Modeling, Identification and Control, Vol. 38, No. 4, 2017, pp.

185–197, ISSN 1890–1328

PD/PID controller tuning based on model

approximations: Model reduction of some unstable
and higher order nonlinear models
Christer Dalen, David Di Ruscio

University College of Southeast Norway, P.O. Box 203, N-3901 Porsgrunn, Norway. E-mail:
[email protected]

Abstract

A model reduction technique based on optimization theory is presented, where a possible higher order
system/model is approximated with an unstable DIPTD model by using only step response data. The
DIPTD model is used to tune PD/PID controllers for the underlying possible higher order system. Nu-
merous examples are used to illustrate the theory, i.e. both linear and nonlinear models. The Pareto
Optimal controller is used as a reference controller.

Keywords: PD and PID controllers, tuning, double integrating system, time delay, maximum time delay
error, relative time delay margin, robustness, performance, Pareto Optimal

1 Introduction to develop the simplified (D)IPTD (IPTD or DIPTD)

model between a control input, u, and output variable,
y = hp (s)u, where
Many unstable plants and higher order systems may
be controlled sufficiently well with PD/PID feedback e−τ s
hp (s) = K , (1)
controllers (p. 162, Silva et al. (2005)). One way of s
designing such control systems is to use a detailed first- e−τ s
principles model of the higher order and possible un- hp (s) = K 2 , (2)
s
stable system. The resulting models often contain a where, K, is the gain- velocity or acceleration, i.o. and,
large number of states, integrators and also double in- τ , is the time delay.
tegrators. This is at least the case in Mariner vessel Hence, the approximating model parameters, K and
models. Based on this higher order model there is for τ , are simply obtained by minimizing the difference be-
the moment no simple way to design a PID controller. tween the output of a step response of the higher order,
In this paper, it is proposed to approximate a Single possibly unstable and nonlinear model or system, and
Input Single Output (SISO) part of the higher order, a step response of the simplified (D)IPTD model ap-
possibly unstable and nonlinear Multiple Input Multi- proximation in Eqs. (1) or (2), i.o. Notice, that the
ple Output (MIMO) system, with an Integrating Plus proposed method is a part of Prediction Error Methods
Time Delay (IPTD) model or Double IPTD (DIPTD) (Ljung (1999)). Real observations of the underlying
model. Initially, only PID controller tuning was consid- real system may also be used. Hence, instead of us-
ered, but for a shot at completion, we will also include ing a first-principles model, real process input-output
PI controller tuning. We propose to use a simple step data/observations of the plant are used.
response of the underlying higher order SISO model or By numerical investigations such approximations
system in connection with optimization theory in order may be deduced from a short time interval, t > τ , and

doi:10.4173/mic.2017.4.3 c 2017 Norwegian Society of Automatic Control

 Modeling, Identification and Control

that the (D)IPTD model approximation is sufficient to 2.2 Definitions

design a PI or PD/PID controller used to control and
Consider the following PID controller on ideal/parallel
stabilize the system. The recently published δ-tuning
form
method (Di Ruscio and Dalen (2017)) is used in or-  
der to design the PD/PID controllers, and the earlier 1
hc (s) = Kp 1 + + Td s , (5)
publication Di Ruscio (2010) for PI controllers. Ti s
The contributions in this paper may be itemized as where, Kp , Ti and Td are the proportional constant,
follows: integral time and derivative time, i.o.
In order to compare the different controllers against
• A model reduction technique to approximate each other we will consider indices such as defined in
higher order, possible unstable and nonlinear sys- Åström and Hägglund (1995), Seborg et al. (1989) and
tems, with (D)IPTD models based on optimiza- Skogestad (2003).
tion over a possible short time interval is proposed For measuring performance in a feedback system as
and presented in Sec. 2. in Figure 1, the Integral Absolute Error (IAE) is de-
fined in the following, as
• The simplified (D)IPTD model is used to design Z ∞
adequate PI or PD/PID controllers for sufficiently IAE = |r − y|dt. (6)
controlling and stabilizing the input-output be- 0
havior of the real plant. Furthermore, from Eq. (6) we define:
• IAEr evaluates the performance in case of a step
• Numerous examples, both unstable and higher or-
response in the reference, no disturbance.
der systems, are used to illustrate the proposed
design strategy. The examples are presented in • IAEvu evaluates the performance in case of a
Sec. 3. step input disturbance, with the reference equal
to zero.
All numerical calculations and plotting facilities are
provided by using the MATLAB software, MATLAB • IAEvy evaluates the performance in case of a step
(2016). The rest of this paper is organized as follows. output disturbance, with the reference equal to
In Sec. 2 we propose definitions and the model re- zero. v
duction technique for a (D)IPTD plant. Simulation Plant
examples are presented in Sec. 3. Lastly, discussion H ?
and concluding remarks are given in Sec. 4.
HH
hv (s)

2 Theory r - i - hc (s) u -
hp (s) - i
? y-
−6
2.1 Underlying Model
Consider a continuous time SISO nonlinear state space
model describing the dynamical system, viz. Figure 1: Control feedback system. Plant model,
hp (s), disturbance model, hv (s), and con-
ẋ = f (x, u), (3) troller, hc (s), in Eq. (5). Disturbance, v,
y = g(x), (4) at the input when, hv (s) = hp (s), and at the
output when, hv (s) = 1.
where, x ∈ Rn , is the state vector, u ∈ R, is the control
signal, y ∈ R, is the output vector, and the vector Robustness is quantified as in Garpinger and
functions, x(t0 = 0), is the initial state, f (x, u) ∈ Rn , Hägglund (2014), i.e., Mst = max{Ms , Mt }, where
and, g(x) ∈ R, are assumed Lipschitz continuous. 1 h0 (jω)
In this paper, noise is not considered, i.e. determin- Ms = max | |, Mt = max | |, (7)
ω 1 + h0 (jω) ω 1 + h0 (jω)
istic systems or models are assumed. A case with noise
is suggested to be handled with a proper system iden- where, h0 (s) = hp (s)hc (s), is the loop transfer func-
tification method, e.g. Ljung (1999), Di Ruscio (1996) tion.
and Di Ruscio (2009), thereafter, possible (D)IPTD Input usage is defined as in (Skogestad et al. (2002)),
model approximations may be done for PI or PD/PID hp (jω)
controller tuning. Mks = max | |. (8)
ω 1 + h0 (jω)

186
 Christer Dalen, “PD/PID controller tuning based on model approximations’

2.3 Model Reduction Techniques 0.8

The half rule for model reduction, proposed by Sko- 0.7

gestad (2003) can not be used to approximate unsta- tangent

ble systems. It is proposed to use optimization the- 0.6

ory to approximate a class/some higher order, possible 0.5

unstable and nonlinear models as a (D)IPTD model,
0.4

y
Eq. (1) or (2). Hence, an optimization method is Ex. 3.1, Input step response

used to find the two unknown parameters, i.e. the 0.3

velocity/acceleration gain, K, and the time delay, τ .
However, it was observed in numerical investigations 0.2

that introducing the final time, tf , as a third unknown 0.1 lag, L

would give successful results.
Defining, ρ = [K, τ, tf ], in an optimal mean square 0
0 0.5 1 1.5 2
sense we have that Time
(
Aρ ≤ b
ρ̂ = arg min V (ρ) s.t. , (9) Figure 2: Illustration of the Ziegler’s lag.
ρ lb ≤ ρ ≤ ub

where
Algorithm 2.1 (Simple δ-PI or PD/PID tuning)
1
V (ρ) = (Y (tf ) − Yb (ρ))T G (Y (tf ) − Yb (ρ)),
N (tf )
(10) 1. Gather time series from open loop unit step re-
sponse.
where, Y (tf ) ∈ RN , and, Yb (ρ) ∈ RN , are the in-
put step response time-series for the system/model 2. Obtain the (D)IPTD model approximation by solv-
and the (D)IPTD model, i.o., structured as vectors. ing the optimization problem given in Eq. (9).
N=length(0:h:tf ), i.e. in MATLAB notation, is the
number of samples and, h, is the sampling interval.
3. Use the δ-tuning method for PI or PD/PID con-
Furthermore, in this paper, it is suggested to use the
troller tuning, i.e. Algorithm 6.1 in Di Ruscio
following
(2010) or Algorithm 2.1 (include Eq. (27) for

1 0 0
 
c1
 PID) p. 98 in Di Ruscio and Dalen (2017) for
A = 0 −1 0  , b =  − 2  , h
(11) (D)IPTD i.o.
0 1 −1 −2h
   
c1 c2 2.4 The Pareto Optimal PID Controller
lb =  h2  , ub =  L2  , G = I, (12)
L
2h 2 For quantifying multiple performances, i.e. indices
IAEvu and IAEvy , consider the Pareto Optimal (PO)
where the upper bound on the time delay, τ , and final criterion (PO concept is presented in p. 60 Pareto
time, tf , corresponds to the pre-act time/derivative ac- (1894b))
tion (p. 190, Seborg et al. (1989)) setting, i.e. L2 (p.
764, Ziegler and Nichols (1942)). IAEvy (p) IAEvu (p)
L is, in this paper, defined as Ziegler’s lag. See illus- J(p) = sr o
+ (1 − sr ) o
, (13)
IAEvy IAEvu
tration in Figure 2. In some cases, where Ziegler’s lag
is not possible to identify, we will instead prescribe, tf ,
where, p = [Kp , Ti , Td ], and, sr = 0.5, is the servo-
i.e. we obtain a two-dimensional optimization problem.
regulator parameter, however it was suggested that,
Note that, in Eq. (12), G, is the identity matrix. No-
0 ≤ sr ≤ 1 (p. 10, Di Ruscio (2012)).
tice, that the lower bound on the time delay τ is chosen
according to Shannon’s sampling theorem (Kotelnikov The essential problem in generating the PO J vs.
(1933), Shannon (1949)). The bounds for the gain K, M st vs. Mks trade-off surface, i.e. finding the PO con-

i.e. c1 < c2 , need to be assigned from trial and error. trollers, renders in solving the following optimization
The optimization problem is solved by using problem given by
fmincon (MATLAB (2016)) with default options, i.e.
‘interior-point’ algorithm. p̂ = arg min J(p) s.t. ceq (p) = 0, (14)
p

187
 Modeling, Identification and Control

where is given in row 1 in Table 4. Interestingly, we ob-

tain, θ̂ = [2.12, 2.12], i.e. cf. Di Ruscio and Dalen
ceq = [cs ; ct ], (15) (2017) where, θ̂∗ = [2.24, 2.24] (* means the JP O in
pre
cs = Mst − Mst , (16) Eq. (18) is substituted by δ-optimal J curve, i.e. the
pre optimization problem is solved per Ms instead of the
ct = Mks − Mks , (17)
full range). Hence, considering the reference example,
pre pre
where, Mst , is the prescribed robustness and, Mks ∈ this change in settings can be thought of as rotating
R, is the prescribed input usage. Notice, that this op- the δ-tuning trade-off curve counter-clockwise around
timization problem is the same as considered in Ja- the point (1.54, 1.56), see Figure 3.
hanshahi et al. (2014), i.e. if we were to use Integral
Squared Error instead of IAE. 1.6
Consider a two-dimensional optimization problem 3=[2.12,2,12]
3=[2.24,2.24]
for stable processes, i.e. we generate a PO J vs. Ms 1.55 PO
trade-off curve, which is a similar problem as found in (1.54,1.46)
Balchen (1958) (i.e. if, sr = 0, in Eq. (13)). The in- 1.5
dex presented in Balchen (1958) can be seen common

Performance J
to IAE, however it is based on frequency response of
1.45
the closed loop system. Furthermore, for stable pro-
cesses, we will consider the range, 1.3 ≤ Ms ≤ 2.0, i.e.
1.4
the robust range suggested in Åström and Hägglund
(1995). Note, that we will assume, Ms > Mt , for sta-
ble and, Mt > Ms , unstable processes, which is usually 1.35

true (Skogestad and Postlethwaite (1996), Jahanshahi

et al. (2014)). 1.3
1.5 1.52 1.54 1.56 1.58 1.6
It might be beneficial to measure how close to op- Robustness M
s
timal, i.e. PO, a given tuning method is performing,
hence the following mean square error criterion is in- Figure 3: Reference Example: Performance vs. ro-
troduced bustness, J vs. Ms trade-off curves. Com-
1 paring settings, θ̂ = [2.12, 2.12], and θ̂∗ =
Γm (θ) = (JP O − Jm (θ))T (JP O − Jm (θ)), (18) [2.24, 2.24] vs. the PO curve.
M
where, Jm (θ) ∈ RM , is the performance vector gener-
ated from a given tuning method, m, with correspond-
ing θ (i.e. generally free-parameter vector), JP O ∈ RM , Table 1: Shows the reference controllers which mini-
is the PO-PID performance vector previously solved, mize the indices, IAEvy , and, IAEvu , i.e.
M = length(Jm ). E.g. for δ-tuning, i.e. step 3. in optimal output and input disturbance (ideal)
Alg. 2.1, we may have that, θ = [K, τ, c, γ], however, PD/PID controllers as in Eq. (5), i.o., given
θ, may be fixed as well. This is exercised in the coming a prescribed, Ms = 1.59. DIPTD process
sections and the numerical examples. where, K = 1, τ = 1. The weights, IAEvy o
=
Lastly, we define an optimization problem o
IAEvu = 1, in Eq. (13).
sr p̂ JAlg.2.1 (p̂)
θ̂ = arg min Γm (θ), (19)
θ 1 0.02, ∞, 25.84 4.15
0 0.07, 9.15, 5.41 169.09
where, Γm , is given in Eq. (18).

2.5 Extending settings for δ-tuning

We propose some results for a possible extension of 3 Numerical Examples
Sec 4.4 in Di Ruscio and Dalen (2017), i.e. choosing
δ-tuning PID parameters c and γ. We seek to justify the proposed method, Algorithm
Consider a DIPTD process model, where K = 1 and (Alg.) 2.1, in the coming sections by using a good va-
τ = 1, i.e. θ = [c, γ]. The reference controllers, i.e. riety of models with stable and unstable nature. We
the optimal output PD controller and the optimal in- will first demonstrate the proposed method on linear
put disturbance PID controller are given in Table 1. process examples, thereafter we consider unstable non-
The solution to the optimization problem in Eq. (19) linear models.

188
 Christer Dalen, “PD/PID controller tuning based on model approximations’

Table 2: Settings for the δ-tuning PID parameters c

T =3.39 =
and γ in Alg. 2.1 and Eq. (27) in Di Ruscio c
/1=3.74 Alg. 2.1
and Dalen (2017). row 1) optimized over the 3 6=1.90 =
SIMC
K-SIMC
robust region and 2) over the entire region. PO-PID

The solution of the optimization problem in

Eq. (19). 2.5

Performance, J
Option θ̂ ΓAlg.2.1 (θ̂) Region
6=0.95 =
1 2.12, 2.12 1.8323E-4 1.3 ≤ Ms ≤ 2.0 2
2 2.24, 2.24 7.8248E-4 1.3 ≤ Ms ≤ 2.0 6=0.62 =

T =1.51 = T c=0.73 =
c
1.5

3.1 Linear Examples

/ =2.00
1 1
We start by studying some of the linear process model /1=1.25
examples given in Seborg et al. (2004) or Skogestad 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
(2003). We will compare Alg. 2.1 vs. some exist- Robustness, M
s
ing model-based PID tuning methods, i.e. the Sim-
ple Internal Model Control (SIMC) (Skogestad (2003)) Figure 4: Example 3.1: Performance vs. robustness,
and the Korean-SIMC (K-SIMC) method (Lee et al. J vs. Ms trade-off curves. Comparing Alg.
(2014)). 2.1, SIMC and K-SIMC vs. the PO-PID con-
trollers for the process model at row 1, col. 2
Example 3.1 (Fourth-Order Model) in Table 3. The weights for the PO criterion
Consider the fourth-order process model studied in ex- in Eq. (13) are given in Table 5. The perfor-
ercise 12.7 on p. 331 in Seborg et al. (2004) and Exam- mance measures, i.e. Γm , Eq. (18), are given
ple E5 in Skogestad (2003), i.e. the transfer function in Table 6.
in row 1 and column (col.) 2 in Table 3.
The Second Order Plus Time Delay (SOPTD) model
approximations (sampling interval h = 0.001) for the
SIMC and K-SIMC methods are given in rows 1:2, and T =2.55 =
c
Alg. 2.1
col. 3 in Table 6, i.o. The DIPTD model approxima- 6=1.14 =
SIMC
3 /1=4.27
tion, i.e. step 2 in Alg. 2.1, is given in row 1 and col. K-SIMC
PO-PID
2 in Table 5, where c1 = 1 and c2 = 10 are chosen as
bounds for the gain. 2.5
Performance, J

In this example, Alg. 2.1 is shown superior to the

SIMC and K-SIMC method, viz. rows 1:3 and col. 4 in 2 6=0.61 =
Table 6 shows that that Alg. 2.1 is ΓΓAlg.2.1
SIM C
= 4.0 times 6=0.42 =
T =0.82 =
better than the runner up, SIMC. The corresponding J 1.5
c
T c=0.24 =
vs. Ms trade-off curves are shown in Figure 4. / =2.38
1

/1=1.55
1
Example 3.2 (Bioseperation Process)
A multistage bioseperation process, as motivated in ex-
ercise 7.4 on p. 183 in Seborg et al. (2004), may be 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2

described by the transfer function in Table 3 (row 2, Robustness, M

s

col. 2).
The SOPTD model approximations (h = 0.01) for Figure 5: Example 3.2: Performance vs. robustness,
the SIMC and K-SIMC methods are given in Table 6 J vs. Ms trade-off curves. Comparing Alg.
(rows 4:5, col. 3). The DIPTD model approximation 2.1, SIMC and K-SIMC vs. the PO-PID con-
(c1 = 0.01, c2 = 1) is given in Table 5 (row 2 and col. trollers for the process model at row 2, col.
2). 2 in Table 3. The servo-regulator, sr = 0.5,
Alg. 2.1 is observed to be superior compared to the and the weights for the PO criterion in Eq.
other model based methods, i.e. in Table 6 (rows 4:6, (13) are given in Table 5. The method per-
ΓSIM C
col. 4) it is seen that Alg. 2.1 is ΓAlg.2.1 = 8.5 times formance measures, i.e. Γm in Eq. (18), are
given in Table 6.
better than the runner up, SIMC. The J vs. Ms trade-
off curves are shown in Figure 5.

189
 Modeling, Identification and Control

Example 3.3 (Positive Numerator) times better than the runner up, K-SIMC. The J vs.
Consider the process model with a positive numerator Ms trade-off curves are shown in Figure 7.
as studied in Example 4 in Åström et al. (1998) and
Example E3 Skogestad (2003), i.e. the transfer func- 2.4 T c=1.81 =
Alg. 2.1
tion given in Table 3 (row 3, col. 2). 6=1.53 = SIMC
K-SIMC
The SOPTD model approximations (h = 0.01) for 2.2
PO-PID
the SIMC and K-SIMC method are shown in Table 6
2
(rows 7:8, col. 3) The DIPTD model approximation

Performance, J
(c1 = 1, c2 = 10) is given in Table 5 (row 3, col. 2). 1.8 T c=0.51 =
T c=-0.01 =
Notice, that the gain, K, in the K-SIMC method is
1.6 / =0.55
1
a function of the tuning parameter, i.e., K = f (λ), 6=0.32 =
where, λ, is similar to the tuning parameter, Tc , in 1.4 6=0.10 =
SIMC. /1=0.17
1.2 /1=0.01
Alg. 2.1 gives the best performance, viz. from Ta-
ble 6 (rows 7:9, col. 4) we have that Alg. 2.1 is 1
ΓSIM C
ΓAlg.2.1 = 2.9 times better than the runner up, SIMC.
0.8
The corresponding J vs. Ms trade-off curves are shown 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Robustness, M
in Figure 6. s

/1=6.98
Figure 7: Example 3.4: Performance vs. robustness,
4
T c=2.10 = Alg. 2.1 J vs. Ms trade-off curves. Comparing Alg.
SIMC
3.5
6=2.23 = K-SIMC 2.1, SIMC and K-SIMC vs. the PO-PID con-
PO-PID
trollers for the process model at row 4, col.
3 2 in Table 3. The servo-regulator, sr = 0.5,
and the weights for the PO criterion in Eq.
Performance, J

2.5 6=1.12 = (13) are given in Table 5. The method per-

6=0.71 =
formance measures, i.e. Γm in Eq. (18), are
T =0.74 =
2 c given in Table 6.
T =0.17 =
c
1.5
Example 3.5 (Perfectly Stirred Tank)
1 /1=2.39 This example is taken from Exercise 14.3. on p. 334 in
/1=3.77
Seborg et al. (1989). A perfectly stirred tank heating a
1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 flowing liquid is described by the process model in Table
Robustness, M 3 (row 5, col. 2), i.e. a transfer function from power
s

applied to the heater to the measured temperature.

Figure 6: Example 3.3: Performance vs. robustness, The SOPTD model approximations (h = 0.01) for
J vs. Ms trade-off curves. Comparing Alg. SIMC and K-SIMC methods are given in Table 6 (rows
2.1, SIMC and K-SIMC vs. the PO-PID con- 13:14, col. 3), i.o. The DIPTD model approximation
trollers for the process model at row 3, col. is shown in (c1 = 1, c2 = 10) Table 5 (row 5, col. 2).
2 in Table 3. The servo-regulator, sr = 0.5, Alg. 2.1 has an edge over the other model based tech-
and the weights for the PO criterion in Eq. niques, i.e. in Table 6 (rows 13:15, col. 4) observe that
(13) are given in Table 5. The method per- Alg. 2.1 is ΓΓAlg.2.1
SIM C
= 2.4 times better than the next best,
formance measures, i.e. Γm in Eq. (18), are SIMC. The corresponding J vs. Ms trade-off curves are
given in Table 6. shown in Figure 8.
Example 3.6 (Pipeline-Riser System)
Example 3.4 (Quadruple Poles) The severe-slugging flow regime is a common problem
Consider the transfer function with quadruple poles in at the oil fields and it is characterized by large oscil-
Table 3 (row 4, col. 2). lations in pressure and flow rates. Active control of
The SOPTD model approximations (h = 0.01) for the topside choke is the recommended solution (Yocum
SIMC and K-SIMC methods are given in Table 6 (rows (1973) and Schmidt Z. (1979)).
10:11, col. 3). The DIPTD model approximation (c1 = A schematic model describing a pipeline-riser system
1, c2 = 10) is given in Table 5 (row 4, col. 2). is given in Figure 9. The inflow rates of liquid and gas
Alg. 2.1 is shown superior, i.e. from Table 6 (rows to the system, wl and wg , i.o., are assumed to be in-
10:12, col. 4) we have that Alg. 2.1 is ΓΓK−SIM
Alg.2.1
C
= 13.0 dependent disturbances. Based on experimental data,

190
 Christer Dalen, “PD/PID controller tuning based on model approximations’

4.5 / =8.95 better than IMC-PIDF, however it might be argued that

1
T =3.43 = Alg. 2.1
c
SIMC
a better criterion may exist.
4 6=1.71 =
K-SIMC Converting the Alg. 2.1 (PIDF) controller in row
PO-PID
3.5 6, col. 2 in Table 8, where τ = 6.66, to Kc =
K
Kp = −11.87, Ki = Tip = −1.38 and Kd = Kp Td =
Performance, J

3 −152.77, we see that these are approximately equal to

6=0.86 = the controller proposed in Table 1 row 1 in Jahanshahi
2.5
T c=1.42 = et al. (2014).
6=0.57 =
Note, that we propose to use PO-PIDF controller for
2
prescribed robustness Mt = 1.4 instead of Mt = 1.15,
1.5 T =0.60 =
c
as in Jahanshahi et al. (2014). See row 6, col. 3 in
/1=4.82 / =3.06
Table 8.
1
1
u
1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Robustness, M
s
w
Figure 8: Example 3.5: Performance vs. robustness,
J vs. Ms trade-off curves. Comparing Alg. y
2.1, SIMC and K-SIMC vs. the PO-PID con-
trollers for the process model at row 5, col. wg
2 in Table 3. The servo-regulator, sr = 0.5, wl
and the weights for the PO criterion in Eq.
(13) are given in Table 5. The method per-
formance measures, i.e. Γm in Eq. (18), are
given in Table 6. Figure 9: Example 3.6. Schematic model of the
pipeline-riser, viz. the plant in the feedback
system in Figure 1.
it was shown in Jahanshahi and Skogestad (2013) that
such a pipeline-riser system can be described by the pro-
cess model in Table 3 (row 6, col. 2), i.e. a transfer
function from the topside choke, u, to inlet pressure, y.
The DIPTD model approximation, i.e. step 2 in Alg. Table 3: Examples and the corresponding linear pro-
2.1, where c1 = −10 and c2 = −0.001, are given in cess models for Sec. 3.1.
Table 5 (row 6, col. 2). The IMC PID plus derivative Example Process model, hp (s)
Filter (IMC-PIDF) method in Jahanshahi et al. (2014) 3.1 1
(s+1)(0.2s+1)(0.04s+1)(0.008s+1)
is based on the exact model, in Figure 9.
2
Note, that for this application it is crucial for the dy- 3.2 (5s+1)(3s+1)(s+1)
namics to have a low-pass filter on the derivative action
2(15s+1)
(Jahanshahi et al. (2014)), hence we have the following 3.3 (20s+1)(s+1)(0.1s+1)2
PIDF controller
1
3.4 (s+1)4
1 Td s
hc (s) = Kp (1 + + ), (20) 10
Ti s Tf s + 1 3.5 (s+1)(5s+1)(0.2s+1)

−0.0098(s+0.25)
where Tf = 4, cf. the numerator in the process model. 3.6 s2 −0.04s+0.025
This filter is also added to Alg. 2.1, hence Alg. 2.1
(PIDF) is written in instead.
The trade-off PO surface together with the curves for
Alg. 2.1 (PIDF) and IMC-PIDF are shown in Figure 3.2 Nonlinear Examples
10. See also Figure 11 for equivalent sets of 2D plots. This section contains results on PID controller imple-
Note, that for making Alg. 2.1 (PIDF) competitive mentation on a good variety of nonlinear models with
we have used, c = γ = 4.24, i.e. 2 times our suggested unstable nature. Note, that we are only given open
settings for this case. loop data, hence model reduction techniques as the
Alg. 2.1 (PIDF) was found superior in terms of Γm , half rule method (Skogestad (2003)) can not be used,
viz. Alg. 2.1 (PIDF) is ΓΓAlg.2.1(P
IM C−P IDF
IDF )
= 18.1 times however we will instead use Alg. 2.1 where step 3. is

191
 Modeling, Identification and Control

4 15
Alg. 2.1 (PIDF) IMC-PIDF
3 PO-PIDF PO-PIDF
10
2

J
5
=c=6.67 1

0 0
50 60 70 80 90 100 50 60 70 80 90 100
IMC-PIDF /1=3.16 Mks Mks
Alg. 2.1 (PIDF)
4 15
10 IMC-PIDF
Performance, J

3 PO-PIDF
10
2

J
5 5
1 Alg. 2.1 (PIDF)
60 PO-PIDF
0 0
1.4 1.5 1.6 1.7 1.8 1.15 1.2 1.25 1.3
0 =c=3.19
/1=1.51 80 Mt Mt
1.2
1.4
1.6
1.8 100 Input usage, Mks
Robustness, M 2
t
Figure 11: Example 3.6: J vs. Mks and J vs. Mt
trade-off curve plots. Comparing Alg. 2.1,
Figure 10: Example 3.6: Performance vs. robustness
IMC-PIDF vs. the PO-PID controllers for
vs. input usage, J vs. Mt vs. Mks trade-off
the process model at row 6, col. 2 in Table
surface. Comparing Alg. 2.1, IMC-PIDF
3. col. 1 is Alg. 2.1 (PIDF) and col. 2 is
vs. the PO-PID controllers for the process
IMC-PIDF.
model at row 6, col. 2 in Table 3. The
servo-regulator, sr = 0.5, and the weights
for the PO criterion in Eq. (13) are given
in Table 5.
Table 6: Summarised results from Sec. 3.1. Compar-
ing SIMC, K-SIMC and the proposed Alg.
Table 4: Examples and the corresponding reference 2.1, based on model reduction to SOPTD
controllers, i.e. the optimal input and out- (K, τ, T1 , T2 ) or DIPTD (K, τ ) models. †
put performances for prescribed Mst = 1.59 means that the model approximation is given
or input usage Mks = 100, (i.e. only ‘and’ for in col. 2 in Table 5 for the corresponding
col. 6). example.
Example o
IAEvy o
IAEvu Ex Method Approximation Γm
3.1 0.0847 0.0133 3.1 SIMC 1, 0.0285, 1, 0.2200 0.0628
3.2 1.2877 0.7215 K-SIMC 1, 0.0446, 1, 0.2040 0.0966
3.3 0.1074 0.0250 Alg. 2.1 † 0.0156
3.4 1.9054 1.5851 3.2 SIMC 2, 0.5050, 5, 3.5 0.1899
3.5 0.3020 0.2740 K-SIMC 2, 0.8471, 5, 3.1667 0.4947
3.6 2.0855 0.0380 Alg. 2.1 † 0.0224
3.3 SIMC 1.5, 0.05, 1, 0.15 0.4255
K-SIMC f (λ), 0.05, 1, 0.15 0.5142
Table 5: Results from Sec. 3.1. Step 2 in Alg. 2.1, i.e. Alg. 2.1 † 0.1489
3.4 SIMC 1, 1.5050, 1, 1.5 0.4000
optimal DIPTD model approximation.
K-SIMC 1, 1.6717, 1, 1.5 0.2068
Example ρ̂ V (ρ̂) Alg. 2.1 † 0.0159
3.1 3.9559, 0.0249, 0.0578 1.3515E-8 3.5 SIMC 10, 0.1050, 5, 1.1000 0.2165
3.2 0.0907, 0.3061, 0.6928 2.4642E-7 K-SIMC 10, 0.1854, 5, 1.0200 0.5637
3.3 4.8212, 0.0193, 0.0530 4.0529E-8 Alg. 2.1 † 0.0885
3.4 0.9164, 0.4900, 0.6079 6.4056E-7 3.6 Alg. 2.1 (PIDF) † 0.0187
3.5 1.0205, 0.0602, 0.3295 1.0444E-7 IMC-PIDF Exact model 0.3392
3.6 -0.0934, 0.1181, 0.5243 1.4997E-6

192
 Christer Dalen, “PD/PID controller tuning based on model approximations’

where,
Table 7: Results from Sec. 3.1. Margins organized as
(GM, P M, DM ) for Alg. 2.1 and runner-up, (m1 + m2 ) lg
b = . (25)
i.e. SIMC, with exception of row 4, K-SIMC, Jt + (m1 + m2 ) lg2
and the optimal controller, PO-PID, for pre-
scribed robustness, Mst = 1.59 or input usage In Eq. (25), the moment of inertia of the rod and the
Mks = 50, (i.e. only ‘and’ for col. 6). distance to center of mass are expressed as
Ex Alg. 2.1 Runner-up PO-PID l2 − 3 l lg + 3 lg2
3.1 8.05, 48.55, 0.06 8.81, 46.67, 0.06 8.18, 47.94, 0.06 Jt = m2 + m1 (l − lg )2 , (26)
3
3.2 ∞, 45.82, 0.79 ∞, 45.49, 0.85 ∞, 46.69, 0.71
3.3 ∞, 47.26, 0.05, ∞, 44.81, 0.07 ∞, 47.11, 0.05
l m1 + 2 m2
lg = , (27)
3.4 4.47, 56.22, 1.68 4.76, 58.88, 2.20 4.32, 65.44, 1.91 2 m1 + m2
3.5 ∞, 45.15, 0.18 ∞, 43.49, 0.21 ∞, 45.28, 0.18
3.6 0.25, 45.97, 1.67 0.11, 75.74, 2.49 0.17, 64.30, 2.11 i.o., where l = 0.7 is the length, m2 = 0.044, is
the mass of the pendulum arm (neglecting the load),
m1 = 0.2, is the mass of the load and g = 9.81 is
Table 8: Results from Sec. 3.1. PID controllers orga- the gravitational constant. A schematic model of the
nized as (Kp , Ti , Td ) for Alg. 2.1 and runner- inverted pendulum on a cart is shown in Figure 12.
up, i.e. SIMC, with exception of row 4, K- Furthermore, we define the input-output case as,
SIMC, and the optimal controller, PO-PID, n
for prescribed robustness, Mst = 1.59 or in- u∈R: = u: force exerted on the cart (N),
put usage Mks = 50, (i.e. only ‘and’ for col. n
y∈R: = y: tilt angle (rad).
6).
Ex Alg. 2.1 Runner-up PO-PID
3.1 24.3, 0.32, 0.15 24.7, 0.51, 0.12 24.5, 0.30, 0.14
3.2 5.54, 4.36, 2.05 5.31, 7.17, 1.79 5.56, 3.40, 2.54 m1
3.3 13.2, 0.39, 0.18 11.0, 0.50, 0.10 13.3, 0.24, 0.17
3.4 1.79, 2.41, 1.14 1.38, 3.06, 0.67 1.82, 2.62, 1.33
3.5 4.29, 1.48, 0.70 4.09, 2.12, 0.53 4.27, 1.00, 0.73
3.6 -30.4, 5.49, 1.30 -11.9, 8.59, 12.9 -21.2, 5.09, 5.45
m2 , l y
substituted with SIMC PID tuning, i.e. denoted Alg.
2.1 (SIMC).
We define Total Value (TV) index formulated in dis-
crete time as
u
∞
X
TV = |∆uk |, (21)
k=1
Figure 12: Example 3.7. Schematic model of the in-
where, ∆uk = uk − uk−1 , is the control rate of change, verted pendulum on a cart, i.e. the plant in
and, k, is discrete time. the feedback system in Figure 1. Note, that
arrows may have different meanings, i.e. la-
Example 3.7 (Inverted pendulum-cart system) bels and vector force u.
In this example, we will consider an inverted pendulum
system, which is one of the most used systems for The optimal DIPTD model approximation is given
benchmarking various control strategies. The reason in Table 9 (col. 2).
being that the structure renders as both relatively
simple and rich (Boubaker (2012)).
It may be shown that the dynamics of an inverted Table 9: Example 3.7. Results from step 2 in Alg. 2.1,
pendulum on a cart with mass can be described using i.e. optimal DIPTD model approximation,
the model in Eqs. (3)-(4), where the RHS is substituted where c1 = 0 and c2 = 10.
by the following, Method p̂ V (p̂)
Alg. 2.1 -1.8252, 0.0139, 0.2423 2.3618E-7
f1 = x2 , (22)
f2 = b(g(1 + m1 + m2 ) sin(x1 ) − u cos(x1 )),(23) Zero tilt angle is desired, at Time = 0.1 s, we give the
π
g = x1 , (24) pendulum an output step vvy = 30 rad (6 deg), and at,

193
 Modeling, Identification and Control

Time = 2.0 s, we introduce an input step disturbance where

vvu = 20 N. The time-series are shown in Figure 13. 
1 0 0 0

From Table 10, we see that Alg. 2.1 outperforms Alg. 0 91 −76 cos(x1 ) 0
2.1 (SIMC) in terms of (sum) IAE. Furthermore, we M =   , (30)
0 −76 cos(x1 ) 124.7 0
Alg.2.1 (SIM C)
IAEvu
see that Alg. 2.1 is = 1.4 times better
Alg.2.1
0 0 0 1
IAEvu
in terms of IAEvu , however Alg. 2.1 (SIMC) is found f1 = x2 , (31)
Alg.2.1
IAEvu
Alg.2.1 (SIM C) = 1.2 times better in terms of IAEvy . f2 = 744.8 sin(x1 ) + 48 u, (32)
IAEvy
f3 = 200 u − 76 x22 sin(x1 ), (33)
y:= tilt angle (rad) f4 = x3 , (34)
0.1 Alg. 2.1
Alg. 2.1 (SIMC) g = x1 , (35)
0.05 reference
y

-0.05

-0.1
0 0.5 1 1.5 2 2.5 3 3.5 vu vy
u:= force exerted on the cart (N)
40

20 r u + + y
+ hc (s) + +
0 −
u

-20

-40
0 0.5 1 1.5 2 2.5 3 3.5
Time (s)

Figure 13: Example 3.7. Simulation of the inverted Figure 14: Example 3.8. Feedback system. Segway,
pendulum nonlinear model with PID con- and PID controller, hc (s). The input, u, is
trol. Output and input step disturbance at, the motor drive voltage, y, is the tilt angle,
Time = 0.1 s and, Time = 2.0 s, i.o. Input vu, is input additive disturbance and, vy, is
u is limited. the output additive disturbance.

We will consider the input-output case defined as,

n
u∈R: = u: motor signal (V),
Table 10: Example 3.7. Shows the indices IAEvy , n
IAEvu , IAE (sum) and T V for Alg. 2.1 y∈R: = y: tilt angle (rad),
where δ = 1.6, and Alg. 2.1 (SIMC) where
Tc = 1.5 τ .
Method IAEvy IAEvu IAE TV Table 11: Example 3.8. Results from step 2 in Alg. 2.1,
Alg. 2.1 0.0182 0.0178 0.0360 184.7389 i.e. optimal DIPTD model approximation,
Alg. 2.1 (SIMC) 0.0158 0.0245 0.0403 171.6880 where c1 = 0 and c2 = 10.
Method p̂ V (p̂)
Alg. 2.1 4.3347, 0.0073, 0.2163 2.6528E-7

Example 3.8 (Segway)

The simulation results are shown in Figure 15, where
A commercial example of a two-wheel system is the Seg- π
at Time = 0.1 s, an output disturbance vvy = 18 rad
way, which has been experiencing a growth of popular-
(10 deg) is introduced, and at Time = 1.0 s, we give
ity as an eco-friendly alternative for short journeys.
an input disturbance vvu = 40 V.
The dynamics of a Segway are presented in LIngenieur
The indices given in Table 12 show similarities with
(2005), which we have put on an implicit state space
previous Example 3.7, viz. Alg. 2.1 gives best re-
form, viz. IAEvuAlg.2.1
sult for IAEvu , i.e. Alg.2.1 (SIM C) = 1.5 times
IAEvu
M (x)ẋ = f (x, u), (28) better than
Alg.2.1
Alg. 2.1 (SIMC) and Alg. 2.1 (SIMC)
IAEvy
y = g(x), (29) Alg.2.1 (SIM C) = 1.1 times better in terms of IAEvy .
IAEvy

194
 Christer Dalen, “PD/PID controller tuning based on model approximations’

y:= tilt angle (rad)

0.2
Alg. 2.1
Alg. 2.1 (SIMC)
0.1
reference
vu vy
y

-0.1
0 0.5 1 1.5 2 r u + + y
+ hc (s) + +
u:= motor signal (V)
−
20

-20
u

-40
Figure 16: Example 3.9. Feedback system. Vessel
-60
0 0.5 1 1.5 model and controller hc (s). The input u is
2
Time (s) the commanded rudder angle and the out-
put y is the perturbed yaw angle about zero.
Figure 15: Example 3.8. Simulation of the Segway vu is input additive disturbance and vy is
nonlinear model with PID control. Output the output additive disturbance.
and. input step disturbance at, Time = 0.1
s and, Time = 1.0 s, i.o. Input u is limited.
Table 13: Example 3.9. Results from step 2 in Alg. 2.1,
i.e. optimal DIPTD model approximation,
Table 12: Example 3.8. Shows the indices IAEvy , where c1 = 0 and c2 = 1.
IAEvu , IAE (sum) and T V for Alg. 2.1
Method p̂ V (p̂)
where δ = 1.6, and Alg. 2.1 (SIMC) where
Alg. 2.1 0.0006, 1.7867, 7.3355 1.6228E-7
Tc = 1.5 τ .
Method IAEvy IAEvu IAE TV
π
Alg. 2.1 0.0143 0.0112 0.0255 245.6765 vvy = 60 rad (3 deg), at Time = 250 s, we introduce
Alg. 2.1 (SIMC) 0.0128 0.0169 0.0297 223.5997 an input step disturbance vvu = 2 π9 rad (40 deg).
We observe from Table 14 that Alg. 2.1 has an
edge over Alg. 2.1 (SIMC) in terms of IAEvu , viz.
Example 3.9 (Nomoto Vessel) IAEvuAlg.2.1 (SIM C)

We will end this section by considering a Mariner ves- IAEvuAlg.2.1 = 1.6 times better. Note, the algo-
sel example. DIPTD model approximations for con- rithms show approximately equal result for IAEvy .
troller tuning for Mariner class vessel was demon-
strated in Di Ruscio and Dalen (2017)) to be a suc-
cessful approach. Table 14: Example 3.9. Shows the indices IAEvy ,
In this example we will consider the nonlinear model IAEvu , IAE (sum) and T V for Alg. 2.1
proposed in Son and Nomoto (1982) which describes where δ = 1.6, and Alg. 2.1 (SIMC) where
the motion of a high-speed container vessel of length Tc = 1.5 τ.
175 m. This model has previously been implemented in Method IAEvy IAEvu IAE TV
the MSS GNC MATLAB Toolbox, Fossen and Perez Alg. 2.1 0.9930 0.3782 1.3712 2.9826
(2004). We will consider the feedback system with Alg. 2.1 (SIMC) 0.9931 0.6140 1.6071 3.0588
added disturbances as illustrated in Figure 16.
Here, we are interested in an input-output case de-
fined as,
n 4 Discussion and Concluding
u∈R: = u: commanded rudder angle (rad), Remarks
n
y∈R: = y: yaw angle (rad). The concluding remarks in this paper may be itemized
as follows:
The simulation results of the PID controller imple-
mentation are shown in Figure 17. The scenario is: at • The bounds, i.e. A, b, lb and ub in Eqs. (11)-(12)
Time = 0 s, the reference yaw angle is zero, at Time in Alg. 2.1 are partly ad-hoc based, it might be
= 50 s, we give the vessel an output disturbance step possible to make improvements on these.

195
 Modeling, Identification and Control

y:= yaw angle (rad)

0.06 • Note, that the derivative kicks are rather undesir-
0.04
Alg. 2.1
Alg. 2.1 (SIMC)
able in practical applications. One possible solu-
reference tion is to use a low-pass filter on the derivative
0.02
y

term, as in Example 3.6. If the process is stable,

0 consider choosing a PI controller instead.
-0.02
0 50 100 150 200 250 300 350 400 • It might be argued that, if the process model is
u:= commanded rudder angle (rad)
known it might be the best solution to simply use
0.2 the PO-PID controller.
0

-0.2
u

-0.4
References
-0.6

-0.8 Åström, K. and Hägglund, T. PID Controllers: The-

0 50 100 150 200 250 300 350 400
Time (s)
ory, Design, and Tuning. Instrument Society of
America, 1995.
Figure 17: Example 3.9. Simulation of the nonlin-
Åström, K. J., Panagopoulos, H., and Hägglund,
ear Nomoto Vessel model with PID control.
T. Design of pi controllers based on non-convex
Output and input step disturbance at, Time
optimization. Automatica, 1998. 34(5):585–601.
= 50 s and, Time = 250 s, i.o. Input u is
doi:10.1016/S0005-1098(98)00011-9.
limited.
Balchen, J. A Performance Index for Feedback Con-
trol Systems Based on the Fourier Transform of the
• In Sec. 3.1 we have demonstrated successful imple- Control Deviation. Acta polytechnica Scandinav-
mentations of the proposed model reduction tech- ica: Mathematics and computing machinery series.
nique Alg. 2.1 on a variety of linear models. We Norges tekniske vitenskapsakademi, 1958.
see from the stable part in Sec. 3.1 that Alg.
2.1 has an edge over the model based techniques Boubaker, O. The inverted pendulum: A fundamen-
SIMC and K-SIMC, viz., considering the optimal- tal benchmark in control theory and robotics. 2012.
ity measures, Γm , in Table 6 (row 1:15, col. 4), the pages 1–6. doi:10.1109/ICEELI.2012.6360606.
proposed Alg. 2.1 is at least ΓΓAlg.2.1
SIM C
= 2.4 times
Di Ruscio, D. Combined Deterministic and Stochas-
better than SIMC and K-SIMC.
tic System Identification and Realization: DSR - A
• Note, that in Sec. 3.1 it is shown in Table 6 that Subspace Approach Based on Observations. Model-
ΓK−SIM C
SIMC is at least ΓSIM C = 1.2 times better than ing, Identification and Control, 1996. 17(3):193–230.
K-SIMC, with the exception of Example 3.4. doi:10.4173/mic.1996.3.3.
Di Ruscio, D. Closed and Open Loop Subspace Sys-
• We see in Table 7 in the stable part in Sec. 3.1 that tem Identification of the Kalman Filter. Model-
all the methods for prescribed robustness, Ms = ing, Identification and Control, 2009. 30(2):71–86.
1.59, gives acceptable margins, i.e. GM > 1.7 and doi:10.4173/mic.2009.2.3.
P M > 30 (Seborg et al. (1989)).
Di Ruscio, D. On Tuning PI Controllers for In-
• Sec. 3.2 demonstrated successful implementations tegrating Plus Time Delay Systems. Modeling,
of Alg. 2.1 on unstable nonlinear models. In each Identification and Control, 2010. 31(4):145–164.
example it is seen that Alg. 2.1 gives lower (sum) doi:10.4173/mic.2010.4.3.
IAE indices than Alg. 2.1 (SIMC), where δ = 1.6
and Tc = 1.5 τ were chosen, i.o. Di Ruscio, D. Pi controller tuning based on integrat-
ing plus time delay models: Performance optimal
Furthermore, Alg. 2.1 was shown to be at least
Alg.2.1 (SIM C)
IAEvu
tuning. 2012. In Proceedings of the IASTED Con-
IAEvuAlg.2.1 = 1.4 times better in terms of in- trol and Applications Conference. Crete Greece June
put disturbance IAEvu , however Alg. 2.1 (SIMC) 18-21.
Alg.2.1
IAEvy
was shown at least = 1.2 times Di Ruscio, D. and Dalen, C. Tuning PD and PID Con-
Alg.2.1 (SIM C)
IAEvy
better in terms of output disturbance IAEvy . The trollers for Double Integrating Plus Time Delay Sys-
exception is on the Nomoto vessel in Example 3.9 tems. Modeling, Identification and Control, 2017.
where both IAEvy were approximately equal. 38(2):95–110. doi:10.4173/mic.2017.2.4.

196
 Christer Dalen, “PD/PID controller tuning based on model approximations’

Fossen, T. I. and Perez, T. Marine Systems Simulator Pareto, V. Il massimo di utilit dato dalla libera con-
(MSS). 2004. URL http://www.marinecontrol. correnza. Giornale degli Economisti,luglio, 1894b.
org. pages 48–66.
Garpinger, O. and Hägglund, T. Modeling for optimal Schmidt Z., J. P. . B., Brill. Choking can eliminate
pid design. 2014. pages 6929–6934. Preprints of the severe pipeline slugging. 1979. 312:230–238.
19th World Congress. Seborg, D., Edgar, T., and Mellichamp, D. Process
Jahanshahi, E., Oliveira, V. D., Grimholt, C., and Sko- Dynamics and Control. Number v. 1 in Chemical
gestad, S. A comparison between internal model Engineering Series. Wiley, 1989.
control, optimal pidf and robust controllers for un-
Seborg, D., Edgar, T., and Mellichamp, D. Process
stable flow in risers. IFAC Proceedings Volumes,
dynamics and control. Wiley series in chemical engi-
2014. 47(3):5752 – 5759. doi:10.3182/20140824-6-
neering. Wiley, 2004.
ZA-1003.02381. 19th IFAC World Congress.
Jahanshahi, E. and Skogestad, S. Closed-loop Shannon, C. E. Communication in the presence of
model identification and pid/pi tuning for ro- noise. Proceedings of the IRE, 1949. 37(1):10–21.
bust anti-slug control. IFAC Proceedings Volumes, doi:10.1109/JRPROC.1949.232969.
2013. 46(32):233 – 240. doi:10.3182/20131218-3-IN-
2045.00009. 10th IFAC International Symposium on Silva, G., Datta, A., and Bhattacharyya, S. PID Con-
Dynamics and Control of Process Systems. trollers for Time-Delay Systems. Control Engineer-
ing. Birkhäuser Boston, 2005.
Kotelnikov, V. A. On the transmission capacity of
”ether” and wire in electrocommunications. Proc. Skogestad, S. Simple analytic rules for model reduc-
1st All-Union Conf. Technological Reconstruction of tion and PID controller tuning. Journal of Process
the Commun. Sector and Low- Current Eng, 1933. Control, 2003. 13(13):291–309. doi:10.1016/S0959-
pages 1–19. 1524(02)00062-8.

Lee, J., Cho, W., and Edgar, T. F. Sim- Skogestad, S., Havre, K., and Larsson, T. Control lim-
ple analytic pid controller tuning rules revis- itations for unstable plants. IFAC Proceedings Vol-
ited. Industrial & Engineering Chemistry Research, umes, 2002. 35(1):485 – 490. doi:10.3182/20020721-
2014. 53(13):5038–5047. URL 10.1021/ie4009919, 6-ES-1901.00330. 15th IFAC World Congress.
doi:10.1021/ie4009919.
Skogestad, S. and Postlethwaite, I. Multivariable feed-
LIngenieur, S. I. P. Comportement dynamique
back control: analysis and design. Wiley, 1996.
d’un vehicule auto-balance de type seg-
way. Concours Centrale-Suplec, 2005. URL
Son, K. H. and Nomoto, K. On the Coupled Motion
https://www.concours-centrale-supelec.fr/
of Steering and Rolling of a High Speed Container
CentraleSupelec/2005/PSI/sujets/SI.pdf. In
Ship. Naval Architect of Ocean Engineering, 1982.
french. Accessed 01.05.17.
20:73–83.
Ljung, L. System Identification (2nd ed.): Theory for
the User. Prentice Hall PTR, Upper Saddle River, Yocum, B. Offshore Riser Slug Flow Avoidance: Math-
NJ, USA, 1999. ematical Models for Design and Optimization. Soci-
ety of Petroleum Engineers of AIME, 1973.
MATLAB. Version 9.1.0.441655 (R2016b). The
MathWorks Inc., Natick, Massachusetts, USA, 2016. Ziegler, J. and Nichols, N. B. Optimum settings for
Control System Toolbox, Version 9.3. Optimization automatic controllers. Trans. of the A.S.M.E., 1942.
Toolbox, Version 6.2. 64(64):759–768.

197