Vous êtes sur la page 1sur 5

2011 21st International Conference on Systems Engineering

Modelling of an air handling unit: a


Hammerstein-bilinear model identication approach
I. Zajic, T. Larkowski, M. Sumislawska, K. J. Burnham D. Hill
Control Theory and Applications Centre Abbott Diabetes Care Ltd
Coventry University Witney, Oxfordshire, UK
Coventry, UK Email: dean.hill@abbott.com
Email: zajici@uni.coventry.ac.uk

AbstractThe paper focuses on modelling and system iden- in the identication of such a real-world system. Therefore,
tication of an air handling unit (AHU), which is part of a simplied rst principles model of an AHU is used as a
larger heating ventilation and air conditioning (HVAC) system surrogate for the actual AHU located in ADC UK. Subse-
dedicated for clean room manufacturing. The aim is to use the
model for subsequent control optimisation and further for control
quently, the underlying dynamic relations between the system
system design. The underlying physical relations of an AHU are inputs and outputs are investigated via computer simulation
investigated through a white-box model. Based on the white-box and the discrete-time structured nonlinear model is identied.
model considerations a discrete-time bilinear black-box model The potential shortfall of using rst principles models is the
is proposed. The AHU cooling/heating capacity is altered by need for knowledge concerning the model coefcients, which
means of water valves. The valve introduces a Hammerstein
nonlinearity on the system input, which needs to be additionally
are not commonly available in practice. However, it can be
identied together with the dynamic bilinear model representing assumed that these are at least partially known. In this regard,
the heating/cooling coil of the AHU. The discrete-time model the parameters of an identied discrete-time model are re-
coefcients are then estimated based on real measurements. estimated based on measurements from the plant. In this way
Keywords-air handling unit; bilinear model; Hammerstein any un-modelled AHU behaviour, which is not accommodated
model; HVAC; system identication in the rst principles model, is then reected within the
reduced order discrete-time structured nonlinear model utilised
I. I NTRODUCTION for control analysis.
Abbott Diabetes Care (ADC) UK, an industrial collaborator II. A IR HANDLING UNIT: FIRST PRINCIPLES MODEL
of the Control Theory and Applications Centre, develops
The investigated AHU unit comprises of heating and cooling
and manufactures the blood glucose and ketones test strips,
coil units. Only the cooling coil unit (CCU) is considered
which are designed to assist persons with diabetes, see [1].
in this paper, and this is depicted in Figure 1. A lumped
One of the manufacturing requirements is that the environ-
parameter approach is adopted. The CCU is assumed to
mental conditions during production are highly stable and
behave as a perfectly mixed vessel, therefore the outow water
within dened limits. To achieve this goal ADC UK utilises
temperature is the same as the mean temperature of the water
heating ventilation and air conditioning (HVAC) systems for
content of the coil. The energy balance on the water side of
all clean room production. Investigations have revealed that
the coil, is given by [3]
the performance of the HVAC systems, which make use
of standard proportional-integral and proportional-integral- dTwo (t)
C =mw (t)cw [Twi (t) Two (t)]
derivative (PI/PID) controllers, can be signicantly improved. dt
Consequently, this paper reports on research which focuses on U A [Two (t) Tao (t)] , (1)
increasing the energy efciency of the HVAC systems through where C [J/K] is the overall thermal capacity (sum of water
the analysis and optimisation of control systems with a view and metal body thermal capacities), Two [K] outow water
to subsequent implementation. temperature, mw [kg/s] water mass-ow rate, cw [J/kgK]
The present paper is concerned with the modelling of an specic water thermal capacity, Twi [K] inow water temper-
air handling unit (AHU), being part of the investigated HVAC ature, U [J/m2 K] overall heat transfer coefcient, A [m2 ]
system based in ADC UK, which is described in more detail effective surface of the coil, Tao [K] discharge (outow) air
in [2]. The aim is to use the identied model of the AHU for temperature and t denotes time in [s].
analysis of the temperature control loop. This facilitates the The energy balance on the air side of the coil can be
use of off-line control optimisation procedures as well as the expressed as
testing of the proposed control strategies in simulation without
dTao (t)
disrupting the manufacturing process. Ca =U A [Two (t) Tao (t)]
The underlying physical relations and deep understanding dt
of the investigated system at hand, play an important part ma (t)ca [Tao (t) Tai (t)] . (2)

978-0-7695-4495-3/11 $26.00 2011 IEEE 59


DOI 10.1109/ICSEng.2011.19
True and discretised output
0
Twi , mw

y [ C]
Two 10
u
20

0 1000 2000 3000 4000 5000


Valve stem position
Tai , ma Tao

u []
1
0.5
0
0 1000 2000 3000 4000 5000
On-coil air temperature

Tai [ C]
60
40
Fig. 1. Schematic diagram of the cooling coil unit.
0 1000 2000 3000 4000 5000
Time index (samples)
Fig. 2. The top plot shows simulated continuous-time system (grey solid line)
Here, Ca [J/K] is the air thermal capacity, ma [kg/s] is and discretised output (solid black line). The middle plot shows the fractional
the processed air mass-ow rate, ca [J/kgK] is the air valve stem position, while the bottom plot presents the on-coil air temperature.
The valve authority is Nv = 1 and the sampling interval is Ts = 1 [s].
specic heat capacity and Tai [K] denotes on coil (inow) air
temperature. In order to further simplify the modelling task, it
is assumed that instantaneous heat exchange between the air
and cooling coil occurs, hence the temperature derivative on obtained as
the left hand side of (2) becomes null.
The valve installed on the plant has linear characteristics y(k) = a1 y(k 1) + b1 u(k 1)
in the range of 0, 30% and equal percentage in the range + 1 y(k 1)u(k 1) + b2 Tai (k) + b3 Tai (k 1)
of 30, 100% of the stem position. However, to simplify + 2 Tai (k 1)u(k 1). (4)
the valve model, only a valve having linear characteristics is
assumed. The valve characteristic is dened as [4] The constant model coefcients are denoted as a1 , b1,2,3
and 1,2 . Note the rst product term between the system
u(t)
u(t) =  . (3) states y u. Such a product is well known to be a so-called
u2 (t) (1 Nv ) + N v bilinear nonlinearity and is commonly encountered in heat
transfer processes, see [5]. Also note, that the valve static
The fractional valve stem position, i.e. the control input, is nonlinearity acts on the selected system input u, which may
denoted u = 0, 1 and the effectively applied fractional be characterised by a Hammerstein nonlinearity.
valve stem position is then u = 0, 1 given by (3). This
Equations (1), (2) and (3), describing the CCU, i.e. a
static nonlinearity is caused by the water pressures that are
surrogate system, were implemented in Simulink software.
developed around the installed valve in the pipe network and
Subsequently, the system has been simulated such that Nv =
is characterised by the so-called valve authority Nv = 0, 1
1, the simulation time is N = 5500 [s] and having the
[]. Subsequently, the chilled water mass-ow rate is given
following model parameters: C = 96324, cw = 3759, Twi =
by mw = umw,max , where mw,max [kg/s] is the maximum
273.15 + 5, U = 187.5, A = 21.2, ma = 2.4, ca = 1005,
water mass-ow rate available when the valve is fully opened,
Tai = 25, 60 + 273.15, u = 0.05, 0.8 and mw,max = 2.4.
i.e. u = 1.
The physical units of the above parameters are dened in
Section II. The simulation results are shown in Figure 2.
A. Model discretisation Since the discrete-time model (4) has been obtained directly
The next step in the development of the reduced order from the rst principles model, the discrete-time coefcients
discrete-time model of the CCU is to, either, simulate the are known and are as follows: a1 = 0.98, b1 = 16.39,
continuous-time system governed by (1), (2) and (3), and 1 = 0.09, b2 = 0.62, b3 = 0.62 and 2 = 0.06 for a
then attempt to obtain such a model by means of system sampling interval of Ts = 1. The discrete-time model (4) has
identication, or to directly discretise the continuous-time been also simulated and is represented in Figure 2 by a black
system. The latter option is preferred here, since the governing dashed line. It can be observed, that the discrete-time model
equations of the CCU are relatively simple. The backward Eu- adequately represents the continuous-time CCU model. The t
ler discretisation method is used to approximate the derivative between these two models is given in terms of coefcient of
term in (1), i.e. dTwo (t)/dt  (Two (k) Two (k 1)) /Ts , determination, i.e.
where Ts [s] denotes the sampling time interval and k is the  
||ysim (k) y(k)||2
discrete-time index. The selected inputs to the system are u RT2 = 100 1 [%], (5)
||y(k) E[y(k)]||2
and Tai , and the system output is selected as temperature
difference between the on-coil and discharge air temperature, where ysim is the simulated system output (by the discrete-
i.e. y = Tao Tai . The discretised model of the CCU can be time model) and E[] denotes the mathematical expectation

60
5
True and simulated output compared, a good model t is achieved RT2 = 99.87 [%].
10 Note, that the saturation type nonlinearity (4) is considered to
y [ C]

15 be difcult to be identied together with the bilinear model


20 at the same time, since such a saturation nonlinearity can be
25
partially explained by the bilinear model as well.
0 1000 2000 3000 4000 5000
Time index (samples) III. A IR HANDLING UNIT: REAL SYSTEM CONSIDERATION
Valve characteristics
1 The identication procedure introduced in Section II-B has
been applied to the actual CCU. Having two measured data
u []

0.5 sets, the rst data set (1 ) is used for system identication
only and the second data set (2 ) is used for model validation
0
0 0.2 0.4 0.6 0.8 1
only. With respect to the CCU slow dynamics the sampling
u [] interval is chosen as Ts = 20 [s]. The identication results
are presented in Figures 4 and 5, for both the estimation and
Fig. 3. The top plot shows simulated continuous-time system (grey solid
line) and identied discrete-time model (black dashed line). The bottom plot validation data sets, respectively. A good model t has been
shows the true (black solid line) and identied (black dashed line) valve static achieved with RT2 = 98.35 [%] for 1 and RT2 = 95.55 [%]
characteristics. The valve authority is Nv = 0.1 and the sampling interval is for 2 , respectively. The identied valve characteristics has a
Ts = 1 [s].
S-shape (as expected). The estimated model coefcients are
as follows: a1 = 0.94(0.2 103 ), b1 = 1.39(0.6 103 ),
1 = 0.038(0.5 103 ), b2 = 0.86(3.5 103 ), b3 =
operator. The model t, in terms of RT2 , is then RT2 = 99.99 0.87(3.5 103 ), 2 = 0.16(0.6 103 ), 1 = 1.95, 2 =
[%] as expected. 4.10 and 3 = 3.15, where the standard deviations are given
in brackets.
B. Identication scheme
In practice, the exact shape of the valve static character- IV. C ONCLUSIONS AND FURTHER WORK
istics (Hammerstein model) as well as the parameters of the A simplied rst principles model of an AHU has been used
dynamic discrete-time model (4) are unknown and need to as a surrogate for the actual AHU located at ADC UK. Based
be identied by means of system identication. It is assumed upon this an appropriate nonlinear discrete-time model of the
that the Hammerstein model can be described by a 3rd order AHU has been obtained. A dedicated optimisation scheme
polynomial, i.e. has been used for the parameter estimation of the proposed
discrete-time model. A good model t between the actual plant
u(k) = 1 u(k) + 2 u2 (k) + 3 u3 (k), (6) and the model has been reported.
where 1,2,3 are coefcients to be estimated. It is known that In further work, the developed model will be used for
when u = 1 then u = 1, so that the appropriate scaling of the control optimisation of the PI/PID based temperature control
Hammerstein model is 1 = 1 2 3 . system at ADC UK. Also, such a structured nonlinear model
An optimization strategy is utilised to obtain both the is highly suitable for model based control strategies, where
polynomial coefcients 1,2,3 and the dynamic model coef- effective scheduling of control gains can be automatically
cients a1 , b1,2,3 and 1,2 . Here, at each step, the polynomial achieved. In this manner, the designed control performance
coefcients are optimised and then the dynamic model is iden- is fullled over a pre-selected range, rather than at the point
tied using the new transformed input u. The dynamic model of tuning, as in the case when dealing with linear control
coefcients are estimated utilising the standard instrumental strategies.
variables method, see [6], and the polynomial coefcients are R EFERENCES
obtained such that the following quadratic cost function is
[1] D. Hill, T. Danne, and K. J. Burnham, Modelling and control optimi-
minimised sation of desiccant rotor dehumidication plant within the heating ven-
N
2 tilation and air conditioning systems of a medical device manufacturer,
V = (y(i) ysim (i)) . (7) in Proc. 20th Int. Conf. Systems Engineering, Coventry, UK, 2009, pp.
i=1 207212.
[2] I. Zajic, T. Larkowski, D. Hill, and K. J. Burnham, Temperature model
The surrogate system of the CCU, see (1), (2) and (3) is of clean room manufacturing area for control analysis, in UKACC Int.
simulated having Nv = 0.1, which corresponds to a rather Control Conf., Coventry, UK, 2010, pp. 12511256.
severe case, since in practice a value of Nv of 0.4 or higher [3] C. P. Underwood, Robust control of HVAC plant I: modelling, Building
services engineering research and technology, vol. 21(1), pp. 5361,
would be expected corresponding to an almost linearly shaped 2000.
characteristic curve. A zero mean white measurement noise is [4] , HVAC control systems: Modelling, analysis and design. London:
added to the simulated output yielding a signal-to-noise ratio E.&F.N. Spon, 1999.
[5] S. Martineau, K. J. Burnham, O. C. L. Haas, G. Andrews, and A. Heeley,
of SN R = 20 [dB]. Both, the measured and the simulated Four-term bilinear PID controller applied to an industrial furnace, IFAC
system outputs are given in Figure 3 together with the true J. Control Engineering Practice, vol. 12, pp. 457464, 2003.
and identied valve characteristics. When the noise free sys- [6] T. C. Hsia, System identication: least-squares methods. Lexington
Books, 1977.
tem output and the simulated discrete-time model output are

61
Measured and simulated output (estimation data set)

5
y [ C]

10

0 200 400 600 800 1000 1200 1400 1600


Valve stem position
0.5
u []

0
0 200 400 600 800 1000 1200 1400 1600
On-coil air temperature
28
Tai [ C]

26
24
22
20
0 200 400 600 800 1000 1200 1400 1600
Time index (samples)

Identied valve characteristics


1
u []

0.5

0
0 0.2 0.4 0.6 0.8 1
u []

Fig. 4. The top plot shows measured system output (grey solid line) and simulated output (black dashed line). The middle two plots show the fractional
valve stem position and the on-coil air temperature. The bottom plot shows the identied valve static characteristics. Estimation data set has been used. The
sampling interval is Ts = 20 [s].

62
Measured and simulated output (validation data set)

4
6
y [ C]

8
10
12
14
0 200 400 600 800
Valve stem position
1000 1200 1400

0.4
u []

0.2

0 200 400 600 800 1000 1200 1400


On-coil air temperature
Tai [ C]

30
25
20
0 200 400 600 800 1000 1200 1400
Time index (samples)

Fig. 5. The top plot shows measured system output (grey solid line) and simulated output (black dashed line). The middle plot shows the fractional valve
stem position, while the bottom plot presents the on-coil air temperature. Validation data set has been used. The sampling interval is Ts = 20 [s].

63

Vous aimerez peut-être aussi