International Journal of Heat and Mass Transfer 44 (2001) 1671±1679

www.elsevier.com/locate/ijhmt

Dynamic prediction and control of heat exchangers using

arti®cial neural networks
Gerardo Dõaz, Mihir Sen *, K.T. Yang, Rodney L. McClain
Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN 46556-5637, USA
Received 2 August 1999; received in revised form 30 June 2000

Abstract
We extend the arti®cial neural network (ANN) technique to the simulation of the time-dependent behavior of a heat
exchanger (HX) and use it to control the temperature of air passing over it. The experiments are carried out in a open
loop test facility. First a methodology is proposed for the training and prediction of the dynamic behavior of thermal
systems with heat exchangers. Then an internal model scheme is developed for the control of the over-tube air tem-
perature with two arti®cial neural networks, one to simulate the heat exchanger and another as controller. An integral
control is implemented in parallel with the ®lter of the neural network controller to eliminate a steady-state o€set. The
results are compared with those of standard PI and PID controller. There is less oscillatory behavior with the neural
network controller, which allows the system to reach steady-state operating conditions in regions where the PI and PID
controllers are not able to perform as well. Ó 2001 Elsevier Science Ltd. All rights reserved.

1. Introduction ¯ow geometries, turbulence in the ¯ow, existence of

hydrodynamic and thermal entrance regions, non-uni-
Most simulations of heat exchangers and other form local heat transfer rates and ¯uid temperatures,
components of thermal systems have concentrated on secondary ¯ows in the tube bends, vortices in the
their steady-state behaviors for heat rate predictions neighborhood of the tube-®n junctions, air-side ¯ow
which are required for system design. The dynamic re- development in ®n passages, heat conduction along tube
sponse of these devices, however, is also very important walls, natural convection within the tubes and between
if these devices are to be controlled in any way. For ®ns, and temperature dependence of ¯uid properties [1].
example, and this is the one that will be taken up here, a Thus, even steady-state predictions are not easily made
hot water heat exchanger may be required to provide from a ®rst principles analysis. Dynamic predictions are,
heated air at a pre-set temperature that does not change of course, harder and it was not until recently that dy-
even though the incoming air or the water may vary in namical models started to appear in the literature [2±4].
either ¯ow rate or temperature. In other control appli- Most of them, in order to make the problem more
cations, the heat rate may be the parameter that is set. tractable, rely on assumptions and simpli®cations that
Heat exchangers (HXs) are extremely complex de- are not totally realistic [5±7]. The results thus are qual-
vices for which the prediction of their operation from itative rather that quantitatively exact. Some of the most
®rst principles is virtually impossible. There are a large common assumptions are: lumped thermal conditions,
number of phenomena associated with ¯ow and heat constant ¯uid properties, constant heat transfer coef-
transfer that are perhaps simple to solve singly, but ®cients, constant ¯ow rates, complete transverse mixing
when combined result in a system that is impossible to in the ¯ow, negligible heat conduction in the wall, neg-
compute. Some of these are: complicated heat and ¯uid ligible heat conduction through the ®ns, and negligible
heat capacity of the wall [4]. The models that include
more physics are usually partial di€erential and their
*
Corresponding author. Tel.: +1-219-631-5975; fax: +1-219- time-dependent solutions are computationally intensive
631-8341. and are not suitable for real-time control purposes.
E-mail address: mihir.sen.1@nd.edu (M. Sen). Another diculty is that the performance of a typical

0017-9310/01/$ - see front matter Ó 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 1 7 - 9 3 1 0 ( 0 0 ) 0 0 2 2 8 - 3
1672 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679

Nomenclature Q_ cor heat transfer rate predicted by power-law

correlations [W]
c speci®c heat [J/kg K] t time [s]
C ANN controller Dt time step [s]
f …t† time-dependent forcing function T temperature [°C]
F robustness ®lter x…t† variable in di€erential equation
I integral control x0 initial condition
m_ mass ¯ow rate [kg/s]
M ANN model of the plant Subscripts and superscripts
n order of the ANN approximation a air side
P real process or plant in inlet
Q_ heat transfer rate between ¯uids [W] out outlet
Q_ ANN heat transfer rate predicted by ANN [W] w water side

HX slowly changes over time due to such factors as compared with those of standard PI and PID tech-
fouling that changes the heat transfer characteristics of niques.
surfaces.
Arti®cial neural networks (ANNs) have been used in
recent years to avoid the problems associated with de- 2. Background
terministic approaches, and have been shown to ap-
proximate nonlinear functions up to any desired level of 2.1. Experimental setup
accuracy [8]. They are also less sensitive to noise and
incomplete information than other approaches such as The experimental setup consists of a variable speed
empirical models and correlations. In recent years, the wind tunnel facility located in the Hydronics Laboratory
technique has been applied to many thermal problems at the University of Notre Dame (details are in [22]). A
[1], among them the prediction of the steady state [9] and single-row water-to-air ®n-tube heat exchanger is used
the dynamic behavior of heat exchangers [10±12]. The to obtain static and dynamic measurements. Fig. 1
advantage of using ANNs to simulate thermal processes shows a picture of the experimental facility. There is a
is that, after they are trained, they represent a quick and single water-side circuit which goes back and forth
reliable way of predicting their performance. They can across the face of the heat exchanger. This is a nominal
also be continuously updated. Thus, if we apply this 18 in:  24 in: type T water coil heat exchanger manu-
technique to the problem of simulation and control of factured by Trane. Type-T isolated thermocouples are
HXs, then we obtain an accurate prediction with a short used to measure the inlet and outlet temperature of the
computational time for the simulation which can be used air and water side. The motion of the air in the tunnel is
in an ecient real-time control scheme. due to a fan that is controlled by a variable speed drive
There are several schemes that have been proposed that can be operated manually or automatically from a
for the neural control of nonlinear systems [13±16]. One personal computer. The air speed is measured using a
of these is a method called internal model control (IMC) Pitot tube, located upstream of the heat exchanger, that
[17±19]. This technique has been used for a variety of is connected to a di€erential pressure transducer. The
problems in di€erent areas due to its excellent charac- ®lter and data acquisition board used can obtain
teristics of robustness and stability [20]. The IMC tech- measurements of up to 16 di€erent channels, simulta-
nique using ANNs consists of training a network to neously. The data acquisition board receives informa-
learn the dynamics of a process, after which another tion about inlet and outlet temperatures of both the air
ANN is trained to learn the inverse dynamics so that it and water side, the mass ¯ow rate of water, the air speed
can be used as a nonlinear controller [17,21]. and the time at which the measurements were taken. The
In this work, we use the combined advantages of inlet water temperature is varied by using a heater with a
ANNs and IMC to generate an ecient real-time con- PID-controlled electrical resistance. The water ¯ow rate
trol scheme for a HX installed in a test facility. The HX is modi®ed by an electronic valve so that the percentage
transfers heat from water to air, and the objective is to of opening can be controlled as desired from the
control a single output variable, the outlet air tempera- personal computer. LabVIEW is used to acquire and
ture, by changing a single input variable, the air speed. send data to the experimental system and an interface
The system consists of the HX and the entire water- and was built in C language to simulate the neural networks
air-¯ow subsystems. The results of the neural control are and perform the desired control action. Time-dependent
 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679 1673

Fig. 1. Experimental setup: (a) wind tunnel; (b) heat exchanger.

information regarding the air and water mass ¯ow rates,

m_ a and m_ w , respectively, the air and water inlet tem-
peratures, Tina and Tinw , respectively, and the air and water
a w
outlet temperatures, Tout and Tout , respectively, are
stored.

2.2. Arti®cial neural networks: steady-state simulations

The description, training and operation of ANNs are

available in many recent texts [23]. Input and output
data have to be supplied to the network so that it can be
trained by using an algorithm that can adjust its internal
weights and biases. It can be shown that multilayer
networks are universal approximators capable of ap-
proximating any measurable function to any desired
degree of accuracy [24]. Here, the variables to and from
the ANN are normalized to be within a ‰0:15; 0:85Š
range. Although there are other methods, the back-
propagation algorithm [25] is one of the most common
learning methods used to train ANNs. We use it because
of its well-known adaptation and generalization char-
acteristics even though other algorithms can lead to
more accurate models.
In the steady state, the ANN predicts the heat rate Fig. 2. Steady-state predictions: (a) inputs and output;
under given conditions. The steady-state heat transfer (b) comparison with measurements;  correlations, + ANN
_ was determined from the measured temperatures
rate, Q, (4±5±5±1); dotted lines are 10% deviations.
by
ÿ
ÿ a

Q_ ˆ m_ w cw Tinw ÿ Tout
w
ˆ m_ a ca Tout ÿ Tina ;
be pointed out that some of the measurements had
where ca and cw are the air and water speci®c heats, about 10% error; the predictions merely re¯ect this in-
respectively. Since the air and water sides give slightly herent inaccuracy in the data since they cannot of course
di€erent Q_ (within 10%) an average value is used. be better than the measurements. More about steady-
An ANN, shown schematically in Fig. 2(a), was state predictions for thermal systems using ANNs are in
trained with m_ a , m_ w , Tina and Tinw as inputs and Q_ as Dõaz et al. [9].
output. Fig. 2(b) shows a comparison between the heat
transfer rates obtained with the ANN, Q_ ANN , and those 2.3. Arti®cial neural networks: dynamic simulations
measured, Q._ Predictions obtained with a heat transfer
correlation, Q_ cor , that was found earlier for the same HX For control purposes, it is not enough to have steady-
[22] are also included. It is observed that the ANN state predictions, since the process is really time-depen-
prediction is superior to that of the correlation. It must dent. It is thus of interest to extend the capabilities of the
1674 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679

One important aspect that has to be considered when

modeling the dynamics of a system is its order. We have
to provide values of the relevant variables at previous
instants in time. This is because the ANN is simulating
an di€erential equation of unknown order. The higher
Fig. 3. Training method for dynamic problems. the order, the larger the number of previous instants for
which information must be provided as inputs. Enough
ANN technique to dynamic simulations. This can be past information at previous instants in time that is
done training the ANN providing it with the informa- appropriate for the actual order of the system must be
tion of the dynamic behavior as shown in Fig. 3. In this provided, as shown in Fig. 6(a), where n is the order of
method, no explicit information about time is provided the system. This is experimentally veri®ed in Fig. 6(b)
to the network; the variables involved in the problem are which shows time-dependent predictions of Tout a
for in-
presented at time t ÿ Dt as an input to the network and creasing assumed order of the system. In each exper-
the output corresponds to the variables at time t. iment, the air speed was decreased in ®ve small steps and
then similarly increased in small steps. Both the exper-
imental measurements and the ANN predictions are
3. Dynamic predictions shown; the temperature is in normalized units and the
time is in terms of the sample number s. The prediction
The dynamic ANN developed in the previous section is seen to improve as we go from n ˆ 1 to n ˆ 2, but
is now used to predict the behavior of the HX. The in- there is little observable di€erence between n ˆ 2 and
puts to the network are m_ a , m_ w , Tina and Tinw and the n ˆ 3. These measurements indicate that the order of the
output is Q. _ Fig. 4 shows the comparison between the system, if one has to choose an integer, is probably two
ANN prediction of the water and air outlet tempera- and it is not necessary to assume a higher value.
tures, using the training method and ®ve di€erent
training curves, and the actual measurements for a step
change in the water ¯ow rate. The ®rst dip in the shape 4. Temperature control
w
of the Tout curve is due to the presence of the mass of
water that was located within the HX when the water We now proceed to design an algorithm to control
w
¯ow was shut down. The second dip in Tout is due to the a
Tout . This can be achieved by controlling one or more of
same mass of water but after going through one lap in the variables m_ w , m_ a , Tinw , and Tina . In this work, we
the water circuit. As its initial temperature was close to con®ne ourselves to a single-input±single-output system,
Tina , the heater is not able to raise its temperature to the and for ease of experimentation we have chosen m_ a as
desired value after only one lap of the water circuit. the control variable while keeping the others ®xed. The
However, good predictions are achieved for both the relationship between m_ a and Tout a
is nonlinear and com-
water and air side of the HX. Fig. 5, on the other hand, plicated. Strictly speaking, it is a solution of a partial
shows the comparison between the ANN prediction, di€erential equation in space and time. Some idea of the
using ®ve training curves, and the measurements taken nonlinearity of this equation can be obtained by looking
for a cooling process in which the heater is shut o€. at the steady state, for which a heat balance gives

Fig. 4. Prediction of transient heating: ± experiment; - - ANN (4±5±5±2).

 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679 1675

The slope of the curve changes considerably with m_ a

indicating that the sensitivity of the system depends on
the operating point.
There are other diculties that increase the com-
plexity of the nonlinear control problem. First, the sys-
tem that we are controlling includes not only the HX but
also its associated hardware, i.e. fan, pump, PID-con-
trolled heater and measuring instruments such as a
water ¯ow meter and a pressure transducer. Second,
there is a delay between what happens at the HX and the
a
measurements of Tout since it takes a while for the air to
¯ow from the HX to the point of measurement. As the
air speed slows down this delay is longer and it is harder
to control the air temperature. Finally, there is a gradual
change in the HX characteristics due to fouling e€ects.
ANNs are very well suited for these tasks because they
Fig. 5. Prediction for transient cooling; ANN (3±5±5±2).
can be taught to learn the response of the system.

a 1 _
Tout ˆ Q ‡ Tina ; 4.1. ANNs with IMC
m_ a ca
where Q_ is the heat rate which also depends on m_ a . Fig. 7 IMC consists of having a model of a plant M in
shows the measured function Tout a
…m_ a † for di€erent m_ w . parallel with the real system P, as shown in Fig. 8. The

a
Fig. 6. (a) Training a system of order n. (b) Response of HX treated as a system of di€erent orders. Tout is normalized and s is the
sample number.
1676 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679

were obtained by taking measurements of the system

subject to small increments in the setpoint temperature.
The controller C is obtained by using a synthetic signal
which is the desired value of the air speed. This signal
a
is supplied to M to give a certain value of Tout which is
then supplied as the input to the controller. The
training algorithm adjusts the weights of C to reduce
the error between the synthetic signal and the con-
troller output.
Since the ANNs only provide an approximation to
the behavior of the actual plant, we used a one param-
eter ®lter F, following the suggestion of Nahas et al. [18],
preceding the controller in the forward path to account
for plant-model mismatch. An integral control path I
was also added in parallel with F to help obtain an
o€set-free controlling action (an initial control scheme
w
Fig. 7. Nonlinear relation between m_ a and Tout for di€erent without it failed to provide o€set-free control). There are
mass ¯ow rates of water. ÿ  ÿ 260 kg/s, ÿMÿ 200 kg/s, ÿ  ÿ two constants that have to be chosen by trial and error,
65 kg/s.
the ®rst for the integral controller and the other for the
®lter.
As a large percentage of the controllers that are
currently being used correspond to proportional-integral
and proportional-integral-derivative schemes, standard
PI and PID controllers were used to compare the per-
formance with the ANN controller. This was through a
general purpose LabVIEW subroutine implementing a
PID controller based in the relations developed by
Shinskey [26], the optimal values of the PID constants
being obtained using the tuning method also explained
there. The derivative action can be shut down by means
of setting the corresponding constant equal to zero. In
this way, the same algorithm is used for both the PI and
PID controlling schemes. The two di€erent tests that
were conducted are described below.

4.2. Comparison with PID: step change in setpoint

Fig. 8. General IMC structure plus integral control
The ®rst test was designed to observe the per-
formance of the controller subject to a step change in
a
di€erence between the outputs of P and M is used as the the value of the setpoint temperature Tout . The system
feedback for a controller C that is located in the forward was taken up to a point in which the outlet air tem-
path of the control scheme. The training procedure of perature was near 32°C. The controller was turned on
such a control system using ANNs has two steps: and we waited for 40 s until the temperature remained
· We ®rst train an ANN to learn the dynamics of the within a band of 0:1°C. The setpoint was then in-
process by providing known input and output data creased to 36°C. Both controllers performed well and
sets. This is M. behaved in a similar way when controlling the system
· Then, another ANN is trained to learn the inverse at large values of air speed. However, on approaching
dynamics of the process and to function as a nonlin- the lower end of air speeds, the system became very
ear controller C. It is trained to invert the model M hard to control for two reasons. One is the e€ect of the
instead of trying to learn the inverse dynamics of delay involved, and the other is the high sensitivity of
a
the actual process. By training in this way, we make Tout to m_ a at low air speeds. This test brings the system
sure that we invert the steady-state gain of the model from a very easy-to-control point at 32°C to a hard-to-
so that the o€set can be eliminated. control state at 36°C. The results are shown in Fig. 9.
For our experiments, we trained the plant model M It is seen that, although the ANN controller has a
a
with information related to Tout and m_ a . These data slightly larger overshoot, it presents less oscillations
 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679 1677

Fig. 9. Change in the setpoint temperature: ± ANN; - - PID; -.- PI.

and it is able to bring the system to a stable condition. shut down completely one of the valves on the water side
On the other hand, both PI and PID controllers os- for a short time. Once again, we test the controllers at a
a
cillate signi®cantly more and are not able to bring the state that is hard to control, i.e. with Tout ˆ 36°C and a
a
system to a steady state, but keep Tout within 36°C by low air speed. The PI controller showed the worst
constantly adjusting the air speed. Thus, the ANN performance and is left out of the comparison shown in
controller uses less energy and is more stable by Fig. 10. Fig. 10(a) shows the change in the water ¯ow
keeping the system steady instead of generating an rate which is the disturbance itself; the water ¯ow is shut
oscillatory controlling action. down between t ˆ 40 and 70 s. After the disturbance
pulse, the controller brings the system back to steady
a
state. Figs. 10(b) and (c) show the change in Tout and m_ a ,
4.3. Comparison with PID: disturbance rejection respectively. Once again it is seen that the PID is not
able to bring the system to a steady-state condition while
We now analyze the disturbance rejection capabilities the oscillations of the ANN controller are quickly
of the control system. In this test, a disturbance is ap- damped out. It is seen in Fig. 10(c) that the PID con-
plied to the plant in the form of a pulse in the following troller, in trying to control the temperature, generates an
way. Once the system is at steady-state operation, we oscillatory air speed.
1678 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679

Fig. 10. Disturbance rejection: ± ANN; - - PID.

5. Conclusions the Hydronics Laboratory. G.D. also thanks the Or-

ganization of American States for a PRA Fellowship.
Previous work has demonstrated the usefulness of the
arti®cial neural network technique for the prediction of
References
the steady-state behavior of heat exchangers. In the
present, this technique is extended to the prediction of [1] M. Sen, K.T. Yang, Applications of arti®cial neural
the dynamic behavior of a thermal system which consists networks and genetic algorithms in thermal engineering,
of a heat exchanger working between a closed hot water in: F. Kreith (Ed.), CRC Handbook of Thermal Engineer-
and an open air loop. The dynamic network is then used, ing, 2000, pp. 620±661 (Section 4.24).
in conjunction with internal model control, to control [2] M. Spiga, G. Spiga, Step response of the cross¯ow heat
the temperature of the air coming out of the heat ex- exchanger with ®nite wall capacitance, Int. J. Heat Mass
changer. The tests showed that the present technique Transfer 35 (2) (1992) 559±565.
performed better than conventional PI and PID control [3] S. Kabelac, The transient response of ®nned cross¯ow heat
in certain cases. exchangers, Int. J. Heat Mass Transfer 32 (6) (1989) 1183±
1189.
Neural networks are powerful tools for thermal
[4] W. Roetzel, Y. Xuan, Dynamic Behaviour of Heat
control. They can be trained to simulate the behavior of Exchangers, WIT Press, Boston, 1999.
a dynamical system and they are adaptive. In the present [5] H. Thal-Larsen, Dynamics of heat exchangers and their
work, the network was trained o€-line, but in the future models, ASME J. Basic Eng. (1960) 489±504.
on-line training will be incorporated to enable continu- [6] J.R. Gartner, H.L. Harrison, Dynamic characteristics of
ous learning and adaptation to changing conditions. water-to-air cross¯ow heat exchangers, Trans. ASHRAE
71 (1965) 212±223.
[7] H. Yamashita, R. Izumi, S. Yamaguchi, Analysis of the
dynamic characteristics of cross-¯ow heat exchangers with
Acknowledgements
both ¯uids unmixed, Bull. JSME 21 (153) (1978) 479±485.
[8] R. Hecht-Nielsen, Kolmogorov's mapping neural network
We gratefully acknowledge the support of Mr. D.K. existence theorem, Proceedings of the First International
Dorini of BRDG-TNDR for this and related projects in Conference on Neural Networks, IEEE 3 (1987) 11±13.
 G. Dõaz et al. / International Journal of Heat and Mass Transfer 44 (2001) 1671±1679 1679

[9] G. Dõaz, M. Sen, K.T. Yang, R.L. McClain, Simulation of [17] K.J. Hunt, D. Sbarbaro, Neural networks for nonlinear
heat exchanger performance by arti®cial neural networks, model control, IEE Proceedings-D 138 (5) (1991) 431±438.
Int. J. HVAC and R Res. 5 (3) (1999) 195±208. [18] E.P. Nahas, M.A. Henson, D.E. Seborg, Nonlinear in-
[10] S. Bittanti, L. Piroddi, Nonlinear identi®cation and control ternal model control strategy for neural network models,
of a heat exchanger: a neural network approach, J. Comput. Chem. Eng. 16 (12) (1992) 1039±1057.
Franklin Inst. 334B (1) (1997) 135±153. [19] N. Abe, K. Seki, M. Kanoh, Internal model control for
[11] M. Ayoubi, Dynamic multi-layer perception networks: single tubular heat exchanger system, in: Proceedings of
application to the nonlinear identi®cation and predictive IECON, International Conference on Industrial Electron-
control of a heat exchanger, in: Applications of Neural ics, Control and Instrumentation vol. 2, 1994, pp. 1165±
Adaptive Control Technology, World Scienti®c series in 1170.
Robotics and Intelligent Systems, vol. 17, 1997, pp. 205± [20] M. Morari, E. Za®riou, Robust Process Control, Prentice-
230. Hall, Englewood Cli€s, NJ, 1989.
[12] G. Dõaz, M. Sen, K.T. Yang, R.L. McClain, Use of [21] N. Bath, T.J. Macavoy, Use of neural nets for dynamical
arti®cial neural networks for temperature control, in: modeling and control of chemical process, Comput. Chem.
Proceedings of the Fourth ISHMT/ASME Heat and Mass Eng. 14 (1990) 573±583.
Transfer Conference, Pune, India, 5±7 January 2000. [22] X. Zhao, Performance of a single-row heat exchanger at
[13] L.B. Gutierrez, F.L. Lewis, J.A. Lowe, Implementation of low in-tube rates, M.S. Thesis, Department of Aerospace
a neural network tracking controller for a single ¯exible and Mechanical Engineering, University of Notre Dame,
link: comparison with PD and PID controllers, IEEE Notre Dame, IN, 1995.
Trans. Ind. Electron. 45 (2) (1998) 307±318. [23] S. Haykin, Neural Networks, a Comprehensive Founda-
[14] K.S. Narendra, K. Parthasarathy, Identi®cation and con- tion, Macmillan, New York, 1994.
trol of dynamical systems using neural networks, IEEE [24] K. Hornik, Multilayer feedforward networks are universal
Trans. Neural Networks 1 (1) (1990) 4±27. approximators, Neural Networks 2 (1989) 359±366.
[15] A. Blazina, N. Bolf, Neural network-based feedforward [25] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning
control of two-stage heat exchange process, Proc. IEEE internal representations by error propagation, in: Parallel
Int. Conf. Syst. Man, Cybern. 1 (1997) 25±29. Distributed Processing: Explorations in the Microstruc-
[16] E. Jeanette, K. Assawamartbunlue, P. Curtiss, J.F. tures of Cognition, vol. 1, MIT Press, Cambridge, MA,
Kreider, Experimental results of a predictive neural 1986, pp. 318±362.
network HVAC controller, ASHRAE Trans. 104 (2) [26] F.G. Shinskey, Process-Control Systems, McGraw-Hill,
(1998) 192±197. New York, 1988.