You are on page 1of 6

Proceedings of the 2009 IEEE International Conference on Networking, Sensing and Control, Okayama, Japan, March 26-29, 2009

Two Degree-of-Freedom of Generalized Predictive Control Based on Polynomial Approach Using a Genetic Algorithm
Akira YANOU
Abstract Generalized Predictive Control (GPC) achieves a robust tracking for step-type reference signal by including an integrator in advance. Although author has proposed a design scheme of two degree-of-freedom GPC system which reveals an effect of integral compensation only if there exists modeling error or disturbance, a gain for integral compensation must be selected by trial and error. In this paper, a new scheme of two degree-of-freedom of GPC system based on polynomial approach is obtained by using a genetic algorithm for selection of integral gain.

control based on polynomial approach for single-input single-output systems[8], [9], [10], there is a problem for a gain of integral compensation. That is, a gain for integral compensation, which designs for the disturbance response, must be selected by trial and error. Therefore this paper newly proposes two degree-of-freedom of GPC by using a genetic algorithm[11], [12], [13] for selection of integral gain. By using a genetic algorithm, a gain for integral compensation can be determined. The proposed method is designed by the following steps. First, although GPC strategy proposed by Clarke and others has an integral action by including an integrator in the performance index, it makes the strategy proposed in this paper difcult, because in this paper the amount of the integral action must be calculated analytically on condition that there does not exist a modeling error and a disturbance. Therefore the controller is rst designed with no integral action in the performance index. Second, a new controller is designed by adding an integral action to the controller in the rst step. The effect of its integral action can be designed by a gain introduced in the controller at once. This controller always reveals the integral action and it may have an extra integral compensation because the integral action is merely added to the controller designed in the rst step. In the third step, two degree-of-freedom GPC is obtained by calculating the amount of the integral for the controller designed in the second step and subtracting its amount from the control input designed in the second step. When there is no modeling error and disturbance, its controller generates the control input designed in the rst step, which has no integral compensation. And it reveals the effect of the integral compensation in the case that there exists modeling error or disturbance. Finally, a gain for integral compensation is selected by using a genetic algorithm, that is, the proposed controller is given. When there is neither modelling error nor disturbance, the proposed controller generates the same control input as one derived in the rst design step with no integral compensation. And the proposed controller shows an effect of integral compensation in the case that there exists modelling error or disturbance. This paper is consisted as the followings. In section II, two degree-of-freedom of GPC based on polynomial approach is shown. In section III, a search method of a solution of integral gain is shown by using a genetic algorithm. In section IV, numerical examples are shown to verify the validity of the proposed method. Finally the conclusion is given.

I. INTRODUCTION Generalized Predictive Control (GPC) has been rst proposed by Clarke and others[1]. The controller is derived by performance index, which includes parameters of prediction horizon, control horizon and weighting factor. These parameters are an interval of predicting a behavior of future output based on a nominal model, an interval of calculating optimal future inputs and a parameter on control input respectively. The control signals are derived by minimizing the performance index on future control inputs and re-calculated receding from their horizons at each sampling time. With these features, the control strategy follows the steptype reference signal robustly by including an integrator in the controller and it has been accepted by many of practical engineers and many papers have been proposed[2], [3], [4], [5]. Whereas, if the controlled plant is modeled accurately and there is no step-type disturbance, the controller can track to the step-type reference signal without an integrator in it. And the effect of an integral has the possibility of the change for the worse of the characteristic of the transient response or the increase of the control input. Therefore it is desirable that an effect of an integral compensation appears only if there exists a modeling error or a disturbance. In this paper, this feature is dened as two degree-offreedom system because the characteristic of the output response and the disturbance response can be designed independently, that is, on one hand the characteristic of the output response is designed by minimizing the performance index which includes a control input and a tracking error between the reference signal and the output, on the other hand the characteristic of the response by the disturbance is designed by the gain of the integral compensator. Although many papers have proposed two degree-offreedom optimal servo systems[6], [7] and the author has proposed two degree-of-freedom of generalized predictive
A. 1, YANOU is with School of Eng., Kinki University, takayaumenobe, Higashi-Hiroshima 739-2116, JAPAN

yanou@hiro.kindai.ac.jp

978-1-4244-3492-3/09/$25.00 2009 IEEE

492

II. DESIGN SCHEME OF TWO DOF OF GPC Two degree-of-freedom of GPC is designed by three steps. In the rst step GPC is designed without an integral action in the performance index. The controller given by this step can not achieve the control objective because it does not include an integral action, that is, if there exists modeling error or a step-type disturbance, a plant output by the controller in the rst step can not track to a reference signal. Therefore, in the second step, the integral action is added to the controller given by the rst step. Then because of an integral action, the output by its controller can track to reference signal even if there exists modeling error or a step-type disturbance. But the integral action always acts, there may be an extra integral action, that is, the degradation of the transient response or the increase of the control input may occur. Therefore in the third step, by calculating the extra integral action and subtracting it from the controller in the second step, two degree-of-freedom of GPC controller is given. The controller in the third step appears the effect of an integral compensation only when there exists modeling error or a step-type disturbance. A. PROBLEM STATEMENT Consider a single-input single-output system, A[z
1

where the deviations y (t) and u(t) are dened as follows respectively. y (t) = y(t) y u(t) = u(t) u The prediction for (3) can be derived by the Diophantine equation, 1 = A[z 1 ]Ej [z 1 ] + z j Fj [z 1 ] Ej [z 1 ]B[z 1 ] = Rj [z 1 ] + z j Sj [z 1 ] where Ej [z 1 ] Fj [z 1 ] Rj [z 1 ] Sj [z 1 ] =
j j j = f0 + f1 z 1 + + fn z n

1 + e1 z 1 + + ej1 z (j1)

= r0 + r1 z 1 + + rj1 z (j1)

= sj + sj z 1 + + sj z (m1) 0 1 m1

In order to nd j-ahead prediction y (t + j|t), j-ahead output y (t + j) is derived. y (t + j) where hj (t) u = Rj [z 1 ](t + j km ) + hj (t) Fj [z 1 ](t) + z km Sj [z 1 ](t) y u

]y(t) = z

km

B[z

]u(t)

(1)
1

where y(t) and u(t) denote the output and the input. A[z ] and B[z 1 ] are the n-order and m-order polynomials respectively. A[z 1 ] B[z 1 ] = 1 + a1 z 1 + + an z n = b0 + b1 z 1 + + bm z m 1 e(t) 1 z 1

Because it is assumed that there exists no perturbation, j ahead prediction is derived as y (t + j|t) = y (t + j). Then the vector form of output prediction at each sampling time and the series of future input, and the matrices R and H are dened as follows. Y = RU + H (4) where Y U H R = [y (t + N1 |t) y (t + N2 |t)]T = [(t) u(t + Nu 1)]T u = [hN1 (t) hN2 (t)]T r0 0 0 . .. . = . 0 . rN2 1 r0

The integrator w(t) is given by the following equation, w(t) where = = (2)

A tracking error is a signal between the reference signal r and the output, and given as follows. e(t) = r y(t)

z 1 denotes backward shift operator: z 1 w(t) = w(t 1). The control objective is that the output y(t) tracks the reference signal r. B. GPC WITHOUT INTEGRAL COMPENSATION At rst the prediction y (t) is given for the deviation system of the plant (1). The steady state values y of y(t) and u of u(t) are derived as follows. A[z 1 ]y = z km B[z 1 ]u

[N1 , N2 ] and [1, Nu ] denote the prediction horizon and the control horizon respectively. To simplify, in this paper it is assumed that N1 = km = 1 and N2 = Nu . The performance index for the deviation system (3) is considered under the condition of y (t + j) = y (t + j|t).
N2

=
j=N1

y 2 (t + j) +

Nu j=1

2 (t + j 1) u

And the performance index can be rewritten by the following vector form. J = (RU + H)T (RU + H) + U T U (5) Minimizing the performance index J on U , the control law is derived. U = (RT R + I)1 RT H

From the previous equation of the steady state values, the deviation system of the plant (1) is obtained, A[z 1 ](t) = z km B[z 1 ](t) y u (3)

493

Because this equation is about the deviation system of the plant (1), the control law of the plant (1) is obtained as the following equation. u(t) = H0 (z where Fp [z 1 ] + (1 + z km Sp [z 1 ])K 1 + z km Sp [z 1 ] A[1] Fp [z 1 ] , K= F0 (z 1 ) = 1 + z km Sp [z 1 ] B[1] [pN1 , , pN2 ] = [1, 0, , 0](RT R + I)1 RT H0 (z 1 ) = Fp [z
1 N2 1

where T (z 1 ) is the transfer function of the closed-loop system. T (z 1 ) Tn [z 1 ] Td [z 1 ] = Tn [z 1 ] Td [z 1 ] = z km B[z 1 ]{Fp [z 1 ] +(1 + z km Sp [z 1 ])K} = A[z 1 ](1 + z km Sp [z 1 ]) +z km B[z 1 ]Fp [z 1 ]

)r F0 (z

)y(t)

(6)

If there exists no modeling error, then the tracking error e(t) is given by the following equation. e(t) = (1 T (z 1 ))r (9)

] =
j=N1 N2 j=N1

pj Fj [z 1 ] pj Sj [z 1 ]

Sp [z 1 ] =

Now the integral compensation z(t) is derived by (2) and (9). z(t) = w(t) 1 (1 T (z 1 ))r

C. GPC WITH INTEGRAL COMPENSATION When there exists modelling error or step-type disturbance, the control law (6) can not achieve the control objective and leaves steady state error between output and reference signal because it does not have an integral compensation. Therefore the integral compensation G0 w(t) is added to the control law (6), where we denote an integral action as w(t) and an integral gain as G0 . Then GPC controller with an integral action is derived as follows. u(t) = H0 (z 1 )r F0 (z 1 )y(t) + G0 w(t) (7)

By using the previous z(t), if there exists no modeling error or a step-type disturbance, z(t) is always equal to zero. That is, the effect of the integral compensation does not appear. III. SEARCH BY A GENETIC ALGORITHM In the previous section, the two degree-of-freedom of GPC is given. This section considers to give a solution of integral gain by a genetic algorithm[11], [12], [13]. In the rst step, each genotype of individual G0 represents in binary as strings of n bits. Each phenotype of individual G0 , that is, the integral gain G0 is calculated by converting binary representation to decimal representation and corresponding to a designed range. In the next step, the tness function to search the solution of G0 is dened as integral of square error between the reference signal and the output,
N

The control law derived here can achieve the control objective, that is, the output can follow to the reference signal even if there exists modelling error or step-type disturbance. D. TWO DOF CONTROLLER Because in the control law (7) the integral compensation always acts, it may cause an extra input or a worse change of the transient response. That is, the two degree-of-freedom of GPC is desirable, which has the effect of the integral compensation only if there exists modelling error or disturbance. The two degree-of-freedom of GPC proposed in this paper has the ability to adjust the characteristics of the output response and the disturbance response independently. In this subsection two degree-of-freedom of GPC controller is designed. At rst, the control law is expressed as follows by including an integral action z(t) and a gain G0 . u(t) = H0 (z 1 )r F0 (z 1 )y(t) + G0 z(t) (8) where H0 (z 1 ) and F0 (z 1 ) is the same coefcients as in (6). Second, assuming that there is neither modelling error nor disturbance, the tracking error e(t) = r y(t) is calculated. By this assumption, the control law expressed here becomes the same controller of (6). Then the closedloop system by the control law (6) is obtained, y(t) = T (z 1 )r

f (G0 ) =
t=1

{r yG0 (t)}2

(10)

where yG0 (t) is the plant output using G0 searched by a genetic algorithm. In this paper the solution of G0 is searched by a genetic algorithm so that the tness function f (G0 ) becomes small. Therefore a disturbance response is given in the sense of minimization of f (G0 ). And the genetic algorithm in this paper is calculated as following steps and repeated until a decided number of generations. 1) Initialization of population of genotype of G0 . The number of generation is set to be 1. 2) Evaluation of f (G0 ). It is repeated until a decided number of generations. 3) Creation of population of genotype of next generation by selection, crossover and mutation. 4) The number of generation is added to 1. Go to step 2.

494

IV. NUMERICAL EXAMPLE In this section the numerical examples are shown to verify the validity of the proposed method. Consider the controlled plant, z (1.2 0.1z ) u(t) 1 1.0z 1 + 0.25z 2 Simulation steps are 200, the initial values of the output and the input are assumed to be zero . The disturbance as the white gaussian noise with the variance 2 = 0.03 is added to the controlled plant output and each design parameter is given by follows. y(t) = N1 = 1, N2 = 10, Nu = 10, = 1 The reference signal is a rectangular signal with amplitude 1 and the period of 50 steps. The parameters of the genetic algorithm are given as follows. The number of individuals is set to be 50. The integral gain G0 is represented in binary as strings of 8 bits, and its phenotype is in the designed range. The number of generations is set to be 50. The selection is chosen to roulette rule and the crossover is one-point crossover. The probability of mutation is set to be 0.1. The numerical example gives the minimal value of tness function to 17.532 and the integral gain to 0.1914. Fig.1 shows the integral gain calculated by the genetic algorithm and tness value is shown in g.2. Fig.3 and g.4 show the plant output and input by the integral gain calculated by the genetic algorithm. It is shown that the proposed method can track to the reference signal by G0 calculated by the genetic algorithm. Fig.5 and g.6 show the comparison among the controllers (6), (7) and (8). The controlled plant, simulation steps, the initial values of the plant, design parameters and the integral gain G0 are the same as the previous example. The reference signal is a step-type signal with amplitude 1. The disturbance as a step-type signal with amplitude 0.5 is added to the controlled plant output after 100th step. In g.5 the solid line shows the output by the two degreeof-freedom of GPC (8) proposed in this paper, the dotted line shows the output by the GPC with integral compensation (7) as z(t) = w(t), and the broken line shows the output by the GPC without integral compensation (6). Fig.6 shows the control input and each type of line is the same as Fig.5. In Fig.5 the proposed method can track to the reference signal and the overshoot in the transient response can be controlled less than the controller (7) as z(t) = w(t). In addition the transient response proposed in this paper is the same as the one by (6). It is shown that the effect of integral compensation of the proposed method does not appear if there is no disturbance. And it is also found that the effect of integral compensation of it appears because there exists disturbance after 100th step. That is, two degree-of-freedom of GPC can be designed by the integral gain calculated by the genetic algorithm. V. CONCLUSION In this paper, a selection method of integral gain for two degree-of-freedom of GPC based on polynomial approach is
1 1

0.6

0.5

integral gain

0.4

0.3

0.2

0.1 5 10 15 20 25 generation 30 35 40 45 50

Fig. 1.

Integral gain

17.8

17.75

17.7

17.65
error

17.6

17.55

17.5

17.45

17.4 5 10 15 20 25 generation 30 35 40 45 50

Fig. 2.

Fitness value

given by using a genetic algorithm. And numerical examples are shown to verify the validity of the proposed method. Although the gain for integral compensation must be selected by trial and error in the conventional method, this paper gives its selection method. As future works, in the case that the proposed method is applied to a general plant, it is necessary to consider aged deterioration. That is, the integral gain must be re-calculated periodically or online. There are also an extension to multiinput multi-output systems of the proposed method and to design a self-tuning controller.

495

1.8 1

1.6

1.4 0.5 1.2


plant output plant output

0.8

0.6 -0.5 0.4

-1

0.2

0 20 40 60 80 100 step 120 140 160 180 200 20 40 60 80 100 step 120 140 160 180 200

Fig. 3.
1.5

Plant output
1

Fig. 5.

Plant output

0.8

0.5
plant input plant input

0.6

0.4

0.2 -0.5 0 -1 -0.2 -1.5 20 40 60 80 100 step 120 140 160 180 200 20 40 60 80 100 step 120 140 160 180 200

Fig. 4.

Plant input

Fig. 6.

Plant input

R EFERENCES
[1] D. W. Clarke, C. Mohtadi and P. S. Tuffs, Generalized Predictive Control, Automatica, Vol. 23, No. 2, pp.137-160, 1987 [2] H. Demircioglu and P. J. Gawthrop, Continuoustime Generalized Predictive Control, Automatica, Vol. 27, No. 1, pp.55-74, 1991 [3] B. Kouvaritakis, J. A. Rossiter and A. O. T. Chang, Stable generalised predictive control:an algorithm with guaranteed stability, Proc. IEE, Vol. 139, No. 4, pp.349-362, 1992 [4] E. F. Camacho and C. Bordons, Model Predictive Control in the Process Industry, Springer, 1995 [5] M.Deng, A.Inoue, A.Yanou and S.Okazaki, Stable anti-windup continuous-time generalised predictive control to a process control experimental system, The Journal of the Institute of Measurement and Control, Vol. 40, No. 4, pp.120-123, 2007 [6] Y. Fujisaki and M. Ikeda, Synthesis of Two-degree-of-freedom Optimal Servosystems, Transactions of SICE, Vol. 27, No. 8, pp.907-914, 1991

[7] T. Hagiwara, M. Ichiki, M. Kanaboshi, K. Fukumitsu and M. Araki, Digital Two-degree-of-freedom LQI Servo Systems -Design Method and Its Application to Positioning Control of a Pneumatic Servo Cylinder, Transactions of ISCIE, Vol. 11, No. 2, pp.51-60, 1998 [8] A. Yanou, S. Masuda and A. Inoue, A Design Scheme of TwoDegree-of-Freedom Generalized Predictive Control System in the Polynomial Approach, Preprints of 12th SICE Symposium in Chugoku Branch(In Japanese), pp.98-99, 2003 [9] Akira YANOU, Shiro MASUDA and Akira INOUE, Two Degreeof-Freedom of Self-Tuning Generalized Predictive Control Based on Polynomial Approach, Proceedings of SICE Annual Conference 2005 in Okayama, pp.1187-1190, 2005 [10] Akira YANOU, Shiro MASUDA and Akira INOUE, Two Degreeof-Freedom of Self-Tuning Generalized Predictive Control Based on Polynomial Approach with Computational Savings, Proceedings of 2006 International Conference on Dynamics, Instrumentation and Control, 2006 [11] T. Yamamoto, Y. Mitsukura and M. Kaneda, A Design of PID

496

Controllers Using a Genetic Algorithm, Transaction of SICE, Vol.35, No.4, pp.531-537, 1999 [12] H. Iba, The Basis of Genetic Algorithms(In Japanese), Ohmsha, 1994 [13] R. Ishida, H. Murase and S. Koyama, The Basis and Application of Genetic Algorithms(In Japanese), Morikita Publishing, 1997

497

You might also like