Control Engineering Practice 20 (2012) 355–370

Contents lists available at SciVerse ScienceDirect

Control Engineering Practice

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

Robust optimization-based multi-loop PID controller tuning: A new tool and

its industrial application
Rainer Dittmar a,n, Shabroz Gill b, Harpreet Singh b, Mark Darby c
a
West Coast University of Applied Sciences, Department of Electrical and Information Engineering, Fritz-Thiedemann-Ring 20, 25746 Heide, Germany
b
IPCOS BV, Bosscheweg 135B, 5282 Boxtel, The Netherlands
c
CMiD Solutions, 13106 Dogwood Blossom Trail, Houston, Texas 77065, USA

a r t i c l e i n f o abstract

Article history: Modern process plants are highly integrated and as a result, decentralized PID control loops are often
Received 15 December 2009 strongly interactive. The iterative SISO tuning approach currently used in industry is not only time
Accepted 17 October 2011 consuming, but does also not achieve optimal performance of the inherently multivariable control
Available online 29 November 2011
system. This paper describes a method and a software tool that allows control engineers/technicians to
Keywords: calculate optimal PID controller settings for multi-loop process systems. It requires the identification of
PID controller a full dynamic model of the multivariable system, and uses constrained nonlinear optimization
Multiloop control techniques to find the controller parameters. The solution is tailored to the specific control system
Tuning characteristics and PID algorithm to be used. The methodology has been successfully applied in many industrial
Decentralized control
advanced control projects. The tuning results that have been achieved for interacting PID control loops
Nonlinear optimization
in the stabilizing section of an industrial Gasoline Treatment Unit as well as a Diesel Desulfurization
System identification
plant are presented.
& 2011 Elsevier Ltd. All rights reserved.

1. Introduction A feedback system with n decentralized controllers is shown in

Fig. 1. In multivariable notation, the controller C(s) is a diagonal
One of the most important challenges facing the process industry matrix given by
today is optimizing the operation of complex units, without com- 2 3
C 1 ðsÞ    0
promising the safety and integrity of the process equipment. Process 6
CðsÞ ¼ 4 0 & 0 7 5 ð1Þ
complexity has increased significantly over the past two decades
0    C n ðsÞ
due to increased level of heat integration and the use of recycle
streams. In addition, the need for increased process flexibility to deal The process transfer function matrix G(s) is
with changing raw materials and alternate energy sources, as well as 2 3
G11 ðsÞ    G1n ðsÞ
the need to adapt quickly to fluctuating throughput and quality
6 0 & 0 7
targets, often means that the process dynamics will vary signifi- GðsÞ ¼ 4 5 ð2Þ
cantly over time and with operating point. The basic regulatory Gn1 ðsÞ    Gnn ðsÞ
control layer of process plants almost always consists of a large
The vector y(t)ARn denotes the process outputs (controlled
number of decentralized SISO PID controllers. Although this
variables), u(t)ARn is the vector of manipulated variables, and
approach is intrinsically inadequate for multivariable processes, it
r(t)ARn the vector of reference signals or setpoints.
is a common practice in industry for decades now, and for good
The industrial practice of single-loop PID controller tuning is still
reasons: single-loop PID controllers are often effective and easy to
dominated by manual trial-and-error tuning. If tuning rules are used
implement, and decentralized structures are failure-tolerant. How-
at all, it is the ‘‘classical’’ ones like ZieglerNichols, CohenCoon,
ever, due to the situation described above, the interactions between
ChienHrones Reswick or l-Tuning, which are based on simplified
these controllers are becoming more important, and tuning these
low order process models, and do not explicitely consider stability
control loops for good performance and adequate robustness is
robustness issues, therefore often being inadequate in modern
becoming a more and more challenging task.
process units with more complex dynamics. Modern identification
techniques are not typically used to develop models for PID tuning.
Substantial progress has been achieved in PID controller design
n
Corresponding author. Tel.: þ49 481 8555 325; fax: þ 49 481 8555 301. during the last decade (see, for example, Aström & Hägglund, 2006;
E-mail address: dittmar@fh-westkueste.de (R. Dittmar). Kristiansson & Lennartson, 2006; Skogestad, 2003), but not yet

0967-0661/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2011.10.011
356 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370
adequately adressed in further education and industrial practice. In
addition, many tuning rules assume that all PID controller equations
work as described in the textbooks, when in fact there is substantial
variation between the different vendors. Different PID controller
structures result due to the use of either the parallel or the serial
form, using the control error or the PV by the Proportional (P) and
Derivative (D) terms, alternative implementations of the derivative
filter and others. Tuning SISO PID controllers in a multivariable
environment is usually done in a time-consuming sequential and
iterative way, starting with the most important loops, and heuristic
detuning in case the interactions are significant. The results in terms
of performance and robustness strongly depend on the particular
application and on the experience of the personnel involved.
For a long time, vendors of automation systems such as Dis-
tributed Control Systems (DCS) and Programmable Logic Controllers
(PLC) have been offering PID self-tuning functionality (tuning on
demand). Unfortunately, they have found only limited application.
This is also true for model based PID controller tuning software
provided by the same or third-party vendors. Moreover, in most
cases these tools are restricted to single-loop tuning applications,
and do not support multi-loop tuning (Li, Ang & Chong, 2006).
Examples are RaPID (Espinosa Oviedo, Boelen & van Overschee,
2006; Van Overschee & de Moor, 2000) and TaiJi PID (Zhu, 2004). In
TiaJi PID, plant models are identified from open-loop or closed-loop
tests using high-order ARX models and reduced to low-order
transfer function models. PID controller parameters are calculated
using IMC-based tuning rules. In the current release, tuning is
restricted to single-loop applications. However, the user can simu-
late the multivariable system controlled by independently tuned
single-loop PID controllers. In RaPID, the process model is identified
using a combination of prediction error methods and subspace
identification. Controller settings are calculated based on optimiza-
tion with robustness and noise amplification constraints.
The design of interacting PID controllers in a multivariable
environment is not a new topic in the process control literature.
At least three research directions can be identified: (1) reduction
of controller interactions by proper pairing of the manipulated
and control variables, (2) design of decoupling networks, and
(3) consideration of MIMO interactions in decentralized controller
tuning. This paper is a contribution to the solution of the third
problem. In the past, many design methods have been developed
for multi-loop tuning. They can be classified into
1. detuning methods;
2. sequential loop closing methods;
3. independent design methods;
4. relay-feedback auto-tuning methods; and
5. optimization methods.
In detuning methods, each controller is first designed based on
the corresponding diagonal element of the process transfer
function matrix while ignoring the interactions from other loops.
The controllers are then detuned to take the interactions into
account. The price to be paid for the reduced interaction is more
sluggish behavior of the individual PID loops. The most popular
detuning method called ‘‘biggest log modulus tuning’’ (BLT) was
developed by Luyben (1986). In its original version, individual PI
loops are first independently tuned by the Ziegler  Nichols (ZN)
rules. Then, for multiloop tuning, all ZN controller gains are
divided by a common detuning factor F 41, and the ZN reset
times are multiplied by the same factor. This detuning factor can
be calculated based on the following considerations. The char-
acteristic equation of the multivariable closed-loop system is
det½I þGðjoÞCðjoÞ ¼ 0
The scalar function
WðjoÞ ¼ 1 þdet½I þ GðjoÞCðjoÞ ð4Þ
can be plotted in the complex plane as a function of frequency o.
The closer W(jo) approaches the critical point (  1,0), the closer
the multivariable system is to closed-loop instability. The expres-
sion W/(1 þW) will be similar to the closed-loop servo transfer
function CG/(1 þCG) in the SISO case. Luyben has shown that the
multivariable system gives reasonable responses for setpoint
changes and load disturbances, if F is selected to make the so-
called ‘‘biggest log modulus’’ in the frequency range
  
 WðjoÞ 
Lmax ¼ max LðjoÞ ¼ max 20log  ð5Þ
o o 1 þ WðjoÞ
equal to 2n (n being the number of loops to be tuned). This value
for F can be found using an iterative procedure. This method has
later been extended to PID controllers, and also to compensate for
possible asymmetries in multi-loop interactions (Lee & Edgar,
2006; Monica, Yu & Luyben, 1988). Its application requires a
dynamic MIMO process model. In Chien, Huang and Yang (1999),
the detuning factor F is set to F¼ 1/RGA(lii), e.g. to the diagonal
elements of the relative gain array, for systems with 0oRGA(lii)o1.
For loops with RGA(lii)41, the controller parameters remain
unchanged. If RGA(lii)o0, the recommendation is to switch the
MVCV pairings. In contrast to Luyben’s method, dynamic models
are required for the diagonal elements of the process Gii(s) only,
while static gains for the off-diagonal elements Gij,i a j(0) are suffi-
cient. In Xiong, Cai, He and He (2006), not only the static gains, but
also the ultimate frequencies ou,ij are used to define the detuning
factors.
In the sequential loop closing method (Hovd & Skogestad,
1994), the loops are closed one after the other, usually starting
with the fastest loop. The dynamic interaction of this loop is
considered when closing the next loop, and so on. The disadvan-
tage of this approach is that the overall result depends on the
order of loop closing and on the method used for individual
controller design, and that iterations may be necessary.
In independent design, loop interactions, robust performance,
and stability are considered first, and each controller is then
designed independently of each other. Hovd and Skogestad
(1993) make use of the m-interaction measure and the structured
singular value to develop bounds for the design of the individual
loops. Chen and Seborg (2003) used multivariable Nyquist stabi-
lity analysis to develop stability regions for SISO PI controllers,
and proposed a tuning method, which guarantees closed-loop
stability of the decentralized MIMO control system. Huang, Jeng,
Chiang and Pan (2003) developed a method for multi-loop PI/PID
controller tuning based on the effective open-loop transmission
from ui to yi and phase/gain margin specifications. Their method
assumes that the effective open-loop processes can be approxi-
mated by FOPDT or SOPDT models and is practically restricted to
TITO systems.
If no analytical process model exists, tuning parameters may be
calculated based on the multivariable generalization of relay-feed-
back autotuning methods, which are described, for example, in
Campestrini, Filho and Bazanella (2009), Halevi, Palmor and Efrati
(1997), and Yu (2007). Since the execution of sequential or simulta-
neous relay-feedback experiments under industrial conditions is
difficult due to noisy signals, drift, and long duration in case of larger
time constants, this approach has found limited application.
Optimization methods have also been used to address the
decentralized control problem. Sourlas and Manousiouthakis
(1995) developed a benchmark for the best achievable decentra-
lized performance. Trierweiler, Müller and Engell (2000) and
Pegel and Engell (2001) describe a generalization of the direct
ð3Þ synthesis method to multivariable systems. First, considering the
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370
nonminimum phase part of the process, the attainable closed-
loop performance is determined. Based on this, an ideal controller
is synthesized, which usually is too complex to be realized in
practice. Therefore, this ideal controller is approximated by a
simpler one where structure (for example, diagonal) and order
(for example, PI/PID) are specified by the user. The approximation
is done by solving a convex optimization problem. Here, the
objective function is to minimize the deviation between the
closed-loop frequency responses achievable with the ideal and
the simplified controllers. The controller parameters of the simpler
controller are the decision variables of this optimization problem. In
Bao, Forbes and McLellan (1999), the tuning problem is formulated
in the HN robust control framework, and a numerical optimization
procedure is proposed to solve the design problem. The structure
constraints (e.g. use of decentralized PID controllers) are converted
into matrix inequalities.
The use of genetic algorithms for multi-loop PID controller
tuning is described in (Farag & Werner, 2006) for the ALSTOM
gasifier benchmark problem and (Sumana & Venkateswarlu, 2010)
for a simulated reactive distillation column. Vlachos, Williams and
Gomm (2000) have shown that the multiloop PID tuning optimiza-
tion problem is non-convex and multimodal. Genetic algorithms as
global search methods are particularly useful to solve this type of
nonlinear optimization problems. The advantage of using optimiza-
tion procedures and a full dynamic process model is that this leads
to less conservative designs, and that overall nominal stability is
naturally achieved.
All multi-loop tuning methods described above share the
common characteristic that only very few industrial applications
have been reported, and that they have not yet been included in a
commercially available software package aimed for use by control
technicians or engineers.
The contribution of this paper is threefold. First, it presents
a method for multi-loop controller tuning, which combines
multivariable system identification to build a (n  n) model of the
uncontrolled system, and nonlinear constrained optimization to
calculate optimum PID controller settings in a multivariable system.
Since the design procedure is aimed at a direct implementation
on a distributed control system or PLC, it is restricted to multiloop
SISO PID controllers. This means no other controller structures can
be selected, and decoupler design is not a part of this approach.
Second, it introduces a software tool ‘‘AptiTuneTM’’, eliminating the
need for industrial practitioners to have a deep understanding of
identification and control theory, to enable them to formulate
Fig. 1. Decentralized multi-loop PID control system.
357
process-specific design requirements, and to deliver a tuning solu-
tion that is tailored to the specific target automation system (DCS or
PLC). Third, the paper presents experience gained in applying the
new method and tool to industrial-scale oil refinery plants. In
particular, a (2  2) application in the stabilizer section of an
industrial Gasoline Treatment Unit (GTU), and a (3  3) application
in a Diesel Desulfurization Unit (DDU) will be described.
2. Method and tool for multi-loop tuning
2.1. Identification of the MIMO process model
The first step of model based multi-loop tuning is to develop a
dynamic model of the multivariable process with n inputs and n
outputs, which are the outputs (ui) and process variables (yi) of
the PID controllers shown in Fig. 1.
Different methods and software tools can be used to develop
such a model. Fortunately, multivariable linear system identification
packages are now in common usage with Model Predictive Control
design and application, and there is already a widespread experi-
ence in using them. So it is natural to leverage this knowledge to
identify MIMO models for multi-loop PID tuning. Examples are the
identification packages of DMCplus (Aspen Technology), Profit
Controller (Honeywell, see MacArthur & Zhan, 2007) or INCA
(IPCOS). An overview of commercially available MPC software is
presented in Dittmar and Pfeiffer (2006). An alternative is the use of
stand-alone multivariable identification packages such as TaiJi ID
(Zhu, 2006) or ADAPTx (Larimore, 2000).
For test data collection, whenever possible all PID controllers
to be tuned are switched into manual mode, and a series of output
steps of different durations and amplitudes is executed. Four to
six steps with duration varying between 10% and 100% of the
desired closed-loop settling time are usually sufficient. If a test
signal generator is available for pseudo-random binary sequence
(PRBS) or generalized binary noise (GBN), then an automated test
may be used as an alternative. Both types of plant tests can be
performed in sequential or in time-saving simultanous mode. In
some cases, clean step responses to controller output changes can
be found in historical data as well. If one or more PID controllers
cannot be switched to manual mode, then the loop can be kept in
automatic mode and multiple setpoint steps can be made. In this
case, an identification method that gives unbiased results such as
high-order ARX (Zhu, 2006), variations of subspace identification
358 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370

(Juricek, Seborg & Larimore, 2002), or the projection method  minimal damping or maximum decay ratio
described in Forsell & Ljung (2000) will be required. DRi rDRi,max i ¼ 1:::n ð9Þ
The ‘‘AptiTuneTM’’ software does not provide an integrated
identification package, but multivariable FIR models created by where the decay ratio DR denotes the ratio of the second to the
identification tools from different MPC packages can be imported. first overshoot in the process variable after a setpoint change
Alternatively, the user can specify a transfer function matrix manu-  maximal measurement noise amplification
ally. In the next step, the MIMO FIR model is approximated by a NAi r NAi,max i ¼ 1. . .n ð10Þ
smooth, higher-order linear state-space model with explicit dead-
where NAi ¼ varðui Þ=varðyi Þ denotes the ratios of the controller
time. This approximation is not based on the raw or pre-processed
output to process value variances.
plant test data, but on a model-to-model fit. The state-space model is
constructed using a variant of the Ho Kalman algorithm (Ho &
 combined process gain and deadtime safety margins KSM and
TDSM, expressed as ratios between the maximum expected and
Kalman, 1966). The user can affect the approximation accuracy by
the nominal process gains Kp,max and Kp,nom, and time delays
specifying the model order search range and indirectly the smooth-
TDmax and TDnom, respectively
ness of the resulting model. While creating the state-space model,
the diagonal model curves are given more preference than the off- K pi,max =K pi,nom rKSMi TDi,max =TDi,nom r TDSMi i ¼ 1. . .n
diagonal models. As a result, diagonal models normally have higher ð11Þ
order than the off-diagonal ones and consequently fit the original FIR
model curves more accurately. The step responses calculated based
on the state-space models are graphically displayed.
 maximum/minimum limits of the controller parameters, e.g.
If it is possible to do a closed-loop setpoint step test (or if K ci,min r K ci r K ci,max , T ri,min r T ri r T ri,max
historical data contain a clear SP step, as it is often the case), a T di,min r T di r T di,max i ¼ 1. . .n ð12Þ
practical way of validating the process model is to simulate the
closed loop behavior of the control system with the actual PID
controller parameters currently entered on the DCS, and to For level buffering controllers, the control objective is to slow
compare the simulation results with historical plant data. If the down the level loop as much as possible, so the downstream flow
observed responses using the old tuning are similar to the changes are as smooth as possible. In this case, the first four
simulated responses, then one can conclude that the model is constraints can be replaced by
sufficiently accurate for loop tuning purposes.
 the maximum deviation in the process variable away from
setpoint (the difference between the level setpoint and the
2.2. Calculation of optimal PID controller parameters closest alarm limit) for the worst case disturbance, and
 the minimum closed-loop settling time after a level
The PID controller parameters (controller gains Kc,i, reset times Tr,i disturbance.
and derivative time constants Td,i) are calculated by solving numeri-
cally the following nonlinear constrained optimization problem:
By careful specification of the constraints, the user can
min J tailor the tuning to exactly meet individual process-specific
KP i, TN i , TV i
ð6Þ requirements.
g j ðK P i , T r i , T d i Þ r0 i ¼ 1:::n, j ¼ 1:::m
The problem defined above is a non-convex, multimodal,
constrained nonlinear optimization problem with a non-smooth
where J denotes the objective function and gj are constraints. The
objective function. The degree of its difficulty depends on the
objective function J is a weighted sum of three terms J ¼ J 1 þ aJ2 þ bJ3 ,
number of control loops involved, the order of the process
which assess different aspects of the control loop performance. The
RT  models, and the number and nature of the inequality constraints.
first part J1 ¼ 0 f  yðtÞyr ðtÞ  dt refers to the Integrated Absolute
For starting the numerical optimization, initial controller
Error (IAE) criterion for setpoint tracking. Here, the error is defined
parameter values have to be selected. For this purpose, the user
as the difference between the PV and a user-defined first order
can choose to use the actual DCS values or values calculated by
reference trajectory yr(t) connecting the actual PV and the setpoint.
the Cohen Coon tuning rule (for individual controller tuning
By specifying the time constant of the trajectory, the user can affect
assuming a SISO model). If the initial solution is infeasible, a
the speed of the response to setpoint changes. The second part
RT   feasible starting point can be searched by applying a global search
J 2 ¼ 0 f  wðtÞyðtÞ  dt denotes the IAE
 for an input step disturbance.
RT algorithm. For this purpose, either a time-consuming brute force
Finally, the third term J 3 ¼ 0 f DuðtÞðtÞ reflects the control effort. By
grid search in the entire parameter space or a genetic algorithm
setting the weighting coefficients a and b, the user can balance a
(GA) can be selected. The genetic algorithm code has been
compromise between the different performance objectives. Another
developed based on a related CODE project open source library
design parameter allows the user to weight the performance of the n
(Kirillov, 2006). In the next step, the search of the optimal
SISO control loops against the necessary degree of decoupling
controller parameters is started by employing a gradient-free
between them.
direct search method similar to the Nelder Mead simplex algo-
For each control loop, the user can specify one or more
rithm. In each step of the optimization, the PID controlled multi-
nonlinear inequality constraints gj r0 from the following list:
variable system is simulated (reaction to setpoint changes and
input step disturbances, and to measurement noise) and viola-
 maximum controller output deviation after setpoint changes,
tions of the constraints may be detected. Constraints are then
Dui ðkÞ r Dui,max or Dui ðkÞDui,max r0, i ¼ 1. . .n ð7Þ enforced using quadratic penalty functions to force the search
inside the feasible region. Since the simplex algorithm is not
 maximum overshoot in the process variable after setpoint suited to find global optima, one can rerun the optimization
changes starting with different initial values. If there are still doubts about
the optimum, a GA global search can be performed again.
OSi r OSi,max i ¼ 1. . .n ð8Þ
Regarding robustness against plant-model mismatch, the
with OSi ¼ ðyi,max yi ð1ÞÞ=Dr i  100% design intent is to build in substantial safety margins to preserve
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370 359

the stability of the control loop when inevitably changes in the

process dynamics occur in the future, e.g., with operating point.
Most uncertainty approaches in the literature are based on HN
formulations. Such techniques assume a particular form of an
uncertainty description and do not easily allow nonlinear con-
straints of the form indicated above. In practice, it is not known
ahead of time how the specific model parameters will vary over
time or how process changes will modify process gains and time
constants. One might think it would be best to optimize for the
worst combination of model errors. However, this cannot only
lead to conservative tuning parameters, but also to a situation
where elements of the relative gain array (RGA) switch signs,
which would be an unusual situation in real-world applications.
Such a situation would indicate that a different loop pairing
should be used, but this would occur only in special cases. In
practice, it can be expected that the diagonal models will have the
largest gains (compared to the non-diagonal ones), a good signal-
to-noise ratio, and also the most accurate gains (usually 10,y,20%
error at the time of the plant tests for identification). According to
the approach in this paper, the combined gain and time delay
safety margins specified as constraints for optimization apply
only to changes in one individual loop at a time. But the design is
Fig. 2. Robustness plot.
not done for a combination of many small (e.g. 10%) errors in
multiple model parameters. Instead, a rather large combined
change in gain and dead time (of 200  300%) is assumed. In
practice, such large safety margins usually ensure that the
stability of the loop is maintained for moderate (realistic) changes parameters for a specific realization of the PID controller equation
to multiple diagonal models, as well as for a combination of for specific commercial control system hardware. This is very
moderate perturbations in multiple diagonal and off-diagonal important from a practical point of view. The user can select
model parameters. The approach taken can further be motivated between different control algorithms of widespread DCS systems
in the frequency domain. It is possible to tune a MIMO system by such as Honeywell, Emerson DeltaV, Foxboro I/A, ABB and several
sequentially tuning a series of SISO loops (but of course this is others. For example, six different versions of the PID algorithms
inefficient). From the standpoint of a specific SISO loop, process are available for the Emerson DCS, for which the optimization
changes or model errors (anywhere in the system) manifest results may be quite different. Optimal controller parameters can
themselves as changes in the amplitude (gain) and phase of the also be calculated for P, I only and PI controllers.
Bode plot. So one can think of lumping all potential model errors After the optimizer has converged and optimal controller
in terms of gain and dead time (which is directly related to the parameters have been found, the design process will be finished
Bode phase plot). For this reason, changes in time constants can by simulation of the dynamic behavior of the control system. It is
also be ignored as they manifest themselves on the Bode plot as useful to study different scenarios: setpoint tracking, input dis-
gain and phase errors too. turbance rejection, and noise attenuation. Since the ‘‘integrity’’ of
The instability boundary for the complete MIMO system is the controlled MIMO system is not considered in the optimization
determined by varying both the model gain and deadtime for loop algorithm, the user should study the system behavior if one or
i, keeping the models associated with other loops (iaj) and all more controllers are put into manual mode, or a controller is lost
other off-diagonal models constant. In the implementation, the due to actuator saturation or component failure.
deadtime of a diagonal model is varied from zero to a sufficiently
large number. The process gain is searched and the instability
point is found where the largest eigenvalue of the system matrix 3. Industrial applications
A crosses the unit circle. The constrained optimization algorithm
ensures that the instability line is not violated for the specified 3.1. Gasoline treatment unit (GTU)
combination of gain and deadtime safety margins.
Once the optimization routine has finished, robustness plots The method and software tool described above have been used
are calculated, and the user can evaluate the closed loop response successfully in a number of advanced control projects. A good
via simulation for different combinations of deadtime and gain example is the stabilizing section of a GTU process, where
error. A robustness plot example is shown in Fig. 2. It shows the improving the PID controller tuning was a prerequisite for
thick purple stability limit in a process gain ratio/dead time successful MPC design and implementation. The Process and
graph. The gain ratio is defined as the actual divided by the Instrumentation Diagram (P&ID ) for the stabilizing column of
nominal process gain. A recommended minimum stability region the GTU process is shown in Fig. 3.
is shown shadowed. Here, the nominal model lies at the center The objective of the stabilizer column is to remove hydrogen
(gain ratio 1, nominal time delay), the vertices are constructed and methane dissolved in the petrol (mostly C5) stream. The
based on robustness constraints entered by the user. In the plot column is essentially a degassing drum with trays for improved
shown, the gain ratio may vary between 0.5 and 2, and the dead separation. The current PID control scheme is somewhat uncon-
time in the range of (0.5,y,2) times its nominal value. The ventional in that the PID loop pairing is the ‘‘wrong’’ way round: a
stability limit should not intersect this region. tray temperature close to the top of the column is controlled
In contrast to some PID controller tuning software available (even though the product specification is on the bottoms stream)
mainly for teaching and training purposes, the ‘‘AptiTuneTM’’ tool using the overhead vent valve for temperature control. There is no
not only calculates ‘‘generic’’ PID controller parameters, but reflux drum, and feed comes in close to the top, providing the
360 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370

Fig. 3. Simplified P&ID of the GTU process stabilizer column.

Fig. 4. Step responses of the dynamic model (Legend: TC.PV¼Tray 6 Temperature; PC ¼Column Pressure; TC.OP¼ Vent Valve; PC.OP ¼Steam Flow SP).

internal reflux stream. The column pressure controller cascades to
the steam flow SP on the reboiler. One reason for the unconven-
tional PID loop pairing is that controlling pressure with the
reboiler duty ensures that the pressure is less likely to go
high and lift the safety valve. If the PID was paired the other
way around, there is a chance for the pressure to go high if
the overhead valve saturates before the column runs out of
reboiler duty.
The drum level controller sets the valve position directly
without the benefit of a cascaded flow controller. Any change in
either upstream or downstream pressure affects the feed flow PV,
which then affects the temperature and pressure in the column.
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370 361

Fig. 5. Response of the (2  2) system to SP step changes (Legend: PV is shown in red, OP in blue). (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)

Fig. 6. Response of the GTU (2  2) system to load disturbances.

362 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370

Due to the coupling between the three loops, it is really not
feasible to tune one loop without consideration of the remaining
two control loops. In the actual project, LC9803 has been tuned
separately as averaging level controller, and PC9876 as well as
TC9854 using the multi-loop tuning software.
Good results were obtained by step testing condensate
flow setpoint and vent valve, fitting a (2  2) model and then
calculating moderately fast but well damped tuning that takes
the strong off-diagonal interaction into account. Here, the loops
were tuned for fast SP tracking but high robustness margins.

Fig. 7. Robustness plot (gain and deadtime stability margins) for the GTU (2  2) problem.

PCA9876 BEFORE TUNING

12
10 Frequency
Normal distribution
Frequency

8
6
4
2
0
−5 −4 −3 −2 −1 0 1 2 3 4
PCA9876 AFTER TUNING
25
Frequency
20
Normal distribution
Frequency

15

10

0
−5 −4 −3 −2 −1 0 1 2 3 4
PCA9876 SP−PV error deviation

Fig. 8. Control error histograms for PCA9876 before and after retuning.
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370 363

The standard deviations of the two PVs are now substantially The gain for the tray temperature vs. vent valve varies very
better than before and both loops are tracking their setpoints substantially depending on operating point (from almost zero
well. when the top of the column is perfectly pure, to almost infinite

TC9854 BEFORE TUNING

20
Frequency
15 Normal distribution
Frequency

10

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
TC9876 AFTER TUNING
20
Frequency
15 Normal distribution
Frequency

10

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
TC9854 SP−PV error deviation

Fig. 9. Control error histograms for TC9854 before and after retuning.

Fig. 10. Simplified P&ID of an oil refinery Diesel Desulfurization Unit.

364 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370

when the column goes rapidly off-spec). The best model was
selected, which could be identified by keeping the tray tempera-
ture in the correct range with the column pressure at the nominal
operating point. The step responses of the dynamic model are
shown in Fig. 4.
Note that the total state-space model order is 18. Fig. 5 shows
the optimized PV responses for a step change in both SPs.
Note that PV overshoot is very low and that damping is
exceptionally good. The OP value for the vent has a peak value
that is almost the same as the steady state value, and the loop has
about the same rise time in closed loop as compared to open loop
(a speed-up factor of about 1). The pressure loop is about two
times faster in closed loop compared to open loop. The load
disturbance response is shown in Fig. 6.
PRBS test: Controller outputs
70
65
60
55
OP [%]
Note that the PV and OP responses have very good damping,
with peak OP values that are very similar to the steady state
values. This will ensure good damping on the actual process unit
even when the process gain varies significantly. Gain and dead
time (stability) margins are very good as well, see Fig. 7.
A sensitivity analysis on the tray temperature loop shows that
the gain of the process will have to increase by three times and
the dead time will have to increase by another 12 s before the
damping of the loops is unacceptable. Instability occurs at a
process gain increase of more than 20 times. A dead time error
of more than 2 min is needed to reach instability, and there is no
process mechanism for this to occur while the TC9854 is in the
active range. These margins are very safe.
In order to compare the performance of the loops before
and after retuning, a week of normal operating data before
re-tuning have been collected, and one week of normal operat-
ing data after the re-tuning work was concluded. From these
large data sets, histograms were calculated to show the distribu-
tion of the control error (SP-PV). For process reasons, the loops
should be able to withstand changes in process dynamics. As a
result, some loops (not shown here) were intentionally slowed
down, and of course, their probability distributions will be wider
than before. However, this compromise is all for a good cause as it
will ensure that the loops remain operational for the years
to come.

50 The performance of the pressure control loop PC9876 is

compared in Fig. 8.
45
It is clear from the two histograms shown above that the
40 variability in the PV has reduced by about a factor of four. This is
a big improvement in performance, yet this could be accomplished
35 TC002OP without compromising the robustness characteristics of the loop.
TC001OP The pink trace shows the best fit for a normal (Gaussian) distribu-
30 PC001OP
tion to the data, assuming a zero mean value. The estimated
standard deviation s reduced from 1.1 to 0.28.
0 200 400 600 800 1000 1200 1400 1600
The histograms for temperature loop TC9854 are compared in
Time
Fig. 9.
Fig. 11. Simultaneous step test trends (Legend: Top: TC002.OP, middle: TC001.OP, The standard deviation reduced from 1.3 down to 0.5, so by a
bottom: PC001.OP). factor of almost three. To be honest, these good results are partly

PRBS test: controller PVs

140
TC002PV

120

100

80
0 200 400 600 800 1000 1200 1400 1600 1800
200
TC001PV

150

100

50
0 200 400 600 800 1000 1200 1400 1600 1800
120
PC001PV

100

80

60
0 200 400 600 800 1000 1200 1400 1600 1800
Time

Fig. 12. DDU step responses (Legend: Top: PC001.PV, middle: TC001.PV, bottom: TC002.PV).
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370
due to the very slow initial tuning of some loops at the beginning
of the project.
3.2. Diesel desulfurization unit (DDU)
The second example refers to an oil refinery DDU where
retuning the basic regulatory loops was also one of the first steps
of an advanced control project. Its simplified P&ID is shown in
Fig. 10.
The feed (diesel plus make-up hydrogen) flows through a
furnace, and is vaporized and heated to the required elevated
temperature before entering the fixed-bed catalyst reactor where
the hydro-desulfurization reaction takes place. The reaction
product consists of a hydrated hydrocarbon mixture and hydro-
gen sulfide. After cooling down, the reaction product enters
a stripper where the hydrogen sulfide is stripped out and leaves
the stripper overhead together with light hydrocarbons (methane
ybutane). The bottoms product is the desulfurized diesel. The
overhead product flows to an amine contactor where the
Fig. 13. Step responses of the DDU state space model.
365
hydrogen sulfide is removed. The remaining light hydrocarbons
are mixed with fresh fuel gas and used to heat the furnace and the
stripper reboiler.
The PID control loops PC001 (unit backpressure), TC001
(reactor inlet temperature), and TC002 (stripper column tempera-
ture) are interactive and were hard to tune independently. The
interaction is caused mainly by PC001, which manipulates the
fresh fuel gas import, changing the heating value of the combus-
tion gas at the furnace and the reboiler, and thus affecting the
temperature controllers TC001 and TC002. On the other hand,
both temperature controllers manipulate the fuel gas consump-
tion and thereby the pressure PC001.
In the past, the three controllers were tuned on a SISO basis as
PI controllers. Analysis of the control behavior showed that the
pressure controller was tuned too aggressively leading to an
oscillatory behavior of the loop PC001, particularly in case of
disturbances caused by the temperature controllers. On the other
hand, TC001 was tuned very slowly, mainly to attenuate the
impact of PC001 on TC001, resulting in a sluggish response. TC002
366 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370

was characterized by a sustained oscillation with a period of process model. The tests were done simultaneously in closed loop
50,y,60 min. Since these controllers were candidates for MVs of utilizing the OPMCHLM parameter of the Honeywell TDC3000
an MPC controller to be implemented as part of the APC project, it DCS PID controllers. This parameter can be used to generate a
was decided to retune them on a multi-loop basis using staircase signal, which is added to the PID controller output. In
‘‘AptituneTM’’. addition, the operators were asked to make setpoint changes of
The first step was to perform plant tests in order to get varying step size in order to change the average process values.
information-rich data to build a multivariable (3  3) dynamic Since the settling time was in the range of 90 min, a total of one
day test data was required for this application. Intermediate
identification results were used to check the progress in terms
Table 1 of model accuracy reached, and these results were used to modify
Old and new PID controller settings for the Diesel Desulfurization Unit. the step width durations (frequency content) of the test signals.
Figs. 11 and 12 present selected trends of the process input
Loop Old, single-loop settings New, multi-loop settings
(controller output) and output (pressure and temperature) sig-
KP Tr (min) Td KP Tr (min) Td (min) nals, respectively (note that the Y scales have been shifted from
actual).
PC001 4.25 2.0 0 1.2 1.05 0.55 Based on the test data, a (3  3) model was developed and
TC001 1.8 1.9 0 7.2 1.9 0.48
TC002 4.0 5.0 0 5.05 13.6 0.32
approximated by subspace identification. Fig. 13 presents the step
responses generated by the state space model. Note that the

Fig. 14. Simulated setpoint and load disturbance step responses, robustness plots.
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370
model matrix is fully dense, and that most of the responses
cannot be easily approximated by simple first or second order
plus deadtime models.
Next, new optimal settings for the three controllers were
calculated based on the following design specifications:
 Honeywell TDC3000 PID equation C, e.g. integral part acting on
the control error, P and D parts acting on the PV (to avoid
proportional/derivative kicks)
 PC001: closed loop settling time 3 min, 10% maximum over-
shoot, good damping, moderate output movement after
setpoint change
Fig. 15. Trends for TC002 before and after multi-loop tuning (Legend: green: SP,
red: PV, blue: OP). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
Fig. 16. Robustness plot for PC001 with old and new tuning parameters.
367
 TC001: little overshoot, good damping, hard constraint for OP
movement after setpoint change to avoid too sharp kicks of the
combustion fuel flow
 TC002: no overshoot, avoid temperature cycling, good robust-
ness against process gain variation
 Sufficient stability margins (for example, for TC001 increasing
both the gain and/or the deadtime by a factor of two should
not lead to instability)
In Table 1, the old PI settings based on single-loop tuning are
compared with the results of the multi-loop controller parameter
optimization.
Fig. 14 presents an overview of the tuning results by simula-
tion. In the left and right columns, setpoint and load disturbance
step responses for TC001, TC002 and PC001 are shown. The
middle column contains the individual gain/deadtime robustness
plots, demonstrating that all controller settings meet the robust-
ness specifications. The decoupling results (not shown here) were
satisfactory as well.
The tuning results were placed in service without fine-tuning
in October 2008 and were kept unchanged since then. Fig. 15
presents 12 h trends for TC002 before and after multi-loop tuning.
The improvement with regard to temperature cycling can be seen
by visual inspection.
The robustness of loop PC001 using previous and new con-
troller parameters is shown in Fig. 16. While the old tuning was
not robust, e.g. the stability limit crosses the shadowed minimum
stability region, the new settimgs lead to sufficient robustness.
Additional information is provided in Darby and Harmse
(2009).
3.3. Discussion
The steady-state and dynamic Relative Gain Arrays (RGA), and
the multivariable RHP zeros of the GTU and DDU applications are
provided in the Appendix. In both cases, centralized multi-loop
tuning is justified. The closed-loop performance of the two
control systems is limited by two MIMO RHP zeros, which
368 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370

constrain the achievable bandwidths (Skogestad & Postlethwaite, 1.4

2005). The RGA analysis indicates that a decentralized tuning λ11
approach might be adequate. But one has to keep in mind that PID
λ12
loop tuning in industry is done in most cases with a trial-and-error 1.2
approach, or based on a single step-test with e.g. FOPTD approxima-
tion and some tuning rule. In particular in the DDU case, higher
order responses can clearly be seen in the step responses. Only 1

RGA elements magnitude

if one has an accurate model and calculates robust tuning
parameters, SISO tuning will give satisfying results in an MIMO
environment, and only if the RGA is not poor. How much robust- 0.8
ness is actually needed is not known before a multivariable simula-
tion is done, and the controller settings are re-optimized based
on that. If SISO tuning is made ‘‘too robust’’, loop response will 0.6
be too sluggish. Multi-loop tuning results can be used with more
confidence than settings calculated using a single-loop tuning
method. Time saving is an additional advantage. For example, in 0.4
the DDU case identification modeling and controller tuning was
completed in 2 h.
0.2

4. Conclusions
0
The following conclusions can be made. A MIMO model-based 10-4 10-3 10-2 10-1 100
approach can be used to successfully tune multiple PID loops that Frequency (rad/s)
interact strongly. If the open loop model is moderately accurate,
then ‘‘one-shot’’ tuning is achievable and the simulated and Fig. A1. Magnitude of the RGA elements for the GTU application.
observed OP and PV responses will be almost identical. The use
of sufficiently large gain and dead time robustness margins
ensures that the loop will remain stable and well damped even Table A1
if the process is nonlinear. This also helps protect against Multivariable RHP zeros of the GTU application.
inaccurate model identification results. The ability to impose
hard constraints on damping ratio, maximum PV overshoot, and RHP zero z1 ¼ 0:0464 z2 ¼ 0:0554
the maximum OP value ensures that the final design is safe from a    
Input direction 0:6837 0:063
process point of view. PID tuning rules cannot achieve this. uz ¼
0:7297
uz ¼
0:998
Meanwhile, the ‘‘AptituneTM’’ tool has been used in more than Output direction

0:7605
 
0:1774

yz ¼ yz ¼
30 industrial APC projects, where in most cases (2  2) and (3  3) 0:6494 0:9841
problems have been solved. Not only control engineers, but also
technicians can use ‘‘AptituneTM’’ given that some training was
provided. The development of the multivariable process model
should be left to an advanced process control engineer. closed-loop performance is limited by two MIMO RHP zeros,
which have been calculated from a reduced 6th order model
(Matlab function balred). They are listed together with their input
5. Appendix A and output directions in Table A1.

Matlab files containing the dynamic process models for

the Gasoline Treatment Unit (GTU) and the Diesel Desulfuri- 5.2. DDU application
zation Unit (DDU) applications are available for download
from the Web, please contact one of the authors for details. The DDU model relates the outputs of TC001, TC002, and
The models are provided as discrete time linear state-space PC001 to the process values of these controllers and consists of
models. 3 inputs, 3 outputs, and 41 states. The sampling time is 30 s. The
steady-state process gain and RGA matrices are
5.1. GTU application 2 3
3:0774 0:8921 2:3627
6 7
K GTU ¼ 4 0:2229 1:1354 0:7275 5, and
The GTU model relates the outputs of TC9854 and PC9876 to
0:4455 0:2518 0:8996
the process values of these controllers and consists of 2 inputs,
2 3
2 outputs and 18 states. The sampling time is 12 s. The steady- 0:7205 0:0214 0:2581
state process gain and RGA matrices are 6 7
RGAGTU ¼ 4 0:0090 0:8432 0:1658 5
   
0:3960 1:7255 0:8854 0:1146 0:2885 0:1354 0:5762
K GTU ¼ and RGAGTU ¼
0:0585 1:9713 0:1146 0:8854
In this case, pairing decision is obvious from technological
The steady-state RGA analysis supports the unusual pairing, considerations, but is also reflected in the diagonal steady-state
which was mentioned earlier in the text and justified there using RGA values. Fig. A2 presents the dynamic RGAs of the DDU
process-safety related arguments. In Fig. A1, the RGA element application.
magnitudes are presented as functions of frequency. Based on the dynamic RGA analysis, interaction is medium over a
Over a wide frequency range, interaction is weak, but it is wide frequency region. The closed-loop performance is limited by
higher in the region of the closed-loop system bandwidth. The two MIMO RHP zeros, which are listed together with their respective
 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370 369

1 0.03 0.4

0.9 0.02

λ11

λ12

λ13
0.2
0.8 0.01

0.7 0 0
10-4 10-3 10-2 10-4 10-3 10-2 10-4 10-3 10-2
0.03 0.86 0.18

0.02 0.17
λ21

λ22

λ23
0.84
0.01 0.16

0 0.82 0.15
10-4 10-3 10-2 10-4 10-3 10-2 10-4 10-3 10-2
0.4 0.2 1

0.8
λ31

λ32

λ33
0.2 0.15
0.6

0 0.1 0.4
10-4 10-3 10-2 10-4 10-3 10-2 10-4 10-3 10-2
Frequency (rad/s)

Fig. A2. Magnitude of the RGA elements for the DDU application.

Table A2 Halevi, Y., Palmor, Z. J., & Efrati, T. (1997). Automated tuning of decentralized PID
Multivariable RHP zeros of the GTU application. controllers for MIMO processes. J. Process Control, 7(2), 119–128.
Ho, B. L., & Kalman, R. E. (1966). Effective construction of linear state-variable
RHP zero z1 ¼ 0:0164 z2 ¼ 0:1258 methods from input-output functions. Regelungstechnik, 14(12), 545–548.
Hovd, M., & Skogestad, S. (1993). Improved independent design of robust
2 3 2 3 decentralized controllers. J. Process Control, 3(1), 43–51.
Input direction 0:9910 0:2583
6 7 6 0:7316 7 Hovd, M., & Skogestad, S. (1994). Sequential design of decentralized controllers.
uz ¼ 4 0:1045 5 uz ¼ 4 5
Automatica, 30(10), 1601–1607.
0:0835 0:6309 Huang, H.-P., Jeng, J. C., Chiang, C.-H., & Pan, W. (2003). A direct method for multi-
2 3 2 3
Output direction 0:9731 0:7210 loop PI/PID controller design. J. Process Control, 13, 769–786.
6 0:0216 7 6 7 Juricek, B., Seborg, D. E., & Larimore, W. E. (2002). Identification of multivariable,
yz ¼ 4 5 yz ¼ 4 0:5307 5
0:2292 0:4456 linear, dynamic models: comparing regression and subspace techniques. Ind.
Eng. Chem. Res., 41, 2185–2203.
Kirillov, A. (2006). Evolution computations in Cþþ. URL /http://www.codeproject.
com/KB/recipes/aforge_genetic.aspxS, accessed on December 2, 2010.
Kristiansson, B., & Lennartson, B. (2006). Robust tuning of PI and PID controllers.
IEEE Control Syst. Mag., 26(1), 55–69.
input/output directions in Table A2. Like in the GTU case, they have
Larimore, W. E. (2000). The ADAPTx software for automated multivariable system
been calculated from a reduced 6th order model. identification. In: Proceedings of the IFAC symposium on system identification,
Santa Barbara, June 21–23, vol. 2 (pp. 693–698).
Lee, J., & Edgar, T. F. (2006). Multiloop PI/PID control system improvement via
References adjusting the dominat pole or the peak amplitude ratio. Chem. Eng. Sci., 61,
1658–1666.
Aström, K., & Hägglund, T. (2006). Advanced PID control. Research Triangle Park: Li, Y., Ang, K. H., & Chong, G. (2006). Patents, software and hardware for PID
ISA. control. IEEE Control Syst. Mag., 26(1), 42–54.
Bao, J., Forbes, J. F., & McLellan, P. J. (1999). Robust multiloop PID controller design: Luyben, W. L. (1986). A simple method for tuning SISO controllers in multivariable
a successive semidefinite programming approach. Ind. Eng. Chem. Res., 38, systems. Ind. Eng. Chem. Proc. Des. Devel., 25. 654 ff.
3407–3419. MacArthur, W. J., & Zhan, C. (2007). A practical global multi-stage method for fully
Campestrini, L., Filho, L. C. S., & Bazanella, A. S. (2009). Tuning of multivariable automated closed-loop identification of industrial processes. J. Process Control,
decentralized controllers through the ultimate point method. IEEE Trans. 17(10), 770–786.
Control Syst. Technol., 17(6), 1270–1281. Monica, T. J., Yu, C.-C., & Luyben, W. L. (1988). Improved multiloop single-input,
Chen, D., & Seborg, D. E. (2003). Design of decentralized PI control systems based single-output (SISO) controllers for multivariable processes. Ind. Eng. Chem.
on Nyquist stability analysis. J. Process Control, 13(1), 27–39. Res., 27, 969–973.
Chien, I.-L., Huang, H.-P., & Yang, J.-C. (1999). A simple multiloop tuning method Pegel, S., Engell, S. (2001). Multivariable PID controller design via approximation of the
for PID controllers with no proportional kick. Ind. Eng. Chem. Res., 38, attainable performance. In Proceedings of the IEEE IECON’01, Denver (pp. 724–730).
1456–1468. Skogestad, S. (2003). Simple analytic rules for model reduction and PID controller
Darby, M., Harmse, M. (2009). Is it time for a new PID tuner? AIChE Spring National tuning. J. Process Control, 13, 291–309.
Meeting, Tampa, April 26–30, 2009. Skogestad, S., & Postlethwaite, I. (2005). Multivariable feedback control—Analysis
Dittmar, R., & Pfeiffer, B.-M. (2006). Industrial application of model predictive and design ((2nd ed). Chichester: John Wiley & Sons.
control (in German). Automatisierungstechnik, 54(12), 590–601. Sourlas, D. D., & Manousiouthakis, V. (1995). Best achievable decentralized
Espinosa Oviedo, J. J., Boelen, T., & van Overschee, P. (2006). Robust advanced PID performance. IEEE Trans. Automatic Control, 40(11), 1858–1871.
control. IEEE Control Syst. Mag., 26(1), 15–19. Sumana, C., & Venkateswarlu, C. (2010). Genetically tuned decentralized PI
Farag, A., & Werner, H. (2006). Structure selection and tuning of multi-variable PID controllers for composition control of reactive distillation. Ind. Eng. Chem.
controllers for an industrial benchmark problem. IEE Proc. Control Theory Appl., Res., 49(3), 1297–1311.
153(3), 262–267. Trierweiler, J. O., Müller, R., & Engell, S. (2000). Multivariable low order structured-
Forssell, U., & Ljung, L. (2000). A projection method for closed-loop identification. controller design by frequency response approximation. Braz. J. Chem. Eng.,
IEEE Trans. Automatic Control, 45(11), 2101–2106. 17(4–7), 793–807.
370 R. Dittmar et al. / Control Engineering Practice 20 (2012) 355–370
Van Overschee, P., de Moor, B. (2000). RAPID: the end of heuristic PID tuning. In
Proceedings of the IFAC workshop on past, present and future of PID control.
Terassa, Spain.
Vlachos, C., Williams, D., & Gomm, J. B. (2000). Genetic approach to decentralized
PI controller tuning for multivariable processes. IEE Proc. Control Theory Appl.,
146(1), 58–64.
Xiong, Q., Cai, W.-J., He, M.-J., & He, M. (2006). Decentralized control system design
for multivariable processes—A novel method based on relative gain array. Ind.
Eng. Chem. Res., 45, 2769–2776.
Yu, C. C. (2007). Autotuning of PID controllers—A relay feedback approach. London:
Springer.
Zhu, Y. (2004). Robust PID tuning using closed-loop identification. Preprints of the
International Symposium on Advanced Control of Chemical Processes ADCHEM
2004, 1, 165–170.
Zhu, Y. (2006). System identification for process control: recent experience
and outlook. In Proceedings of the IFAC symposium on system identification.
Newcastle, Australia.