Theory and Design of PID Controller
Lei GUO, Cheng ZHAO
Institute of Systems Science, AMSS,Chinese Academy of Sciences, Beijing
Hangzhou, April 22, 2017
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 1 / 44
Outline
1 Overview of PID Control
2 Mathematical Formulation
3 Theory and Design of PID
4 Concluding Remarks
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 2 / 44
A brief history of PID
Proportional feedback in the form of a centrifugal governor was usedto regulate the speed of windmills around 1750.
In 1788 James Watt used a similar system for speed control of steamengines.
The first mathematical analysis of a steam engine with a governor wasmade by Maxwell in 1868.
One of the earliest examples of a PID-type controller was intuitivelydeveloped by Elmer Sperry in 1911.
It was not until 1922 that PID controllers were analytically developedby N.Minorsky for automatic ship steering.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 3 / 44
The Impact of PID
Despite of the remarkable progresses of modern control theory over thepast half a century, the classical PID controller is still the most widelyused ones in engineering systems today.
As an example, 95% control loops are of PID type in process control,and most loops are actually PI control(Astrom and Hagglund,1995).
In 2016, IFAC publicized a survey conducted by a ”Pilot” IndustryCommittee launched by IFAC and chaired by Tariq Samad. The surveyshows that the PID control has much higher impact rating than other12 advanced control technologies, and ”we still have nothing compareswith PID”, see,
http://blog.ifac-control.org/
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 4 / 44
The structure of PID
Linear feedback structure of the form “present-past-future”:
u(t) = kpe(t) + ki
∫ t
0e(s)ds + kd
de(t)
dt
where e(t) = y∗ − y(t).
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 5 / 44
Why the PID so effective?
It is simple, model-free and easy-to-use.
It can eliminate steady state offsets via the integral action.
It can anticipate the tendency through the derivative action.
It has strong robustness w.r.t both system uncertainties and controllerparameters.
The well-known Newton’s law plays a fundamental role in modelingphysical systems.
... ... ...
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 6 / 44
Fundamental Theoretical Problems
How to properly design the PID parameters?
How to guarantee the desired control performance?
What is the maximum capability of PID feedback?
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 7 / 44
Classical Design Methods
Ziegler-Nichols method
Two classical methods for determining the parameters of PID controllerswere presented by Ziegler and Nichols in 1942. These methods are stillwidely used, either in their original form or in some modification.
It is based on some features of the process dynamics extracted fromexperiments, conducted by either the step response method or the fre-quency response method, for linear time-invariant systems.
The PID controller u(t) = K
(e(t) + 1
Ti
∫ t0 e(τ)dτ + Td
de(t)dt
).
See: Ziegler J G, Nichols N B,1942.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 8 / 44
Other Methods
Many other methods including tuning and adaptation for the design of thePID parameters have also been proposed but mainly for linear systems.
References
Astrom K J, Hagglund T. ( 1995,2006)
Blanchini F, Lepschy A, Miani S, et al. (2004)
Hara S, Iwasaki T, Shiokata D. (2006)
Ho M T, Lin C Y. (2003)
Keel LH, Bhattacharyya S P. (2008)
Killingworth N J, Kristic M.(2006)
O’Dwyer. (2006)
Silva G J, Datta A, Bhattacharyya S P.(2005)
Soylemez M T, Munro N, Baki H.(2003)
......
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 9 / 44
Understanding PID:Uncertainty, nonlinearity and feedback
To understand PID, we have to face with uncertainties and nonlineari-ties, because they always exist in practical systems.
Basic questions: why it is so powerful? how much uncertainty can itdeal with?
As pointed out by Astrom and Hagglund( 1995,2006), better under-standing of PID control may improve its widespread practice, and socontribute to better product quality.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 10 / 44
Description of Uncertainty
Uncertainty is mathematically described by a set F , either parametricor functional.
The control of uncertain systems is by definition the control of allpossible systems related to this set.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 11 / 44
The Maximum Capability of Feedback
Consider the following control system:
yt+1 = f (yt) + ut + wt+1, y0 ∈ R
with f ∈ FL where
FL = f : R→ R∣∣supx 6=y
|f (x)− f (y)||x − y |
≤ L
L: Serves as a measure of uncertainty
Theorem(Xie-Guo,2000). The above class of uncertain nonlinear dynamicalsystems described by FL is globally stabilizable by the feedback mechanismif and only if
L < 32 +√
2
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 12 / 44
Framework of PID Theory
Following a similar theoretical framework as the investigation of the maxi-mum capability of the feedback mechanism:
The maximum capability of feedback is defined by the largest possibleclass of nonlinear functions that can be dealt with by the feedbackmechanism.
The size of the uncertain functional class is characterized by the corre-sponding Lipschitz constant.
See:
Xie L L, Guo L. IEEE Trans Automat Control, 2000.
Guo L. Plenary Lecture at the 19th IFAC World Congress, Cape Town,2014.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 13 / 44
Mathematical Formulation
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 14 / 44
Mathematical Formulation
Background
The Newton’s second law plays a fundamental role in modeling dy-namical systems of the physical world, which is actually a second orderordinary differential equation of the position of a moving body.
PID control is sufficient for processes where the dominant dynamics areof the second order. For such processes there are no benefits gained byusing a more complex controller. (Astrom K J, Hagglund T. 1995)
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 15 / 44
Mathematical formulation
Consider a moving body of unit mass in R which is regraded as acontrolled system.
x(t), v(t), a(t) are its position, velocity and acceleration at the timeinstant t
Assume that the external forces acting on the body consist of f and u.
f = f (x , v , t) is a nonlinear function of the position x , velocity v andtime t and u is the control force.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 16 / 44
The equation of motion
ma(t) = f (x(t), v(t), t) + u(t)
Objective:
To understand when and how the PID controller can guarantee that theposition converges to a given constant reference value y∗ for any initialposition and any initial velocity.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 17 / 44
State space equation
Denote x1(t) = x(t) and x2(t) = dx(t)dt =
.x(t), then the state space equa-
tion of this basic mechanic system under PID control is
.x1 = x2.x2 = f (x1, x2, t) + u(t)
u(t) = kpe(t) + ki∫ t0 e(s)ds + kd
de(t)dt
(1)
where x1(0), x2(0) ∈ R and e(t) = y∗ − x1(t).
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 18 / 44
Theory and Design of PID
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 19 / 44
The Class of Uncertain Functions
Define a functional class:
FL1,L2 =
f ∈ C 1(R2 × R+)
∣∣∣∣ ∂f
∂x1≤ L1, |
∂f
∂x2| ≤ L2,∀x1, x2 ∈ R, ∀t ∈ R+
where L1 and L2 are positive constants, and C 1(R2 × R+) denotes thespace of all functions from R2 × R+ to R which are locally Lipschitz in(x1, x2) uniformly in t and piecewise continuous in t, with continuous partialderivatives with respect to (x1, x2).
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 20 / 44
The Parameter Manifold
Denote(kp, k i , kd) = (kp − L1, ki , kd − L2)
and introduce
Ωpid =
kpkikd
∣∣∣∣kp > 0, k i > 0, kpkd > k i + L2
√k i (kd + 2L2)
(2)
Theorem 1: Assume that f ∈ FL1,L2 and that f (y , 0, t) = f (y , 0, 0) for allt ∈ R+ and y ∈ R. Then, whenever (kp, ki , kd) ∈ Ωpid , the PID controlledsystem (1) will satisfy
limt→∞
x1(t) = y∗, limt→∞
x2(t) = 0
for any (x1(0), x2(0)) ∈ R2 and any setpoint y∗ ∈ R.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 21 / 44
An Illustration: L1 = 5 and L2 = 5
The set Ωpid when restricted to the domain 0 ≤ kp, ki , kd ≤ 50.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 22 / 44
Remarks
The selection of the PID parameters has wide flexibility(Ωpid is open and unbounded).
Theorem 1 gives a global convergence result.
The selection of the PID parameters does not depend on the initialstates and the setpoint y∗.
PID controller has strong robustness with respect to uncertain nonlinearfunctions and to the selection of parameters.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 23 / 44
Remarks
L1 and L2 represent the “anti-stiffness” coefficient and the “anti-damping”coefficient of the nonlinear system, respectively.
For any kp > L1 and kd > L2, we have (kp, ki , kd) ∈ Ωpid for allsufficiently small ki > 0.
The results can be generalized by replacing the conditions on the partialderivatives with Lipschitz-like properties.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 24 / 44
Necessity of Parameter Manifold
If we have more constrains on the unknown function f (x1, x2, t), such as f
is independent of t and ∂2f∂x22
= 0, then we can find a larger and necessary
parameter manifold to stabilize the system.
Examples:
f is of the form f (x1, x2, t) = a(x1) + b(x1)x2.
f is merely a function of the variable x1, i.e., the open-loop system isconservative.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 25 / 44
Definitions
Let us introduce the following functional class,
GL1,L2 =
f ∈ C 2(R2)
∣∣∣∣ ∂f
∂x1≤ L1,
∂f
∂x2≤ L2,
∂2f
∂x22
= 0, ∀x1, x2 ∈ R,
where L1 > 0, L2 > 0 are constants and C 2(R2) is the space of twicecontinuously differentiable functions from R2 to R.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 26 / 44
Proposition 1
Assume that f ∈ GL1,L2 does not depend on time t. Then for any f ∈ GL1,L2
and any setpoint y∗ ∈ R, the closed-loop system (1) satisfies
limt→∞
x1(t) = y∗ limt→∞
x2(t) = 0
if and only if the PID parameters (kp, ki , kd) lie in the following 3-dimensionalmanifold:
Ω′pid =
kpkikd
∣∣∣∣kp > 0, k i > 0, kpkd > k i
(3)
where (kp, k i , kd) = (kp − L1, ki , kd − L2).
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 27 / 44
An Illustration: L1 = 5 and L2 = 5:
The set Ω′pid when restricted to the domain 0 ≤ kp, ki , kd ≤ 50.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 28 / 44
PD control
When (y∗, 0) is an equilibrium point of the open-loop systems, i.e. f (y∗, 0) =0, the I-term is not necessary for regulation.Define a functional class FL1,L2,y∗ as follows,
f ∈ C 1(R2)
∣∣∣∣ ∂f
∂x1≤ L1,
∂f
∂x2≤ L2, ∀x1, x2, f (y∗, 0) = 0
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 29 / 44
Theroem 2: Consider the PD controlled system.x1 = x2.x2 = f (x1, x2) + u(t)
u(t) = kpe(t) + kdde(t)dt
(4)
where the unknown f ∈ FL1,L2,y∗ . Then for any f ∈ FL1,L2,y∗ , we have
limt→∞
x1(t) = y∗, limt→∞
x2(t) = 0
if and only if the PD parameters (kp, kd) lie in the following 2-dimensionalmanifold:
Ωpd =
(kp, kd)
∣∣∣∣kp > L1, kd > L2
. (5)
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 30 / 44
A Generalization
The next theorem is a generalization of Theorem 2, where the second statevariable is not the derivative of the first in general.
Consider the following uncertain nonlinear system with unknown f = (f1, f2) ∈C 1(R2 → R2),
.x1 = f1(x1, x2).x2 = f2(x1, x2) + u(t)
u(t) = kpe(t) + kdde(t)dt
(6)
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 31 / 44
Definition
Define a functional class GL1,L2,y∗ ⊂ C 1(R2 → R2) as follows,f =
(f1f2
) ∣∣∣∣ ∂f1∂x2
> 0,−(∂f1∂x2
)−1 det(Df ) ≤ L1, (∂f1∂x2
)−1tr(Df ) ≤ L2, f (y∗, 0) = 0
,
where det(Df ) is the determinant of the Jacobian matrix of f defined by
Df =
∂f1∂x1
∂f1∂x2
∂f2∂x1
∂f2∂x2
and tr(Df ) is the trace of Df defined by
∂f1∂x1
+∂f2∂x2
.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 32 / 44
Theorem
Let the unknown f ∈ GL1,L2,y∗ , and u(t) is the PD control:
u(t) = kpe(t) + kd.e(t), e(t) = y∗ − x1(t).
Then for any f ∈ GL1,L2,y∗ , the closed-loop system satisfies
limt→∞
x1(t) = y∗ limt→∞
x2(t) = 0
for any initial value (x1(0), x2(0)) ∈ R2 if and only if the PD parameters
(kp, kd) ∈ Ωpd =
(kp, kd)∣∣kp > L1, kd > L2
.
Remark. If f1(x1, x2) = x2, then the functional class GL1,L2,y∗ reduces toFL1,L2,y∗ .
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 33 / 44
Markus-Yamabe Conjecture(or Jacobian Conjecture)
Let f ∈ C 1(Rn,Rn), f (0) = 0. Consider the following n-dimensional au-tonomous differential equation,
.x = f (x)
If for any x ∈ Rn, the eigenvalues of the Jacobian matrix ∂f (x)∂x of f at x
have negative real parts, then it is conjectured that the zero solution of thedifferential equation is globally asymptotically stable.
Markus-Yamabe Theorem: The above conjecture is true for n = 2.
References
Markus L, Yamabe H. Osaka Math J, 1960.
Feßler R, Ann Polon Math, 1995.
Chen P N, He J X, Qin H S. Acta Math Sin, 2001.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 34 / 44
First order systems
Finally, it is worth mentioning that for first order systems, PI control issufficient. The next proposition gives a rigorous description. Define
FL = f ∈ H(R×R+) : |f (x , t)−f (y , t)| ≤ L|x−y |, ∀x , y ∈ R, ∀t ∈ R+,
where L > 0 is a constant and H(R × R+) is the space of functions fromR× R+ to R, which are piecewise continuous in the second variable t.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 35 / 44
Proposition 2
Consider the following first order nonlinear system
.x = f (x , t) + u
where the unknown f ∈ FL and the PI control is defined by:
u(t) = kpe(t) + ki
∫ t
0e(s)ds.
Then for any f ∈ FL and any setpoint y∗ satisfying f (y∗, t) = f (y∗, 0)for all t ∈ R+, the closed-loop system is globally stable and satisfieslimt→∞ x(t) = y∗ if and only if the PI parameters lie in the following2-dimensional manifold:
Ωpi = (kp, ki ) ∈ R2∣∣kp > L, ki > 0.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 36 / 44
Uncertainty in control channel
If we only know the upper bound M of the mass of the moving body, then thecontrol channel would contain an unknown parameter, say b, where b = 1
mis an unknown positive constant with a known lower bound b = 1
M > 0.We assume the unknown disturbance F is proportional to the mass m.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 37 / 44
Proposition 3
In this case, the state space equation in Theorem 1 under PID control is.x1 = x2.x2 = f (x1, x2, t) + bu(t)
u(t) = kpe(t) + ki∫ t0 e(s)ds + kd
de(t)dt
(7)
where unknown f ∈ FL1,L2 and unknown b ≥ b > 0. Then for any L1, L2 >0, the closed-loop system will satisfy
limt→∞
x1(t) = y∗, limt→∞
x2(t) = 0
for any initial value (x1(0), x2(0)) and any constant setpoint y∗ ∈ R if theparameters (bkp, bki , bkd) ∈ Ωpid .
Remark All the above results remain to be true as long as (bkp, bki , bkd)are chosen from the corresponding manifolds.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 38 / 44
Concluding Remarks
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 39 / 44
What we have done
We have presented a mathematical theory together with a design methodfor the well-known PID controller of a basic class of second order non-linear uncertain dynamical systems.
We have investigated several related issues including global stabilizationand asymptotic regulation.
The PID design rules given in this paper is quite simple and is almostnecessary for global stabilization.
Both our theory and design methods demonstrate that the PID con-troller is indeed quite robust with respect to both the design parametersand the nonlinear uncertainties.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 40 / 44
The limitations of PID
The above theoretical results may not be true in the following cases:
The nonlinearity has a growth rate “faster” than linear growth. Forexample, f (x1, x2) = (x2
1 + x22 )δ, where δ > 1
2 .
Systems described by differential equations of order ≥ 3..x1 = x2
· · ·.xn = f (x1, · · · , xn) + kpe(t) + ki
∫ t0 e(s)ds + kd
.e(t)
(8)
even if f (x1, · · · , xn) is linear, known.
ReferenceZhao C, Guo L, 2017. To appear in 2017 IFAC World Congress.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 41 / 44
Some generalizations
Similar results can also be established for PID controlled nonlinear un-certain stochastic systems.
Any n−dimensional nonlinear uncertain system of the form.x1 = x2.x2 = f (x1, x2) + u
u = kpe(t) + ki∫ t0 e(s)ds + kd
.e(t)
can be stabilized globally by PID, as long as the nonlinearity satisfiesa global Lipschitz condition.
References
Cong X R, Guo L, 2017. Submitted to 56th IEEE-CDC, 2017.
Zhao C, Guo L, 2017. To appear in 2017 IFAC World Congress.
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 42 / 44
Some future problems
To extend the results and methods on PID to more general nonlinearuncertain systems.
To improve the existing results on control of uncertain nonlinear sys-tems in the literature, by either improving the structure of PID or usingthe analytical methods developed here.
To investigate under what additional conditions, the Jacobian Conjec-ture is also true for high-dimensional systems.
To consider more complicated situations such as time-delayed inputsand sampled-data PID controllers under a prescribed sampling rate,and to connect the related boundaries established for the maximumcapability of the general feedback mechanism.
.........
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 43 / 44
THANK YOU!
(This lecture is mainly based on the authors’ paper published bySCIENCE CHINA-Information Sciences, Feb.2017)
Lei GUO, Cheng ZHAO (AMSS) Theory and Design of PID Controller 2017 44 / 44