m2
1.2s + 1 Tout, v2 22 oC
The proposed greenhouse climate control structure with wout, v3 0.004 kg/m3
decoupling using the Matausek modified Smith predictor, y1-ref ramp change: 25 oC to 20 oC at t = 2000 s
and internal model to estimate undelay output variables, is y2-ref ramp change: 0.019 kg/m3 to 0.013 kg/m3 at t = 2000s
presented in Fig. 5.
V 2800 m3
For performance test comparison, a greenhouse climate
x1(0) 25 oC
control structure with decoupling employing other fitted
controller as PID controller tuned by genetic algorithms is x2(0) 0.019 kg/m3
presented in Fig. 6. Λ 2257 J/g

– 81 –
E. H. Gurban and G.-D. Andreescu • Comparison of Modified Smith Predictor and PID Controller Tuned by Genetic Algorithms…

and humidity channels [5], are reported. These results


were obtained by employing a proposed objective cost
function and by using in the initial population the PID
controllers tuned by four PID tuning methods. A diurnal
to nocturnal (dusk) transient scenario has been used.
The GA objective cost function contains three integral
penalty terms for: absolute value of normalized input
error, square of normalized command variable, and
actuator saturation duration. Because the three terms are
considered to have the same importance, thus weight
factors (a, b) are used to respect this fact.
The cost function has the following expression:
-a-
J1 = fθ ( τ ) + a ⋅ gθ ( τ ) + b ⋅ pθ ( τ ) (9)
The penalty term for the input error is:
τ
y1 _ ref ( t ) − y1( t )
fθ ( τ ) =
∫ Δθ
dt , (10)
2000
where Δθ is the difference between initial and final
reference set-point.
The penalty term for the command variable is:
τ 2
1 ⎛ u (t) ⎞
gθ ( τ ) =
τ − 2000 ∫ ⎜ 1
⎜ u1 _ final
⎟ dt .

(11)
-b- 2000 ⎝ ⎠
The penalty term for actuator saturation duration is:
τ
⎧0, if no saturation occurs
pθ ( t ) = ∫ s ( t ) dt , sθ ( t ) = ⎨⎩1, if saturation occurs
2000
θ (12)

The GA PID controller Kp (1 + 1/sTi + Td s) has the


parameters Kp = 0.0427, Ti = 60.05 and Td = 14.37.
The system responses to ramp inputs, considering the
PID controllers obtained by using GA [5] and the
modified Smith predictor, are presented in Fig. 7.
As expected when using the Smith predictor, the state
variables follow the set-points, but the system outputs
-c- (temperature and humidity) have a 30 s lag time. For the
Smith predictor structure, the system outputs are without
overshoot and the two commands u1, u2 do not show
shock/sudden changes. This is translated in an improved
actuator lifetime.
B. Parameter Variations
To obtain the integrator behavior of the greenhouse
climate process with decoupling, the feedback-
feedforward linearization and decoupling procedure [3] is
used. To compute the commands u1, u2 (4.1-4.3), and also
to obtain the undelayed system states x1, x2, accurate
identification of the process parameters is required.
-d- In [14] it was tested by simulation the system behavior
considering the wrong estimation of solar radiation (Si),
Figure 7. System responses for diurnal to nocturnal transient (dusk
regime) using PID controllers tuned by genetic algorithms and the heat transfer coefficient (UA) and shading and leaf area
modified Smith predictor index coefficient (α). The simulation results show that the
wrong estimation of the UA heat transfer coefficient had
The comparative simulation results obtained for the the greatest impact in system response degradation
greenhouse temperature and humidity loops by employing (maximum overshot and settling time). Therefore, in the
two fitted controllers, i.e., the modified Smith predictor next simulation scenario the system robustness is tested
and PID controllers tuned by genetic algorithms (GA) considering only UA parameter uncertainty.
(Fig. 6), are presented. The GAs prove to be a useful tool In order to test the system robustness, a 50% variation
for PID tuning with suboptimal results. In this paper, just of the UA parameter is employed. The same simulation
the results obtained by using GAs to determine the scenario, diurnal to nocturnal (dusk) transient regime, as
parameter sets of the two PID controllers for temperature in the previous section is used (Table I).

– 82 –
9th IEEE International Symposium on Applied Computational Intelligence and Informatics • May 15-17, 2014 • Timişoara, Romania

used by employing ramp set-point changes for greenhouse


temperature and humidity references.
The system is tested in nominal conditions and with
50% UA variation for robustness. Based on simulation
results, a comparison on the control system performances
for the temperature and humidity loops are given by
employing the modified Smith predictor and by using PID
controllers tuned by genetic algorithms. The modified
Smith predictor proves to be a viable solution for the
greenhouse climate control system with decoupling for the
following reasons: reduced settling time, no overshoot,
and the both commands u1, u2 do not show shock/sudden
changes. This is translated in an improved actuator
lifetime.

ACKNOWLEDGMENT
This work was partially supported by the strategic grant
POSDRU 107/1.5/S/77265 (2010) within the Sectoral
Operational Program for Human Resources Development
2007-2013, Romania, co-financed by the European Social
Fund - Investing in people.

REFERENCES
[1] L. D. Albright, K. G. Arvanitis, and A. E. Drysdale,
“Environmental control for plants on Earth and in space,” IEEE
Control Systems, vol. 21, no. 5, pp. 28–47, Oct. 2001.
Figure 8. System responses: a) greenhouse air temperature, b) [2] G. D. Pasgianos, K. G. Arvanitis, P. Polycarpou, and N. Sigrimis,
greenhouse air absolute humidity, for diurnal to nocturnal transient “A nonlinear feedback technique for greenhouse environmental
(dusk regime) using PID tuned by genetic algorithms and the modified control,” Computers and Electronics in Agriculture, vol. 40, no. 1-
Smith predictor considering a 50% UA parameter variation 3, pp. 153–177, Oct. 2003.
[3] A. Isidori, Nonlinear Control Systems, 3rd Ed., Springer, 1995.
Fig. 8 shows minor differences between the nominal [4] J.-J. Slotine and W. Li, Applied Nonlinear Control, Prentice-Hall,
system responses and the system responses in the case a Englewood Cliffs, New Jersey, 1991.
50% UA parameter variation. This conclusion is sustained [5] E. H. Gurban, T.-L. Dragomir, and G.-D. Andreescu, "Greenhouse
climate control enhancement by using genetic algorithms,"
also by the response quality indicators: maximum Journal of Control Engineering and Applied Informatics, in press.
overshoot and settling time (Table II). In the case of
[6] M. R. Matausek and A. D. Micic, “On the modified Smith
control structure employing PID tuned by GA, the system predictor for controlling a process with an integrator and long
response follows the reference value, but a small dead-time,” IEEE Transactions on Automatic Control, vol. 44, no.
overshoot is observed (2%-4%). By using the modified 8, pp. 1603–1606, Aug. 1999.
Smith predictor in control structure, the two set-points for [7] E. H. Gurban and G-D. Andreescu, “Comparison study of PID
temperature and humidity are followed with a 30 s delay, controller tuning for greenhouse climate with feedback-
but there is no overshoot, and the settling time is reduced feedforward linearization and decoupling,” Proc. 16th
International Conference on System Theory, Control and
in comparison with the PID tuned by GA. Computing (ICSTCC 2012), Sinaia, pp. 1–6, Oct. 2012.
TABLE II. [8] A. Visioli and Q.-C. Zhong, Control of Integral Processes with
SYSTEM PERFORMANCE INDICATORS FOR TEMPERATURE AND Dead Time, Springer, 2011.
HUMIDITY SETPOINT RAMP CHANGE FOR: 1) SMITH PREDICTOR, 2) PID [9] O. J. M. Smith, Feedback Control Systems, McGraw-Hill, New
TUNED BY GA, 3) SMITH PREDICTOR WITH 50% UA UNCERTAINTY, 4) York, 1958.
PID TUNED BY GA WITH 50% UA UNCERTAINTY [10] I.-L. Chien, S. C. Peng, and J. H. Liu, “Simple control method for
Humidity loop Temperature loop response integrating processes with long deadtime,” Journal of Process
response performance performance Control, vol. 12, no. 3, pp. 391–404, April 2002.
No. [11] K. Watanabe, E. Nobuyama, and A. Kojima, “Recent advances in
Max. Settling Max. Settling time
overshoot time(s) overshoot (s) control of time delay systems—a tutorial review,” Proc. IEEE
1 0 379 0 366 International Conference on Decision and Control, Kobe, Japan,
vol. 2, pp. 2083–2089, Dec. 1996.
2 3.8% 424 2.3% 409
[12] Z. Guanghui, Q. Feng, and S. Huihe, “Robust tuning method for
3 0 380 0 377 modified Smith predictor,” Journal of Systems Engineering and
4 3% 428 2.5% 411 Electronics, vol. 18, no. 1, pp. 89–94, March 2007.
[13] W. D. Zhang and Y. X. Sun, “Modified Smith predictor for
V. CONCLUSIONS controlling integrator/time delay process,” Industrial &
Engineering Chemistry Research, vol. 35, no. 8, pp. 2769–2772,
A solution based on a modified Smith predictor for the 1996.
greenhouse temperature and humidity control with [14] E. H. Gurban and G.-D. Andreescu, “Employing 2DoF PID
decoupling is developed. To test this control structure, a controllers to improve greenhouse climate system robustness,”
diurnal to nocturnal (dusk) transient regime scenario is Proc. IEEE International Conference on System Science and
Engineering (ICSSE 2013), Budapest, pp. 93–98, July 2013.

– 83 –

S-ar putea să vă placă și