See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/252033980

Tuning method for second-order active disturbance rejection control

Article · January 2011

CITATIONS READS

103 3,448

4 authors, including:

Donghai Li Zhiqiang Gao

Tsinghua University Cleveland State University
111 PUBLICATIONS 1,195 CITATIONS 220 PUBLICATIONS 13,370 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Zhiqiang Gao on 15 January 2015.

The user has requested enhancement of the downloaded file.

Proceedings of the 30th Chinese Control Conference
July 22-24, 2011, Yantai, China

Tuning Method for Second-order Active Disturbance Rejection Control

CHEN Xing1, LI Donghai1, GAO Zhiqiang2, WANG Chuanfeng1
1 State Key Lab of Power Systems, Dept of Thermal Engineering, Tsinghua University, Beijing 100084, P. R. China
E-mail: [email protected]
2 The Center for Advanced Control Technologies, Department of Electrical and Computer Engineering, Cleveland State University,
Cleveland, OH 44115, USA.
E-mail: [email protected]

Abstract: A simple tuning method for second-order active disturbance rejection control (ADRC) that achieves high performance
and good robustness for a wide range of processes is presented. ADRC is a novel control strategy whose exciting performance
has been shown by literatures. The proposed method makes ADRC become easy to tune and more practical. Once the desired
settling time is given, only one parameter need to be tuned during the design procedure, which can be found by monotonously
increasing. Examples are given to show the effectiveness and flexibility of the method. Simulations demonstrate that
second-order ADRC can handles processes with various characteristics, including low- and high-order, large dead time,
non-minimum phase, unstable and distributed parameter systems.
Key Words: Second-order ADRC, One Parameter Tuning Method, Monte-Carlo Experiment

Besides the difficulty in tuning, there is another obstacle in

1 Introduction the implementation of ADRC. That is, the order of ADRC is
generally varied with the controlled plant. Much work has
Proportional-integral-derivative (PID) controllers are been done in the cases of n-order ADRC applying to n-order
widely used in the process control industry [1]. The main plants. However, the order of practical plants exists from
reason is its relatively simple structure, which can be easily one to infinite, even unknown or time-varying. Therefore,
understood and implemented in practice. Another reason is it’s necessary to study how a fixed-order ADRC can deal
that users do not have to change the structure of PID, only with the real processes, and how to tune its parameters. To
re-tuning of its parameters is necessary, while handling the best of our knowledge, however, little has been done
different kinds of problem. However, many PID controllers about this. Yao[12] discussed the case of second-order ADRC,
are poorly tuned and limited in performance, especially but the controlled objects were limits to third-order, and that
while dealing with dynamic uncertainties. no specific tuning rule was provided therein.
To overcome the limitations of PID, the active disturbance In this paper, we present a simple tuning method for
rejection controller (ADRC), a novel controller, was second-order ADRC that demonstrates promising results in
developed by Han [2, 3]. The basic idea of ADRC is to use an controlling an extensive class of linear processes, including
extended state observer (ESO) to estimate the internal and low- and high-order, large dead time, non-minimum phase,
external disturbances in real time. Then, through disturbance unstable and distributed parameter systems. Only one
rejection, the originally complex and uncertain plant parameter needs to be tuned in this method under the
dynamics is reduced to a simple cascade integral plant, condition that the desired settling time is given. To verify
which can be easily controlled. Due to its strong robustness the inherent robustness of ADRC, the Monte-Carlo
and disturbance rejection, ADRC has been successfully method[13] is used. Simulations show that both high
applied in many fields [4-6]. performance and good robustness are achieved.
However, the tuning procedure of ADRC is very The remainder of this paper is organized as follows. Section
complicated due to its large number of parameters. The 2 formulates the design problem, in which the second-order
tuning is usually relied on the human experiences. Some ADRC is presented. A result about stability of ADRC
artificial intelligence approaches were adopted to regularize system is discussed in section 3, which is an important hint
the ADRC parameters (see, for example, Tong[7] and Shi[8]), for the tuning. Then, the specific tuning method is given in
but the algorithm was complex and the computational cost section 4. In section 5, the method is applied to several
was large. To simplify the problem, Gao[9] used linear gains examples in simulation. Finally, section 6 offers concluding
in place of the original nonlinear gain in ADRC. Thus, the remarks.
number of parameters was reduced obviously. This made
the tuning more realistic. Researches showed that the linear 2 Problem Formulation
ADRC still achieved high performance and good robustness
(see, for example, Sun[10] and Miklosovic[11]). The
2.1 Process Description
discussions in this paper are limited to the linear case.
The active disturbance rejection concept has been applied to
problems of different kinds, including single-input-
single-output (SISO), as well as multi-input-multi-output.
This work is supported by National Nature Science Foundation under (MIMO), plants that are nonlinear, time-varying, and most
Grant : 51076071 of all, uncertain. In this paper, however, the process is

6322
assumed to be linear, time invariant, and specified by a 
y f  u0  z3 | u0 (6)
transfer function:
Then, substituting (4) into (6) yields the closed-loop
y (s) dynamic characteristic:
G p ( s, p ) (1)
u (s)

y  kd y  k p y k p ysp (7)
where p denotes the parameters of system. The description
Taking the Laplace transform yields the close-loop transfer
covers finite dimensional systems with time delays and function:
infinite dimensional systems described by linear partial
differential equations. To verify the robustness of ADRC, y( s) kp
we assume that p may have a 10% change, that is, Gd ( s ) (8)
ysp ( s ) s  kd s  k p
2

p  [0.9p 0 ,1.1p 0 ] , where p 0 is the nominal value of p .

Remarks:
2.2 Second-order ADRC a. There are six tuning parameters in second-order ADRC,
The structure of second-order ADRC is illustrated in Fig. 1. namely control parameters B, k p and kd , and observer
parameters E1 , E 2 and E 3 .

b. In the frame of ADRC, the external disturbances and

internal uncertainty can be estimated and canceled in real
time. This is the reason that ADRC is model-independent
and inherent robust.

Fig.1 Structure of second-order ADRC 2.3 Design Objective

Gp is the process to be controlled. Two external signals act The design objective of this paper is to determine the six
on the control loop, namely set point ysp and load parameters in ADRC so that the system behaves well with
respect to changes in the two signals ysp and d , as well as in
disturbances d . An extended state observer (ESO) is used to
estimate the external disturbances d and internal uncertainty the process model Gp .Hence, the specification will express
(such as parameters perturbation in Gp ) in real time. The requirements on
inputs of ESO are control signal u and process output y . < Set point response. The system asymptotically tracks
The outputs of ESO are z1 , z2 and z3 . Then, a simple control stepwise set point changes, and the settling time and
law is applied, in which k p , kd and B are the control overshoot V% are smaller than desired value.
parameters. < Load disturbance response. The process output y comes
The ESO in second-order ADRC is back to set point ysp quickly and the impact of

z2  E1 ( y  z1 )  disturbance is low.

­ z1
° < Robustness with respect to model uncertainties. That is,
® z2 z3  E 2 ( y  z1 )  Bu (2)
° z E 3 ( y  z1 ) 
for a 10% change in parameter p of the process, the
¯ 3 system is still stable and the control performance is still
good.
where E1 , E 2 and E 3 are observer parameters to be
determined. If Gp , somehow, can be approximated by a 3 Stability and the Parameter B
second order system model structure:
Converting the ADRC equations to frequency domain, the

y f (t , y, y , Z )  Bu (3) closed-loop transfer function Gcl ( s ) of the control system is
readily available:
and the observer (2) is well tuned, z1 , z2 , z3 will track y , y
and f respectively. The term f represents the combined G p ( s )k p ( s 3  E1 s 2  E 2 s  E 3 )
Gcl ( s ) (9)
effect of the internal dynamics and external disturbances Z . BA1 ( s )  G p ( s ) A2 ( s )
Note that, there is no need to know the accurate mathematic
description of f in model (3). where

The control law in second-order ADRC is A1 ( s ) s 3  ( E1  kd ) s 2  ( E1kd  E 2  k p ) s 

u0 k p ( ysp  z1 )  kd z2 (4) A2 ( s ) ( E 3  E 2 kd  E1k p ) s 2  ( E3 kd  E 2 k p ) s  E3 k p

From this transfer function, the stability analysis will

u (u0  z3 ) / B (5)
proceed.
With z3 | f obtained from the ESO, (5) reduces (3) to an We learned from experiments that, in the case of a
approximate double integral plant: second-order ADRC dealing with a plant whose order is

6323
higher than two, the parameter B is important concerning to Zc2
the stability of system. To study how the parameter B Gd ( s ) (12)
( s  Zc ) 2
influence the stability of system, we simplify the problem by
making E1 , E 2 , E3 a function of Zo and k p , kd a function As the settling time t s and the overshoot V % is the main
of Zc , as proposed in Gao[9]: dynamic performance index, and the overshoot of (12) is
constantly zero, t s becomes the only factor taken into
s  E1 s  E 2 s  E 3
3 2
( s  Zo ) 3

(10) account while choosing Zc . Therefore, the relation between

s  kd s  k p
2
( s  Zc ) 2
ts and Z c is derived subsequently.
such that there are only three parameters left in (9),that is,
B, Zc and Zo .Then, a search program is used to determine Under a unit step change in input signal, the output of (12)
with Laplace transform is
the region in Zo  Zc plane where the closed-loop system is
stable. And this is repeated for different B . The results for Zc2 1 Zc 1
Y (s)   (13)
an example of G p ( s ) 1/( s  1)3 are shown in Fig. 2, where s ( s  Zc ) 2 s ( s  Zc ) 2 s  Zc
the area to the lower-left side of the curve is the stable region.
(See appendix 1 for more examples). Thus, the unit step response of system is

y (t ) 1  Zc teZc t  e Zc t 1  (1  Zc t )e Zc t (14)

According to the definition of setting time

y (t s )  y ( f ) ' (15)

where y (f) 1 and '  2% ,we have

(1  Zc ts )e Zc ts 0.02 (16)

which yields to
ts 5.85 / Zc (17)

Once the setting time is given, the parameter Z c can be

determined by (17). Thus, the desired dynamic characteristic
Fig.2 Stability in Zo  Zc plane of G p ( s ) 1/( s  1)3
of systems is decided. However, the actual output, somehow,
Results show that as B increases within certain bound, the can’t act the desired dynamic characteristic exactly.
stable area in Zo  Zc plane is expanded. This is an Therefore, some margins are considered to make sure the
important hint for the tuning method discussing design is dependable. In this paper, Z c is determined as
subsequently. following
Remark: Zc | 10 / ts (18)
c. In the tuning procedure, B is chosen to make sure the Then, k p and kd can be computed from Eq. (11).
closed-loop system is stable. It must be large enough so
the right control parameters, which are determined to
achieve the desired performance, lie in the stable region. 4.2 Tuning of Observer Parameters E1 , E 2 and E 3
On the other hand, B can’t be too large, otherwise the
For tuning simplicity, Gao[9] suggested the observer gains,
control signal u is weak and the system becomes slow.
E1 , E 2 and E 3 ,can be chosen as following
Therefore, a trade-off has to be made between stability
(robustness) and response speed.
E1 3Zo , E 2 3Zo2 , E 3 Zo3 (19)
4 Tuning Method where Zo is denoted as observer bandwidth. This make all
three of the observer poles be placed at Zo . The larger the
4.1 Tuning of Control Parameters k p and kd Zo , the sooner the disturbance is observed by ESO and
cancelled by the controller. But, according to the analysis in
Note from (8) that k p and kd decide the closed-loop
section 3, Zo cannot be too large or it will lie out of the
transfer function. By setting stable region, especially when the order of the controlled
kp Zc2 , kd 2Zc (11) plan is higher than two. Therefore, to make the ESO work
well in the case of small Zo , it is necessary to find a new
where Zc is denoted as control bandwidth by Gao[9].Thus, way to tune the observer parameters.
the desired transfer function becomes:

6324
To this end, the transfer function between z3 ( s) and f ( s) known in a practical problem)
is derived (See appendix 2 for details): (2) Let Zc | 10 / ts* , compute k p and kd from Eq. (11) .
z3 ( s ) E3 (3) Let Zo 4Zc , and k 4 , compute E1 , E 2 and E 3
(20)
f ( s) E3  E 2 s  E1 s 2  s 3 from Eq. (23).
Due to the following reasons: (4) Monotonously increase the value of parameter B from
(1) in the actual control situation, low and middle a small value, until the dynamic performance is
frequencies are much more important than high satisfactory.
frequencies; (5) Verify the robustness of controller by Monte-Carlo
(2) in general, the coefficients of low and middle simulation. (As all the control parameters were decided
frequencies ( E 3 and E 2 ) are much more larger than from the former steps, we let the plant coefficients p
the coefficients of high frequencies ( E1 and 1); have a r 10% stochastic change, simulate the close-loop
system and record the result. By repeating it, 200 times
only the first two terms in the denominator of Eq. (20) are in this paper, we believe the results cover most of the
often sufficient to describe the character that z3 tracks f . case under parameters perturbation. Then, we can see
Thus, we have whether the controller is robust or not).

z3 ( s ) k Note that, since the desired setting time is given, only one
| (21) parameter B , need to be tuned.
f (s) s  k
Remark
where k { E 3 / E 2 .
d. For a wild range of process with the desired setting time is
With the knowledge of first-order systems, we can get from larger than 1 second, we make k { 4 in step (3). If the
Eq. (21) that the larger the k { E 3 / E 2 , the sooner the ESO. desired setting time is smaller than 1 second, we can
Furthermore, in accordance with the definition of 2% setting adjust the value of k in step (3) from Eq. (22) to enhance
time, we can get the speed that z3 tracks f . There is an example in section
5 for illustration.
Tt | 4 / k (22)
e. It is no doubt that there are always a conflict between
where Tt is defined as the time for z3 tracks f . In general, performance and robustness. The conflict in ADRC is not
the tracking time Tt of ESO should be smaller than the as strong as in many other controllers, which make ADRC
advanced. Therefore, the result in step (5) is generally
desired setting time t s of system. satisfactory. However, if a better robustness is expected,
we can go on increasing B in step (4), making a trade-off
Now, we go back to the Eq. (19), find that it results in between robustness and performance.
k Zo / 3 , thus the value of k is limited by Zo (note that,
in many cases, Zo must be in a small value to make the 5 Simulation Examples
system stable). To free k from Zo , we improve Gao’s We shall now look at some examples and demonstrate the
method as following: use of the method. Comparisons of both performance and
robustness will be made with Panagopoulos and Åström’s
E1 3Zo , E 2 3Zo2 , E 3 k E2 (23) two degrees of freedom PID (2DOF-PID)[14, 15]. The
following transfer functions have been considered
where k is a constant which can be conveniently determined
by the characteristic of the controlled process. For example, e  ps 1  ps
G1 ( s ) , p 5G2 ( s ) ,p 2
if the desired setting time t s is larger than 1 second, we can ( s  1) 3
( s  1)3
choose k 4 , which make Tt 1s according to Eq. (22). 1
G3 ( s ) , p 1
s ( s  1)3
Once k , the key parameter of ESO, is determined, we
1
choose Zo using a rule of thumb: G4 ( s ) ,
( s  p1 )(1  p2 s )(1  p3 s)(1  p4 s )
Zo 4Zc (24) p1 1, p2 0.2, p3 0.04, p4 0.008
1 1
Then E1 , E 2 and E 3 can be computed from Eq. (23). G5 ( s ) , p 1G6 ( s ) ,p 1
(s  p) 5
( s  1) 6
p
4.3 Tuning Procedure G7 ( s ) , p 4 G8 ( s ) e  p s , p 1
( s  p)( s  1)
The above development can be summarized into the
following procedure: The first two models describe process of 3-order. G1 models
*
(1) Get the desired setting time t . (we assume that t is * a process with long dead time and G2 is a non-minimum
s s

6325
phase process. Model G3 and G4 are 4-order process, in times and with right half-plane zeros, unstable and of
which G3 is an integrating process and G4 is a process with distributed parameters. It was shown that both satisfactory
performance and robustness can be obtained.
four different poles. Model G5 and G6 represent processes
of 5-order and 6-order respectively. Model G7 is considered
to show that the proposed method can also be used for
unstable systems. G8 is a distributed parameter system,
whose dynamic is described by partial differential equations
in nature. Models G1  G8 which represent processes with
large variations in process dynamic, are included to
demonstrate the wide applicability of the design procedure.
Firstly, we make model G1 as an example to illustrate the Fig.3 Comparison between ADRC (solid line) and 2DOF-PID
(dashed line), showing step response followed by load disturbance
design procedure specifically. As the desired setting time of lose loop system
ts* | 30 s ! 1s ,we have Zc 10 / ts* 0.33 | 0.4 .Then
Zo 4Zc 1.6 , k 4 . Thus, E1 , E 2 , E 3 and k p , kd can be
obtained from Eq. (11) and Eq. (23). The last tuning
parameter B is monotonously increased by a step of 1 from
B 1 . We can find that the dynamic performance is
satisfactory at B 3 . The tuning parameters we get can be
summarized as [ B, Zc , Zo , k ] [3, 0.4,1.6, 4] . The tuning
procedures of the other processes are similar. Note that,
model G8 is different from others for it’s desired setting
Fig.4 Comparison between ADRC (upside) and 2DOF-PID
time is smaller that 1s. Therefore, according to Eq. (22), we
(downside), showing the distribution of t s  V % under
increase the value of k to 20.
parameters perturbation
Fig. 3 shows the responses to changes in set point and load.
The details of the design calculations and simulations are Table1 Properties of ADRC and 2DOF-PID for system G1  G8
summarized in Table 1. We can see that the proposed
method yields a faster and smoother respond when
compared to Panagopoulos and Åström’s 2DOF-PID (There
were two sets of controller given in Panagopoulos and
Åström’s paper, and we choose the one which yields a faster
respond to compare). Especially, the overshoot of ADRC is
quite smaller. Furthermore, the control signal of ADRC is
smaller. Note that, the control problem of model G7 and G8
were not presented in Panagopoulos[14]. Therefore, the
ADRC controller obtained of them is compared to the
corresponding 2DOF-PI controller in Åström [15]. Then, the *The design parameters of ADRC and 2DOF-PID are
resulting performance of ADRC is great superior. [ B, Zc , Zo , k ] and [b, k p , ki , kd ] respectively.
The Monte-Carlo method was adopted to verify the
robustness of system. For a r10% stochastic change in References
p of models G1  G8 , we do simulations applying the same
[1] K. J. Åström, T. Hagglund The future of PID control.
controller designed for nominal system, and record the
Control Engineering Practice, 2001, 9(11):1163Ͳ1175.
values of ts  V % . This is repeated 200 times, and the
[2] J.HanAutoͲdisturbance rejection control and its applications.
results are shown in Fig. 4. The more concentrated the Control and Decision, 1998, 13(1): 19Ͳ23.
distribution of ts  V % , the better the robustness. The [3] Z Gao, Y. Huang, J. Han. An alternative paradigm for
bounds of the distribution, [ts , ts ] and [V  %, V  %] , are control system design. Proceedings of the IEEE Conference
listed in Table 1. We can see from Fig. 4 and Table 1 that, on Decision and Control, Orlando, FL, 2001: 4578Ͳ4585.
the robustness of ADRC is superior. [4] J. F. Pan, N. C. Cheung, J. M. Yang. AutoͲdisturbance
rejection controller for novel planar switched reluctance
motor. IEE ProceedingsϋElectric Power Applications, 2006,
6 Conclusions 153(2):307Ͳ316.
[5] H. Huanpao, L. Wang, J. Han, F. Gao, Y. Lin. A new
A simple tuning method for second-order ADRC is synthesis method for unit coordinated control system in
presented. Once the desired settling time is given, only one thermal power plant Ͳ ADRC control scheme. 2004
International Conference on Power System Technology ϋ
parameter need to be tuned. The method has been applied to
POWERCON (IEEE Cat. No.04EX902), Singapore, 2004:
a variety of systems, low- and high-order, with long dead 133Ͳ138.

6326
[6] J. B. Su, W. B. Qiu, H. Y. Ma, P. Y. Woo. CalibrationͲfree
robotic eyeͲhand coordination based on an auto
disturbanceͲrejection controller. IEEE Transactions on
Robotics and Automation, 2004, 20(5): 899Ͳ907.
[7] C. Tong, L. Wang, X. Yin, J. Dong. Finishing width control
system based on active disturbance rejection control and
chaos optimization. Proceedings of the World Congress on
Intelligent Control and Automation (WCICA), Dalian, China,
2006: 7751Ͳ7755.
[8] Y. L. Shi, C. Z. Hou. AutoͲdisturbanceͲrejection controller
design based on RBF neural networks. Intelligent Control and
Automation, 2006, 344: 500Ͳ505.
[9] Z. Gao. Scaling and BandwidthͲParameterization based
Controller Tuning. Proceedings of the American Control
Conference, Denver, CO, United States, 2003: 4989Ͳ4996. Fig.A1 Stable region of models Ga  Gd
[10] B. Sun, Z. Gao. A DSPͲbased active disturbance rejection
control design for a 1ͲkW HͲbridge DCͲDC power converter.
IEEE Transactions on Industrial Electronics, 2005, 52(5): A.2 Transfer Function Between z3 ( s ) and f ( s ) !
1271Ͳ1277.
Rewrite the Eq. (2) and Eq. (3) as following
[11] R. Miklosovic, Z. Gao. A robust twoͲdegreeͲofͲfreedom
control design technique and its practical application. z1 z2  E1 ( y  z1 ) (A1.1)
Conference Record ϋ IAS Annual Meeting (IEEE Industry z2 z3  E 2 ( y  z1 )  Bu (A1.2)
Applications Society), Seattle, WA, United States, 2004:
1495Ͳ1502.
z3 E 3 ( y  z1 ) (A1.3)
[12] X. Yao, Q. Wang, W. Liu, R. Liu. TwoͲorder ADRC control 
y f  Bu (A2)
for general industrial plants. Control Engineering China, Combine Eq. (A1.2) and Eq.(A2),and cancel the term
2002, 9(5): 59Ͳ62. Bu yielding
[13] I. Beichl, F. Sullivan. Monte Carlo methods. Computing in
f z3  E 2 ( y  z1 )  (  y  z2 ) (A3)
Science and Engineering, 2006, 8(2): 7Ͳ8.
[14] H. Panagopoulos, K. J. Åström, T. Hagglund. Design of PID Taking the Laplace transform of Eq.(A1.1) and Eq.(A3),we
controllers based on constrained optimization. IEE have
ProceedingsϋControl Theory and Applications, 2002, 149 sz1 ( s ) z2 ( s )  E1[ y ( s )  z1 ( s )] (A4)
(1): 32Ͳ40.
[15] K. J. Åström, H. Panagopoulos, T. Hagglund. Design of PI
f ( s) z3 ( s )  E 2 [ y ( s)  z1 ( s)]  [ s 2 y ( s)  sz2 ( s)] (A5)
controllers based on nonͲconvex optimization. Automatica, Combine Eq.(A4) and Eq.(A5), and cancel the term z2 ( s ) ,
1998, 34(5): 585Ͳ601.
yielding
Appendices f ( s) z3 ( s )  E 2 [ y ( s)  z1 ( s)]  E1 s[ y ( s)  z1 ( s)] 
(A6)
s 2 [ y ( s)  z1 ( s)]
A.1 Examples for Stable Region
Taking the Laplace transform of Eq.(A1.3),we have
2 s  1 z3 ( s ) s E 3 [ y ( s )  z1 ( s )] (A7)
(a ) Ga ( s ) 
( s  1)3 Times s to both sides of Eq. (A6), then substitute Eq. (A7)
1 into it, yielding
(b) Gb ( s )
( s  1)(1  0.2 s )(1  0.04s )(1  0.008s ) f ( s ) s ( E 3  E 2 s  E1 s 2  s 3 )[ y ( s )  z1 ( s )] (A8)
1 Divide Eq. (A7) by Eq. (A8), yielding
(c ) Gc ( s )
( s  1)5 z1 ( s ) E3
(A9)
100 1 0.5 f ( s ) E 3  E 2 s  E1 s 2  s 3
(d ) Gd ( s ) (  )
( s  1) s  1 s  0.05
2

All these results show that, as B increases within certain

bound, the stable area in Zo  Zc !plane is expanded. In fact,
for some special processes, there exists a constant value B0,
and the above law is just correct when B increases within the
range of [0, B0]. For example, in the case of Gd , the stable
area is no longer expanded, but contracted, while B>5.
However, in our tuning procedure, B shall not be too large,
and we consider that the law is always correct.

6327

View publication stats