Journal of Process Control 24 (2014) 479–486

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

Set-based adaptive estimation for a class of nonlinear systems with

time-varying parameters
Samandeep Dahliwal a , Martin Guay b,∗
a
Hatch and Associates, Sudbury, ON, Canada
b
Department of Chemical Engineering, Queen’s University, Kingston, ON, Canada

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, an adaptive estimation technique is proposed for the estimation of time-varying parameters
Received 28 December 2012 for a class of continuous-time nonlinear system. A set-based adaptive estimation is used to estimate the
Received in revised form time-varying parameters along with an uncertainty set. The proposed method is such that the uncertainty
22 November 2013
set update is guaranteed to contain the true value of the parameters. Unlike existing techniques that rely
Accepted 26 November 2013
on the use of polynomial approximations of the time-varying behaviour of the parameters, the proposed
Available online 20 February 2014
technique does require a functional representation of the time-varying behaviour of the parameter esti-
mates. A simulation example and a building systems estimation example are considered to illustrate the
Keywords:
Nonlinear estimation
developed procedure and ascertain the theoretical results.
Adaptive estimation © 2013 Elsevier Ltd. All rights reserved.
Time-varying parameters

1. Introduction a local polynomial expansion of the time-varying parameters. In

The problem of parameter estimation has been of consider-
able interest during the last two decades. The vast majority of
the literature on adaptive estimation and adaptive control relies
on the assumption that the parameters of the system are slowly
time-varying or constant. In practice however, the time-varying
behaviour of the process parameters may be of significant impor-
tance. Time-varying parameters can arise from uncertain complex
mechanisms not captured by the process model. They can also arise
from model-plant mismatch or unmeasured inputs that affect the
process dynamics.
Evidently, the ability to estimate such time-varying behaviour
can have a significant impact on control system performance. Sev-
eral researchers have considered the problem of estimation of
time-varying parameters. In [18], an estimation routine is proposed
for a class of nonlinear systems with time-varying parameters. A
Taylor series expansion of the time-varying parameters is consid-
ered. The time varying behaviour is captured by applying a local
polynomial expansion and estimating the (locally) constant Tay-
lor series coefficients using a standard least-squares approach. The
authors demonstrate the convergence of the parameter estimates
to a neighbourhood of the true parameters within a given reset-
ting period. An related approach is proposed in [5] where a direct
least-squares based method is used to estimate the coefficients of
∗ Corresponding author. Tel.: +1 613 533 2788; fax: +1 613 533 6637.
E-mail address: martin.guay@chee.queensu.ca (M. Guay).
[7], a sliding mode approach is proposed where convergence of
the parameter estimates to the true estimates is guaranteed. The
technique relies on an estimate of the sign of the parameter esti-
mation error which may be very difficult to obtain in practice. An
alternative approach was proposed in [12] for the estimation of
time-varying parameters using a dynamic inversion approach. In
this approach, it is shown that inversion technique can be useful in
the design of reliable parameter estimation techniques subject to
the assumption that the regressor vector is uniformly full rank.
Many researchers have considered the problem of unknown
input estimation which is very closely related to problem of esti-
mation of time-varying parameters. Combined state and unknown
input estimation for a class of nonlinear systems is proposed in [6].
High-gain observer methods are proposed in [13]. A sliding-mode
observer design approach is proposed in [14] for the estimation
of unknown inputs. In contrast to the estimation of time-varying
parameters, the estimation of unknown inputs rely on often restric-
tive structural assumptions which are related to the observability of
the inputs from the available measurement. In parameter estima-
tion, this structural information would in general be limited to an
identifiability requirement or a persistency of excitation condition.
The analysis of nonlinear systems with time-varying parame-
ters has been considered in [3]. In this paper, the authors propose an
adaptive observer canonical form and propose an adaptive observer
or this class of systems. It is claimed that the estimation of time-
varying parameters can be performed subject to a bound on the rate
of change of the parameters. A related result is presented in [10]
where an adaptive observer is proposed for a class of nonlinear

0959-1524/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jprocont.2013.11.015
480 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486

systems with time-varying parameters and bounded disturbances. where (t) is a disturbance that depends on the dynamics of the
As in [3], estimation of the time-varying parameters is possible measurement noise ω(t) and the difference between the internal
subject to a bound on the derivative of the parameters. state z and its measurement x. This term can be added as a param-
The present paper is a generalization of the identification eter to be estimated if it meets the Assumptions stated above. In
scheme presented in [2] to the problem of time-varying parameter most situations, one can assume that the measurement noise ω
estimation. We propose a set-based parameter estimation tech- is small relative to the magnitude of the state measurement, i.e.,
nique that provides simultaneously estimates of the parameters z(t) ≈ x(t).
along with an estimate of the uncertainty of the parameters. The If, on the other hand, the system is stated in terms of the internal
approach provides conditions that ensure the convergence of the state variables:
parameter to a neighbourhood of the true value. In addition, we
develop a set-update algorithm that updates the uncertainty set to ż = f (z) + g(z)(t)
guarantee non-exclusion of the true parameter estimate. x =z+ω
This paper is organized as follows. The problem description is
then an adaptive filter for the estimation of z and  is required. The
given in Section 2. The parameter estimation routine and uncer-
design of such nonlinear filters is beyond the scope of the proposed
tainty set adaptation are presented in Section 3. Two simulation
approach.
examples are presented in Section 4 followed by brief conclusions
in Section 5.
3. Parameter and uncertainty set estimation
2. Problem description
3.1. Parameter estimation
Consider a nonlinear system
Let K be an n by n matrix chosen such that K + KT > I. Let the state

p
predictor model for (1) be chosen as
ẋ = f (x) + gi (x)p (t) = f (x) + g(x)(t) (1)
i=1 x̂˙ = f (x) + g(x)(t) ˆ˙
ˆ + Ke + wT (t), (2)
where f(x) and gi (x) (i = 1, . . ., n) are sufficiently smooth vector val- T T
ẇ = −Kw + g(x), w(t0 ) = 0. (3)
ued functions, g(x) = [g1 (x), . . ., gp (x)] ∈ Rn×p , x ∈ Rn is the state,
(t) ∈ Rp is an unknown time varying bounded parameter vector. where x̂ ∈ Rn
is the vector of predicted state variables, w ∈ Rn×p
The following assumptions are made about (1). is the regression matrix, e = x − x̂ is the state estimation error, ˆ ∈
˙
Assumption 1. The state trajectories x(t) evolve on a compact Rp are the parameter estimates and ˆ is the parameter estimation
subset X of Rn . update law to be defined below. By (2) and (3), the error dynamics
are given by:
Assumption 2. The unknown parameter vector (t) is such that
˙
(t) < c1 , where c1 is a positive constant. ė = g(x)(t) ˆ˙
˜ − Ke − wT (t) (4)
Assumption 3. The parameter vector (t) is assumed to be ˜
Next, we define an auxiliary variable  = e − wT (t). The  dynam-
uniquely identifiable and lies in a known compact set 0 = B( 0 , ics are given by:
z ), where  0 = (0) is a nominal parameter value and z > 0 is the
˙
˙ = −K − wT (t), (t0 ) = e(t0 ) (5)
radius of the parameter uncertainty set.
Remark 4. Identifiability is a standard requirement in adaptive ˙ we define an
Since the rate of change of the unknown parameter ,
estimation. It states that for any two output trajectories y(t,  1 ) and estimate of  with dynamics given by:
y(t,  2 ) corresponding to two parameter values  1 and  2 , respec-
ˆ˙ = −K 
 ˆ (6)
tively, the following property can be guaranteed:
˜ =−
We define the  estimation error  ˆ with dynamics
y(t, 1 ) = y(t, 2 ), ∀t≥t0 ⇒ 1 = 2 .
˜˙ = −K 
 ˙
˜ − wT (t), ˜ 0 ) = 0.
(t (7)
Assumption 5. The state variables are available for measurement
in continuous time. In the following, a bound on  ˜ is required. Such a bound can
be generated as follows.
The objective of this paper is provide an estimation of the Let us first note that the initial estimate (0)
˜ = (0) − (0)
ˆ = 0.
unknown time-varying parameters along with an estimate of their Next, consider the Lyapunov function V = (1/2) ˜ T .
˜ The rate of
uncertainty. change of V is given by:
Remark 6. The class of nonlinear system considered does not  
V˙ ˜T K
= ˙
˜ − wT (t) ˜T K
= − ˜ T wT ˙
˜ +
explicitly take into account the effect of measurement noise and
disturbances. However, the time-varying parameters can be inter- (8)
k1 T 1 ˙T ˙
preted as process disturbances that are subject to a bound on their ˜T K
≤ − ˜+ ˜ wwT 
 ˜+  
2 2k1
time derivative. If the disturbances are such that they do not meet
Assumption 3 then one must assume that the disturbances have a where k1 is some positive constant.
known bound as in [2]. The implication is that the disturbances will Thus if one assigns K as
increase the size of the parameter uncertainty region.
k2 k1
In situations where the system is subject to measurement noise, K= I+ wwT
2 2
one must write the measured state variables as x = z + ω(t) where
ω(t) is a bounded signal with zero mean. The dynamics (1) must be for some positive constant k2 with I, the identity matrix, the
interpreted in terms of the noisy state measurement x. That is, it is inequality above becomes:
assumed that one can write:
1 ˙ 2 c2
V˙ ≤ −k2 V +  ≤ −k2 V + 1 (9)
ẋ = ż + ω̇ = f (x) + g(x)(t) + (t) 2k1 2k1
 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486 481

Since (0) = (0)

ˆ = e(0), it follows that (0)
˜ = 0. The inequality 3.2. Parameter set adaptation
(9) yields:
 t
The analysis from the last section indicates that it will not be
−k2 t −k2 (t−)
c12 c12 ˜
possible in general to guarantee that lim (t) = 0. However, one
V (t) ≤ e V (0) + e d ≤ t→∞
2k1 2k1 k2
0 can always guarantee that the parameter estimation error belongs
or to a neighbourhood of the origin. In this section, we propose an
c1 algorithm that yields an estimate of this neighbourhood which we
(t)
˜ ≤  ∀t≥0. call the uncertainty set. The uncertainty set is taken to be a ball of
k1 k2 radius z centred at the origin.
The proposed parameter estimation scheme can now be defined. Using the analysis from the previous section, an update law that
Let  ∈ Rn ×n be generated from measures the worst-case progress of the parameter identifier in the
presence of a disturbance is given by
˙ = wwT − kT  + ıI,
 (t0 ) = ˛I
0, (10)
Vz (t)
where ˛, kT and ı are positive constants to be assigned. z (t) = (16)
4min ((t))
Based on (2), (3) and (6), one considers the parameter update
2
law as proposed in [1] is given by Vz(t) (0) = 4max ((0)) (z0 ) (17)

˙ −1 1
 = −−1 wwT −1 + kT −1 − ı−T −T , −1 (t0 ) = I, (11) 1

˛ V̇z (t) = −k4 Vz (t) − (e − )

T
ˆ + z2 + Vz
ˆ (e − )
  2 2 0
ˆ˙
(t) = proj −1 (w(e − ) ˆ (t)
ˆ −
), ˆ ˆ 0 ) =  0 ∈ 0 ,
, (t (12)
c12
+ max ((t)) (18)
2k3
ˆ
where
is a positive constant and Proj {, (t)} denotes a Lipschitz
projection operator. The projection [8] has the following proper-
ties: c12
V̇z (t) = −k2 Vz (t) + , Vz (0) = 0. (19)
T 2k1
ˆ
−Proj{, (t)} ˜
(t) ˜
≤ −T (t), (13)
where Vz (t) represent the solutions of the ordinary differential
ˆ 0 ) ∈ 0 ⇒ (t)
(t ˆ ∈ , ∀t≥t0 (14) equation (18) with the initial condition (17), and Vz (t) is the solu-
tion of the initial value problem (19).
ˆ
where   B((t), ˆ and z are the parameter estimate
z ), where (t)  The parameter uncertainty set, defined by the ball B(ˆ c , zc ) is
and set radius found at the latest set update respectively. updated using the parameter update law (12) and the error bound
(16) according to the following algorithm:
Assumption 7. There exists constants ˛1 > 0 and T > 0 such that
 t+T Algorithm 1.
w()w()T d≥˛1 I (15)
t ˆ i−1 ) = ˆ 0
(1) Initialize z (ti−1 ) = z0 , (t
∀t > 0. (2) At time ti , update

Theorem 1. Let Assumption 7 be met by the trajectories of (1). The
 ˆ i ), (ti )),
identifier composed of (11) and the parameter update law (12) is such
˜ ˆ
that the estimation error (t) = (t) − (t) is bounded and such that
˜
lim (t) ≤ (3.) Iterate back to step 2, incrementing i = i + 1.
t→∞
where > 0 is a strictly positive constant.
Proof. The proof is given in Appendix A. 
Remark 9. Theorem 1 outlines the conditions required for the
parameter estimates to converge to the unknown time-varying
parameters. The results demonstrate that the proposed estimation
technique has the property to converge to a neighbourhood of the
parameter estimates that is proportional to the rate of change of
the unknown parameters. In situations where the anticipated rate
of change of the parameters is large, one can increase the gain of
the estimation routine kT to compensate. The proof of the theorem
also demonstrates that this action of the gain kT will depend on
the PE condition (15). Since it is difficult to assess this condition
in real-time, the ability to increase kT may be limited. In the next
section, we provide an algorithm that can be used to monitor the
uncertainty associated with parameter estimates in real-time. The
uncertainty incorporates various factors that lead to uncertainty
estimates including external excitation signals and uncertainty in
the rate of change of the parameters. The estimation of the uncer-
tainty set can be used to handle the inherent limitations of the
estimation technique by mitigating all the contributing factors.
((t if z(ti ) ≤ z (ti−1 )−c1 |(ti+1 − ti )| − (ˆ i+1 − ˆ i )
ˆ  =
,
ˆ i−1 ), (ti−1 )),
((t otherwise
Algorithm 1 ensures that  is only updated when the value of
z has decreased by an amount which guarantees a contraction of
the set. Moreover, the evolution of z as given in (16) ensures non-
exclusion of (t) as shown in the following lemma.
Lemma 1. The evolution of  = B(0, z ) under (11), (16) and Algo-
rithm 1 is such that
(1)  ∈ (t0 ) ⇒  ∈ (t) ∀ t ≥ t0
(2) (t2 ) ⊆ (t1 ), t0 ≤ t1 ≤ t2
Proof. The proof is given in Appendix B 
Remark 12. This last lemma provides a generalization of the result
of [2] for time-varying parameters. It guarantees that the param-
eter uncertainty set estimate is forward invariant. Starting from a
known uncertainty set, the set update algorithm ensures that the
true value of the parameter is always contained in a ball of known
radius centred at the current value of the parameter estimates. The
only tuning parameter required are the upper bound on the rate
of change of the unknown parameters and the radius of the initial
parameter uncertainty region.
482 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486

4.5 2
zθ
Error
1.8
4

1.6
3.5
1.4

3
1.2

2.5 1
θ

zθ
0.8
2

0.6
1.5
0.4

1
0.2
θ
1
θ
2
0.5 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time Time

Fig. 1. Plot of the parameter estimates along with the true parameter values. Fig. 2. Upper bound of the estimation error norm ˜ and the uncertainty set radius
z as a function of time.

4. Simulation example
ceiling) and mass of air is assigned a uniform temperature whose
4.1. Example 1 dynamic behaviour is modeled as due to conductive or convective
heat transfer [4]. Hence, for a reasonably complex model, many
To illustrate the effectiveness of the proposed method, we con- temperatures are available for measurement.
sider the following system The reduced order model of the temperature dynamics is based
on mass and energy balances around the entire building envelope
ẋ1 = x2 and around each zone. Under the thermal-network framework,
an effective modeling technique is the “3R2C” approximation. The
ẋ2 = −1 (t)x2 − 102 (t)arctan(900x2 ) + 10u
“3R” component refers to the three resistances to heat transfer
where  2 (t) = 3 cos(0.5 t) and between adjacent zones. Resistances (1) and (2) are due to the
⎧ combined effects of two convective heat transfer stages between
⎨ 2 + sin(t) 0 ≤ t < 40 surfaces and air, and resistance (3) originates from the conductive
1 (t) = 2 + sin(40) + sin(2(t − 40) 40 ≤ t < 80 heat transfer through solid surfaces. The term “2C” refers to the dis-
⎩ cretization of walls and other surfaces into two separate entities,
2 + sin(40) + sin(80) t≥80
each with its own uniform temperature and thermal capacitance.
For simulation purposes, it is assumed that x1 (0) = 1 and x2 (0) = −8 Neglecting water evaporation, exogenous air infiltration, and
and u = 100 sin(20 t). The parameter estimates are plotted along interzone air mixing, an energy balance around zone j is given by:
with their true value in Fig. 1. The parameter estimates follow 
Nsurf,j
dT zone,j
the true unknown parameters effectively. The uncertainty radius is Mair,j Cp,air = ṁsa Cp,air (Tsa − Tzone,j ) + Qint,j + Qsurf,i ,
compared to the norm of the parameter estimation error in Fig. 2. As dt
i=1
expected, the norm of the parameter estimation error is always less (20)
than the uncertainty radius. This confirms that the true unknown
uncertain parameters are contained in the uncertainty set. where Tzone,j (K) and Mair,j (kg) are zone j’s air temperature and mass
of air, Cp,air (J/kg K) is the specific heat capacity of air (assumed con-
stant), ṁsa (kg/s) and Tsa (K) are the mass flow rate and temperature
4.2. Building systems example
of air supplied by the FCU, Qint,j (W) is zone j’s internal load, Qsurf,i
We consider the dynamical model of a one-floor office build-
ing represented by a thermal network model, a popular approach
for modeling heat transfer in the building systems literature
[9,11,15–17]. The building system, whose floor plan is shown in
Fig. 3, contains seven zones that are defined based on the avail-
ability of local actuators. In this example, it is assumed that the
office is subject to a warm climate. As a result, only air condition-
ing (cooling) is required to maintain a comfortable atmosphere for
the building’s occupants. Air conditioning in each zone is accom-
plished via fan coil units (FCUs), which supply cool air to the zones.
The white, unlabeled spaces in Fig. 3, are uncontrolled areas whose
temperatures are considered to be disturbances. Surrounding the
building is a corridor whose temperature is also uncontrolled.
Fig. 3. A floor plan of the one-floor office building considered in this chapter. Zones
Further disturbances include the ground temperature, ceiling tem-
are defined based on the availability of an actuator. White spaces and the area
perature, and internal load (which is due to occupants, lights, etc.). surrounding the figure are uncontrolled areas whose temperature is considered a
Under the thermal network formulation, each surface (wall, floor, disturbance.
 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486 483

(W) is the heat exchanged to/from the room via the various sur- 38
faces (walls, ceiling, ground), and Nsurf,j is the number of surfaces
adjacent to zone j’s air mass. For a surface adjacent to zone j, the 36
rate of change of its temperature is given by
in out − T in
Tsurf,i 34
dT surf,i in surf,i
Cs,i = hj As,i (Tzone,j − Tsurf,i )+ , (21)
dt Rcon,i
32

Tz on e i
in
where Tsurf,i (K) is the ith surface adjacent to zone j’s mass of air. The
out (K), is that of the surface adjacent to the ith sur-
temperature Tsurf,i 30

face, hj (W/m2 K) is the convective heat transfer coefficient between Zone 1

Zone 2
the surface and zone j’s air (assumed constant), Rcon,i (K/W) is the 28
Zone 3
conductive resistance between two adjacent surfaces, and As,i (m2 ) Zone 4
is the area of the surface. The thermal resistance and capacitance Zone 5
26
Zone 6
are defined by Zone 7

wi i Cp,i wi As,i 24
Rcon,i = , Cs,i = , (22) 0 2 4 6 8 10 12 14 16 18 20
ki As,i 2 t (Days)

where wi (m), and ki (W/(m K)) are the thickness and thermal con-
Fig. 4. State trajectories over 20 days.
ductivity of surface i respectively, i (kg/m3 ) is the density of the
surface’s material, and Cp,i (J/kg K) is the specific heat capacity of
the surface’s material. The thermal-network parameters used are
0.35
based on those determined during on-line estimation by the con-
trol systems research group at the United Technologies Research
0.3
Center (UTRC) for the particular building.
The disturbances that affect the indoor temperature of the office
building consist of: 0.25

• a corridor temperature, Tcorr 0.2

θi ( t )

• a ground temperature, Tgr ,

• a ceiling temperature, Tceil , 0.15
• internal loads Qint,j , j = {1, . . ., 7}.
0.1
The estimation problem considered in this paper is based on the
following parameterization of the model equations (20): 0.05
dT zone,j
Mair,j Cp,air = ṁsa Cp,air (Tsa − Tzone,j ) + j (t) (23)
dt 0
0 2 4 6 8 10 12 14 16 18 20
where  j (t) represents the contribution of the internal zone load t (Days)
and the effect of surfaces in each zone. We use the set-based
estimation routine presented in this paper to estimate the seven Fig. 5. Plot of the heat loads estimates for each zone along with the true values.
parameters. Refer to the legend in Fig. 4 that identifies the uncertain parameters in each zone.

The internal load is assumed to be time-varying with


1 + 2 sin(2 t)) if t < 10 or t≥ 13 0.35
Q (J/s) =
0 otherwise
0.3
θ -Fulllines
where the units of t are given in days (for simplicity).
The objective is to estimate the corresponding effect on the tem- θ -Dottedlines
perature dynamics in the form of the time varying estimation of  i , 0.25
i = 1, . . ., 7. The initial estimate for each parameter are set to 0 and
the initial uncertainty radius is set to 2. Initial conditions of all tem- 0.2
peratures are set to 25 ◦ C. The flow of air into each zone ṁsa (kg/s)
θi ( t )

is chosen such that ṁsa Cp,air Tsa = 1 (J/s).

0.15
The simulation results are shown in Figs. 4–7. The temperature
in each zone is shown in Fig. 4. The parameter estimates are plotted
along with their true value in Fig. 5. Since the two are indistin- 0.1

guishable on the time scale considered, Fig. 6 shows the same plot
over the interval t ∈ [0, 1]. The parameter estimates, which are the 0.05
unknown heat loads in each zone, follow the true values effectively.
The uncertainty radius is compared to the norm of the parameter 0
estimation error in Fig. 7. As expected, the norm of the parame- 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

ter estimation error is always less than the uncertainty radius. This
confirms that the true unknown uncertain parameters are always
contained in the uncertainty set.
To ensure that practical viability of the proposed parameter esti-
mation framework, we verify its performance in the presence of
t (secs)
Fig. 6. Plot of the heat loads estimates for each zone along with the true values
over the interval t ∈ [0, 1]. Refer to the legend in Fig. 4 that identifies the uncertain
parameters in each zone.
484 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486

2 0.5

1.8 0.45

1.6 0.4

1.4 0.35

1.2 0.3

1 0.25

0.8 0.2

0.6 0.15

0.4 0.1

0.2 0.05

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0

Fig. 7. Upper bound of the estimation error norm ˜ and the uncertainty set radius Fig. 10. Upper bound of the estimation error norm ˜ and the uncertainty set radius
z as a function of time. z as a function of time in the presence of noise.

measurement and/or process disturbances. We repeat the simula-

45 tion study in the presence of noise in each zone temperature. The
noise term (t), sampled from a standard normal distribution with
zero mean and standard deviation 0.01, enters the dynamics of the
40 zone temperatures additively. That is,

dT zone,j
Mair,j Cp,air = ṁsa Cp,air (Tsa − Tzone,j ) + j (t)
35
dt
+ 0.01Mair,j Cp,air (t). (24)

30
Fig. 8 shows the corresponding zone 1 temperature. The parameter
estimates for zone 1 along with the true value is shown in Fig. 9.
The noise has a noticeable effect on the parameter estimates but the
parameter estimates remain close to the true values. Furthermore,
25
if one considers the bounds computed by the set-based algorithm
as shown in Fig. 10, we notice that the variations of ˆ 1 are within the
expected theoretical limits imposed by the set update. Overall, the
20
0 0.5 1 1.5 2 2.5 3 simulation results demonstrate that the technique performs very
Time (Days) well in the presence of external disturbances.

Fig. 8. Temperature for zone 1 in the presence of exogenous noise.

5. Conclusion

An adaptive estimation technique is proposed for the estimation

0.25 of time-varying parameters for a class of continuous-time nonlin-
ear system. A set-based adaptive estimation is used to estimate
the time-varying parameters along with an uncertainty set. The
0.2 proposed method is such that the uncertainty set update is guar-
anteed to contain the true value of the parameters. Unlike existing
techniques that rely on the use of polynomial approximations of
0.15 the time-varying behaviour of the parameters, the proposed tech-
nique does require a functional representation of the time-varying
behaviour of the parameter estimates. The techniques yields guar-
0.1 anteed bounds for the uncertainty on the parameter estimates that
can be utilized for the monitoring and robust control of uncertain
systems.
0.05

Appendix A. Proof of Theorem 1

0
0 0.5 1 1.5 2 2.5 3 Consider the Lyapunov function,

1˜ T ˜
Fig. 9. Parameter estimate for the heat load in zone 1 in the presence of noise. V(t)
˜ = (t) (t)
2
 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486 485

it follows from (11) that As a result, (A.2) yields:

˜˙
˜ T (t)(t) 1˜ T ˜ 1˜ T ˜ + ı ˜ T .
˜ 1 2 2 c2

V̇(t)
˜ = (t) − kT (t) (t) + (t) wwT (t) V̇(t)
˜ ≤ −k4 V˜ − (e − )
T
ˆ (e − )
ˆ + c + 1 + z2 . (A.3)
2 2 2 2 2k3 1 2k1 2 0
(A.1)
 t  
1
2
e−k4 (t−)
2
˜ V˜ (t) ≤ e−k4 t V˜ (0) + + c12 + z )d
If ones substitutes for (12) and wT (t) =e−
˜ − ,
ˆ it follows that, 0
2k3 2k1 2 0
invoking the properties projection algorithm,  1

2
2
≤ e−k4 t V˜ (0) + + c12 + z .
˜ ˙T 1˜ T ˜ ˜ T w(e − ) 2k3 2k1 2 0
V̇(t)
˜ ≤ (t) (t)(t) − kT (t) (t) − (t) ˆ
2
Taking the limit as t→ ∞, one gets
1˜ T
+ (t) ˜ + ı ˜ T ˜ +
˜ T ˆ
wwT (t)  1

2
2 2 2
lim V˜ (t) ≤ + c12 + z
t→∞ 2k3 2k1 2 0
˜ T (t)(t)
˙ kT ˜ T ˜ ı
≤ (t) − (t) (t) + ˜ T ˜ or
2 2
1  1
 c2
2
T
−(e − 
ˆ − )
˜ (e − )
ˆ + e −  ˜ 2 +
˜ T ˆ
ˆ −  ˜
lim (t) ≤
2
+ 1
+ z = .
2 t→∞ k3 2k1 1 2 1 0
˜ T (t)(t)
˙ 1˜ T ˜ ı
≤ (t) − kT (t) (t) + ˜ T ˜ As a result, the parameter estimation error enters a small
2 2 neighbourhood of the origin approximated by a ball of radius
1 T 1 T parameterized by the adjustable positive gains
, k1 and k3 .
− (e − )
ˆ (e − )
ˆ +  ˜ +
˜ T ˆ
˜ 
2 2

˜ T (t)(t)
˙ 1˜ T ˜ ı Appendix B. Proof of Lemma 1
≤ (t) − kT (t) (t) + ˜ T ˜
2 2
1 T 1 T (1) By definition, we can establish that V˜ (t0 ) ≤ Vz (t0 ). From (9), it
− (e − )
ˆ (e − )
ˆ +  ˜ −
˜ T ˜ +
˜ T 
˜  follows that
2 2
Next, one sets
= ı to obtain: c12
V̇˜ (t) ≤ −k2 V˜ (t) + .
1˜ T ˜ ı 2k1
V̇(t)
˜
˜ T (t)(t)
≤ (t) ˙ − kT (t) (t) − ˜ T ˜
2 2
By the comparison lemma, the solution of the differential equa-
1 T 1 T
− (e − )
ˆ (e − )
ˆ +  ˜ + ı˜ T 
˜  tion (19) guarantees that V˜ (t) ≤ Vz (t)∀t≥t0 .
2 2
Similarly, we know that V˜ (t0 ) ≤ Vz (t0 ) (by definition). It fol-
By Young’s inequality on the first term, one obtains, lows from (A.2) that
k3 ˜ T 1 ˙T

V̇(t)
˜ ≤ (t) (t)˜ + ˙
 (t)(t) + T  1 T

2 2k3 2 V̇˜ (t) ≤ −k4 V˜ (t) − (e − ) ˆ + z2

ˆ (e − )
2 2 0
1˜ T ˜ − 1 (e − ) T 1 T
−kT (t) (t)(t) ˆ (e − )
ˆ + ˜ 
˜ c12
2 2 2 +Vz + max ((t)) .
2k3
where k3 is some positive constant. It follows that if one lets
kT = k3 + k4 for a positive constant k4 then
Hence, by the comparison lemma, the solution of the initial
k4 ˜ T value problem (18), (17) is such that
V̇(t)
˜ ≤ − (t) ˜ − 1 (e − )
(t)(t)
T
ˆ (e − )

ˆ + T 
2 2 2
(A.2) V˜ (t) ≤ Vz (t) ∀t≥t0 (B.1)
1 ˙T ˙ 1 T
+  (t)(t) + ˜ 
˜ 1˜ T ˜
2k3 2 and since V(t)
˜ = 2
(t) (t), it follows that
Next, we claim the boundedness of the matrix (t) as follows. By Vz (t)
integration, one gets: ˜
(t)2
≤ = z2 (t) ∀t≥t0 . (B.2)
4min ((t))
 t
(t) = e−kT t (0) + e−kT (t−) (w()w()T +
I)d
˜
Hence, if (t) ˜
∈ (t0 ), then (t) ∈ B(0, z (t)), ∀t≥t0 .
 0 (2) If (ti+1 )  (ti ), then
t
≥ e−kT (t−) (w()w()T +
I)d≥e−kT T (˛1 +
)I = 1 I sup ˜ i )≥z (ti )
(t (B.3)

t−T ˜ )∈(t )
(ti i+1

By the boundedness of w(t), one can also write,

Now, one can write:
 t
(t) ≤ (0) + (ˇ2 +
) e−kT (t−) dI ≤ ˛I + (ˇ2 +
)I = 2 I. sup ˜ i )
(t
0 ˜ )∈(t
(ti i+1

As a result, we get that: ≤ sup ˜ i ) + (t

(t ˜ i ) − (t
˜ i+1 )
˜
(ti+1 )∈(ti+1
1 I ≤ (t) ≤ 2 I ˆ i+1 ) − (t
≤ z (ti+1 ) + (t ˆ i ) + (ti+1 ) − (ti )
and ˆ i+1 ) − (t
ˆ i ) + |c1 (ti+1 − ti )|
≤ z (ti+1 ) + (t
−1
2−1 I ≤ (t) ≤ 1−1 I. ≤ z (ti )
486 S. Dahliwal, M. Guay / Journal of Process Control 24 (2014) 479–486
which contradicts (B.3). Hence,  update guarantees
(ti+1 ) ⊆ (ti ). And  is held constant over update intervals
 ∈ (ti , ti+1 ).
References
[1] V. Adetola, M. Guay, Robust adaptive mpc for systems with exogeneous distur-
bances, in: American Control Conference, St. Louis, MO, 2009.
[2] V. Adetola, M. Guay, Robust adaptive mpc for constrained uncertain nonlinear
systems, Int. J. Adapt. Contr. Signal Process. (2010).
[3] G. Bastin, M. Gevers, Stable adaptive observers for nonlinear time varying sys-
tems, IEEE Trans. Automat. Contr. 33 (1988) 650–658.
[4] Y.A. Çengel, Heat and Mass Transfer: A Practical Approach, 3rd ed., McGraw-Hill
Education (India) Pvt Limited, India, 2006.
[5] J. Chen, G. Zhang, Z. Li, On-line parameter estimation for a class of time-varying
continuous systems with bounded disturbances, Int. J. Adapt. Contr. Signal
Process. 25 (2011) 18–32.
[6] Q. Ha, H. Trinh, State and input simultaneous estimation for a class of nonlinear
systems, Automatica 40 (2004) 1779–1785.
[7] G. Kenné, T. Ahmed-Ali, F. Lamnabhi-Lagarrigue, A. Arzandé, Nonlinear systems
time-varying parameter estimation: application to induction motors, Electr.
Power Syst. Res. 78 (2008) 1881–1888.
[8] M. Krstic, I. Kanellakopoulos, P. Kokotovic, Nonlinear and Adaptive Control
Design, Wiley, Toronto, 1995.
[9] K. Lee, J.E. Braun, Model-based demand-limiting control of building thermal
mass, Build. Environ. 43 (2008) 1633–1646.
[10] R. Marino, G.L. Santosuosso, P. Tomei, Robust adaptive observers for nonlinear
systems with bounded disturbances, IEEE Trans. Automat. Contr. 46 (6) (2001)
967–972.
[11] Z. O’Neill, S. Narayanan, R. Brahme, Model-based thermal load estimation in
buildings, in: Proceedings of SimBuild, 4th National Conference of IBPSA-USA,
New York, 2010, August, pp. 474–481.
[12] S. Tatiraju, M. Soroush, Parameter estimator design with application to a chem-
ical reactor, Ind. Eng. Chem. Res. 37 (2) (1998) 455–463.
[13] M. Triki, M. Farza, T. Maatoug, M. M’Saad, Y. Koubaa, B. Dahhou, Unknown
inputs estimation for a class of nonlinear systems, Int. J. Sci. Tech. Automat.
Contr. Comput. Eng. 4 (1) (2010) 1218–1229.
[14] K. Veluvolu, Y. Soh, Multiple sliding model observers and unknown input esti-
mations for Lipschitz systems, Int. J. Adapt. Contr. Signal Process. 21 (2011)
1322–1340.
[15] S. Wang, X. Xu, Simplified building model for transient thermal performance
estimation using GA-based parameter identification, Int. J. Therm. Sci. 45 (2006)
419–432.
[16] X. Xu, S. Wang, A simplified dynamic model for existing buildings using CTF
and thermal network models, Int. J. Therm. Sci. 47 (2008) 1249–1262.
[17] Y. Zhang, V.I. Hanby, Model-based control of renewable energy systems in
buildings, HVAC&R Res. 12 (3a) (2006) 739–760.
[18] Y. Zhu, P. Pagilla, Adaptive estimation of time-varying parameters in linearly
parametrized systems, J. Dyn. Syst. Meas. Contr. 128 (2006) 691–695.