Design and Performance Comparison of Variable

Parameter Nonlinear PID Controller and Genetic

Algorithm Based PID Controller

Mehmet Korkmaz, Ömer Aydoğdu, Hüseyin Doğan

Department of Electrical and Electronics Engineering
Selcuk University
Konya-Turkey
[email protected], [email protected], [email protected]

Abstract— In this study, design and performance comparison of Genetic algorithms have placed a much stronger emphasis
variable parameter nonlinear PID (NL-PID) and Genetic than their counter-parts on global, as opposed to local, search
Algorithm (GA) based PID controller are achieved. To begin with and optimization. They could look into not only in local optima
the proposed method, an error function depending on the system but also in global points. GAs were initially formulated as
input and output are defined to determine variable coefficients of combinatory search algorithms, which required discretization
the nonlinear PID controller. A new type non linear PID of the search space in order to be applied to problems involving
controller is designed by using defined error function. By this real decision variables. Theoretical arguments based on the
way, the nonlinear PID controller changes its own parameters building-block hypothesis and the Schema Theorem [9-10] also
over time according to the output response. Secondly, genetic
appeared to support discretization.
algorithm based PID controller are defined to performance
comparison of the proposed NL-PID controller and Ziegler- In this paper, GA based PID parameters and NL-PID
Nichols PID controller. Simulation results show that the effects of parameters allied to change of error are analyzed and nonlinear
the PID controllers; nonlinear, GA based and Ziegler-Nichols. functions of Proportional (P), Integral (I) and Derivative (D)
depending on error and GA-PID parameters are presented
Keywords- genetic algorithm; variable parameter PID; Ziegler- respectively.
Nichols method; gaussian error function (erf).

II. DESIGN OF PID CONTROLLER, PERFORMANCE INDICES

I. INTRODUCTION
AND GENETIC ALGORITHMS
PID control schemes based on the classical control theory
A general body of a PID control system is shown in Fig.1,
have been widely used for various industrial control systems
where it can be seen that in a PID controller, the error signal
for a long time [1]. Generally, it is simply to determine
e(t) is used to generate the proportional, integral, and derivative
parameters and easy to apply them. Thus, PID controller is the
actions, with the resulting signals weighted and summed to
most common form of feedback. The controllers consist of in
form the control signal u(t) applied to the plant model. A
many different forms.
mathematical description of the PID controller is,
However when the system is more complex traditional PID
parameters do not efficiency to control system. So this reason,
nonlinear PID (NL-PID) controller is thought where the ⎡ 1 de(t ) ⎤
parameters are depending on system error amount. A nonlinear u (t ) = K p ⎢e(t ) + ∫ e(t )dt + Td ⎥ (1)
⎣ Ti dt ⎦
combination can provide additional degrees of freedom to
achieve much improved system performance [2-5]. Practically
all PID controllers made today are based on microprocessors. where u(t) is control input to the plant model, e(t) is error
This has given opportunities to provide additional features like which is difference between actual output (y(t)) and reference
automatic tuning, gain scheduling, and continuous adaptation input (r(t)), Kp is proportional gain, Ti is integral time constant
[1]. So, to improve the control quality, many scholars use and Td is derivative time constant.
nonlinear characteristics to modify traditional linear PID Due to system response is dependent on controller
controller in recent years [6]. The change characteristics of the parameters settling of these parameters is significant. It is
nonlinear functions accord with the ideal change process of the presented different approaches to the determination of the
parameters, so the nonlinear PID (NL-PID) can achieve both constants in the literature by researchers. In addition this,
good static and dynamic performances and improve the control nonlinear or optimization algorithms based convergence are
quality [7,8]. used.

978-1-4673-1448-0/12/$31.00 ©2012 IEEE

 Crossover: In this process, two kinds of chromosomes are
combined to obtain new chromosome. After that new
chromosomes can be better than both parents.
Mutation: It is a genetic operator that alters one or more
gene values in a chromosome from its initial state. This can
result in entirely new gene values being added to the gene pool.
This step is also important by the view of preventing the
population local optimal points. Genetic algorithm steps are
like this:
Figure 1. A typical PID control system
• Randomly generate certain amount of initial
population and calculate the fitness of each
A. PID Controller Design Methods individual.
Despite there are many design methods for PID controllers,
the most widely used design methods in the literature are • Select populations according to certain rules.
Ziegler-Nichols rules, Chien-Hrones-Reswick PID tuning • Mate the population selected in a certain probability
algorithm, Cohen-Coon tuning algorithm, Wang-Juang-Chan and generate new creatures.
tuning formulae.
• Investigate the optimal conditions, at the end of the
The Ziegler-Nichols design method which was presented in maximum iteration get the minimum values if they
mid-20th century is the most popular methods used in process provide bounds.
control to determine the parameters of a PID controller. One of
the important specialties of this system guarantees the Here, the objective function value is determined by means
stabilities [11]. Ziegler-Nichols tuning rule is the first such of mean square of error criteria (MSE).
effort to provide a practical approach to tune a PID controller.
Ziegler Nichols method (Z-N) is useful for plants of which C. Performance Indices
mathematical models are unknown or difficult to obtain. There In order to select the best controller, we define a cost
are two ways of implementing Ziegler-Nichols tuning rules and function. The cost function mainly derives on how the
controller parameters get from the Z-N empirical tables [12]. controller reacts to a given disturbance. There are many of cost
The Chien-Hrones-Reswick (CHR) PID tuning method functions. In fact, we can define in infinitive criterion. It is a
emphasizes the set point regulation or disturbance rejection. In quantitative measure of the performance of a system and is
addition one qualitative specification on the response speed and chosen so that emphasis is given to the important system
overshoot can be accommodated. Compared with the specifications. There are many type of performance indices is
traditional Ziegler-Nichols tuning formula, the CHR method described in literature.
uses the time constant T of the plant explicitly [13]. A performance index is criterion measures that are based on
Another tuning method of the PID controller is the Cohen– the integral of some function of the control error and on
Coon tuning formula. This method is used a rule table that is possibly other variables (such as time). Having the smaller
obtained empirically like as Ziegler-Nichols rule table. value of the integral criterion provides the better the
performance of the control loop. In this study IAE and ISE
Based on the optimum ITAE criterion, the tuning algorithm indices is used.
proposed by Wang, Juang and Chan is a simple and efficient
method for selecting the PID parameters. For example, IAE integrates the absolute error over time. It
does not add weight to any of the errors in a systems response.
B. Genetic Algorithms It tends to produce slower response than ISE optimal systems,
but usually with less sustained oscillation.
A GA is an optimization technique in which the solution
space is searched by generating a population of candidate The integral of the square error (ISE) penalizes for large
individuals to find best values [14]. This process is similar to errors more than for small errors. It tends to eliminate large
natural evolution of biological individuals. These are generally errors quickly. The response often has a smaller initial
having global search capability, better robustness and not overshoot, but the cycle does not decay rapidly. Basically
depending on initial conditions. These algorithms present several small peaks are tolerated to reduce the magnitude of the
excellent global optimization points to solve system first peak. This type of behavior is usually not desired in
optimization problem [15]. process loops [16].
Generally, GAs consist of three fundamental operators:
reproduction, crossover and mutation. III. PROPOSED METHODS
Considering the above classical methods, they present poor
Reproduction: It is a process by which the strings with robustness, high overlapping, late rise time etc. Thinking of
larger fitness values can produce accordingly with higher fixed parameters causes this result in steady state and
probabilities large numbers of their copies in the new temporary state. In defiance of these explanations, better
generation. system response is occurred when the PID parameters are
looked into depending on error function owing to they are Minimum and maximum value of Kp is a1 and (a1+a2)
variable. Correspondingly, system responds better than respectively. All parameters are positive real constant that
traditional PID controller methods when the PID fixed depends on an error function.
parameters determine with genetic algorithms based on
objective functions or designing them as a nonlinear form. The derivative gain parameter (Kd) provides to reduce
oscillation and exceeding. Counter to this, causes the slowdown
Considering the step response of a common control system the system response.
we need to decrease overlapping and oscillation, accelerate the
system response and initialize the steady state error. For these In a similar vein Kd can be formulized in the equation of the
conditions the parameters can be analyzed like this way. (6) (Fig. 2b).

The proportional gain (Kp) contribute to accelerate system

response, decrease the settling time however increase the K d = c1 + f (e) * c 2 (3)
oscillation and in the large values system becomes unstable. As
a simple we can model the proportional gain parameter (Kp) by Similarly, range of Kd is from c3 to (c3 + c4) and both of
this formula (Fig. 2a). value is real positive constant.
Integral gain parameter (Ki) yields to zeroize the steady
K p = a1 + f (e) * a 2 (2) state error but for the big values of this parameter induces the
oscillation and higher overshoot (Fig. 2c).

Figure 2. Curves of Kp, Kd and Ki to Absolute of Error

In respect of explanations above, it is formulized this way: In this study, the absolute of error was used to prevent the
negative values of PID parameters. Therefore it was utilized
from only positive values of the function.
K i = b1 − f (e) * b2 (4)
IV. SIMULATION RESULTS
As it thought from this analysis, we may provide the quick The control methods mentioned above was examined with
response system without overlapping and short settling time the classical common method, Ziegler – Nichols and
basically if the parameters set like above depending on the comparison of some results can be seen on Table I. These
change of error with an appropriate function. methods and Z-N PID values were compared to some criterion
It is benefitted from “error function” (also called Gaussian and implementing methods values were better than Ziegler –
error function) to determine NL-PID variable coefficients (see Nichols method.
in Fig.3). In this study it was benefited from the below Matlab-
Simulink model in Fig. 4. For this simulation two different
x plants model tested. The coefficients (ai, bi, ci) shown in the
2
∫ Table II were obtained for both application. Steady state error
2
erf ( x) = e −t dt (5)
∏0 and performance indices were measured via the block diagram
and system response curves were figured for both. Two kind of
plant was used in this paper which is third order process and
fourth order system (Eq. 6, 7). Output responses and changing
of NL-PID parameters were showed in Fig. 5.a - 5.b. and Fig.
6.a - 6.b for both two plant model respectively.
1 (6)
G1 ( s ) =
s( s 2 + s + 1)( s + 2)
1 (7)
G2 ( s) =
Figure 3. Error Function s ( s + 2)( s + 4)
 Figure 4. Matlab Simulink Mode

TABLE I. COMPRAMISION VALUES BETWEEN IMPLEMENTED METHODS

First Plant ess IAE ISE
Z-N PID -0.0052 6.071 3.516
NLPID -0.0490 4.791 2.417
GA-PID -0.0150 3.569 2.388
Second Plant
Z-N PID -0.00148 1.655 0.5768
NLPID -0.01800 1.180 0.5440
GA-PID -0.00131 0.788 0.5768

TABLE II. COEFFICIENTS FOR THE NONLINEAR PID CONTROLLER PLANTS

a1 a2 b1 b2 c1 c2
First
Plant 0.933 0.233 0.061 0.061 0.897 0.897

Second
Plant 28.8 7.2 6.49 6.49 7.97 7.97
(a)

V. CONCLUSION
In spite of the fact that controllers designed by the Ziegler-
Nichols rules give a good performance, they create poor
robustness and high exceeding. It is obvious that in case of few
parameter changes of the plant led to decline of the
performance of the conventional PID controller drastically.
Thus, it is not enough to control process dynamics swimmingly
although it is a good start to tune PID parameters. Therefore, in
this paper to slacken the disfavor of classical tuning methods,
nonlinear PID approach and PID parameters based on GA were
searched. It was thought error function to determine of
nonlinear coefficients of the NL-PID. On the other hand, the
coefficients of GA-PID parameters were set with idea of the
objective function, called mean square error (MSE). As a result
of all these above, the methods mentioned in this paper were
compared considering the performance indices values and
steady state error. The methods presented gave lower exceeds,
(b) short settling time and better performance indices than that of
classical form for the third and fourth order systems.
Especially, genetic algorithm which converge the minimal
Figure 5. Output Response for Plants Respectively a) Plant-1 b) Plant-2
points used to obtain the best PID controller parameters. The REFERENCES
main advantage of it was to find optimal points of PID [1] K.J. Äström, T. Hägglund, Advanced PID Control, ISA-The
parameters and as it seen in the graphics, GA-PID results were Instrumentation, Systems, and Automation Society, 2006.
the best ones in both plants. [2] B. Armstrong, D. Neevel, T. Kusik, “New results in NPID control:
tracking, integral control, friction compensation and experimental
results”, IEEE Trans. Control Syst. Technol., vol. 9(2), pp. 399 – 406,
2001.
[3] B. Armstrong, B.A. Wade, “Nonlinear PID control with partial state
knowledge: damping without derivatives”, Int. J. Robotics Research,
vol. 19(8) , pp. 715 – 731, 2000.
[4] W.H. Chen, D.J. Balance, P.J. Gawthrop, J.J. Gribble, J. O’Reilly,
“Nonlinear PID predictive controller”, IEE Proc. Control Theory Appl.,
vol. 146 (6), pp. 603 – 611, 1999.
[5] J.Q. Han, “Nonlinear PID controller”, Acta Automatica Sinica, vol.
20(4), pp. 487 – 490, 1994.
[6] W. Wang, J.T. Zhang, T.Y. Chai, “A survey of advanced PID parameter
tuning methods”, Acta Automatica Sinica, vol. 26(3), pp. 347-355, 2000.
[7] J.J. Gu, Y.J. Zhang, D.M. Gao, “Application of Nonlinear PID
Controller in Main Steam Temperature Control”, Asia Pacific Power
(a) and Energy Engineering Conference, pp. 1-5, Wuhan, Chine, 2009.
[8] O. Aydogdu , M. Korkmaz “A Simple Approach to Design of Variable
Parameter Nonlinear PID Controller” International Conference on
Electrical Engineering and Applications, pp. 81-85, Chennai, India,
2011.
[9] J.H. Holland, Adaptation in Natural and Artificial Systems, The Uni-
versity of Michigan Press, Ann Arbor, 1975.
[10] D.E. Goldberg, Genetic Algorithms in Search, Optimization and
Machine Learning, Addison-Wesley, Massachusetts, 1989.
[11] K. Ogata, Modern Control Engineering, 5th ed., Prentice Hall, New
Jersey, 2010.
[12] J.G. Ziegler, N.B. Nichols, “Optimum Settings for Automatic
Controllers”, Trans. ASME, vol. 64, pp. 759-768, 1942.
[13] X. Dingyu, Q.C. Yang, P.A. Derek, Linear Feedback Control Analysis
and Design with MATLAB, SIAM Press, Philadelphia, 2007.
(b)
[14] J.Y. Wu, Y.K. Chung, “Real-Coded Genetic Algorithm for Solving
Generalized Polynomial Programming Problems”, Journal of Advanced
Figure 6. Change of NL-PID Parameters a) Plant-1 b) Plant-2 Computational Intelligence and Intelligent Informatics, vol.11(4), pp.
358-364, 2007.
[15] Z.W. Ping, D.Y. Chao, Y.D. Zhou, “Small Unmanned Helicopter
ACKNOWLEDGMENT Longitudinal Control PID Parameter Optimization Based On Genetic
Algorithm”, ICACTE 2010, vol. 6, pp. 142-145, Chengdu, Chine, 2010.
This work was supported by Selçuk University scientific
[16] C. Smith, Practical Process Control: Tuning And Troubleshooting, John
research projects coordinate (BAP). Wiley & Sons, New-Jersey, 2009.