Automatica, Vol. 25, No. 1, pp. 33-40, 1989 0005-1098189 $3.00 + 0.

00
Printed in Great Britain. PergamonPresspk
1989 International Federation of Automatic Control

Comparison of an Auto-tuned PID Regulator

and an Adaptive Predictive Control System
on an Industrial Bleach Plant
G U Y A. D U M O N T , t J U A N M. MARTIN-SANCHEZ$ and
CHRISTOS C. ZERVOS§

A PID-based scheme with linearizing block and gain scheduling, and an

adaptive-predictive controller are compared on the same industrial bleach
plant. Both behave well, and results suggest a hybrid of the two schemes,
for example, on-line improvement of the linearizing block and of the gain
scheduling.
Key Words--Adaptive control; identification; orthonormal functions; PID control; predictive control;
process control; pulp industry; self-adjusting systems.

Abstraet--A PID-based scheme and an adaptive-predictive process, gain scheduling and linearizing blocks to
controller are compared on the same industrial bleach plant. increase the robustness. On the other hand,
The PID scheme is auto-tuned and uses a linearizing block
and gain scheduling to compensate for sensor nonlinearities adaptive control schemes tend to look at the
and production rate changes, while APCS treats those as plant as a black-box and the estimator gains the
unknown perturbations. Both schemes behave well. It shows knowledge on-line. How do these two different
that, when nonlinearities are known and changes in process
dynamics are predictable, gain scheduling combined with a approaches compare? The opportunity was
well-tuned PID regulator behave remarkably well. At the seized to compare a well-tuned PID-based
same time, it shows that adaptive control can handle the scheme with a black-box adaptive controller on
same situation with less tailoring.
an industrial bleach plant. Of course it is
recognized that these results are very much case
dependent, but some of the lessons gained from
INTRODUCTION this work can easily extend to other industrial
VERY OFTEN, adaptive controllers are compared processes. The first section of this paper
with PID controllers, with the result that describes the dynamics of the chlorination stage
superior performance is claimed. However, of a bleach plant, as well as the industrial control
often these comparisons are unfair in the sense scheme in place in that particular plant. Then we
that the PID scheme used in the comparison describe briefly the auto-tuner used on the
might not correspond to a true industrial design. industrial scheme and the adaptive-predictive
For example, industrial control schemes fre- control scheme (APCS) used for comparison.
quently use first principle knowledge of the Finally, we present and discuss the experimental
results.
* Received 21 January 1986; revised 28 February 1987;
revised 15 November 1987; revised 24 April 1988; received in BLEACH PLANT DYNAMICS AND CONTROL
final form 4 June 1988. The original version of this paper was
presented at the IFAC Workshop on Adaptive Control of
Chemical Processes which was held in Frankfurt, F.R.G. Process description
during October 1985. The Published Proceedings of this The objective of bleaching is to increase the
IFAC meeting may be ordered from: Pergamon Press plc, brightness of pulp to make it suitable for use in
Headington Hill Hall, Oxford OX3 0BW, England. This
paper was recommended for publication in revised form by printing or tissue papers by removing the
Associate Editor Y. Arkun under the direction of Editor H. colouring components, primarily the residual
Austin Spang III.
t Pulp and Paper Research Institute of Canada and
lignin from the cooking stage. This is generally
Department of Electrical Engineering, University of British achieved in a multistage plant, using expensive
Columbia, Vancouver, B.C., Canada. chemicals, such as chlorine and chlorine dioxide.
$ Escuela de Minas, Universidad Politrcnica de Madrid
and SCAP Europa, S.A., Madrid, Spain.
Any bleach plant control system must be able to
§Department of Electrical Engineering, University of produce high brightness pulp consistently,
British Columbia, Vancouver, B.C., Canada. without degrading the wood fibres and at
A U T O 25:1-C
33
34 G . A . DUMONTet al.

ORP2 temperature in Kelvin:

I ]- ©
dILl] -
0.535 exp ( - 250/r)[c121[L,l
] dt
ATION / WASHER
PULP ~ TOWER ,[ ORP1 d[L2] _
0.0107 exp (-250/T)[CI2][L21.
dt

FIG. 1. Simplified diagram of the chlorination stage. If [Lo] denotes the initial pulp lignin concentra-
tion, then initially [L1]=0.5[Lo] and [Lz]=
minimal chemical consumption. The first and 0.3[Lo]. Note that 20% of the lignin in the pulp
most critical stage in a typical bleach plant is the does not react with chlorine. Thus, at any time
chlorination stage, depicted schematically in Fig. the lignin remaining in the pulp is 0.2[Lo] +
1. The stock is pumped at low, controlled [L1] + [Lz]. The chlorine consumption rate can
consistency (fibre concentration) to a mixer be assumed at 1.37 g [C12] per g lignin removed,
where chlorine gas is added. The reaction is then thus, if c denotes the consistency of the incoming
completed in the chlorination tower. In modern pulp
bleach plants, a small amount of chlorine dioxide
proportional to the chlorine flow is also added. d[Cl2]
dt = 1"37(~t~] + ~t2])c"
Chlorine reacts with lignin by substitution and by
oxidation as follows. Figure 2 shows the evolution with time of the
pulp lignin content and of the chlorine
Substitution concentration under typical conditions. Beyond
conventional instrumentation, sensors for a
C12 + lignin---* lignin-Cl + HCI. bleach plant control fall in two classes: those to
measure the chlorine residual and those to
Oxidation measure the pulp colour or brightness. In the
first class, the ORP sensor measures the
C12 + lignin--* oxidized lignin + 2HCI. Oxidation-Reduction Potential between a calo-
mel and a platinum electrode suspended in the
The substitution reaction is quite rapid, pulp. This voltage V is a function of the chlorine
requiring only 5 min for completion, whereas the concentration [Clz] as per the following equation:
oxidation is rather slow and takes place in the
chlorination tower. As chlorine oxidizes cellu- V = 1172 + 29.5 log [C12]
lose as well as lignin, an excess of it will seriously
with V in mV and [C12] in m o l l - ' . The
degrade the pulp. It is important to maintain the
more recent polarographic method involves a
amount of chlorine consumed close to that
third platinum electrode, used to generate a
required for chlorination and oxidation of lignin
current to maintain a constant voltage between
so as to minimize cellulose degradation. The
the other two electrodes. The significant
reactions are affected by temperatures, con-
advantage of this approach is the fact that this
centrations, pH and time. The main pertur- current is directly proportional to the chlorine
bations are the variations in the incoming lignin concentration. The second class of sensors
content of the pulp and the black liquor consists of optical sensors to measure the colour
carryover. After the cooking stage, the pulp is or brightness of the pulp, the idea being that the
washed in order to separate the fibres from the delignification of the pulp produces a change in
reacted cooking chemicals (black liquor), which colour. All optical sensors are also affected by
contain a significant amount of dissolved lignin. other factors, and thus must be used with great
Thus, any black liquor carried over with the pulp
care.
to the bleach plant will consume chlorine and There are primarily two points in a chlori-
affect the operation of the chlorination stage. nation stage where those sensors can be installed
Process dynamics to some benefits. Sensors after the tower give
A kinetic model based on the existence of the information on how well the chlorination stage is
two above reactions has been proposed (Ackert functioning, but due to the long delay and poor
et al., 1975). It assumes that two first-order mixing characteristics of the tower, it is not
reactions are taking place in parallel and is adequate for feedback control. This is why on
described by the following equations where [L] modern units, sensors are also located between
denotes concentration of lignin in percent, [CIz] the mixer and the tower or at the bottom of the
denotes concentration of chlorine in g l -~, t tower. Depending on where the sensor is
denotes the time in seconds and T denotes the installed, the time delay from the chlorine
 Bleach plant PID and APCS control 35

0
0 200 400 600 600 1000 1200 1400 1600 1800 2000

Time i n seconds

0.3 ' I

.= ........................................................... .............................................................................................

0.1 [ .......................................................................... ~.................................................... ~....................... ........... , ...............................

0 i A
200 400 600 800 1000 1200 1400 1600 1800 2000

Time i n seconds

FIG. 2. R e a c t i o n k i n e t i c s .

application point to the sensor is between 30 s ORP1 is given by the ORP2 control loop. The
and 2 min. dead times from the E l 2 application point to the
ORP1 and the ORP2 are 2 and 30min,
Chlorination stage control respectively. The sampling periods are 30 s and
Modern chlorination stage control systems 12min for the ORP1 and ORP2 loop,
combine the use of a polarographic residual respectively. Note the presence of the linearizing
sensor and of an optical brightness sensor both block to compensate for the nonlinear charac-
located on a sample line between the mixer and teristic of the ORP sensor. In addition, as the
the tower. Those systems generally provide output of the ORP1 controller is the
superior performance. However, because of the chlorine/fibre ratio, it is then multiplied by the
higher cost of the sensors involved, many pulp fibre flow, estimated from flow and concen-
mills still simply use ORP sensors located both at tration measurements. This provides gain sched-
the bottom of and after the tower. This is the uling, since in essence it multiplies the regulator
case of the mill being studied, as shown in Fig. 1. gain by the fibre feedrate. Despite the presence
The idea is that the pre-tower ORP sensor, of time delays, as in most plants the regulators
ORP1 will indicate any change in the lignin used are all PID.
content of the incoming pulp and will help
achieve the target residual after the tower, as THE AUTOMATIC PID TUNER
indicated by ORP2. In such a case, the standard The bleach plant is controlled by means of the
control scheme is depicted in block-diagram Process Control and Monitoring System (PMCS)
form on Fig. 3. It consists of three cascaded software package implemented on a PDPll-44
loops, the innermost one being the chlorine flow computer running under the RSX-11M operating
control loop whose setpoint is given by the system. The basic control block provided with
ORP1 control loop. Further, the setpoint for the this system is a digital PID, derived from the

CI 2 FLOW RESIDUAL CI 2
I ~ ~ ORP 2

CI2 I FLOW SETPOINT ~ [

r ................................................... q
',~="1 PRODUCTION --1-]
I[ " I'~"f-E
i
i [i
i -i

1 I BLOCK CO.T.OLL .
......................................................
i
jji
j
ORP1 SETPOINT ORP 2 SETPOINT
FIG. 3. T h e r e s i d u a l c o n t r o l l o o p .
36 G . A . DUMONTet al.

idealized continuous PID using the rectangular the error response is described as
integration method, and with a minimum
sampling interval of 30 s. The tuner is imple- E(s) = ~ aiti(s ).
i=0
mented on the system so that it can access any
loop. It is an off-line, closed-loop tuning method A simple least-squares estimation scheme is
whose main characteristic is that it describes the used to identify the coefficients ai. The free
closed-loop plant dynamics as a series of parameter p is chosen as indicated above. It is
Laguerre functions. For more details, see shown in Zervos et al. (1985) that the
Dumont et al. (1985) and Zervos et al. (1985). estimates are unbiased.
The Laguerre functions form a complete, Determination of the tuning constants. The
orthogonal set in L2 and are described by the criterion used to tune the controller will be the
following transfer functions where s is the Integral-Square-Error (ISE). Due to orthog-
Laplace transform operator and p > 0 is a free onality of the Laguerre functions, straight-
parameter used for time scaling (Lee, 1960): forward application of Parseval's theorem
shows that the ISE is simply the sum of the
squares of the Laguerre gains. As shown in
_ (s - p ) ' Zervos et al. (1985), a prediction of the
ti(s)- t P) ( ; ~ _ p - ~ l i=0 ..... N- 1.
changes in Laguerre gains due to changes in
tuning constants is achieved by means of a
Because of the completeness of this set, any polynomial long division. The Hooke and
L2[O, oo) stable linear system can be synthesized Jeeves pattern search method is then used to
as find the optimum tuning constants. The user
has the option of specifying a desired
closed-loop response.
H(s) = ~, aiLi(s).
i=0
ADAPTIVE-PREDICTIVE CONTROL SYSTEM
The advantage of this approach is that no AS witnessed by the recent literature on
assumption is made about the structure of the process control, predictive control methods have
system. In particular, the dead time is become rather popular in the process control
approximated by the Laguerre functions in a community, see for instance Cutler and Rama-
manner similar to the use of Pad6 approximants. ker (1980). Although the idea of predictive
The choice of the free parameter p is not crucial. control goes back about 20y, it is in
For delay free systems, it should be chosen such Martin-Sanchez (1976a, b) that its use in the
that the first Laguerre filter has about the same context of adaptive control was first proposed.
crossover frequency as the plant. For plants with The idea was further exploited by De Keyser
long dead times, the similarity to Pad6 and Van Cauwenberghe (1985) for extended
approximants can be used to choose p. prediction self-adaptive control, by Ydstie
The tuning method consists of three steps, an (1982,1984) for extended horizon adaptive
identification experiment, the computation of control and by .~str6m (1986) for long prediction
the coefficients of the Laguerre series, and the horizon adaptive control. The underlying prin-
search for better tuning constants, summarized ciple behind all those methods is that the
as follows. controller aims at making the output equal to a
desired value at some instant several sampling
• Identification experiment. The identification intervals beyond the time delay. Because of this,
experiment is performed in a closed loop by at every sampling interval, a sequence of future
perturbing the setpoint using a Pseudo- control signals has to be computed, either
Random-Binary-Signal (PRBS). The only attempting at the same time to minimize the
knowledge required about the plant are control energy, or assuming constant input
estimates of the process noise level and during the prediction horizon. The adaptive-
settling time in order to choose the amplitude predictive control system (APCS) as defined by
and sampling interval of the PRBS. This Martin-Sanchez (1976a, b) fits that latter cate-
information is easily obtained by examination gory. Although a sequence of control signals is
of the closed-loop step response. computed at each step, only the first is applied,
• Computation of the Laguerre gains. As the and the process is repeated at the next sampling
closed-loop system is assumed to be stable, the interval. APCS uses a predictor of the form
error response to a step change lies in L2. For
practical reasons, the Laguerre series is y(k + i l k ) = A(q-1)y(k +1 I k)
truncated, and thus the Laplace transform of +B(q-1)u(k+jlk) j=l,...,d (1)
 Bleach plant PID and APCS control 37

1074

1073

1072

1o71

~ 1070

1069

1068

1067 i L i i L ~ A L

0 2 4 6 6 10 12 14 16 18 20

T i m e in m i n u t e s

FIG. 4. Step response with Ziegler-Nichols tuned PID controller.

where prediction error is used to stop parameter

A ( q -1) = alq -1 + . . . + anq -n adaptation. Also, in order to avoid biased
B(q-1) = blq-1 +... + b,,q-'n estimation and control, the d-incremental
approach akin to that of Clarke et al. (1983) is
y(k+l-ilk)=y(k+l-i) i=l,n used (Martin-Sanchez and Shah, 1984). In
applying APCS to the bleach plant, the
u(k+l-ilk)=u(k+l-i) i=l,m
linearizing block and the gain scheduling scheme
and d is the prediction horizon. Assuming were purposely discarded, thus the process
constant future input during the prediction output y(t) was the raw ORP signal and the
horizon, i.e. u ( k +j I k)=u(k I k) for ] = process input was the chlorine flow setpoint.
1. . . . . d, and by recursive application of the APCS was implemented on a LSIll/23 micro-
equation above, one obtains computer under the RT-11 operating system,
communicating with the process computer via a
y ( k + d I k) = Fa(q-1)y(k + 1)
RS232-C serial link.
+ G d ( q - 1 ) u ( k + 1) + g d l u ( k I k)
with EXPERIMENTAL RESULTS
F d ( q - ' ) = f d l q - ' + ' ' " + fdnq-"
Gd(q -1) = gdEq -2 + . •. + gdmq -m Auto-tuned P I D
In order to do a fair comparison with the
or, in condensed form existing situation, the PID regulator was first
y ( k + d J k ) = O~(k)C,(k) + galU(k ] k) tuned using the relay control method of ,~str6m
and H~igglund (1984) with Ziegler-Nichols
where tuning rules. Figure 4 shows the resulting
O~(k) = [fal(k), . . . , fdn(k), gd2(k), . . . , gd,,,(k)] closed-loop step response, obtained with a
proportional gain of 20 and a reset time of
• T(k)---- [ y ( k ) , . . . , y ( k + l - n ) , 10 min.
u ( k - 1) . . . . , u ( k + 1 - m)]. For the identification experiment, four periods
of a 64 point PRBS with a sampling interval of
The control u ( k I k) is then computed so that 30 s, and peak-to-peak amplitude of 6 mV were
y ( k + d I k) equals a desired trajectory value sent to the setpoint of the loop. For investigation
ya(k + d I k), i.e. purposes, the identification was then performed
u ( k ) -- u ( k I k ) = gd~(k) for values of p ranging from 0.1 to 1, without
noticeable degradation in the quality of the
× [yd(k + d i g ) - O~(k)~(k)]. estimated model, as compared with a third-order
In an adaptive mode, the coefficients ai and bi A R M A X model estimated using the maximum-
in equation (1) for j = 1 are updated through a likelihood method. Sixteen Laguerre functions
recursive least-squares estimation scheme. In were used in the model, although in all cases
order to increase robustness, a dead zone on the only the first eight gains or so were really
38 G . A . DUMONTet al.

1055

1054

1053

"" I052
\,
\\
\
1051

i
1050

1049
2 4 6 8 10 12 14 16 18 20

Time i n m i n u t e s

FIG, 5. Step response with auto-tuned PID controller.

significant. Using the ISE criterion a propor- APCS under more demanding conditions, no
tional gain of 25 and a reset time of 8 min were feedforward from the production rate was
then obtained. Figure 5 shows the resulting allowed. Of course, for continuous industrial
closed-loop step response. It is seen that the operation, this feedforward compensation should
response time is reduced to about 5 from about be included. In all experiments, the prediction
15rain with the Ziegler-Nichols tuning con- horizon d was set to 5. Two sets of experiments
stants. Figure 6 shows a 3 h period during which, were performed, one with a sampling interval of
except for a minor change, the setpoint is 30s, pure delay of 2 and n = m - - 2 , the other
constant but two major changes in the fibre flow one with sampling interval of I min, no delay
occur. Despite the presence of the gain and n = 2, m = 3. The number of b parameters
scheduling, the production rate change results in allows for uncertainty in the time delay. To
a non-negligible transient. However once this initialize the model parameters, the estimator
transient has elapsed, the regulatory control is was run while the loop was under PID control.
satisfactory. Figure 7 shows the closed-loop response
obtained with APCS, using the second-order
Adaptive-predictive control model. The step response compares favorably
The APCS was implemented on the bleach with Fig. 5. Note the increase in production
plant for a 10 day trial period. In order to test around 50min and the resulting change in

1044 .... /. ~ io,~4

1042 / "-.. .... :" ......... - ,o42
1040 "~'.7"..~7:.*~ . . . . . . -' ~.7.~" -" ~040
.,(.

3400 34OO

i3000 3000

2600 2600

~1. 2200 - 2200

600

550
L-..,, f oo
'~" 5 0 0

450 I " I I I I I I I 1450

0.15 1.0 15 2.0 2 5 30
T i m e (hours}
FIG. 6. Setpoint and load changes with auto-tuned PID.
 Bleach plant PID and APCS control 39

1054 q I O54 Fig. 9, which compares the autocovariances for

....... z.'2";._ ~ . . ~ ~
the two schemes, APCS is closer to minimum
~" 1052 ..... " '111052
variance, with an ORP1 standard deviation of
! 0.20 vs 0.33mV for the PID. This is the
OC 1050 ""•"" '" '/ 1 1050
0
advantage of the adaptive controller, which
- - - -'- Q - ' ' ~ :"::" . . . . . . . "" 1 1048 performs dead time compensation. However, the
@
regulatory behaviour of the PID is more than
acceptable from the mill point of view. The
surprising fact is that, even without gain
28OO
scheduling, APCS handles production rate
changes better than the PID scheme. The two
e reasons for this are the fact that APCS
L
590 ......,,.........,..-"2 590
compensates for the time delay, and the fact that
the consistency measurement used in the gain
570 570
• .../ scheduling scheme is less than satisfactory.
.."••,...••......../ .............. ....•• : 550 The advantages of the PID scheme are its
simplicity and robustness, since the tuning is
530 I I I I I I I .uk30
done in batch and only on demand. Its main
I 1~ ........... ..-,. ........ .,'1 , ~1.1
| 1.ol- o ''-"__2: ""~., . . . . . . . . :--11.o weakness is its inability to compensate for dead
| o.o - - - " :' o.o time or to adapt to changing dynamics.
T - - "
OAk, . . . . . . . . . . . . . . . . . . . . . . .
.......... " T.1 0.1 However, after several months of operations,
! !'!I "~0~1 the performance of the automatically tuned PID
- T ; T ; t. . L. . 1 I -~ -0.2 had not deteriorated, indicating fairly time-
0 15 30 45 60 invariant dynamics. This is further confirmed by
Time (mln) the APCS experiments, where due to lack of
FIt3.7. Setpoint and load changes with APCS. gain scheduling, the parameter estimates seem to
change significantly only in case of a production
parameter estimates. Figure 8 should be rate change. Most of the time the prediction
compared with Fig. 6, as the setpoint is constant error is within the dead zone and thus APCS,
during a long period while the production rate although adaptive behaves like a gain scheduler.
changes significantly. Note that the adaptive This indicates that in this particular case,
control scheme handles the production rate continuous adaptation is not required, and as the
changes fairly well. Experiments with the third- main source of dynamic change is known, gain
order model gave very similar results. scheduling is sutticient. The advantage of the
APCS approach is that it provides the same
Discussion performance as the fixed scheme while requiring
For the servo problem, the performances of less a priori knowledge. Our recommendation
the two schemes appear similar. As seen from for this particular plant would be to use a fixed

1058 1058

I
1o88L "'" ................. ............
"" 1056

0 1 ../'.. 1054
4~

~,(X)

~eoo

2200

a2

.oo r_J ~--'~-~

150o

55o i550
o
',\ 500

450 l I I I I I I I 1 I I I I I 450
0 0.5 1.0 1.5 2.0 2,5 3.0 3.5
Time (hours)

FIG. 8. Setpoint and load changes with APCS.

40 G . A . DUMONTet al.

O.l~

0.1

0.08

o* 0.08

cJ 0.04

o iii ......

-0.02 , , L , ~ J J ~
0 1 2 3 4 5 8 7 8 9 10

Time in minutes
Pro. 9. Autocovariance functions for the ORP1 signal under PID and APCS
control.

controller with dead time compensation and gain ,~str6m, K. J. (1986). Adaptive control with a long
scheduling. An adaptive feature might be kept in , prediction horizon. 25th IEEE CDC, Athens, Greece.
Astr6m, K. J. and T. Hiigglund (1984). Automatic tuning of
order to improve both the linearizing block and simple regulators with specifications of phase and
the gain scheduling. This suggests a hybrid of the amplitude margins. Automatica, 20, 645-651.
two previous methods, based on the sound Clarke, D. W., A. J. Hodgson and P. S. Tufts (1983). Offset
problems and k-incremental predictors in self-tuning
engineering principle that available process control. Proc. lEE, 130, 217-225.
knowledge should not be discarded even if an Cutler, C. R. and B. L. Ramaker (1980). Dynamic Matrix
adaptive controller is used. As for chlorination Controlma computer control algorithm. 1980 Joint
Automatic Control Conf., San Francisco, California.
stage control in general, we would recommend De Keyser, R. M. and A. R. Van Cauwenberghe (1985).
the now accepted scheme based on combined Extended prediction self-adaptive control. 7th IFAC
brightness and residual sensors described earlier. Symp. on Identification and System Parameter Estimation,
York, U.K.
Dumont, G. A., C. C. Zervos and P. R. Brlanger (1985).
Automatic tuning of industrial PID controllers. American
CONCLUSIONS Control Conf., Boston, MA, U.S.A.
These results show that, when n o n l i n e a r i t i e s Lee, Y. W. (1960). Statistical Theory of Communications.
are known and changes in process dynamics are Wiley, New York.
Martin-Sanchez, J. M. (1976a). Adaptive predictive control
predictable, gain scheduling combined with a system. U.S.A. Patent No. 4,197,576.
well-tuned PID regulator behave remarkably Martin-Sanchez, J. M. (1976b). A new solution to adaptive
well. At the same time, it shows that adaptive control. Proc. 1EEE, 64, 1209-1218.
Martin-Sanchez, J. M. and S. L. Shah (1984). Multivariable
control can handle the same situation with less adaptive predictive control system of a binary distillation
tailoring. The best solution might be a hybrid of column. Automatica, 20, 607-620.
the two schemes, e.g. on-line improvement of Ydstie, B. E. (1982). Robust adaptive control of chemical
processes. Ph.D. thesis, Imperial College, University of
the linearizing block and of the gain scheduling. London
Ydstie, B. E. (1984). Extended horizon adaptive control. 9th
IFAC World Congress, Budapest, Hungary.
REFERENCES Zervos, C. C., P. R. Brlanger and G. A. Dumont (1985).
Ackert, J. E., D. D. Koch and L. L. Edwards (1975). Tuning of PID controllers using orthonormal functions.
Displacement chlorination of kraft pulps---an experimental IFAC Workshop on Adaptive Control of Chemical
study and comparison of models. Tappi, ~ , 141-145. Processes, Frankfurt/Main, F.R.G.