Discrete-Time Fractional-Order PID Controller Defi
Discrete-Time Fractional-Order PID Controller Defi
net/publication/262030240
CITATIONS READS
4 1,009
3 authors, including:
SEE PROFILE
All content following this page was uploaded by Farshad Merrikh Bayat on 21 May 2015.
Abstract
In some of the complicated control problems we have to use the controllers
that apply nonlocal operators to the error signal to generate the control. Cur-
rently, the most famous controller with nonlocal operators is the fractional-
order PID (FOPID). Commonly, after tuning the parameters of FOPID con-
troller, its transfer function is discretized (for realization purposes) using the
so-called generating function. This discretization is the origin of some er-
rors and unexpected results in feedback systems. It may even happen that
the controller obtained by discretizing a FOPID controller works worse than
a directly-tuned discrete-time classical PID controller. Moreover, FOPID
controllers cannot directly be applied to the processes modeled by, e.g., the
ARMA or ARMAX model. The aim of this paper is to propose a discrete-
time version of the FOPID controller and discuss on its properties and ap-
plications. Similar to the FOPID controller, the proposed structure applies
nonlocal operators (with adjustable memory length) to the error signal. Two
methods for tuning the parameters of the proposed controller are developed
and it is shown that the proposed controller has the capacity of solving com-
plicated control problems with a high performance.
Keywords: fractional-order PID, discrete-time controller, ARMA model,
ARMAX, long memory, tuning.
1
Please be advised that this a preprint submitted to a journal for possible publication
and it may be subjected to some changes during the review process.
where m is a positive integer such that m − 1 < µ ≤ m and Γ(.) is the well
known gamma function. Other definitions for fractional differential operators
2
can also be found in the literature [17]. Note that according to (2) and (3),
fractional integral and derivative are nonlocal operators which apply the past
values of f to determine Dt−λ f (t) and Dtµ f (t). Hence, unlike classical PID
controllers, the derivative term of the FOPID controller is actually a nonlocal
operator acts on the error signal. Taking the Laplace transform from both
sides of (1) leads to the following transfer function for FOPID controller [16]:
U(s)
C(s) = = kp + ki s−λ + kd sµ . (4)
E(s)
Here it is worth to mention that unlike the classical PID controllers,
currently there is no direct definition available for discrete-time FOPID con-
trollers. However, for realization purposes, it is common practice to first de-
sign the (continuous-time) FOPID controller and then approximate it with
a discrete-time system. This approximation is often performed by using the
so-called generating function [18]-[21]. In this technique the Laplace variable
s in (4) is substituted with a certain function of z (using, e.g., the Tustin
transform) and then the power series expansion (PSE) of the resulted ex-
pression is obtained in terms of z. Finally, since any practical discrete-time
system must necessarily use a limited memory, the resulted PSE is truncated.
This approach suffers from many drawbacks (see the discussions below), the
source of all is the unavoidable mismatch between the frequency response of
continuous and discrete transfer functions. The aim of this paper is to pro-
pose a long-memory discrete-time PID (LDPID) controller which removes
the limitations of the above-mentioned discretization scheme and still has
the high performance of FOPID controllers.
There are many good reasons for defining and using LDPID controllers.
First of all, it is very common practice to model a real-world continuous-time
process by a (discrete) transfer function in the z variable. For example, such
a model is obtained when an unknown process is identified using ARMA or
ARMAX model. Obviously, in dealing with such a discrete-time model of
process it is more reasonable to directly design a discrete-time controller as
well (instead of designing a continuous-time controller and then approximat-
ing it with a discrete-time one). Another good reason for developing LDPID
controllers is that even when both the controller and process are continuous-
time, the controller is more likely to be realized using digital microprocessors.
As a classical fact, some unwanted effects may occur in the feedback system
when the continuous-time controller is replaced with an approximate discrete-
time controller (for example, the phase margin is decreased according to the
3
effect of sample-and-hold and, in addition, the feedback system may become
unstable according to the effect of truncation).
Another important point in relation to the classical approximation meth-
ods is that FOPID controllers have some features which are not preserved
after approximating them with discrete transfer functions. For example, one
important feature of every FOPID controller is the so-called long memory
principle, which is lost after approximating it with an integer-order trans-
fer function. Another property that is lost after approximation is the op-
timality of controller. In fact, the optimal controller designed for a certain
continuous-time process will no longer be optimal after approximating it with
a discrete-time controller.
In addition to the above-mentioned points, it should also be noted that
the derivative term of the FOPID controller given in (4) cannot exactly be
realized in practice since it is a non-causal operator. In fact this term should
always be considered in series with a low-pass filter to be realizable (more
precisely, the practical FOPID controller is actually a six degrees of freedom
controller). The proposed LDPID controller will remove this difficulty.
The rest of this paper is organized as the following. In Section 2 we
introduce the proposed LDPID controller and develop two methods for its
tuning. Four simulated examples and an experimental study are presented
in Section 3. Finally, Section 4 concludes the paper.
4
where α , ωc / tan(ωc T /2) and ωc is the gain crossover frequency of the
open-loop transfer function. (Here we have applied Tustin method since
it is more accurate compared to other transforms such as backward and
forward difference. Section 2.3 makes a connection between different possible
transforms.) To proceed, we need to calculate the PSE of the expression in
the right hand-side of (5), which cannot be performed without determining
its region of convergence (ROC) in the complex z-plane. Note that this
expression is actually a multi-valued function of z which has two branch
points at z = 1 and z = −1 located on the unit circle. Mathematically, the
corresponding branch cut (BC) can be considered either inside or outside the
unit circle as shown in Fig. 1. Clearly, the choice of BC affects the ROC
and consequently, causality of the resulted system. Similar to the classical
results [22], here considering the BC inside the unit circle and the ROC as
|z| > 1 yields a causal system as it is desired. Hence, assuming w = z −1 and
|z| > 1 the PSE of the expression in the right hand-side of (5) is obtained as
the following:
µ ∞
µ 1−w µ
X
α × =α fk (µ)w k , |w| < 1, (6)
1+w
k=0
where µ
dk
1 1−w
fk (µ) = × . (7)
k! dw k 1+w w=0
where again the coefficients fk (µ) are calculated from (7). It can be shown
(using Maple) that the first few coefficients in (8) are as the following:
5
Figure 1: Two possible branch cuts for the multi-valued function given in (5).
4 6 8 4 46 2
f6 (µ) = µ + µ + µ , ...
45 9 45
Note that (8) holds for both the positive and negative values of µ. Hence,
one may try to expand the integral term of (4) in a similar manner and
arrive at an equation like (8) in λ, but the problem with such an expansion
is that the resulted series does not have infinite DC gain (considering the
fact that any infinite series must be truncated in practice), which is essential
for tracking the step command without steady-state error. In order to find
a series approximation for s−λ in terms of z −1 which has infinite DC gain,
first we write it as s−λ = (1/s) × s1−λ and then apply the prewarped Tustin
method to it. Applying this technique yields
∞
1 + z −1 X
s−λ = α−λ fk (1 − λ)z −k , |z| > 1, (9)
1 − z −1 k=0
where
Kp = k p , Kd = k d α µ , Ki = ki α−λ . (11)
The only problem with (10) is that infinite number of memory units are
needed for its realization, and consequently the computational cost is in-
creased by increasing the time. In other words, in practice the upper bound
of sigmas in (10) cannot be considered equal to infinity. Restricting the
6
number of memory units to M, the following formula is proposed for the
Mth-order LDPID controller:
M M
X
−k 1 + z −1 X
Cd (z) = Kp + Kd fk (µ)z + Ki fk (1 − λ)z −k . (12)
1 − z −1
k=0 k=0
In the rest of this paper whenever we refer to the LDPID controller, a system
with transfer function (12) is under consideration. Figure 2 shows the block
diagram of the proposed LDPID controller where C/D and D/C stand for the
ideal continuous-to-discrete-time and discrete-to-continuous-time converters,
respectively [23]. Obviously, in practice the C/D is realized using a sample-
and-hold (or an A/D converter) and the output of the adder in Fig. 2 can
directly be applied to the process. For simulation in Matlab, the C/D is
modeled with a sample-and-hold, and the D/C is simply omitted. It will
be shown later that one advantage of the proposed LDPID controller is that
it can easily cope with the unwanted effects caused by the sample-and-hold
used in practice.
It is important to note that the proposed LDPID controller as defined
in (12) actually has six parameters to tune. Considering the fact that the
continuous-time and discrete-time FOPID controllers (as defined in (4) and
(12), respectively) have different number of parameters to tune, and taking
into account the effect of sample-and-hold, it is evident that in practice the
parameters of these two controllers cannot simply be related according to
(11). In other words, the transfer function of a certain LDPID controller is
not, in general, obtained by applying the Tustin transform to any FOPID
controller. It concludes that the parameters of the proposed LDPID con-
troller should be tuned directly. For this purpose, a certain value can be
assigned to M and then the value of other parameters be calculated such
that a predetermined set of objectives is met. The other possible approach
is to consider M as a tuning parameter and then find the values of Kp , Kd ,
Ki , µ, λ, and M such that the objectives under consideration are met. The
first approach is used in the rest of this paper.
7
Figure 2: Block diagram of the M th-order LDPID to be applied instead of the continuous-
time FOPID
8
that any floating point multiplication is much more time consuming than
any summation, the computational cost of the proposed controller can be
approximated by the number of multiplications, which is equal to 2M + 6 at
each sampling period. So, in order to determine the suitable value of M first
we should determine the suitable sampling period of system and then choose
the value of M such that the digital processor can perform at least 2M + 6
floating point multiplications at each sampling period.
After determining the value of M, the values of the remaining five parame-
ters of controller should be determined such that the following five conditions
are satisfied simultaneously (P (s) is the process transfer function, which is
located in a standard unity feedback system in series with the LDPID con-
troller):
holds for the desired ωc , where P (s) is the process transfer function.
Cd (ejω )P (jω)
≤ A dB ω ≥ ωt rad/s, (17)
1 + Cd (ejω )P (jω)
9
• The feedback system rejects the disturbance, which is achieved by sat-
isfying the inequality:
1
≤ B dB ω ≤ ωs rad/s, (18)
1 + Cd (ejω )P (jω)
where B and ωs are desired constants.
Similar to [24], in this paper (14) is considered as the main object of opti-
mization and (15)-(18) are considered as the corresponding constrains. More
precisely, we have applied the genetic algorithm to find the values of Kp ,
Kd , Ki , µ, and λ in (12) such that ||Cd (ejωc )P (jωc )| − 1| is minimized and
simultaneously the equality constraints (15) and (16), and inequality con-
straints (17) and (18) are fulfilled. Clearly, it may happen that the above
mentioned optimization problem does not have any solution, but even ap-
proximate solutions (which violate the constrains to some extent) are useful
in practice. Numerical simulations performed by authors show that the ge-
netic algorithm toolbox of Matlab can effectively solve such a complicated
constrained optimization problem in a relatively short time (see Examples 1
of Section 3), however the application of this tuning strategy is restricted to
relatively simple processes.
The second method that can be used for tuning the parameters of LDPID
controller is to assign a certain value to M and then calculate the values of
Kp , Kd , Ki , µ, and λ in (12) such that an integral performance index (e.g.,
the IAE or ISE performance index corresponding to the tracking error of step
command) is minimized. In this method, the genetic algorithm (or any other
meta-heuristic optimization algorithm such as PSO) can be used to search
the five-dimensional space to find the optimal solution. For this purpose,
the corresponding integral performance index should be considered as the
object of minimization and the equation obtained by equating the number
of unstable poles of the closed-loop system to zero can be considered as
the constraint of optimization. This equality constraint is mandatory when
the integral performance index is evaluated in the frequency domain (using
Parseval’s theorem) since in this case the stability issue is not involved in the
cost function, i.e., a set of parameters that minimize the integral performance
index in frequency domain may lead to an unstable feedback system.
2.3. Generalization
The coefficients fk (µ) and fk (1 − λ) of the proposed LDPID controller
(12) are calculated from (7). The question that may arise at this point is:
10
Can one propose another reasonable method instead of (7) for calculation
of these coefficients? To provide an answer for this question first recall that
any classical PID controller (either of continuous or discrete type) combines a
proportion, derivative and integral of the error signal to generate the control.
It motivates us first to study the general behavior of discrete-time derivative
operators. Then based on these results we can conclude that any alternating-
sign function of k provides us with a natural definition for fk (µ) to be used
in (12). Finally, we briefly extend the results to discrete-time integrators.
Suppose that we want to approximate the time derivative of the function
e(t) only by using its samples ei = e(ti ) (i = 1, . . . , n) where T = ti − ti−1 .
Mathematically, this task can be performed by using one of the following
formulas [25]:
ei+1 − ei
ė(ti ) = + O(T ), (19)
T
ei − ei−1
ė(ti ) = + O(T ), (20)
T
ei+1 − ei−1
ė(ti ) = + O(T 2), (21)
2T
which are called forward, backward, and centered difference approximations,
respectively. In each case one can increase the accuracy of the resulted time
derivative by using larger number of sample points. For example, backward
difference approximations with second and third order errors are obtained as
the following [25]:
3ei − 4ei−1 + ei−2
ė(ti ) = + O(T 2 ), (22)
2T
11ei − 18ei−1 + 9ei−2 − 2ei−3
ė(ti ) = + O(T 3). (23)
6T
The general formula for approximating the time derivative of a certain func-
tion with nth-order error using the backward difference method is as the
following [26]:
0
1 X B,1
ė(ti ) = g ek+i + O(T n ), (24)
T k=−n k,n
B,1
where the coefficients gk,n are calculated from the following iterative proce-
dure: n
B,1
X B,1
g0,n = (1/j), g−1,n = −n, (25)
j=1
11
B,1 B,1
g−k,n = −g−k+1,n (k − 1)(n − k + 1)/k 2 , k = 2, . . . , n. (26)
Very similar formulas can also be found in [26] for forward and centered dif-
ference approximations. The intersection point of all of these finite-difference
formulas is that they approximate the time derivative of a function by the
weighted sum of its samples, where these weights change sign decussately
as the k is increased. For example, according to (26) it is obvious that
B,1
the weights gk,n in (24) are positive for k = 0, −2, −4, . . . and negative for
k = −1, −3, −5, . . ..
Equation (24) leads us to the fact that in digital control system design one
can use the following nth-order difference equation to calculate the derivative
of error signal from its samples with an arbitrary precision:
n
1 X B,1
edot [i] ≈ g−k,n e[i − k], (27)
T
k=0
where e[i] and edot [i] stand for the samples of error signal and its derivative,
B,1
respectively, and the alternating-sign weights g−k,n are defined similar to
(25) and (26) (note that the same discussion goes for forward and centered
difference approximations, however, the backward difference approximation
is advantageous in the way that it leads to a causal relation). Recall that
according to (12) the proposed LDPID controller relates the samples of error
signal to its derivative through the following difference equation
M
X
edot [i] = Kd fk (µ)e[i − k], (28)
k=0
where the weights fk (µ) are shown in Fig. 3 for three different values of µ.
As it can be observed in this figure the weights fk (µ) in (28) change sign
B,1
decussately similar to the weights g−k,n in (27). It means that the derivative
term of the proposed LDPID controller also has the property of subtracting
many two successive (positive-weighted) samples of the error signal and then
forming the cumulative sum of results. However, the difference equation
(28) is, compared to (27), advantageous in the way that one can adjust the
amplitude of weights simply by changing the value of µ (note that in case
B,1
of using (27) we have no control on the amplitude of the weights g−k,n for
the given n). For this reason, it is not surprising if the derivative action
of the LDPID controller as given in (28) reduces to a classical discrete-time
12
differentiator for some µ and M. For example, assuming µ = 0.5 and M = 1
equation (28) yields
1
X
edot [i] = Kd fk (0.5)e[i − k] = Kd (e[i] − e[i − 1]), (29)
k=0
Ke−50s
P (s) = , (30)
433.33s + 1
13
1
f (0.5)
0
k
−1
0 5 10 15
2
f (1)
0
k
−2
0 5 10 15
20
fk(1.5)
−20
0 5 10 15
k
where the nominal value of K is equal to 3.13 and it may vary between
2.75 and 3.75. Our aim here is to design a LDPID controller such that the
following objectives are achieved simultaneously:
• ωc = 0.008 rad/s,
• ϕm = 60◦ ,
• |S(jω)| ≤ −20 dB, ∀ω ≤ ωs = 0.001 rad/s,
• |T (jω)| ≤ −20 dB, ∀ω ≥ ωt = 10 rad/s,
• robustness to variations in the gain of the process.
Following the design procedure presented in Section 2 the following LDPID
controller is obtained (assuming M = 15) using the genetic algorithm toolbox
of Matlab:
15 15
X
−k 1 + z −1 X
Cd (z) = 3.059 + 0.384 fk (1.228)z + 0.059 −1
fk (0.45)z −k .
k=0
1 − z k=0
(31)
Note that the above controller is actually a very good suboptimal solution
which approximately satisfies all requirements of the problem. Figure 4 shows
14
50
Magnitude (dB)
0
−50
−100
−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1
200
100
Phase (deg)
−100
−200
−3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1
log ω
the Bode plot of Cd (esT )G(s). In this figure the gain crossover frequency is
approximately equal to 0.008 rad/s and the phase plot is almost flat around
this frequency, as it is expected (note that log10 0.008 ≈ −2.1). Moreover, the
phase margin is slightly larger than 60◦ . Figure 5 shows the Bode magnitude
plots of the sensitivity and complementary sensitivity function. Using this
figure it can be easily verified that S(s) and T (s) satisfy the requirements of
problem. Figure 6 shows the unit step response of the corresponding feedback
system for three different values of K. As it can be observed, the overshoot
of response is almost constant for different values of K, which is the direct
consequent of the flatness of the Bode phase plot of open-loop transfer func-
tion at frequencies around ωc .
15
Magnitude of S(s)
20
Magnitude (dB)
0
−20
−40
−60
−4 −3 −2 −1 0 1 2
Magnitude of T(s)
50
Magnitude (dB)
−50
−100
−4 −3 −2 −1 0 1 2
log ω
This process can effectively be controlled with the following PID controller,
which is obtained by trial and error:
0.1
C(s) = 1.1 +
+ 0.4s. (33)
s
Figure 7 shows the unit step response of the feedback system in this case.
As it can be observed, the command following is quite satisfactory.
Now let us examine the performance of this feedback system when con-
troller (33) is discretized and then applied. In this example the gain crossover
frequency of the open-loop system is ωc ≈ 0.21 rad/s and the phase margin
is about 60◦ . Applying prewarped Tustin method to (33) (assuming that the
sampling period is equal to T = 0.1 s) leads to the following discrete-time
controller:
1 + z −1 1 − z −1
Cd1 (z) = 1.1 + 0.005 + 8 . (34)
1 − z −1 1 + z −1
Figure 8 shows the Bode plots of C(s)P (s) and Cd1 (esT )P (s) (the latter cor-
responds to the open-loop transfer function when the discrete-time controller
is applied in series with ideal C/D and D/C converters). This figure clearly
shows that the closed-loop system with discrete-time controller is unstable,
while we observed that the original continuous-time feedback system is stable
with a satisfactory phase margin (recall that any spike in the Bode magni-
tude plot of Cd1 (esT )P (s) in Fig. 8 corresponds to an encirclement around -1
16
1.4
Process gain=3.13
Process gain=3.75
1.2 Process gain=2.75
0.6
0.4
0.2
0
0 500 1000 1500 2000
Time (s)
Figure 6: Unit step response of the feedback system when the LDPID controller is applied,
corresponding to Example 1.
The main advantage of this controller over the one given in (33) is that the
unwanted effects caused by using zero-order hold are also taken into account
during the controller design and consequently it is ready to be realized using
17
1.2
0.8
Process output
0.6
0.4
0.2
PID controller
LDPID controller
0
0 10 20 30 40 50 60
Time (s)
microprocessors. Figure 7 also shows the unit step response of the closed-loop
system when the LDPID controller (35) is applied. As it can be observed,
the LDPID controller results in a satisfactory transient response.
18
50
Spikes in the Bode magnitude plot
Magnitude (dB)
0
−100
−1 −0.5 0 0.5 1 1.5 2 2.5 3
200
100
Phase (deg)
−100
−200
−1 −0.5 0 0.5 1 1.5 2 2.5 3
log ω
Figure 8: Bode plots of C(s)P (s) and Cd1 (esT )P (s), corresponding to Example 2.
19
1.5
Process output
0.5
0
0 10 20 30 40 50
Time (s)
20
2.5
Process output
1.5
1
proposed LDPID
0.5
2DOF PID (tuned using AMIGO method)
0
0 5 10 15 20
1.5
Control variable
0.5
−0.5
0 5 10 15 20
Time (s)
controller can compete the 2DOF PID controller tuned by using the AMIGO
method. However, the proposed controller is advantageous in the way that
it is in the discrete-time form and ready for realization.
21
taking into account the function of PWM it is obvious that the system is
actually nonlinear. More precisely, the negative control signals generated by
controller are simply neglected by PWM. Note also that applying a kind of
anti-windup technique is also mandatory for digital realization of integrators.
At different working conditions each box can (very approximately) be
modelled by the following first-order uncertain transfer function
K
P (s) = , 14 < K < 34, 380 < T < 570. (40)
1 + sT
(In the above transfer function the input is duty cycle and the output is
temperature in ◦ C). The nominal process model is also considered as P (s) =
32/(1 + 425s) (note that the probability of occurring different uncertainties
is not the same). Our aim here is to design a controller which leads to ωc = 1
rad/s, ϕm ≥ 75◦ , ωt = 10 rad/s, ωs = 0.1 rad/s, and A = B = −20 dB
(see Section 2.2). Moreover, according to the high uncertainty in the process
model it is highly desired that the open-loop phase plot be as flat as possible
at frequencies around ωc .
Following the procedure presented in Section 2.2 the genetic algorithm
leads to the following PID and LDPID controllers (T = 0.1s):
1.187
C(s) = 7.937 + − 0.935s, (41)
s
5 5
X
−k 1 + z −1 X
Cd (z) = 7.109 + 0.711 fk (0.077)z + 0.750 fk (0.415)z −k .
k=0
1 − z −1 k=0
(42)
Figure 11 shows the Bode phase and magnitude plots of C(s)P (s) and
Cd (esT )P (s) when the nominal process model is considered. In this example
PID and LDPID controllers lead to phase margins equal to 73.3◦ and 91◦ ,
respectively. As it can be observed in this figure the LDPID controller has
perfectly satisfied the design requests. Especially, unlike the PID controller,
the proposed LDPID has led to a very flat curve in the phase plot (for more
than 2 decades), which is highly desired in dealing with the uncertain process
under consideration. Using trivial simulations it can be easily verified that
both controllers lead to very similar and satisfactory time-domain responses
when the nominal linear process model is considered.
22
150
PID
Magnitude (dB)
100
LDPID
50
−50
−4 −3 −2 −1 0 1 2
−50
Phase (deg)
−100
−150
−200
−4 −3 −2 −1 0 1 2
log ω
Figure 11: Bode plots of C(s)P (s) and Cd (esT )P (s), corresponding to Example 5.
Both of the above mentioned controllers are realized using the digital
system shown in Figure 12 (the classical PID is discretized using the Tustin
method with pre-wrapping beforehand). Here it is worth to mention that
since the 1-wire output of DS18B20 digital sensor cannot directly be con-
nected to PC, a kind of transducer is needed. The RaspberryPi in Fig. 12 is
used for this purpose. The input of optocoupler is connected to the software-
generated PWM with frequency 50Hz. It was observed that in dealing with
LDPID controller calculation of each control signal takes about 2ms in prac-
tice.
The time-domain responses of the practical closed-loop system when PID
and LDPID controllers are applied (assuming that the reference temperature
is equal to 28◦ C) are shown in Figs. 13 and 14, respectively in five different
conditions. Note that the curves with similar colors in Figs. 13 and 14 are
obtained under exactly the same conditions in practice (i.e., exactly the same
DC voltages across the heating wire, the same materials in the box, etc.).
Note also that the system itself has an initial condition and the vertical axis
in Figs. 13 and 14 begins from 20◦ C. Figure 13 clearly shows the superiority
of the proposed LDPID controller. In fact, the maximum overshoot (in the
worst case) caused by PID and LDPID controllers is equal to 23.43% and
6.1%, respectively. Moreover, fluctuations in the response are settled down
much faster when LDPID is applied.
23
Figure 12: The digital system used to realize the controllers of Example 5.
4. Conclusions
In this paper we proposed a new formulation for PID controllers, which
can be considered as the discrete-time equivalent of fractional-order PIDs.
Experimental and numerical examples were also presented which verified the
fact that the proposed controller is capable of solving some difficult control
24
35
30
Temperature ( C)
o
25
20
0 5 10 15 20 25 30 35
Time (min)
Figure 13: Step responses of the practical closed-loop system when LDPID is applied,
corresponding to Example 5.
References
[1] K.J. Åström, T. Hägglund, Advanced PID Control, ISA-The Instrumen-
tation, Systems, and Automation Society, 2006.
[3] K. Ogata, Discrete-Time Control Systems, 2nd ed., Prentice Hall, 1995.
[4] W.-D. Chang, R.-C. Hwang, J.-G. Hsieh, A self-tuning PID control for a
class of nonlinear systems based on the Lyapunov approach, J. Process
Contr. 12(2) (2002) 233-242.
[5] C.-M. Lin, M.-C. Li, A.-B. Ting, M.-H. Lin, A robust self-learning PID
control system design for nonlinear systems using a particle swarm op-
timization algorithm, Int. J. Mach. Learn. Cyber. 2(4) (2011) 225-234.
25
35
30
Temperature ( C)
o
25
20
0 5 10 15 20 25 30 35
Time (min)
Figure 14: Step responses of the practical closed-loop system when classical PID is applied,
corresponding to Example 5.
[7] M.A. Johnson, PID Control: New Identification and Design Methods,
Springer-Verlag, 2005.
[8] Q.G. Wang, T.H. Lee, H.W. Fung, Q.Bi, Y. Zhang, PID tuning for
improved performance, IEEE T. Contr. Syst. T. 7(4) (1999) 457-65.
[10] F. Merrikh-Bayat, General rules for optimal tuning the PIλ Dµ controllers
with application to first-order plus time delay processes, Can. J. Chem.
Eng. 90(6) (2012) 1400-1410.
[11] Y.X. Su, D. Sun, B.Y. Duan, Design of an enhanced nonlinear PID
controller, Mechatronics 15(8) (2005) 1005-1024.
[13] W.H. Chen, D.J. Balance, P.J. Gawthrop, J.J. Gribble, J.O. O’Reilly,
26
Nonlinear PID predictive controller, IEE Proc. Control Theory Appl.
146(6) (1999) 603-11.
[15] T.P. Bucklaew, C.-S. Liu, A new nonlinear gain structure for PD-type
controllers in robotic applications, J. Robot. Syst. 16(11) (1999) 627-49.
[19] B.M. Vinagre, I. Petras, P. Merchan, and L. Dorcak, Two digital real-
ization of fractional controllers: application to temperature control of a
solid, in: Proc. of the European Control Conference (ECC2001), Porto,
Portugal, pp. 1764-1767, 2001.
[20] Y.-Q. Chen and K.L. Moore, Discretization schemes for fractional-order
differentiators and integrators, IEEE T. Circuits-I 49(3) (2002) 363-367.
[22] A.V. Oppenheim, A.S. Willsky, and S.H. Nawab, Signals and Systems,
2nd ed., Prentice Hall, 1996.
[24] C.A. Monje, B.M. Vinagre, V. Feliu, and Y.-Q. Chen, Tuning and auto-
tuning of fractional order controllers for industry applications, Control
Eng. Pract. 16 (2008) 798-812.
27
[25] R.W. Hornbeck, Numerical Methods, Quantum Publishers, New York,
1975.
[26] I.R. Khan, Ryoji Ohba, Closed-form expressions for the finite difference
approximations of first and higher derivatives based on Taylor series, J.
Comput. Appl. Math. 107 (1999) 179-193.
[27] B.-S. Chen, T.-Y. Yang, Robust optimal model matching control design
for flexible manipulators, J. Dyn. Syst.-T. ASME 115 (1993) 173-178.
28