You are on page 1of 17

ARTICLE IN PRESS

Control Engineering Practice 17 (2009) 1157–1173

Contents lists available at ScienceDirect

Control Engineering Practice
journal homepage: www.elsevier.com/locate/conengprac

An adaptive temperature control law for a solar furnace
B. Andrade da Costa , J.M. Lemos
INESC-ID/IST/TU Lisbon, R. Alves Redol 9, 1000-029 Lisboa, Portugal

a r t i c l e in fo

abstract

Article history:
Received 14 August 2008
Accepted 7 May 2009
Available online 5 June 2009

This paper describes the development of an adaptive control law based on the exact feedback
linearization and Lyapunov adaptation of the process dynamics applied to a solar furnace. The algorithm
resulting from these underlying design principles is approximated in order to relate it with an adaptive
PI controller with feedforward. The controller is tested on a 6 kW solar furnace model that represents a
plant installed at the Odeillo Processes Materials and Solar Energy Laboratory (Oriental Pyrenees in the
South of France). The adaptive features allow to tackle the problems posed by knowledge uncertainty
about furnace dynamics. It is concluded that the specifications related to material testing are met.
& 2009 Elsevier Ltd. All rights reserved.

Keywords:
Solar furnace
Thermal stress tests
Nonlinear control
Exact linearization
Lyapunov method

1. Introduction
Nowadays there is a huge concern about the use of energy
sources that contribute to the climatic change problem. To
mitigate the problem, the exploitation of solar energy has been
considered in several ways, by converting it to thermal and
electric energy, using either thermal solar fields or photovoltaic
systems/panels. Solar energy is also used in other fields such
as the synthesis of high-temperature materials, and in material
testing (Fernandes et al., 2006), thermal stress (Fernandes,
Amaral, Rosa, & Shohoji, 2000), or to recreate the conditions of
the reentry of spacecrafts in Earth’s atmosphere. It is worth to
mention that the development of the Odeillo solar plant was
motivated by research on materials to be used in space rockets,
guided missiles and nuclear plants (Boyle, 1996), thereby avoiding
the problems of testing them with direct radioactive elements.
All those applications of solar energy systems pose interesting
problems from the automatic control point of view (Berenguel,
Camacho, Garcia-Martin, & Rubio, 1999; Lemos, 2006), such as the
presence of disturbances, non-linearities, and variable delays.
Solar furnaces concentrate solar energy in a limited area,
around the focus of a concentration mirror, or Fresnel lens, and
allow to attain high temperatures. Despite its interest, there is a
scarce number of references on the specific topic of solar furnace
control. A major exception is the work of Berenguel et al. (1999)
where modeling and control of a 20 kW furnace located at
Plataforma Solar de Almeria (southern Spain) is presented. The
work reported therein includes several types of PID with gain 
Corresponding author. Tel.: +351 213100259; fax: +351 218417499.

E-mail addresses: bac@comp.ist.utl.pt (B. Andrade da Costa), jlml@inesc-id.pt
(J.M. Lemos).
0967-0661/$ - see front matter & 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2009.05.001

scheduling and a self-tuning controller. In Paradkar and Feliachi
(2002) a controller is proposed to compensate disturbances in a
solar furnace, and in Lacasa, Berenguel, and Yebra (2006) a
controller based in fuzzy logic is evaluated in a solar furnace for
copper sintering. In Garcia-Gabin, Zambrano, and Camacho (2009)
a sliding mode predictive controller is evaluated and applied to a
solar air conditioning plant, and in Kojima, Taniwaki, and Okiami
(2008) the problem of positioning a flexible solar array is
addressed. In Costa et al. (2008a) experimental results are
described with a PI for temperature control in the same furnace
considered here.
In this paper the exact feedback linearization method together
with Lyapunov adaptation (Slotine & Li, 1991) is used to design an
adaptive controller for controlling the temperature of a sample in
a solar furnace (Costa et al., 2008b). The prototype is the 6 kW
solar furnace of Odeillo solar complex. In addition to the
algorithm yielded by the direct application of these techniques,
a modified version is also presented. This has the advantage of
having a structure that is comparable to an adaptive PI controller
with feedforward, thereby rendering the commissioning easier.
The contributions of this paper consist in the application of the
above-mentioned algorithms, in order to develop a new modified
structure, and its demonstration by simulation in a realistic
furnace model.
This paper is organized as follows. Section 2 describes the solar
furnace plant, in particular the thermal model subsystem. Section 3
describes the design of the adaptive controller using the exact
linearization method. Simulation results are shown and discussed.
Section 4 describes the modification of the adaptive controller and
results are presented. Section 5 presents the stability analysis of
the closed loop system with the modified adaptive controller.
Section 6 draws conclusions.

clouds. 2 shows two data records taken in (2003/05/03) and (2006/05/25). Parabolic concentrator The parabolic concentrator is built with small hexagonal mirrors that direct the solar energy to the location where the sample to be tested is placed. operates in closed-loop control and follows the movement of the Sun with accuracy. in such a way that the parabolic concentrator located at the top of the building. The physical aspect of the shutter. 1.5 cm centered at the focus receives a Fig. for practical purposes.2. location. 3. The closed-loop mode is selected whenever the measured direct Sun’s power is higher that 300 W=m2 . 800 Day 06−05−26 600 400 200 0 0 0. receives the Sun beam always along the same direction. In the shutter subsystem there are two aspects to be considered: the static function sfs ð:Þ that describes the steadystate relation between the power available before the shutter and the power available at the focus. The solar direct radiation is not constant but can exhibit some changes due to air moisture. The subsystem motor-blades is controlled with a Digivex controller. 2. It is made of four subsystems: the heliostat system. Fig. Lemos / Control Engineering Practice 17 (2009) 1157–1173 2.5 time [s] 3 3. Assuming that the shutter is not present. and y0 the minimal angle value below which there is no power at the . it may be considered that a circle with radius of 2.5 4 4. The quantity of energy reflected by the heliostat’s mirror (Fig. dust and clouds. 3. Fig. The maximum available power at the focus is 6 kW. 3) is made of 10 moveable aluminum blades with a thickness of 2. size and material were selected in order to yield a fast time response. The focus has a diameter of 6 cm. the angle of the shutter. Plant description Direct Solar Radiation. The flux inside the area of the focus is not uniform but. Day 03−05−03 1000 Irradiance [W/m2] The plant to control is the 6 kW solar furnace of the Odeillo Process Materials and Solar Energy Laboratory. Test tube in the focus with a SiC sample (center).5 m. (8h00m00s−17h59m59s) 1200 2. the parabolic mirror and its flux distribution. The position of the sample can be adjusted by manually commanding the position of the supporting arm. With the heliostat operating in closed-loop control. uniform flux.1. The size of the focus can be changed by moving the supporting arm in the up–down direction.5 x 104 Fig. it may be assumed that the tracking of the Sun’s position is perfect and there is no need to consider the effects of heliostat’s dynamics in the temperature of the sample. Fig.5 1 1. 1. This allows the focus to be in the same place during operation of the solar furnace. Examples show disturbances due to moisture in air. J. Fig. inside the laboratory.5 2 2. 2. and receives 95% of the power concentrated by the parabolic mirror.3. Heliostat The heliostat. which receives positional commands (reference signal) to position the blades. that can be moved in the North–South. Below 300 W=m2 the furnace does not operate. dust. 1) depends on its area and also on the cleanliness of the mirror. West–East and up–down directions using the operating console. Direct solar radiation evolution with time. Shutter of the 6 kW solar furnace (top).0 mm. Andrade da Costa. the reflected power can vary from 85% to 90% of the available Sun’s power. The commands can be sent using the operating console or by a personal computer through the data acquisition system InstruNet. Depending on time and dust level. and during the day and season. Fig. which will decrease with time due to dust deposition. The shutter of the solar furnace is able to quickly change the incident power on the sample.ARTICLE IN PRESS 1158 B. an important factor for the purpose of temperature control. 2. and the dynamics of the shutter. The blades are moved by a brushless motor (from the Parvex manufacturer) using a gear mechanism. Shutter The shutter (Fig. The static function sfs ð:Þ depends on y. 1. Schematic of the Odeillo 6 kW solar furnace. the shutter system and the temperature control system.M. It has a circular like shape with a radius of 0. the flux on the focus is approximated by a Gaussian function.

For higher temperatures. The parameters a1 . 4.1 s of rise time. where a sample of SiC material is considered (C . parabolic mirror and position of the supporting arm. a2 and a3 are temperature dependent and are defined by the following equations: a1 ¼ ðT s ÞsAs C p ðT s Þm . The power absorption of the sample is modeled by as As g f Gs ðtÞsfs ðus ðtÞÞ. with a 5% overshoot and 0. As ðm2 Þ: Sample’s radiation incident area Lc (m): Characteristic length 3:29  103 Eq. that does not give reliable readings at lower temperatures. The value of y0 is 251. 3 shows the apparatus. Using a geometric analysis. Lemos / Control Engineering Practice 17 (2009) 1157–1173 focus. In similar cases the usual procedure to tackle the controller design problem is to obtain linearized models (Berenguel et al.engel. 3. a sampling time of 0. Andrade da Costa. As shown in Fig. The SiC emissivity in the temperature range of 300–1500 K is described by ðT s Þ ¼ 0:87  2:222  105 ðT s  600Þ (4) The SiC specific heat is described in Table 2. The factor g f describes the concentration power of the solar furnace and depends on the heliostat. This model will be used in computer simulation tests despite it does not provide a good description in the 300–800 K range. The shutter operates in closed-loop control and is able to move the blades to its target position in 0.... 300 675 1400 400 880 Model Output 600 1050 800 1135 1000 1195 → 1200 1000 Temp. where the tip of the thermocouple is in contact in the lower side of the SiC sample. The experimental command signal of the shutter and the solar flux was supplied to the temperature model and the model output was compared with the experimental temperature measurements. 6. a B thermocouple.M. (1999). This dynamics is much faster than the temperature dynamics. between 800 and 1300 K. The convection factor is hconv ðT s . this can be explained in part by the type of thermocouple used. where T e ¼300 K. (4) 5:67  108 7:068  104 7:50  103 Eq. such as the one made in Berenguel et al. Fig. the static function is described by cosðy0 þ us ðtÞð90  y0 Þ=100Þ sfs ðus ðtÞÞ ¼ 1  cosðy0 Þ (1) Table 2 Specific heat of SiC as a function of temperature. 2008a). The following model is used to describe the temperature behavior of the sample T s ðtÞ (in Kelvin) dT s ðtÞ ¼ a1 ½T 4s ðtÞ  T 4e ðtÞ  a2 ½T s ðtÞ  T e ðtÞ þ a3 Gs ðtÞsfs ðus ðtÞÞ dt (2) where T e ðtÞ (in K) is the temperature of the environment that contributes to losses by radiation and convection. The output from the simulated model using a sequence of steps in the shutter command is shown in Fig. For digital implementations of the temperature controller. change with temperature T s according to Table 3. T (K) C p ðTÞ with 0ous ðtÞ  100. The temperature model of the sample is developed based on an energy balance. and it was selected after the temperature model validation with experimental results. 4 and 5. a2 and a3 . The parameters a1 . 2008b). 600 400 200 0 500 1000 1500 Time [seg] 2000 2500 Fig. T e Þ ¼ 1:32ððT s  T e Þ=Lc Þ0:25 (5) The sample’s solar absorption factor as is modeled by as ¼ 0:09 þ ðT s  867:5Þ  2:2086  104 8 0:055 if as  0:055 > > as ðT s Þ ¼ < as > > : 0:9 if 0:055oas  0:9 (6) if as 40:9 Table 1 Furnace and sample parameters values. Parameter: description Value r ðkg m3 Þ: Density of the material 3:10  103 Table 2 1 C p ðJkg K1 Þ: Material specific heat m (kg): Mass of the sample : Emissivity of the material s ðWm2 K4 Þ: Stefan–Boltzmann const. [K] 2.ARTICLE IN PRESS B. this means that the digital temperature controller can be seen as having a continuous-time behavior and the shutter can be described by its static function sfs ð:Þ only. 2003). T e ÞAs . It can be concluded that the temperature model has different local dynamics which are dependent on the temperature level. the blades are not in the horizontal position but the sample has no Sun power.5 s is used (Costa et al. The function sfs ð:Þ describes the nonlinear contribution of the shutter to the incident power on the sample. The temperature model was validated1 using temperature experimental data obtained from a SiC sample test with a PI controller (Costa et al. 100] and represents the command signal of the shutter.4. The results are shown in Figs. 1999) and 1 Note: using the experimental data. .6 Gs ðW=m2 Þ: Max.5 s. Temperature model of the sample 1159 ← 800 Measured Temp. C p ðT s Þm a3 ¼ as As g f C p ðT s Þm (3) The parameters in (3) are described in Table 1. Comparing the temperature model output with the measured temperature in a SiC test. (5) hconv ðWm2 K1 Þ: Convection factor as : Sample’s solar absorption factor g f : Furnace gain Eq. the model provides a good description of the experimental results. the temperature dependence of as was modeled by a linear equation. solar flux 1:0  103 The sample is assumed to be of SiC. J. The signal us ðtÞ is in the range [0. (6) 1713. a2 ¼ hconv ðT s . note that there is a large mismatch at lower temperatures.

2. Experiment signals: command of the shutter. To simplify the notation the derivative operator dðÞ=dt will be represented . a~ 1 ðtÞ. a^ 3 ought to be adapted in such a way that es ðtÞ converges to zero. To avoid that problem. the estimates a^ 1 . This corresponds to the actual control law used to manipulate the shutter: 3. such as a bank of PID controllers with a switching mechanism to select the adequate controller. 2 (7) T_ R ðtÞ þ K L es ðtÞ þ a^ 1 ½T 4s ðtÞ  T 4e ðtÞ a^ 2 ½T s ðtÞ  T e ðtÞ þ a^ 3 Gs ðtÞ a^ 3 Gs ðtÞ (10) _ by ðÞ. a~ 3 ðtÞÞ that jointly encompasses control and estimated states is chosen " # 3   X 1 2 1 2 a~ i ðtÞ (12) es ðtÞ þ Vðes ðtÞ. the transformed model becomes an integrator and may be stabilized with the control law vðtÞ ¼ K L es ðtÞ. T (K) a1 ðTÞ a2 ðTÞ a3 ðTÞ 300 1:584  1011 0 3:000  101 400 1:212  1011 3:466  103 2:303  102 600 1:010  1011 3:823  103 1:930  102 800 9:299  1012 4:019  103 2:438  102 12 3 2 1000 8:786  10 4:152  10 3:678  10 e_ s ðtÞ ¼ T_ R ðtÞ þ a1 ½T 4s ðtÞ  T 4e ðtÞ þ a2 ½T s ðtÞ  T e ðtÞ  a3 Gs ðtÞsfs ðus ðtÞÞ (8) In order to use the exact linearization method. a candidate Lyapunov function Vðes ðtÞ. a~ 2 ðtÞ. Eq.ARTICLE IN PRESS 1160 B. a feedback linearization control law is obtained. 8i 2 f1. where T R ðtÞ represents the temperature reference and T s ðtÞ represents the temperature of the sample. a~ 3 ðtÞÞ ¼ gi 2 i¼1 with adjustable parameters gi 40. Under perfect knowledge of the parameters (a^ i . a virtual manipulated variable vðtÞ is defined by the change of variable vðtÞ ¼ T_ R ðtÞ þ a^ 1 ½T 4s ðtÞ  T 4e ðtÞ þ a^ 2 ½T s ðtÞ  T e ðtÞ  a^ 3 Gs ðtÞsfs ðus ðtÞÞ (9) use them to design local controllers. a~ 1 ðtÞ. To tackle the problem. 2. Computing the time derivative2 of the tracking error yields Table 3 Process model parameters for SiC as a function of temperature. To develop the adaptive controller. us ðtÞ and solar flux Gs ðtÞ used in the temperature model evaluation. the presence of parameter errors a~ 1 . a~ 2 and a~ 3 may generate an unstable closed loop.M. Lemos / Control Engineering Practice 17 (2009) 1157–1173 100 us(t) [%] 80 60 40 20 0 0 500 1000 1500 Time [seg] 2000 2500 0 500 1000 1500 Time [seg] 2000 2500 Gs(t) [Wm−1] 900 800 700 600 500 400 Fig. In the present case an adaptive control technique based on a feedback exact linearization is used in the controller design. 5. a^ 2 . the parameters entering linearly on the model are estimated using a Lyapunov method. 3g. 3g). based on the physical model Eq. where K L 40 is a tuning parameter. The adaptive controller us ðtÞ ¼ s1 fs ½wðtÞ The design of the adaptive controller comprises two main steps: wðtÞ ¼  first. 8i 2 f1. However. with ai ¼ a^ i þ a~ i .  then. J. (2) is used in conjunction with the tracking error es ðtÞ defined by D es ðtÞ ¼ T R ðtÞ  T s ðtÞ (11) D where estimates a^ i of ai are used. a~ 2 ðtÞ. The motivation to explore this approach is to use the information about the process. Andrade da Costa. namely the structure of the temperature model. (2). the static characteristic of the shutter and the direct measurements of the Sun’s radiation to compensate the radiation disturbances.

1. Andrade da Costa. Steps responses of the temperature model with Gs ðtÞ ¼ 600 W=m2 . . In the simulations. 7 and 8. corresponding to a possible nominal working point. To guide the selection of these parameters. a^ 2 ðt 0 Þ. and considering that a~^ i ¼ a_^ i for i 2 f1. 8i 2 f1. 2.M. ai Þ  0. 3g yields the following adaptation laws: a_^ 1 ðtÞ ¼ g1 ½T 4s ðtÞ  T 4e ðtÞes ðtÞ a_^ 2 ðtÞ ¼ g2 ½T s ðtÞ  T e ðtÞes ðtÞ (15) a_^ 3 ðtÞ ¼ g3 Gs ðtÞwðtÞes ðtÞ (16) algorithm. g2 and g3 . (ii) the temperatures T s ðtÞ and T e ðtÞ are assumed to be near 800 and 300 K. A standard argument based on that results in Vðe LaSalle’s invariance principle allows to conclude that es ðtÞ approaches zero and a~ i . the integral form of the parameter adaptation laws is used with the following assumptions: (i) a 50% error is assumed for the initial estimates. To set up the adaptive 3.ARTICLE IN PRESS B. g3 ¼ 2:0  105 . g1 ðT 4s ja^ 1 ðtÞ  a^ 1 ðt 0 Þj 1:25  1023  ¼ R      t 4 R t  T e Þ t0 es ðtÞ dt  t0 es ðtÞ dt (17) g2 ja^ 2 ðtÞ  a^ 2 ðt 0 Þj 4:0  106  ¼ R  R   t   t ðT s  T e Þ t0 es ðtÞ dt  t0 es ðtÞ dt (18) g3 ja^ 3 ðtÞ  a^ 3 ðt 0 Þj 1:5  105  ¼ R  R   t   t Gs  t0 wðtÞes ðtÞ dt  t0 wðtÞes ðtÞ dt (19) 3. The a^ 1 ðt 0 Þ. The results are shown in Figs. a^ 3 ðt 0 Þ ¼ 2:3  102 and the K L ¼ 10. a~ 2 and a~ 3 . 6.1. At the beginning of the simulation the controller was able to track the reference signal. The parameters a1 . These results are now used to guide the selection of the g1 . It is then possible to obtain (14) _ s ðtÞ. One problem with adaptive controllers is that they usually have knob parameters that must be selected and they affect the change rate of estimates. g2 ¼ 1:0  107 . which is the case with g1 . seven parameters must be specified. the term T_ R ðtÞ of Eq. Simulation I In this simulation the adaptation gains are g1 ¼ 1:0  1024 . 3g are bounded.1. a^ 2 ðt 0 Þ ¼ 4:0  103 . a^ 3 ðt0 Þ were selected to be near the process parameter values for T s ¼800 K. 2. a2 and a3 are assumed to be constant in this derivation. g2 and g3 which are very small! The performance of the adaptive controller is evaluated by computer simulations. this means it will act as an error source when the reference is not constant. All the three computer simulations use a^ 1 ðt 0 Þ ¼ 9:2 1012 . respectively. (8) yields dVðes ðtÞ. Three simulation results are shown with different adaptation gains. Computing the time derivative of and using Eq. Lemos / Control Engineering Practice 17 (2009) 1157–1173 1161 Temperature 1400 Ts(t) [K] 1200 1000 800 600 400 200 0 200 400 600 time [s] 800 1000 1200 800 1000 1200 Shutter position 100 us (t) [%] 80 60 40 20 0 0 200 400 600 time [s] Fig. this means that the parameters should vary slowly when compared with T s ðtÞ. J. a~ i ðtÞÞ ¼  K L e2s ðtÞ dt   1 þ a~ 1 ðT 4s ðtÞ  T 4e ðtÞÞes ðtÞ þ a_~ 1 ðtÞ g1   1 þ a~ 2 ðT s ðtÞ  T e ðtÞÞes ðtÞ þ a_~ 2 ðtÞ g2   1 _ þ a~ 3 Gs ðtÞwðtÞes ðtÞ þ a~ 3 ðtÞ g3 (13) Equating to zero the terms that are multiplying a~ 1 . Computer simulation results where 0  wð:Þ  1:0 and Gs is assumed to be 800 Wm2 . the difference is in the adaptation gains. In practice. (10) is set to zero.

a^ 3 ðtÞ recover to the values prior to the control signal saturation. g2 ¼ 1:0  107 . To use the adaptive controller in practice the above points must be solved.1. 12 and 13.1 was avoided by selecting a very small adaptation gain value for g3.5 s.3 show that the controller has good performance but that depends on the correct selection of the adaptation gains g1 . The results show that the adaptive controller has a good performance. g2 ¼ 1:0  107 . the problem reported in Section 3.1 for the behavior of a^ 3 ðtÞ. The parameter estimate a^ 3 ðtÞ can be considered constant. The adaptive controller was able to track the temperature reference. However. Note that the shutter controller is able to move the blades to the target position in 0. J. control signal and the simulated Gs ðtÞ.4. 3 Note that wðtÞ ¼ sfs ðus ðtÞÞ. g3 ¼ 1:0  1010 . Simulation I. (16). Process output and reference signal. The effect of the saturation signal causes a huge tracking error es ðtÞ. couple with higher adaptation gains cause the divergence of a^ 1 ðtÞ and a^ 2 ðtÞ. and the adaptation of a^ 3 ðtÞ is blocked.1. Simulation III The adaptation gains are now selected to g1 ¼ 1:0  1022 . Simulation II The adaptation gains are now selected to g1 ¼ 1:0  1024 .1. The simulation results are shown in Figs. Fig. The wrong signal value of a^ 3 ðtÞ has the consequence that the control signals us ðtÞ and wðtÞ are set to zero3 due to the constraints imposed by the operating limits of the shutter. a^ 2 ðtÞ. g2 and g3 . . The simulation results are shown in Figs. 9–11. The aim is to avoid the problem of having a negative value on a^ 3 ðtÞ.2. Note that the a^ 2 ðtÞ. The adaptation parameters g1 . After this event the control signal is blocked at zero and the controller is not able to track the reference signal. g3 values are very small and it was difficult to select them. that can cause stability problems.3.M. It is difficult to incorporate an anti-windup mechanism in the control law defined by Eqs.1. 10 is a zoom to show the results of the temperature and shutter aperture. 3. the estimate of a3 starts to decrease and becomes negative.1. g3 ¼ 1:0  1010 .1. Comments The simulation results presented in Sections 3. Andrade da Costa.1–3. 3. but the estimates a^ 2 ðtÞ. The tracking error es ðtÞ being positive starts to increase and according to Eq. see Section 3. a^ 3 ðtÞ. But at time near 50 s the control signal becomes saturated at its maximum value. Lemos / Control Engineering Practice 17 (2009) 1157–1173 RefT(t) and Ts(t) [K] Temperature 1200 Ref (t) → T 1000 800 600 400 200 0 50 100 150 200 250 time [s] 300 350 400 450 500 0 50 100 150 200 250 time [s] 300 350 400 450 500 350 400 450 500 us (t) [%] 100 50 0 Simulation of Solar Power variability Gs(t) [W/m2] 1000 800 600 400 0 50 100 150 200 250 time [s] 300 Fig. There are also the following points to be mentioned:  The saturated control signal causes problems in the adaptation   of a^ 1 ðtÞ. but adaptation gains must be very small to achieve the good results. 7. (10) and (11). a^ 3 ðtÞ are affected by the control signal saturation.1. g2 . In practical the adaptation laws must include a mechanism to overcome the control signal saturation. which can cause a misbehavior of the controller. Closed-loop control. 3. The aim is to test ‘‘higher’’ adaptation gain values for g1 and g2 .ARTICLE IN PRESS 1162 B. note that g3 was multiplied by a factor of 5:0  106 . which is by freezing a^ 3 ðtÞ to its initial value.

(25) R _ tÞ dt ¼ f ðtÞgðtÞ Using now the relationship given by f ðtÞgð R_ f ðtÞgðtÞ dt. The modified adaptation law of the parameter estimates Z a^ 1 ðtÞ ¼ a^ 1 ðt 0 Þ þ g1 t0 t ½T s ðtÞ  T e ðtÞes ðtÞ dt (21) Gs ðtÞwðtÞes ðtÞ dt (22) t0 t t0 To prevent a^ 3 ðtÞ to assume a negative or null value this variable will be kept constant and equal to a^ 3 ðt 0 Þ40. Modification of the adaptive controller for practical use With these assumptions Eqs. Note that a^ 3 is given by Eq. From Eqs. Evolution of a^ 1 ðtÞ. note that es ðtÞ ¼ T R ðtÞ  T s ðtÞ. in Eqs. (23) and (24) are written as a^ 1 ðtÞ ¼ a^ 1 ðt 0 Þ þ g1 ½T 4R ðtÞ  T 4e ðtÞ  4g1 T_ R Z t t0 T 3R ðtÞ Z t Z t t0 Z t  es ðtÞ dt t0 Z (26) t es ðtÞ dt t0 es ðt1 Þ dt1 dt (27) t0 Considering now the terms with the slope of the temperature reference signal T_ R . and a^ 3 ðtÞ. (20)–(22) are modified to Due to the difficulties described in Section 3. (3) and a value can be computed for a^ 3 ðt 0 Þ from the size and the properties of the material to be tested.5 0 0 50 100 150 200 250 300 350 400 450 500 300 350 400 450 500 300 350 400 450 500 time [s] 4 α2 (t) 3 2 1 0 0 50 100 150 200 x 10−3 250 time [s] α3 (t) 10 5 0 −5 0 50 100 150 200 250 time [s] Fig. (20) and (21) in the case that the output of the process is near the reference signal and the closed-loop system is stable. Simulation I. Lemos / Control Engineering Practice 17 (2009) 1157–1173 1163 Controller gains − Estimates of α1. a^ 2 ðtÞ.1. Andrade da Costa.4 an analysis of the adaptation equations of a^ 1 ðtÞ. α2. Closed-loop control. α3 x 10−8 α1 (t) 1 0. Additionally the following assumptions are made:  The environment temperature T e ðtÞ can be considered constant and is smaller than T R ðtÞ and T s ðtÞ. J. This assumption is used to approximate the adaptation laws Eqs. Eqs. (14)–(16) the adaptation equations of a^ i ðtÞ are Z t a^ 1 ðtÞ ¼ a^ 1 ðt0 Þ þ g1 ½T 4s ðtÞ  T 4e ðtÞes ðtÞ dt (20) a^ 2 ðtÞ ¼ a^ 2 ðt0 Þ þ g2 a^ 3 ðtÞ ¼ a^ 3 ðt0 Þ  g3 Z a^ 2 ðtÞ ¼ a^ 2 ðt 0 Þ þ g2 Z t t0 Z t ½T 4R þ ðtÞ  T 4e ðtÞes ðtÞ dt (23) ½T R ðtÞ  T e ðtÞes ðtÞ dt (24) t0 a^ 3 ðtÞ ¼ a^ 3 ðt 0 Þ 4. This will be details next. a^ 2 ðtÞ and a^ 3 ðtÞ is done and a modification is proposed.ARTICLE IN PRESS B. 8.M. The aim is to obtain an implementation of the adaptive controller suitable to be used in practice. then the following parameter adaptation laws are obtained:  The kes ðtÞk is assumed to be small such that T s ðtÞ T R ðtÞ. (26) and (27) and assuming that they can be discarded when compared with the other terms. t es ðt1 Þ dt1 dt t0 a^ 2 ðtÞ ¼ a^ 2 ðt 0 Þ þ g2 ½T R ðtÞ  T e ðtÞ  g2 T_ R Z a^ 1 ðtÞ ¼ a^ 1 ðt 0 Þ þ g1 ½T 4R ðtÞ  T 4e ðtÞ a^ 2 ðtÞ ¼ a^ 2 ðt 0 Þ þ g2 ½T R ðtÞ  T e ðtÞ a^ 3 ðtÞ ¼ a^ 3 ðt 0 Þ Z t Z t0 t es ðtÞ dt (28) es ðtÞ dt (29) t0 (30) .1. The reference signal T R ðtÞ is assumed to be a constant signal or a ramp signal with a ‘‘small’’ slope T_ R . 4.

2 and the controller has a simpler structure that can be easily interpreted. Process output and reference signal. (33)–(35) show the explicit relationship between the process parameters and the controller parameters.2 a redefinition of g1 and g2 is presented and it is shown that only one parameter. with adaptive gains. In Section 5.M. a^ 2 ðt 0 Þ ¼ 4:0 103 .2. 4. K I ðtÞ. (33) can also be redefined to improve the numeric representation. K P ðtÞ. Note that the control law depends only on the initial estimates a^ 1 ðt 0 Þ. The stability analysis is addressed in Section 5. Lemos / Control Engineering Practice 17 (2009) 1157–1173 RefT(t) and Ts(t) [K] Temperature 1200 1000 800 600 400 200 0 50 100 150 200 250 300 350 400 450 500 300 350 400 450 500 400 450 500 time [s] us (t) [%] 100 50 0 0 50 100 150 200 250 time [s] Simulation of Solar Power variability Gs(t) [W/m2] 1000 800 600 400 0 50 100 150 200 250 time [s] 300 350 Fig. 9. (28)–(30) in Eq. g2 ¼ 1:0  107 . Note that the parameters in Eq. which can be modified to have an anti-windup mechanism. and the adaptation gains g1 ¼ 1:0  1024 . . with the equality corresponding to have a double real pole for the closed-loop assuming a perfect knowledge of the process parameters.3. a^ 3 ðt 0 Þ ¼ 2:4  102 . Closed-loop control. Andrade da Costa. control signal and the simulated Gs ðtÞ. a^ 2 ðt 0 Þ and a^ 3 ðt 0 Þ can be computed from physical data using Eq. K L ¼ 10. These controller parameter values are the same as in the simulation described in Section 3.ARTICLE IN PRESS 1164 B. and only g1 and g2 and K L are the controller ‘‘knob’’ parameters. The results are shown in Figs. The aim is to compare the results of the modified adaptive controller with the unmodified adaptive controller in the same test conditions. J. by using a constant based on the temperature reference mean value. the results are similar as the results described in Section 3. it seems that the proposed controller has too many parameters to be tuned. and Section 5. Using algebraic manipulation the modified control law assumes the following structure: us ðtÞ ¼ s1 fs ½wðtÞ  Z wðtÞ ¼ U o ðtÞ þ K P ðtÞ es ðtÞ þ K I ðtÞ (31)  t es ðtÞ dt (32) t0 D U o ðtÞ ¼ D K P ðtÞ ¼ D K I ðtÞ ¼ a^ 1 ðt 0 Þ½T 4s ðtÞ  T 4e ðtÞ þ a^ 2 ðt0 Þ½T s ðtÞ  T e ðtÞ a^ 3 ðt0 ÞGs ðtÞ (33) KL a^ 3 ðt 0 ÞGs ðtÞ g1 ½T 4R ðtÞ  T 4e ðtÞ½T 4s ðtÞ  T 4e ðtÞ KL (34) þ present case the estimates a^ 1 ðt 0 Þ. it must be emphasized that Eqs.6 demonstrates that the parameter g can be selected as g  ðK L =2Þ2 . (3). The modified control law The modified control law is obtained by substituting Eqs. This simplifies the parameter selection and improves the numerical representation of variables and constants in a digital computer with limited resources. Simulation II. Comparing this controller with a standard PI controller. 14 and 15. a^ 2 ðt 0 Þ. The modified adaptive controller has a good performance.1. the adaptation gains g1 and g2 and K L .1. In the 4. in practice to design a PID controller. a^ 3 ðt0 Þ. and there is only one integrator. the process parameters must be known. (11). a^ 1 ðt 0 Þ ¼ 9:2  1012 . Simulation results with the modified controller The modified adaptive controller was simulated with the following parameter values. However. g2 ½T R ðtÞ  T e ðtÞ½T s ðtÞ  T e ðtÞ KL (35) This new control law may be interpreted as being a PI controller with a feed-forward term U o ðtÞ.2. g. is needed.

(36) includes the process dynamics and the modified control law. Using Eq. Simulation II. Eq. 10. This problem was not so easily resolved with the first adaptive controller. The aim is to find the equilibrium points of system Eq. J. Note that in both simulations the controllers do not have antiwindup reset mechanisms. The stability analysis is based on Lyapunov’s theory with the assumptions described in Section 4. x1 ðtÞ ¼ t0 es ðtÞ dt and x2 ðtÞ ¼ es ðtÞ. Lemos / Control Engineering Practice 17 (2009) 1157–1173 1165 RefT(t) and Ts(t) [K] Temperature 1200 1000 800 600 200 220 240 260 time [s] 280 300 320 220 240 260 time [s] 280 300 320 300 320 us (t) [%] 100 50 0 200 Simulation of Solar Power variability Gs(t) [W/m2] 900 800 700 600 500 200 220 240 260 time [s] 280 Fig. Equilibrium points D Rt D Defining two state variables. 5. . (31)–(35).1. Andrade da Costa. xe2 Þ given by 8 <x : e1 DTe GTe ¼ ðT R ðtÞ  T e ðtÞÞ (38) xe2 ¼ 0 with D DTe ¼ a1    a^ 3 ðt0 Þ a^ 1 ðt0 Þ ðT 4R ðtÞ  T 4e ðtÞÞ a^ 3 ðt0 Þ a^ 2 ðt0 Þ þ a2   a3 a1 a3 a2 ðT R ðtÞ  T e ðtÞÞ  (39) 4 In order to keep equations as simple as possible the variable T s ðtÞ is not substituted by T s ðtÞ ¼ T R ðtÞ  es ðtÞ. Stability analysis of the modified adaptive controller This section presents the stability analysis of the closed-loop system using the modified adaptive controller. Zoom of Fig. then the tracking error dynamics is described by des ðtÞ a3 a3 K e ðtÞ þ ða1  a^ 1 ðt 0 ÞÞ ðT 4 ðtÞ  T 4e ðtÞÞ ¼  dt a^ 3 ðt0 Þ L s a^ 3 ðt 0 Þ s   Z t a3 a3 þ a2  a^ 2 ðt 0 Þ Lðes ðtÞÞ es ðtÞ dt ðT s ðtÞ  T e Þ  a^ 3 ðt0 Þ a^ 3 ðt0 Þ t0 (36) 5. Closed-loop control. (36) can be written in a state space representation. To address the saturation of the control signal an anti-windup mechanism must be incorporated and the stability analysis will be more complex but can be addressed using techniques described in Glattfelder and Schaufelberger (2003).1. With the new controller it is easy to tackle the control saturation by including an anti-windup reset mechanism in the integrator. (8) that describes the dynamics of the tracking error as a function of the control signal. (36) and to demonstrate their stability. which means the stability analysis is local. which yields an equilibrium point xe ¼ ðxe1 . Eqs. and the equations that define the modified control law. It is also assumed that the control signal is not saturated.M. with4 D Lðes ðtÞÞ ¼ g1 ½T 4R ðtÞ  T 4e ðtÞ½T 4s ðtÞ  T 4e ðtÞ þ g2 ½T R ðtÞðtÞ  T e ðtÞ½T s ðtÞ  T e ðtÞ (37) Eq.ARTICLE IN PRESS B. 9 to show the temperature and shutter aperture.

α3 x 10−11 5 α1 (t) 0 −5 −10 0 50 100 150 200 250 time [s] 300 350 400 450 500 50 100 150 200 250 time [s] 300 350 400 450 500 50 100 150 200 250 time [s] 300 350 400 450 500 x 10−3 5 α2 (t) 0 −5 −10 0 x 10−3 α3 (t) 10 9.995 9. Evolution of a^ 1 ðtÞ. g1 ¼ g12ðT 4R ðtÞ  T 4e ðtÞÞ2 g2 ¼ g12ðT R ðtÞ  T e ðtÞÞ2 (42) (43) with g40 being the new positive adjustable parameter which is common to both g1 and g2 .ARTICLE IN PRESS 1166 B. (36) and (37) the parameters g1 and g2 are only used as a weight on the integration of es ðtÞ. Lemos / Control Engineering Practice 17 (2009) 1157–1173 Controller gains − Estimates of α1. (41) at the equilibrium point xe ¼ ð0. for i 2 f1. (39) ðT 4R ðtÞ  T 4e ðtÞÞ=ðT R ðtÞ  T e ðtÞÞ ¼ ðT 2R ðtÞþ (44) The proposed control algorithm is now written as us ðtÞ ¼ s1 fs ½wðtÞ  Z wðtÞ ¼ U o ðtÞ þ K P ðtÞ es ðtÞ þ K I ðtÞ  t es ðtÞ dt (45) (46) t0 D U o ðtÞ ¼ a^ 1 ðt0 Þ½T 4s ðtÞ  T 4e ðtÞ þ a^ 2 ðt 0 Þ½T s ðtÞ  T e ðtÞ a^ 3 ðt0 ÞGs ðtÞ KL D K P ðtÞ ¼ D K I ðtÞ ¼ a^ 3 ðt0 ÞGs ðtÞ g T 4s ðtÞ  T 4e ðtÞ 2K L T 4R ðtÞ  T 4e ðtÞ (47) (48) þ T s ðtÞ  T e ðtÞ T R ðtÞ  T e ðtÞ ! Note that if T s ðtÞ T R ðtÞ then K I ðtÞ ¼ g=2K L . α2. Redefinition of g1 and g2 Looking at Eqs. however. a^ 2 ðt0 Þ ¼ a2min and a^ 3 ðt0 Þ ¼ a3max . Andrade da Costa. Closed-loop control. and5 parameters the following definition is proposed to g1 and g2 : D GTe ¼ g1 ½T 4R ðtÞ  T 4e ðtÞ2 þ g2 ½T R ðtÞ  T e ðtÞ2 (40) It is worth to mention that the equilibrium point corresponds to a null tracking error but the value at the integrator ðxe1 Þ can be positive. (36) is simplified to Z t des ðtÞ es ðtÞ dt (41) ¼ K L es ðtÞ  Lðes ðtÞÞ dt t0 The linearization of Eq. If. If the ideal case is ^ 0 Þ ¼ a1 . 0Þ generates a second order ordinary equation with two poles that can be selected by choosing K L and by a combination of the two parameters. 3g then xe1 can be always positive if a^ 1 ðt0 Þ ¼ a1min . Simulation II. With this definition of g1 and g2 . g1 and g2 .2. a^ 2 ðtÞ.M. bounds are known such that aimin  ai  aimax . depending on the process parameter initial estimates. (40) is simplified to D GTe ¼ g 5.99 0 Fig. 2. null or negative. (49) . J. To reduce the number of adjustable 5 Note that T 2e ðtÞÞðT R ðtÞ þ T e ðtÞÞ. Eq. in Eq. considered with the initial estimates selected as a1 ðt ^ ^ a2 ðt0 Þ ¼ a2 and a3 ðt0 Þ ¼ a3 then Eq. and a^ 3 ðtÞ. 11.

T R ðtÞ  T e ðtÞ (55) (52) 8 See footnote 7. 12. See footnote 7.3. x2 ðtÞÞ _  0. T e ðtÞ. Closed-loop control. Simulation III. (50). J. it is possible to obtain    _ ¼  x2 ðtÞ K L þ a1 a^ 3 ðt 0 Þ  a^ 1 ðt 0 Þ ðT 2 ðtÞ þ T 2 ðtÞÞðT 2 ðtÞ þ T 2 ðtÞÞ SðtÞ 2 R s R s a3 a1   a^ 3 ðt0 Þ a^ 2 ðt0 Þ g  x1 ðtÞ  þa2 a3 a2 2ðT R ðtÞ  T e ðtÞÞ " #) 2 2 ðT R ðtÞ þ T s ðtÞÞðT R ðtÞ þ T s ðtÞÞ þ1 (51)  ðT 2R ðtÞ þ T 2e ðtÞÞðT R ðtÞ þ T e ðtÞÞ that can be rewritten as7  _ ¼ x2 ðtÞ K L þ Dðx2 ðtÞÞ  SðtÞ 2 g ðT R ðtÞ  T e ðtÞÞ Yðx2 ðtÞÞx1 ðtÞ  " # 1 ðT 2R ðtÞ þ T 2s ðtÞÞðT R ðtÞ þ T s ðtÞÞ þ1 Yðx2 ðtÞÞ ¼ 2 ðT 2R ðtÞ þ T 2e ðtÞÞðT R ðtÞ þ T e ðtÞÞ D (54) Eq. From a practical point of view T R ðtÞ. Lemos / Control Engineering Practice 17 (2009) 1157–1173 1167 RefT(t) and Ts(t) [K] Temperature 1200 1000 800 600 400 200 0 50 100 150 200 250 time [s] 300 350 400 450 500 0 50 100 150 200 250 time [s] 300 350 400 450 500 350 400 450 500 us (t) [%] 100 50 0 Simulation of Solar Power variability Gs(t) [W/m2] 1000 900 800 700 600 0 50 100 150 200 250 time [s] 300 Fig. and using Eq. (52) delimited by fg is positive. (38). 10 Note that 0  us ðtÞ  umax which implies that T e ðtÞ  T s ðtÞ  T max . control signal and Gs ðtÞ.10 Imposing this condition the stability zone is defined by 8 > < x ðtÞoK L ðT R ðtÞ  T e ðtÞÞ þ Dðx2 ðtÞÞ ðT R ðtÞ  T e ðtÞÞ 1 g Yðx2 ðtÞÞ g Yðx2 ðtÞÞ > : 8x2 ðtÞ 2 ½T R ðtÞ  T max .M. 7 Note that T s ðtÞ ¼ T R ðtÞ  x2 ðtÞ. Andrade da Costa. 9 6 Note that in this analysis T R ðtÞ and T e ðtÞ are assumed to be constant possibly evolving by steps but with T R ðtÞ4T e ðtÞ. . Process output and reference signal. This can be obtained if the expression in the plane where SðtÞ right side of Eq. T s ðtÞ have positive values because these signals are measured in Kelvin.ARTICLE IN PRESS B. Stability region where8 The stability region is investigated using the following Lyapunov’s function candidate6: Dðx2 ðtÞÞ ¼ a1 SðtÞ ¼ 1 1 a^ ðt Þ gðx1 ðtÞ  xe1 Þ2 þ 3 0 ðx2 ðtÞ  xe2 Þ2 2 2 a3 D (50)   a^ 3 ðt0 Þ a^ 1 ðt0 Þ 2 ðT R ðtÞ þ T 2s ðtÞÞ  a3 a1   a^ 3 ðt0 Þ a^ 2 ðt0 Þ  ðT R ðtÞ þ T s ðtÞÞ þ a2 a3 a2 (53) and9 Computing the time derivative of Eq. (52) can now be used to find the region in the ðx1 ðtÞ. 5. The reference signal T R ðtÞ must be selected according to T e ðtÞoT R ðtÞ  T max .

is given by xe ¼ ð0.5 0 −0. J. a^2 ðt 0 Þ.4. a3 then the equilibrium point xe ¼ ðxe1 . T R ðtÞ  T e ðtÞ The function Yðx2 ðtÞÞ (Eq. Stability of the equilibrium point xe . xe2 Þ inside the region defined by Eq. a2 . 16. T R ðtÞ  T e ðtÞ Note that if ðDðx2 ðtÞÞ=DTe Þð1=Yðx2 ðtÞÞÞ41. when using process parameter estimates When the initial parameter estimates a^1 ðt 0 Þ. Two cases are now analyzed. This is illustrated in Table 4 where the function ðT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞ is evaluated in three points. Simulation III. The function ðK L =gÞðT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞ is a continuous and positive function of x2 ðtÞ for 8x2 ðtÞ 2 ½T R ðtÞ  T max . with perfect knowledge of the process parameter values With perfect knowledge of the values of a1 . T R ðtÞ  T e ðtÞ then the equilibrium point in included in the stability region.M. can be rewritten in a new form which include the coordinate xe1 as shown in (58). x2 ðtÞÞ plane that delimits the stability region.5 −1 x 10−3 α3 (t) 10 9. 5. α3 x 10−8 1 α1 (t) 0. ðDðx2 ðtÞÞ=DTe Þð1=Yðx2 ðtÞÞÞ  1 then the term ðK L =gÞððT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞÞ (in particular K L =g) must be adjusted to increase the stability region in order to contain the equilibrium point xe . the part containing the process parameter estimates. The controller gains K L 40 and g40 have the effect of modifying the size of the stability region.5. (54)) depends only on x2 ðtÞ. is not located at the origin of the ðx1 ðtÞ.5 0 −0. the first one considers a perfect knowledge of the process parameter values a1 .ARTICLE IN PRESS 1168 B. (55). . For simplicity at x2 ðtÞ ¼ 0 it is assumed that T R ðtÞbT e ðtÞ. This is also illustrated in Fig. with xe1 different from null. a2 and a3 the equilibrium point.5 −1 0 50 100 150 200 250 time [s] 300 350 400 450 500 0 50 100 150 200 250 time [s] 300 350 400 450 500 50 100 150 200 250 time [s] 300 350 400 450 500 α2 (t) 0. Evolution of a^ 1 ðtÞ.99 0 Fig. x2 ðtÞÞ plane. Eq. T R ðtÞ  T e ðtÞ b2 define the boundary line in the ðx1 ðtÞ. Lemos / Control Engineering Practice 17 (2009) 1157–1173 Controller gains − Estimates of α1. however. (58) are analyzed to find their contribution to the size of the stability region. and a^ 3 ðtÞ. 5. T R ðtÞ  T e ðtÞ and it does not have singular points. To do that the right side of (55). 13. For this case the equations that define the stability region. If. Stability of the equilibrium point xe . α2. 8x2 ðtÞ 2 ½T R ðtÞ  T max . a2 and a3 and the other case considers the estimates of the process parameters. Andrade da Costa. (38)–(40) yields 8 1 > < x ðtÞoK L ðT R ðtÞ  T e ðtÞÞ þ x Dðx2 ðtÞÞ 1 e1 g Yðx2 ðtÞÞ DTe Yðx2 ðtÞÞ (58) > : 8x2 ðtÞ 2 ½T R ðtÞ  T max . The several terms and factors of Eq. The main conclusion is that the equilibrium point xe ¼ ð0. xb2 ðtÞÞ such that 8 > < x ðtÞ ¼ K L ðT R ðtÞ  T e ðtÞÞ þ Dðx2 ðtÞÞ ðT R ðtÞ  T e ðtÞÞ b1 g Yðx2 ðtÞÞ g Yðx2 ðtÞÞ (56) > : 8x ðtÞ 2 ½T R ðtÞ  T max . is multiplied by ðxe1 Þðx1 e1 Þ. Closed-loop control. 0Þ. and for x2 ðtÞ ¼ T R ðtÞ  T max it is assumed that T max ¼ 10T R ðtÞ. (55). It does not depend on the controller parameters and on the process estimates. (38). a^3 ðt 0 Þ are different from a1 . 0Þ and the condition in (55) is simplified to 8 > < x1 ðtÞoK L ðT R ðtÞ  T e ðtÞÞ g Yðx2 ðtÞÞ (57) > : 8x2 ðtÞ 2 ½T R ðtÞ  T max . a^ 2 ðtÞ.995 9. The closed-loop system to be stable must have the equilibrium point ðxe1 . Using Eqs. 0Þ is inside the stability region and therefore it is stable. This means that the line formed by the points ðxb1 ðtÞ.

4. that it is a continuous function always positive and have values greater of equal to 1. R es ðtÞ dt. 14. control signal and Gs ðtÞ. Closed-loop control. In this analysis the relative magnitude of a1 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 1 ðt0 Þ=a1 Þ and a2 ðða^ 3 ðt0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ and their signs are considered. More at x2 ðtÞ ¼ 0 the function ðDð0Þ=DTe Þð1=Yð0ÞÞ41. The term ðK L =gÞððT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞÞ was analyzed in Section 5.5. 15. Error Integral Table 4 Evaluation of ðT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞ. x2 ðtÞÞ plane. Evolution of the Closed-loop control. 100 x2 ðtÞ ðT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞ T R ðtÞ  T e ðtÞ 0 ðT R ðtÞ  T e ðtÞÞ 2 5ðT R ðtÞ  T e ðtÞÞ if T R bT e ðtÞ T R ðtÞ  T max 1 556ðT R ðtÞ 50 ∫ es(t)dt 0  T e ðtÞÞ if T max ¼ 10T R ðtÞ −50 5. as shown in Table 5. Andrade da Costa. Assuming that a2 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ is dominant If a2 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ is dominant.1. (59) Evaluating now ðDðx2 ðtÞÞ=DTe Þð1=Yðx2 ðtÞÞÞ it can be concluded. It can be used to enlarge the stability region. Process output and reference signal. Assuming that a1 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 1 ðt 0 Þ=a1 Þ is dominant If a1 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 1 ðt 0 Þ=a1 Þ is dominant.M.2. then Dðx2 ðtÞÞ=DTe is simplified to Dðx2 ðtÞÞ ¼1 DTe (60) . J. Without loss of generality it is assumed that xe ¼ ðxe1 .ARTICLE IN PRESS B.5. then Dðx2 ðtÞÞ=DTe is simplified to −100 Dðx2 ðtÞÞ ðT 2R ðtÞ þ T 2s ðtÞÞðT R ðtÞ þ T s ðtÞÞ ¼ 2 DTe ðT R ðtÞ þ T 2e ðtÞÞðT R ðtÞ þ T e ðtÞÞ −150 −200 0 50 100 150 200 250 300 350 400 450 500 time [s] Fig. Simulation of the modified adaptive controller. Simulation of the modified adaptive controller. xe2 Þ is located at the right side of the ðx1 ðtÞ. Lemos / Control Engineering Practice 17 (2009) 1157–1173 1169 RefT(t) and Ts(t) [K] Temperature 1200 1000 800 600 400 200 0 50 100 150 200 250 time [s] 300 350 400 450 500 0 50 100 150 200 250 time [s] 300 350 400 450 500 350 400 450 500 us (t) [%] 100 50 0 Simulation of Solar Power variability Gs(t) [W/m2] 1000 800 600 400 0 50 100 150 200 250 time [s] 300 Fig. The conclusion is that the equilibrium point xe is included in the stability region for K L 40 and g40. 5.

To compensate this situation K L =g must be selected according to the following condition which is a generalization of Eq.25 s is included and used to filter T s ðtÞ þ ZðtÞ. (66) and (67) yields if T R bT e ðtÞ if T max ¼ 10T R ðtÞ K L ðT R ðtÞ  T e ðtÞÞ 1 þ xe1 g Yð0Þ Yð0Þ (61) r 1. first order filter with time constant 1.3. but does not define a strategy on how to select K L and g.   ¼1 @x1 @x2 x xe e @f 2 ðx1 .M. Simulation results This section presents simulation results with the modified adaptive control described by Eqs. (62):   Dð0Þ Yð0Þ  KL DTe (63) 4xe1 ðT R ðtÞ  T e ðtÞÞ g vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u   !2  @f 2 ðx1 . this is used to expose the control algorithm to the control signal saturation . x2 Þ a3 a^ 3 ðt0 Þ a^ 1 ðt0 Þ ¼   K L þ a1  a3 a1 @x2 a^ 3 ðt0 Þ xe Table 6 Evaluation of 1=Yðx2 ðtÞÞ. In order to include the equilibrium point xe in the stability zone.ARTICLE IN PRESS 1170 B.5 enables the definition of conditions on K L =g to have a stable system. x2 Þ @f 1 ðx1 . Lemos / Control Engineering Practice 17 (2009) 1157–1173 x2(t)=TR(t)-Te(t) 5. x2 Þ a3  ¼  a^ ðt Þ g @x1 3 0 x e    @f 2 ðx1 . Selection of the controller parameters K L and g x2(t) P1 Boundary Lines Stability Region P2 x1(t) " dxðtÞ ¼ KL(TR(t)-TR(t))  Θ(x2(t)) Equilibrium Point The stability analysis presented in Sections 5. x2 Þ= @x1 Þjxe ¼ 0 to have a double real pole. This can be concluded by inspecting ðDðx2 ðtÞÞ=DTe Þ as a function of x2 ðtÞ. T R ðtÞ  T max @f 1 ðx1 . x2 Þ=@x2 Þjxe Þ2  4ð@f 2 ðx1 . x2 Þ  4 2    @x2 @x2 @x1 xe xe xe 2 (69) A necessary condition on K L to have a stable linearized system is that ð@f 2 ðx1 . x2 Þ þ u @f ðx1 . There is the possibility that the sign of ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ forces ðDðx2 ðtÞÞ=DTe Þo0 for x2 ðtÞoT r ðtÞ  T e ðtÞ. a^ 3 ðt 0 Þ ¼ a3 those conditions imposed a relation between K L and g given by which defines a minimum value on K L =g such that ðYð0Þ  1Þ ðT R ðtÞ  T e ðtÞÞ (67) (68) In this case ðDðx2 ðtÞÞ=DTe Þð1=Yðx2 ðtÞÞÞ ¼ 1=Yðx2 ðtÞÞ. x2 Þ 5 @x2 Calculating the partial derivatives at the equilibrium point and performing some algebraic manipulation yields Table 5 Evaluation of ðDðx2 ðtÞÞ=DTe Þð1=Yðx2 ðtÞÞÞ. J. These two filters act as unmodeled dynamics. The dynamics of the shutter is simulated by a first order filter with a time constant of 0. In order to tackle the selection of K L and g a linearization of Eq. the other condition on g is needed and can be selected by setting ðð@f 2 ðx1 . White noise ZðtÞ with limited bandwidth and power value of 20 is added to the process output T s ðtÞ. (45)–(49). The temperature reference signal T R ðtÞ is chosen to evolve by steps.7. Assuming that a2 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ and a2 ðða^ 3 ðt0 Þ=a3 Þ  a^ 2 ðt0 Þ=a2 Þ have different signs Assuming that a2 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ and a2 ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt0 Þ=a2 Þ have different signs. A second. except at x2 ðtÞ ¼ T R ðtÞ  T e ðtÞ. then the worst case is when ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þo0 and ðða^ 3 ðt 0 Þ=a3 Þ  a^ 2 ðt 0 Þ=a2 Þ40. x2 ðtÞ ðT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞ T R ðtÞ  T e ðtÞ 0 1 41 85 if T R bT e ðtÞ 41 2 if T max bT R ðtÞ ðT R ðtÞ  T e ðtÞÞ=Yðx2 ðtÞÞ T R ðtÞ  T e ðtÞ 0 1 o1 25 T R ðtÞ  T max o1 1 556 Computing the roots of the characteristic polynomial of Eq. the following condition must be fulfilled: 4xe1 (66) ðT 2R ðtÞ þ T 2e ðtÞÞðT R ðtÞ þ T e ðtÞÞ #   a^ 3 ðt0 Þ a^ 2 ðt0 Þ 4T 34 ðtÞ  þa2 a3 a2 ðT 2R ðtÞ þ T 2e ðtÞÞðT R ðtÞ þ T e ðtÞÞ x2 ðtÞ KL (65) xe   @f 1 ðx1 . 16. x2 Þ ¼ 0.12 s. The perturbed state vector is defined as x1 ðtÞ  xe1 # (64) x2 ðtÞ  0 The dynamics of the linearized system is represented by P3 x2(t)=TR(t)-Tmax 2 Fig. x2 Þ t @f 2 ðx1 . In the special case with a^ 1 ðt 0 Þ ¼ a1 .4 and 5. Andrade da Costa. and using Eqs. This is illustrated in Table 6. as a remark. a^ 2 ðt 0 Þ ¼ a2 .2 ¼ g (62) 5. (65).6. The function 1=Yðx2 ðtÞÞ have values lower than 1. Stability region in the ideal case (perfect knowledge of the a1 . x2 Þ=@x2 Þjxe o0. x2 Þ @x1 g¼  KL 2 2 (70) 5. xe1 o 3 @f 1 ðx1 .5. a2 and a3 values). x2 Þ 6 @x1 6 dx_ ðtÞ ¼ 6 4 @f 2 ðx1 . (36) at the equilibrium point is performed. the unmodeled dynamics was not considered in controller design and in the stability analysis. x2 Þ 7 @x2 7 7 dxðtÞ @f 2 ðx1 .

(70). use of the anti-windup mechanism. The controller is able to track the temperature reference signal but the control signal is too oscillatory and has a huge amplitude. The main conclusion is that the modified adaptive controller shows a good performance and is easier to use. use of the anti-windup mechanism. The results shown in Figs. The control signal us ðtÞ is now more smooth than in the previous computer simulations. a^ 1 ðt 0 Þ ¼ 1:2  1011 .M. J. K L ¼ 0:2 and g ¼ 1:0  102 . The control algorithm is able to deal with the process input saturation. a^ 2 ðt 0 Þ. this can damage the shutter. control signal and Gs ðtÞ. This can be interpreted as a combination of noise amplification and unmodeled dynamics ‘‘excitation’’ caused by a huge value of K L . the aim is to simulate the lack of knowledge about the process parameter values at higher temperatures. The value of K L was selected based on the level of noise present at the process output ZðtÞ and the allowed level of shutter oscillation induced by the feedback of ZðtÞ. The following values are used in the computer simulations. Integral Error 500 400 ∫ es(t)dt 300 200 100 0 –100 0 50 100 150 200 250 300 350 400 450 500 Fig.ARTICLE IN PRESS B. and also prevents the damage of the shutter. Andrade da Costa. a^ 3 ðt 0 Þ were selected to be near the process parameter values at T s ¼ 400 K. Simulation of the simplified adaptive controller. a^ 3 ðt 0 Þ ¼ 2:3  102 . Lemos / Control Engineering Practice 17 (2009) 1157–1173 1171 TR(t) and Ts(t) [K] Temperature 1200 1000 800 600 400 200 0 50 100 150 200 250 300 350 400 450 500 300 350 400 450 500 350 400 450 500 time [s] us (t) [%] 100 50 0 0 50 100 150 200 250 time [s] Simulation of Solar Power variability Gs(t) [W/m2] 1000 800 600 400 0 50 100 150 200 250 time [s] 300 Fig. the gain was increased by a factor of 5. problem. Closed-loop control. 17. a^ 2 ðt 0 Þ ¼ 3:41:2 103 . Process output and reference signal. Simulation of the simplified adaptive controller. The following anti-windup mechanism ( x_ 1 ðtÞ ¼ T R ðtÞ  T s ðtÞ if uðtÞ is not saturated otherwise x_ 1 ðtÞ ¼ K L x1 ðtÞ is used to deal with the control signal saturation. This also contributes to reduce the effects of the presence of unmodeled dynamics that can cause stability problems. 18. . R es ðtÞ dt. 19 and 20 were obtained using K L ¼ 1 and g ¼ 0:25. The simulation results are shown in Figs. The value of g was computed from Eq. Evolution of the Closed-loop control. 17 and 18. due to the selection of a lower gain for K L. (71) The controller parameters a^ 1 ðt 0 Þ.

The new version can be interpreted as a PI controller with adaptive parameters and a feed-forward term. control signal and Gs ðtÞ. The . an analysis of the adaptive controller was done. which can be 15 ∫ es(t)dt a problem. use of the anti-windup mechanism. Guerra from the Materials and Engineering Department/ IST/TU Lisbon—Portugal. The evaluation of the simplified control with computer simulation demonstrated that it has a good performance. Conclusions This paper explores exact linearization and the Lyapunov methods to design an adaptive controller for a solar furnace. Closed-loop control. With this new controller version it is easier to understand how to select the controller parameter values. J. Simulation of the simplified adaptive controller. Process output and reference signal. and to Prof. L. R. 20. The experimental data used for the simulations were obtained under the SolControl project. After development and evaluation of the adaptive controller the following points can be drawn: Integral Error 25 20  The controller performed well in computer simulations. and a modified version was developed. rities present in the plant and disturbances associated to solar radiation. 6. G. use of the anti-windup mechanism. 5 0 –5 0 50 100 150 200 250 300 350 400 450 500 Fig. R es ðtÞ dt. 10 To overcome the problems described above. 19.M. Andrade da Costa. Simulation of the simplified adaptive controller. which was financed by EU under the SolFace Contract number RITA-CT-2003-507091.  The controller has seven parameters to be tuned. Evolution of the Closed-loop control. Lemos / Control Engineering Practice 17 (2009) 1157–1173 TR(t) and Ts(t) [K] Temperature 1200 1000 800 600 400 200 0 50 100 150 200 250 300 350 400 450 500 300 350 400 450 500 350 400 450 500 time [s] us (t) [%] 100 50 0 0 50 100 150 200 250 time [s] Simulation of Solar Power variability Gs(t) [W/m2] 1000 800 600 400 0 50 100 150 200 250 300 time [s] Fig.ARTICLE IN PRESS 1172 B. The aim is to develop a controller able to compensate the nonlinea- Acknowledgments The authors thank Dr. Olalde from PROMES-France for its supporting in the SolControl project.

(2006). E. Costa. & Feliachi. F. G. Slotine. J. & Schaufelberger. M. J. L. References Berenguel. Kinetic aspects of reaction between tantalum and carbon material (active carbon or graphite) under solar radiation heating. M. E. Control Engineering Practice. C. Yebra.-J. Lacasa. Olalde. J. Temperature control of a solar furnace. B. A. Germany. E. Amaral. & Rubio. Fernandes. Berlin: Springer.. J. Heat transfer—a practical approach (2nd ed. Applied nonlinear control.. 8–24. J. L. (2002). C. In CONTROLO 2008. Guillot. Lemos / Control Engineering Practice 17 (2009) 1157–1173 control algorithm was developed under the project SFERA. Olalde. Copper sintering in a solar furnace through fuzzy control. J. Zambrano. (2006). Dynamic simulation of stick-slip motion of a flexible solar array. S. In Proceedings of the 2006 IEEE international conference on control applications. 17(6). Camacho. B. Renewable energy—power for a sustainable future.. Rosa L. J. F. C... C. A. Taniwaki. & Shohoji.. Control Engineering Practice.. NJ: Prentice-Hall. . Rosa. 652–663. Badie. (2009). Temperature control of a solar furnace for material testing. In MED08. A. An adaptive temperature control law for a solar furnace. Y. 26. Temperature control of a solar furnace with disturbance accommodating controller. (2003). L. Y.. Garcia-Gabin. G. Englewood Cliffs. A. 523–533. Rosa. Berenguel.. A... F. D. & Martı´nez D. F. Andrade da Costa. F.. N. 37(8). & Shohoji.M.. Granier. J. In Proceedings of the 34th Southeastern symposium on system theory.. W. G.. M. W. IEEE Control Systems. Paradkar. Lemos. N. 80... Weibull statistical analysis of flexure breaking performance for alumina ceramic disks sintered by solar radiation heating. P.. & Camacho. & Fernandes J. Oliveira.ARTICLE IN PRESS B. Control systems with input and output constraints. 19(1). B. 16(6). (June 2008). Y. R. G. H. (1999). (2008a). Rosa L. 724–735. Grant agreement no. Boyle. 1553–1560. & Li. C . & Fernandes J. (2000)... W... Glattfelder. Lemos. New York: McGraw-Hill. 8th Portuguese conference on automatic control. M. G. Sliding mode predictive control of a solar air conditioning plant.. A. Solar Energy.). G. 203–206. & Okiami.engel. M. (1996). Garcia-Martin. A.. 1173 Fernandes. Science Direct—Elsevier.... 2008 Mediterranean conference on control and automation. Ceramics International—Elsevier. J. (2003).. D. M. E.. Lemos. 228296. Kojima. E. (2008b). Davari A.. Guillot. Costa. International Journal of Systems Science. Adaptive control of distributed collector solar fields.. (2006). C. G. Oxford: Oxford University Press (in association with the Open University). M. (1991)..