Control Engineering Practice 21 (2013) 700707

Contents lists available at SciVerse ScienceDirect

Control Engineering Practice

journal homepage: www.elsevier.com/locate/conengprac

Variable-structure PID controller for level process

I. Boiko n
The Petroleum Institute, P.O. Box 2533, Abu Dhabi, United Arab Emirates

a r t i c l e i n f o
Article history: Received 30 March 2011 Accepted 18 April 2012 Available online 9 May 2012 Keywords: Process control Variable-structure systems PID control Optimization Performance evaluation

a b s t r a c t
A variable-structure (VS) PID controller for the level process is proposed. A methodology of analysis of its stability and performance is given. It is proposed that stability of the VS system can be approximately analyzed via the describing function method. The describing function of the VS PID controller is derived. It is shown that the system with the VS PID controller is quasi-linear. Tuning rules for the VS PI controller for the level process are given. It is shown via the theory and simulations presented that, if properly tuned, the VS PI controller has higher performance than the conventional PI controller for the process considered. & 2012 Elsevier Ltd. All rights reserved.

1. Introduction Liquid level control in various tanks and vessels is one of the most common controls in the process industry. It can usually be categorized into (a) the process control in which maintaining the level to a certain set point is the primary objective (boiler steam drums, bottom-product and reux drums of distillation columns just to name a few), and (b) the control in which large level uctuations are allowed and even assumed the case of so called surge vessels which accumulate feedstock from one or more sources and deliver a smooth feed rate. In the latter case the primary control objective is the outow stabilization. In the present paper, only the rst category of level control objectives will be considered. Normally level is controlled by a PI or PID controller, which can be implemented as a part of a distributed control system or locally. The controllers are tuned in accordance with established methods and techniques (Astrom & Hagglund, 1984,1995; Shinskey, 1988; Ziegler & Nichols, 1942). However, in many situations satisfactory performance can hardly be achieved. This happens due to the fact that level process is an integrating process, which in combination with the integral term of the PI/PID controller results in a double integrator in the loop. The presence of the controller integral term is absolutely necessary to ensure zero error in a steady state, and the use of a PI controller in combination with an integrating process usually results in oscillatory transients having low damping. In summary, level process is not as easy to control in terms of providing a good performance as it might seem.

Tel.: 971 2 607 5505; fax: 971 2 607 5194. E-mail address: i.boiko@ieee.org

PID controllers are used more seldom for the considered process than PI controllers because the performance improvement due to introduction of the derivative term is marginal while the derivative term would amplify the measurement noise. Therefore, in the application part of this paper our analysis and design are limited to the case of PI controllers. Yet, the use of a PID controller would differ from the presented analysis only by the tuning rules applied. The variable-structure (VS) control was proposed a few decades ago and was mainly developed as a sliding mode control (Utkin, 1992). There are a number of controllers described in the literature under the name of variable-structure PID controllers, which, however, dene a few different types of control. Let us use the term variable-structure understanding it as a type of the switching control in which switching occurs when the state trajectory moves from one region of the partitioned state space to another as it was described in Utkin (1992). The sliding mode may or may not occur in dependence on the designed switching strategy. Therefore, the following properties of the variable-structure system are assumed. A. It is a switching type of control (can also be viewed as parameters of the controller being changed in a discontinuous manner). B. The controller switching happens in dependence on the value of the state vector. C. The state space is partitioned into a few regions corresponding to a few different controllers, so that i-th controller is selected when the current value of the state vector belongs to the i-th region. It should be noted that various switching strategies are widely used in the control engineering practice. However, most of them are not variable-structure (VS) controllers considering the features given above. The switching usually is organized to provide better performance under varying parameter conditions, which is not much different from the gain-scheduling strategy

0967-0661/$ - see front matter & 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.conengprac.2012.04.004

I. Boiko / Control Engineering Practice 21 (2013) 700707

701

that has been used in practice for decades. Most of the described VS PID controllers utilize a certain switching strategy between a few different PID controllers or continuous change of controller parameters. A continuous change of controller parameters in dependence on the parameters of the input signal is used in Mainardi, Mantovani, Fabbri, and Bonfe (2003). Also, a continuous change of the parameters of the PID controller by neural/fuzzy logic (so that in certain modes the integral gain becomes zero) is proposed in Chen and Chang (1995). In Ismail and Bedwani (2001), Suyitno, Fujikawa, Kobayashi, and Dote (1993), a continuous change of the PID controller parameters is carried out by a fuzzy logic and a genetic algorithm, respectively, to optimize the transient tracking performance. In Balestrino, Biagini, Bolognesi, and Crisostomi (2009), a continuous change of the PID controller parameters is proposed, so that it becomes in fact the controller with nonlinear gains for every control component, which is similar to Shinskey (1988). A VS PID controller with a sliding mode is proposed in Jafarov, Parlakci, and Istefanopulos (2005), in which the proposed control is, in fact, a combination of an integral sliding mode and a conventional PID controls. In the paper (Hodel & Hall, 2001) and subsequent discussion offered in Hodel and Hall (2004) and Mantz and Battista (2004), a VS PID controller for counteracting the integral windup is proposed. The switching between the two controller structures is based on the comparison of the signals before the limiter (saturation nonlinearity) and after the limiter. In Qiu, Yuan, and Wang (2006), a VS PID controller with switches between P, PD and PID structures in dependence on the error magnitude was reported as benecial for reactor temperature control. Similar design was used in Zhang et al. 2010 to improve dynamics of positioning of a telescope. The brief overview given above shows that most of the controllers that are called VS PID are in fact not variable-structure systems (paper by Jafarov et al. (2005) is an exception) but different types of switching control logic or continuously changed controller parameters. In the present paper, a VS controller, in which switching between a few PID controllers occurs in dependence on the state vector value, is proposed. Switching strategy of this controller is in agreement with the denition of the VS system: see Utkin (1992) and comments above. The proposed switching strategy does not account for the changes of the plant (process) parameters. Instead, it assumes constant parameters of the process but provides a type of nonlinear control that allows for the enhancement of the system performance (in comparison with a conventional PID control) via exploiting nonlinear features of the control. It should be noted that the proposed controller does not generate a SM in the system despite the fact that it is a VS controller. Another distinction of the present results from the references given above is that the proposed VS control leads to a solid controller design methodology based on the process model. The design methodology of the VS PI controller is presented. The objective of this paper is, therefore, to develop a VS PI controller for the level process and a methodology of its parameters design or tuning. The paper is organized as follows. At rst the model of a PIcontrolled level process disturbed by ow change is considered. Then the describing function analysis is carried out for VS PID controller, and stability analysis is given. In the following section, the dynamics of the level loop having valve dynamics is analyzed. After that, tuning rules for a VS PI controller for level process are provided. And nally, a simulation example is given.

Fig. 1. The tank level process.

opposite: the inow is uncontrolled, and level control is done via manipulating the outow. Yet, the second situation can be transformed into the rst one via changing the ow signs. Let us assume that inow can manipulated through some linear dynamics, so that in a steady state the inow is proportional to the controller command. (Note: in practice, the valve opening is usually proportional to the controller command, but the ow is not necessarily proportional to the controller command and also depends on the upstream pressure. However, this dependence can be linearized by the use of the ow controller cascaded with the level controller, for example.) Write the equation of the process. _1 x 1 q q a in out 1

where x1 is the level value, qin is the controlled ow to the tank, qout is the uncontrolled ow from the tank, a is the cross-sectional area of the tank (it is assumed the tank has such geometry that a is constant). Let the process be controlled by a PI controller given by the following equation in the Laplace domain.   1 us K 1 es 2 Ts where u is control, K is the controller proportional gain, T is the controller integral time constant, s is the Laplace variable, e is the error (the difference between the level set point and the actual level value). At this point, let us consider that the control u produced by the controller is equal to the inow (no actuator-valve dynamics): qin u, that the outow is zero, and that the set point value is zero, so that e x. Rewrite Eqs. (1) and (2) in the normal form: ( _ 1 x2 x 3 K _ 2 Ta x1 K x a x2 Get rid of time in (3) via dividing the second equation by the rst one and obtain the equations of the state trajectories.   dx2 K 1 x1 1 4 a T x2 dx1 Depending on the parameters K and T of the controller, Eq. (4) can represent either an underdamped (oscillatory) process (Fig. 2) or an overdamped process (Fig. 3), with the origin being focus or node, respectively. To analyze advantages and drawbacks of each of the presented controllers with respect to the level control process, let us dene the control objectives. First, level controller is a regulator: the set point is usually constant; the main objective of the controller is to attenuate (reject) possible disturbances. Second, the only possible disturbance is the change of outow. This change is often an abrupt change due to connection or disconnection of consumers.

2. Simplied model of level process and VS principle The model of the level process can be schematically represented by a tank, which has a controlled inow and uncontrolled outow (Fig. 1). In many cased the actual arrangement is the

702

I. Boiko / Control Engineering Practice 21 (2013) 700707

Fig. 2. Phase portrait of PI controlled level process (underdamped).

Fig. 4. Phase trajectories for instantaneous outow decrease.

Table 1 Settling time and maximum level deviation. T Tf x1(tf) 0.1 0.452 0.252 0.3 0.739 0.379 0.7 1.05 0.495 1.0 1.21 0.546 5.0 2.15 0.762 10.0 2.66 0.835

distance between the point of intersection of the trajectory beginning in (0,x02) and of the horizontal axis, and the origin. This requires the solution of Eq. (3). Consider the case of l1 a l2. The analytical solution can be given as follows: 5 p _ 0, and l1 and where c1 c2 a=K x02 = 14a=TK , x02 x l2 are given above. The powers of the exponents are complex in the case of the underdamped process and real in the case of the overdamped process. Let us denote the time corresponding to the intersection of the horizontal axis by the trajectory tf. This time can be found as a solution of the equation _ t f c1 l1 el1 tf l2 el2 tf 0, x from which time tf can be derived as   1 l1 tf ln , l2 l1 l2 and consequently xt f c1 el1 tf el2 tf , 8 6 x c1 el1 t c2 el2 t ,

Fig. 3. Phase portrait of PI controlled level process (overdamped).

If, for example, the initial state is the equilibrium point (inow is equal to outow) then a decrease of outow would cause an instantaneous change of the state from (0,0) to (0,x02). The same would happen if the outow increased. Therefore, the typical situation that needs to be analyzed is the motion from the point (0,x02). The respective system trajectories for overdamped and underdamped processes are presented in Fig. 4. Third, the control objective is to minimize the effect of this disturbance, which is manifested as level increase (decrease) from the set point. The maximum level deviation corresponds to the distance between the point of intersection of the horizontal axis by the trajectory and the origin (Fig. 4). Fourth, another control objective is to ensure a smooth and possibly non-oscillatory (overdamped) transient. Let us analyze how those objectives (criteria) are related to the controller and process parameters. The closed-loop system is a linear second order system with the following characteristic polynomial: P l l
2

K K l , a Ta

p which has the following roots: l1 K =2a1 14a=KT , and p l2 K =2a1 14a=KT . The damping ratio of the closedp loop system is, therefore, x KT =4a. Let us also analyze how each point (0,x02) is mapped to the horizontal axis in terms of the

A number of calculations of the values of x(tf) for different controller integral time constant values, subject to x02 1, and the unity controller gain (K/a 1), that are depicted in Fig. 4, are presented in Table 1. Table 1 presents the comparison of the transient process characteristics for different values of T and the same values of K/a, which ensures the same initial trajectory slopes as per (4). By changing the controller gain K one can vary the transient characteristics of the process, including the possibility of reducing time tf and the maximum level deviation x(tf). However, rst, that would be an additional way of enhancement of the transient dynamics, and second, the model being considered is a simplied one, which does not account for the existence of the actuator-valve dynamics, and in practice the possibility of the controller gain increase is limited and normally fully utilized. The purpose of the analysis carried out is thus in showing the

I. Boiko / Control Engineering Practice 21 (2013) 700707

703

multiple Lyapunov functions (Branicky, 1998), and the nested Lyapunov functions presented in Adamy and Flemming (2004). Despite possible rigorous proof of stability, the Lyapunov analysis of stability always gives conservative estimates, while for the practice of process control it is always important to have such characteristics as stability margins. It is not possible to obtain any performance measure using the Lyapunov function approach. Therefore, in the present research, the describing function method is going to be used, which despite its approximate nature allows one not to simply prove the fact of stability but also quantitatively compute the stability margins, which can be related to the quality of the transient process. Let us derive the describing function (DF) (Atherton, 1975) of the VS PID controller. The PID controller equations are written below in the so-called expanded form as the following transfer function: W c12 s K p12 K i12 =s K d12 s,
Fig. 5. Phase portrait of level process (variable-structure).

trade-off between the damping properties of the transient process and the degree of the disturbance attenuation. One can see from Table 1 and the phase portraits Figs. 24 that each of those two different transients has some properties that are valuable to the level control process. The underdamped response has the points of intersection of the state trajectories with the real axis that are located closer to the origin than the ones for the underdamped response. Therefore, the maximum deviation of level from the set point due to a disturbance in the form of the instantaneous outow change will be smaller for this type of control. On the other hand, the non-oscillatory transient that corresponds to the overdamped response is much more suitable for the level process than the oscillatory response. Therefore, a tradeoff between the disturbance attenuation and the quality of the transient takes place. The use of the variable structure principle (Utkin, 1992) would, in our opinion, resolve the noted tradeoff allowing one to utilize the advantages of both controllers: the better disturbance attenuation provided by the underdamped control and the quality of the transient of the overdamped controller. If one creates a variable-structure controller such that takes the underdamped control if x1x2 4 0 and overdamped control if x1x2 r 0 then both useful properties of the two controllers would be utilized, and the phase portrait of the system would look as given in Fig. 5. In Fig. 5, the trajectories of the 1st and 3rd quadrants are from Fig. 2, and of the 2nd and 4th quadrants are from Fig. 3. One can see that the points of intersection with the real axis are the same as in Fig. 2, and after that intersection the process becomes overdamped. Therefore, the variable-structure controller given in Fig. 5 utilizes the valuable properties of both controllers considered above.

where Kp is proportional gain, Ki is the integral gain, and Kd is the derivative gain, the subscripts 1 and 2 refer to the rst and second PID controller, respectively. The switching between the rst and second PID controllers is done in accordance with the following switching logic: ( _ 40 W c1 sUss if ss us 10 _ r0 W c2 sUss if ss The switching condition in (10) is obviously equivalent to the _ x1 x2 . Due to additive form of the original one if r 0, as ss controller (9), the VS PID controller can be represented as a sum of the three components propagated through three separate switches (Fig. 6). Assuming that s(t) a sin ot a sinc nd the DF for each of the three components and the DF of the VS PID controller as a sum of the three components. First, nd the DF Np(a) for the proportional component as follows: ReN p a 1 pa
p

1 pa

Z 2p
0

up csincdc Z p
p=2

(Z p=2
0

K p1 asin2 cdc

K p2 asin2 cdc ) K p2 asin2 cdc K p1 K p2 2

Z 3p=2

K p1 asin2 cdc

Z 2p
3p=2

ImNp a

Z 2p 1 up ccoscdc pa 0 (Z Z p p=2 1 K p1 asinccoscdc K p2 asinccoscdc pa 0 p=2 ) Z Z

3p=2 2p

K p1 asinccoscdc

3p=2

K p2 asinccoscdc

K p1 K p2

3. Describing function analysis of VS PID controller Let us now develop a method of analysis and design of the VS PID system more suitable for practice than the presented consideration of the phase portraits. The matter is that in the analysis presented above no valve dynamics were accounted for, which as shown below have signicant effect on system performance. Account of additional dynamics increases the order of the model and makes the phase plane methods unsuitable for analysis. There have been a number of methods of stability analysis for VS systems, based on the Lyapunov theory: piecewise quadratic Lyapunov function (Johansson & Rantzer, 1998),

Fig. 6. VS PID controller.

704

I. Boiko / Control Engineering Practice 21 (2013) 700707

Therefore, the DF of the proportional component is K p1 K p2 K p1 K p2 j Np 2 p 11

written as the following quasi transfer function: NPID s Np Ni0 =s Nd0 s, where N i0 Nd0 K i1 K i2 K K i2 j i1 2 p K d1 K d2 K K d2 j d1 2 p 15 16 14

In the same way, nd the DF Ni(a) for the integral component as follows: Z 2p 1 ReNi a ui csincdc pa ( Z 0 Z p p=2 1 a a K i1 sinccoscdc K i2 sinccoscdc pa o o p=2 0 ) Z 2p Z 3p=2 a a K i1 sinccoscdc K i2 sinccoscdc
p

3p=2

1 K i1 K i2

Z 2p 1 ImNi a ui ccoscdc pa ( Z 0 Z p p=2 1 a a K i1 cos2 cdc K i2 cos2 cdc pa o o p=2 0 ) Z 2p Z 3p=2 a a 1 K i1 K i2 2 2 K i1 cos cdc K i2 cos cdc o o o 2 p 3p=2 Therefore,   1 K i1 K i2 K K i2 j i1 Ni jo 2 p

In formula (14), the DFs Ni0 and Nd0 pertain only to the switching gains without involvement of the integrator and differentiator, respectively. Stability analysis can be carried out on the basis of consideration of the open-loop transfer function, which would comprise the controller quasi transfer function (14) and the plant transfer function. In particular, the Nyquist criterion can be conveniently used for this purpose.

4. Level process model and VS PI controller tuning Consider now the dynamics of the actuator, of the valve and of the ow build-up (due to the valve position change) in the process model. The existence of these dynamics results in some lag in the inow with respect to the controller command. This lag can be relatively precisely modeled by the rst-order plus dead time (FOPDT) dynamics (Astrom & Hagglund, 1995; Boiko & Sayedain, 2010; Shinskey, 1988) given by the following transfer function: W a s ets Tas 1 17

12

Find the DF Nd(a) for the derivative component similarly to the above as follows: Z 2p 1 K K d2 ReNd a ud csincdc o d1 , pa 0 p Z 2p 1 K K d2 ud ccoscdc o d1 ImNd a pa 0 2 Therefore,   K K d2 K K d2 j d1 N d jo d1 2 p

13

Therefore, the detail model of the level system with a PI controller can be presented as the block diagram in Fig. 7. In Fig. 7, the controller parameters K and T are assumed to take two different sets of values depending on the states of the system as follows: ( ( _ 40 _ 40 K p1 if ss K p1 =K i1 if ss K , T _ r0 _ r0 K p2 if ss K p2 =K i2 if ss It was demonstrated above that the controller design should ensure a relatively fast response when the trajectory is in quadrants 1 and 3 of the phase plane x1Ox2, and a relatively slow response when the trajectory is in quadrants 2 and 4. Despite the fact that now the order of the systems is higher than two, partitioning of the state space still can be done as at with account of only the variables x1 and x2 that represent the principal dynamics. The other variables should, however, be accounted for in selecting the optimal controller parameters. If the dynamics of the actuator-valve did not exist then the two optimal PI controllers could be designed separately using the following criteria. In quadrants 1 and 3, the PI controller is to provide a minimal level deviation as a response to the step change of the outow, subject to the constraint of the stability of the closedloop system with the rst (fast) PI controller. (Note: the stability constraint is not necessary but in practical terms it is better to limit the controller gain). In quadrants 2 and 4, the PI controller is to provide minimum settling time of the transient process from the initial condition (x01,0), subject to the constraint on the non-oscillatory character of the transient process.

Analysis of the DF formulas (11)(13) leads to the following important conclusions: A. The VS PID controlled system is quasi-linear, as the describing functions do not depend on the amplitude; B. Through the switching gain mechanism, VS PID controller introduces in the open-loop dynamics either phase lead or phase lag; C. The value of this lead or lag depends only on the ratio of the switched gains; D. If the switched gains in each pair are equal the equations of the VS PID controlled reduce to the equations of the conventional PID controller; E. The DFs for the integral and derivative components are products of two factors: one is from the integrator or differentiator, respectively, which includes the frequency, and the other one is due to the switching gains, which depends on neither the frequency nor the amplitude; F. If K1 4 K2 then the switching gains give phase lead for the proportional component (due to the synchronous switching with s(t), which is the signal at the switch input) and phase lag for the integral and derivative components (due to the phase shift between the signal at the switch input and the switching signal). The DF of the VS PID controller is comprised of the component DF (11)(13), which, considering the quasi-linear property, can be

Fig. 7. Level control system.

I. Boiko / Control Engineering Practice 21 (2013) 700707

705

However, due to the presence of the dynamics of the actuatorvalve the inow always lags with respect to the controller command. In fact, when the command to decrease the valve opening comes to the actuator-valve (which corresponds to the point (x01,0)) the valve still continues to increase the opening due to the lag. This effect distorts the optimal settings of the controller if those were found separately for each partition. In that respect, the solution of the controller parameters optimization problem, when all four parameters are optimized together, while the switching conditions are determined by x1x2 4 0 and x1x2 r 0, is necessary. In this case a criterion can be formulated that would characterize the overall dynamics of the system. It was found that among the available integral performance criteria (IAE, ISE, ITAE, ITSE), the use of the minimal integral time by absolute error (ITAE) criterion allows for obtaining a properly damped transient for the level process, with the criterion given as follows: Z 1 Q ITAE 9x1 t 9t dt -min, 18
0

Table 3 ITAE optimal VS PI controller settings.

t0
K01 T01 K02 T02 ITAE

0.1 2.509 1.221 3.944 91.92 1.48

0.2 1.760 1.417 2.892 4 100.0 2.41

0.5 0.5744 1.75 0.7560 4 100.0 18.41

0.7 0.4841 2.059 0.6193 4 100.0 29.76

1.0 0.3563 2.777 0.4227 36.44 65.11

1.5 0.3242 3.819 0.4193 31.77 111.02

where the system is given by Fig. 7, the external input is the step change of the outow qout (t) 1(t), the error is equal to the negative output as the set point is zero. Let us nd the optimal controller settings for a certain normalized set of process parameters, after which the actual settings for each particular process can be computed by scaling of the normalized solution. Let the normalized transfer function of the open-loop system be as follows:   t0 s 1 e W ol s K 0 1 , 19 T 0 s ss 1 Formulate the optimization problem for the conventional PI controller as the problem of nding such values of K0 and T0 that minimize the cost function (18). The results of the optimization are given in Table 2, where the value of the cost function for the optimal set of the controller parameters is given too. The optimal settings are computed for a number of different values of time delay t0. Now let us solve the same optimization problem for the VS PI controller. The VS PI controller has four parameters: K01, T01 for the controller in quadrants 1 and 3, and K02, T02 for the controller in quadrants 2 and 4. The results of the optimization are given in Table 3, along with the value of the objective function for the optimal set of the controller parameters. Comparison of the ITAE value of Tables 2 and 3 shows that the VS PI controller provides a signicant improvement of the system performance in terms of the cost function values. Fig. 8 provides the plots of the transient response of the two controllers for time delay t0 0.5. The VS PI controller gains are as follows: Kp1 0.5744, Kp2 0.7560, Ki1 0.3282, Ki2 0.0076. The DF (quasi transfer function) of the PI controller is, therefore, as follows: NPI jo 0:6652j0:0578 0:1679j0:1021=jo and the open-loop frequency response is
  0:1679j0:1021 ejt0 o W jo N PI joW p jo 0:6652j0:0578 jojo 1 jo

Fig. 8. Response to step in outow for ITAE optimized system (t0 0.5).

Fig. 9. Nyquist plots for ITAE optimized system (t0 0.5).

The Nyquist plots of the open-loop system are given in Fig. 9. One can notice that in spite of smaller gain and phase margins

Table 2 ITAE optimal PI controller settings.

(see Fig. 9) the VS PI controller provides better transient response (see Fig. 8). Therefore, the VS PI controller is superior over the conventional PI controller in terms of providing more advantageous trade-off between the transient process quality and the quality of disturbance rejection. On the other hand, it may not be as robust as the conventional PI controller with respect to parameter variations, and proper tuning of the VS PI controller must be ensured to make the most of its advantageous features.

t0
K0 T0 ITAE

0.1 2.764 4.744 8.45

0.2 1.584 5.053 15.44

0.5 0.7925 6.153 42.62

0.7 0.6201 6.898 67.43

1.0 0.4803 8.045 116.55

1.5

5. VS controller design methodology

0.3562 9.922 234.87

The VS PI level controller design should include the following steps: state space partitioning and optimal parameter selection

706

I. Boiko / Control Engineering Practice 21 (2013) 700707

for the VS PI controller. State space partitioning also includes selection of some dead zone or boundary layer 2d around the vertical axis in Fig. 5 for noise protection, with the boundary layer width selected to be higher than the noise level. Therefore, the switching logic in implementation should be modied to ( _ 4 0 and jsj 4 d W c1 sUss if ss us : _ r 0 or jsj r d W c2 sUss if ss In the boundary layer, the selected controller is the slow one, so that the control does not respond aggressively to the uctuations induced by noise. Controller parameters selection is based on the data of Table 3 and recalculation formulas (scaling). It is assumed that the model of the process is given by the FOPDT plus integrator dynamics. Considering a, Ta, and t to be known values, the ITAE optimal controller parameters can be computed as follows. At rst the value of relative dead time computed as t0 t/Ta. After that the values of K01, T01, K02, T02 are selected from Table 3 for the calculated value of t0. The in-between values are computed via interpolation. To obtain the formulas for the scaled optimal solution use the substitution s0 T a s in the formula for the transfer function of the open-loop system, where s0 is a scaled Laplace variable:     0 1 ets KT a T a ets =T a W ol s0 1 0 0 0 W ol s K 1 Ts sT a s 1 a Ts s s 1 Therefore, since t0 t/Ta, K0 KTa/a, T0 T/Ta, then K1, T1, K2, T2 are computed as follows: K 12 K 0102 a , T 12 T 0102 T a Ta

Simulations and performance comparison of the optimally designed VS PI and conventional PI controllers for the level process are given below.

6. Simulations and implementation The example below shows the system response to the situation imitating random connections and disconnections of consumers, so that the outow from the tank changes by steps of random amplitude (Fig. 10(a)). Let us consider two types of control for the process given by the transfer function W ol s e0:1s =ss 1. In the rst case, the process is controlled by an ITAE optimized conventional PI controller (Fig. 10(b)), so that the controller transfer function is W PI s 2:7641 1=4:744s. In the second option, the process is controlled by an ITAE optimized VS PI controller (Fig. 10(c)), with the following transfer functions: W VSPI1 s 2:5091 1=1:221s and W VSPI2 s 3:9441 1=91:92s (for the two regions of the state space). The control performance is assessed using the integral absolute error (IAE) criterion because the ITAE criterion cannot be used for this situation due to the time acting as a weighing function, so that the effect of disturbances coming at a later time would be higher, while process is running continuously, and errors in the beginning should have the same weight as the errors in the end of the test. The conventional PI controller produces the value of the criterion IAEPI 42.59, while the VS PI controller gives the criterion value IAEVSPI 23.83, which is signicantly lower than the former. Another comparison of the same two controllers is done for the case of the simultaneous effect of the changes in outow, as given by Fig. 10(a), and the source (pump) discharge pressure change. The latter is simulated as variable valve gain K(t) 1 0.5 sin(0.1t), with minimum and maximum values Kmin 0.5 and Kmax 1.5, respectively, differing by the factor of 3. The results of the simulations for the PI and VS PI controllers are presented in Fig. 11(a) and (b), respectively. In these tests, the conventional PI

Fig. 10. Level response to random step changes in outow for IAE optimized PI control and IAE optimized VS PI control: (a) disturbance (outow increments); (b) level controlled by PI controller; (c) level controlled by VS PI controller.

controller produces the value of the criterion IAEPI 58.20, while the VS PI controller gives the criterion value IAEVSPI 35.62. The proposed VS PI controller was implemented and tested on a water tank level at a synthetic oil production facility and realized as a combination of two PI controllers with additional switching logic in Honeywell TPSs distributed control system. In overall, the control performance is in agreement with the presented theory and simulations. However, due to the character of disturbances, which depend on a number of operational factors

I. Boiko / Control Engineering Practice 21 (2013) 700707

707

is matched by the controller action to ensure a damped transient process of bringing the system output to the set point. The describing function analysis is developed for the VS PID controller. The design methodology for the level process and controller performance assessment are provided.

References
Adamy, J., & Flemming, A. (2004). Soft variable-structure controls: a survey. Automatica, 40(11), 18211844. Astrom, K. J., & Hagglund, T. (1984). Automatic tuning of simple regulators with specication on phase and amplitude margins. Automatica, 20, 645651. Astrom, K. J., & Hagglund, T. (1995). PID Controllers: Theory, Design and Tuning (2nd Edn.). Research Triangle Park, NC: Instrument Society America. Atherton, D. P. (1975). Nonlinear Control Engineering Describing Function Analysis and Design. Workingham, Berks, UK: Van Nostrand Company Limited. Balestrino, A., Biagini, V., Bolognesi, P. & Crisostomi, E. (2009). Advanced variable structure PI controllers. In: Proceedings IEEE Conference on Emerging Technologies and Factory Automation, Mallorca. Boiko, I., & Sayedain, S. (2010). Analysis of dynamic nonlinearity of ow control loop through modied relay feedback test probing. International Journal of Control, 83(12), 25802587. Branicky, M. (1998). Multiple Lyapunov functions and other analysis tools for switched and hybrid systems. IEEE Transactions on Automatic Control, 43(4), 478482. Chen, C.-L., & Chang, F. Y. (1995). Design and analysis of neural/fuzzy variablestructure PID control system. IEE Proceedings. Control System Applications, 143(2). Hodel, A. S., & Hall, C. E. (2001). Variable-structure PID control to prevent integrator windup. IEEE Transactions on Industrial Electronics, 48(2), 442451. Hodel, A. S., & Hall, C. E. (2004). Authors reply to comments on variable-structure PID control to prevent integrator windup. IEEE Transactions on Industrial Electronics, 51(3), 738739. Ismail, O. & Bedwani, W. (2001) Compliant motion control using variable-structure PID control system. Proceedings 2001 IEEE International. Symposium on Intelligent Control. Mexico City, Mexico, (pp. 397401). Jafarov, E. M., Parlakci, M. N. A., & Istefanopulos, Y. (2005). A new variablestructure PID controller design for robot manipulators. IEEE Transactions on Control Systems Technology, 13(1), 122130. Johansson, M., & Rantzer, A. (1998). Computation of piecewise quadratic Lyapunov functions for hybrid systems. IEEE Transactions on Automatic Control, 43(4), 555559. Mainardi, E., Mantovani, C., Fabbri, D. & Bonfe, M. (2003) Variable-structure PID controller for co-operative eye-in-hand/eye-to-hand visual servoing. Proceedings 2003 IEEE Conference. Control. Application., Istanbul, Turkey, (pp. 989 994). Mantz, R. J., & Battista, H. D. (2004). Comments on variable-structure PID control to prevent integrator windup. IEEE Transactions on Industrial Electronics, 51(3), 736738. Qiu, X., Yuan, J., & Wang, Z. (2006). Feedforward variable structural proportionalintegral-derivative for temperature control of polymerase chain reaction. Chinese Journal of Chemical Engineering, 14(2), 200206. Shinskey, F. G. (1988). Process Control Systems Application, Design, and Tuning (3rd Edn.). New York: McGraw-Hill. Suyitno, A., Fujikawa, J., Kobayashi, H., & Dote, Y. (1993). Variable-structured robust controller by fuzzy logic for servomotors. IEEE Tranactions on Industrial Electronics, 40(1). Utkin, V. I. (1992). Sliding Modes in Control and Optimization. Berlin: Springer Verlag. Zhang, B., Li, H., Guo, L., Meng, H., Wang, J., & Yin, Y. (2010). Application of variable structure PID in velocity control for large telescope. Optics and Precision Engineering, 18(7), 16131619. Ziegler, J. G., & Nichols, N. B. (1942). Optimum settings for automatic controllers. Transactions of the American Society of Mechanical Engineers, 64, 759768.

Fig. 11. Level response to random step changes in outow and harmonic changes in discharge pressure for IAE optimized PI control and IAE optimized VS PI control: (a) level controlled by PI controller; (b) level controlled by VS PI controller.

and never repeat, it is hard to create the same conditions for the comparison of the performance of the PI and VS PI controllers in the conditions of a real plant. For that reason the comparison is done by means of simulations.

7. Conclusion A variable-structure PID controller and its application for level control are considered. The proposed VS PID principle has a very simple physical interpretation, which is the application of a relatively aggressive control immediately after the disturbance comes to the process aimed at minimizing its effect, and application of a relatively sluggish control after this disturbance