You are on page 1of 5

Demonstrating effects of a dynamical feedforward

control for the first order system


Peter Ťapák and Mikuláš Huba
Slovak University of Technology in Bratislava
Ilkovičova 3, 812 19 Bratislava, Slovakia
Email: peter.tapak, mikulas.huba@stuba.sk

Abstract—An analysis of different 2DOF PI controller tuning the plant dynamics, the dynamical feedforward control may be
methods in [1], [2] showed significant advantages of a noise atten- applied to unstable plants just under a stabilizing controller R.
uation oriented predictive disturbance observer based filtered PI
Within a reference model concept [3], the stabilizing con-
control (PDO-FPI) against the traditional two degree of freedom
(2DOF) PI control. Analytically, as well as by the numerical troller should act just in situations with a plant-model mis-
performance portrait method it discovered that the potential match, when y 6= wr . But, in a closed loop control an exact
improvements in the disturbance response are significantly higher expression of an ideal output wr as a function of an acting
than those in the setpoint response. This paper shows that the disturbance becomes complex. Furthermore, for unstable and
PDO-FPI does not make use of the whole control potential
integral plants the corrective signal in the red path diverges to
regarding the analysed first order plants with a possible dead
time. The corresponding control structures will be derived by infinity. Therefore, such a scheme is obviously not applicable
augmenting the well known control loop with a reference model to unstable plants. Usually, the corrective path is fully omitted
by an additional disturbance observer for an input disturbance. and the controller R actively compensates disturbances also
The theoretical analysis will be illustrated by experimental in a nominal case. However, it also means that the real output
results.
transients differ from (2) and, in general, they are not known
I. I NTRODUCTION in advance. They will be simply approximated as
Let us start with an ideal setpoint and disturbance feedfor- Wr (s) ≈ Q(s)W (s) (3)
ward control in Fig. 1. Since for majority of plants the required
plant dynamics inversion that would guarantee for all times For the case of an input disturbance feedforward control (3)
the identity y(t) = w(t) is not feasible, one has to consider a the control scheme with a reference model [4], [5] should then
lowpass filter Q with an appropriate relative degree that will be modified according to Fig. 3 (black).
−1
offer a proper filtered plant model inversion QS . For plants In Fig. 4, a direct disturbance feedforward is replaced by a
with a noninvertible model dynamics denoted as S + , the filter disturbance observer based on an inverse plant dynamics [6].
should be composed as Thereby, Qi has to fullfill the same requirements as Q (1).
Q = Q0 S + (1) In a modular approach to disturbance observer based in-
tegral control it may be advantageous to filter by a lowpass
with Q0 denoting an appropriate low pass filter. A lowpass filter not only the disturbance observer signal, but also the
filter Qi representing a band limited dynamics of a disturbance plant output for the stabilizing controller R. Then, the scheme
measurement, or reconstruction, has also to be considered in has to be modified according to Fig. 5.
a disturbance feedforward channel.
This all will slow down real dynamics of the output that II. E XAMPLE
will be given for an open loop control as
Fig. 6 shows transient responses of the controller according
Wr (s) = Q(s)W (s) + S(s)[1 − Qi (s)]Di (s) to Fig. 4 designed for the plant
(2)
⇒ wr (t) 6= w(t)
Y (s) Ks −Td s
Since a pole-zero cancellation is not applicable to unstable S(s) = = e (4)
systems and the ideal output response (2) directly depends on U (s) s+a

Figure 1. Ideal setpoint and input disturbance feedforward control Figure 2. Filtered setpoint and input disturbance feedforward control
with the parameters a = 1/3; Ks = a, Td = 1. The required III. E XPERIMENTAL R ESULTS
filters have been chosen as In this section, experiments will be reported carried out
1 by using the laboratory thermo-optical plant [14]. The light
Q(s) = Qi (s) = e−Td s (5)
1.5s + 1 channel consist of the light bulb which acts as a light source,
i.e. with a time constant equal to one half of the plant time the system input is the bulb voltage[0-100 %], the output is the
constant Tp = 1/a. value from the light intensity sensor filtered with the first order
Under the same requirements, a dead time compensator has low pass filter with the time constant set to Tp = 3 seconds. It
been proposed by using the modified Smith predictor approach corresponds to typical FOPDT model, but the input-to-output
presented in [7], [8]. Thereby, a first order predictor filter has characteristics is strongly non-linear as one can see in Fig. 8.
been considered. All three controllers have been experimentally compared
Finally, the third considered approach corresponds to a by introducing an additional software delay Tda = 42Ts =
2DOF PI control tuned by the triple real dominant pole method 42∗0.02s, which together with the inherent delay of the optical
[9] that allows to tune analytically all required parameters channel Tdi = 0.15s give the total delay Td = 0.99s. The plant
Kc , Ti and the setpoint weighting b. As it was shown in [1], time constant is Tp = 3s. The loop has been linearized by an
it offers a well balanced setpoint and disturbance responses inverse input-output nonlinearity at the controller output [15].
with a sufficient robustness against uncertainties. Transients in Fig. 9 fully confirm the simulation results
By applying these control concepts to the plant (4) without above. For evaluating deviations from ideal strictly one-pulse
control constraints, one gets the responses in Fig. 6. Obviously, (1P) shapes of a plant input u(t) (uT V1 = 0) may under
the setpoint responses of the dead time compensator according denotation um = max{u} be expressed by the uT V1 measure
to [8] are fully equivalent to the presented solution. However, defined as
X
disturbance responses given by the predictive disturbance uT V1 = |ui+1 − ui | − |2um − u∞ − u0 | (6)
observer are obviously faster. i
Differences between the dead time compensator by Normey-
Speed of transients at the plant output may be quantitatively
Rico and Camacho occur in the disturbance response. Due
evaluated by the IAE (Integral of Absolute Error)
to the predictive disturbance observer combined with the Z ∞
stabilizing controller, the disturbance response is considerably IAE = |e(t)| dt (7)
accelerated. Thereby, the new solution may also be applied in 0
a fully continuous implementation to unstable plants, whereas Evaluation of the IAE and uT V1 values in Fig. 10 illus-
the modified Smith predictor may deal with such situations trates data expected on the basis of the analysis made in [15]:
just in a digitized version. Both dead time compensators yield an equivalent dynamics
Another important difference appears in constrained control, of the setpoint responses that is significantly better than that
where the modified Smith predictor typically yields a windup of the traditional PI control. Regarding a setpoint response,
effect. As it was shown in [10], [11], [12], [13], this handicap the controller according to Fig. 4 gives the fastest responses.
could be elliminated by using a 2DOF P control instead of the The increased uT V1 values could be at least partially reduced
stabilizing PI control.
Although the 2DOF PI control could yield acceptable re-
sults, in general it is no more able to guarantee a speed of
setpoit responses comparable with both above mentioned dead
time compensators. Here, it is considered just to illustrate
limits of the traditional PI control.

Figure 4. Control with a reference model extended by a disturbance observer

Figure 3. Control with a reference model modified for a filtered disturbance Figure 5. Filtered control with a reference model and a disturbance observer
feedforward control for an input disturbance
1.4

Constrained Control
1.2 1.4

1 1.2

0.8 1
−−−> y

0.6 NR−CA 0.8

−−−> y
Vi−Vi
DFF−PDO−FPI
0.4 0.6 NR−CA
Vi−Vi
0.2 DFF−PDO−FPI
0.4

0 0.2
0 10 20 30 40 50 60 70 80
−−−> t

0
0 10 20 30 40 50 60 70 80
2 −−−> t
NR−CA
1.8 Constrained Control
Vi−Vi
1.4
DFF−PDO−FPI
1.6
NR−CA
1.4 1.2 Vi−Vi
DFF−PDO−FPI
1.2
1
−−−> u

0.8 0.8
−−−> u

0.6
0.6
0.4
0.4
0.2

0 0.2
0 10 20 30 40 50 60 70 80
−−−> t

0
0 10 20 30 40 50 60 70 80
Figure 6. Transient responses of different controllers - linear case −−−> t

Figure 7. Transient responses of different controllers - constrained control

by considering a filtered controller modification according to


Fig. 5.
The following figures (Fig. 11- 12) show the dependency of
the control quality on the dead time value. It shows the results 80
Input−output characteristic y=f(u)

of the real experiments with various additional dead times. The


70
IAE values of both dead time compensators are similar, the
60
IAE values of traditional PI control are many times higher
and the value rises dramatically with increasing dead time. 50
Measured points of IO characteristics
−−−> y

Linear interpolation of IO characteristics


The uT V1 values of the traditional PI controller are lower 40

for large dead times. The solution proposed by Normey-Rico 30

and Camacho yields better uT V1 values in experiments with 20

relatively small dead time. The disturbance observer based PI 10

controller gives the highest uT V1 values in all setpoint step 0


0 20 40 60 80 100
experiments. −−−> u

The disturbance steps performance was the same by the Figure 8. Input-to-output characteristic of the light channel
means of the uT V1 values. However in all these experiments
the disturbance observer based PI controller gives the best
performance considering the IAE values.
IAE Real experiment results, System output
60 36
Setpoint
NR−CA
NR−CA
Vi−Vi 34 Vi−Vi
50 DFF−PDO−FPI DFF−PDO−FPI

32
40

−−−> Light intensity


30
30
28

20
26

10
24

0 22
Setpoint Up Setpoint Down Disturbance Up Disturbance Down 50 100 150 200 250 300
−−−> t

uTV Real experiment results, Control signal


1
14 40
NR−CA
NR−CA Vi−Vi
12 Vi−Vi DFF−PDO−FPI
35
DFF−PDO−FPI

10
30
−−−> Control signal
8
25
6

20
4

15
2

0 10
Setpoint Up Setpoint Down Disturbance Up Disturbance Down 50 100 150 200 250 300
−−−> t

Figure 10. IAE and uT V1 values of different controllers - real time control Figure 9. Transient responses of different controllers - real time control

Control quality − Setpoint step − IAE − dead time dependancy


300 Control quality − IAE − Disturbance step − dead time dependancy
600
NR−CA
200
−−−> IAE

Vi−Vi NR−CA
400
−−−> IAE

DFF−PDO−FPI Vi−Vi
DFF−PDO−FPI
100
200

0
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
−−−> Dead Time [s] 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Control quality − Setpoint step − uTV1 − dead time dependancy −−−> Dead Time [s]
60 Control quality − uTV1 − Disturbance step − dead time dependancy
60
−−−> uTV1

40
−−−> uTV1

40

20
20
NR−CA
0
0.4 0.6 Vi−Vi0.8 1 1.2 1.4 1.6 1.8 2 NR−CA
DFF−PDO−FPI−−−> Dead Time [s] 0
0.4 0.6 Vi−Vi0.8 1 1.2 1.4 1.6 1.8 2
DFF−PDO−FPI−−−> Dead Time [s]

Figure 11. IAE and uT V1 values dependency on dead time for setpoint
step Figure 12. IAE and uT V1 values dependency on dead time for disturbance
step
IV. C ONCLUSIONS [3] I. M. Horowitz, Synthesis of Feedback Systems. N.York: Academic
Press, 1963.
The paper compares a dead time compensator proposed by [4] K. J. Åström and T. Hägglund, Advanced PID Control. NC: ISA,
an extension of the control loop with a reference model by a Research Triangle Park, 2006.
disturbance observer based on inversion of the plant dynamics [5] A. Visioli, Practical PID Control. London.: Springer, 2006.
[6] M. Huba, L. Rovanova, and L. Marko, “From automatic feedback sys-
with two the best known alternatives used in practice: the tems to moocs,” in Emerging eLearning Technologies and Applications
2DOF PI control by [9] and the modified Smith predictor by (ICETA), 2013 IEEE 11th International Conference on, Oct 2013, pp.
[8], [16]. 151–158.
[7] J. Normey-Rico, J. Guzmán, S. Dormido, M. Berenguel, and E. Cama-
The comparison shows that the new solution gives setpoint cho, “An unified approach for dtc design using interactive tools,” Control
responses equivalent to the previously published solution Engineering Practice, vol. 17, pp. 1234–1244, 2009.
by Normay-Rico and Camacho [8]. However, the stabiliz- [8] J. Normey-Rico and E. Camacho, “Dead-time compensators: A survey,”
Control Engineering Practice, vol. 16, 4, pp. 407–428, 2008.
ing controller combined with the disturbance observer yields [9] M. Vı́tečková and A. Vı́teček, “2DOF PI and PID controllers tuning,”
significantly better disturbance response and guarantees also in 9th IFAC Workshop on Time Delay Systems, vol. 9, Praha, 2010, pp.
improved properties under constrained control. Influence of 343–348.
[10] M. Huba and M. Halas, “New Filtered Smith Predictors for FOPDT
the dead time value on the performance of the all considered Plants. Part 1. First Order Disturbance Filter.” in ERK 2010 - Devetnajsta
controllers was verified by experiments on real plant. The loop mednarodna Elektrotehniska in racunalniska konferenca. Portoroz,
may be extended by introducing an additional filtration to im- Slovenija: IEEE, 2010, pp. 339–342.
[11] M. Huba and M. Halas, “New Filtered Smith Predictors for FOPDT
prove a noise attenuation. The experimental results also show Plants. Part 2. Second Order Disturbance Filters.” in ERK 2010 -
that a significant input plant nonlinearity may successfully Devetnajsta mednarodna Elektrotehniska in racunalniska konferenca.
be compensated by an inverse non-linearity at the controller Portoroz, Slovenija: IEEE, 2010, pp. 343–345.
[12] M. Huba and P. Ťapák, “Experimenting with the Modified Filtered
output. Smith Predictors for FOPDT Plants,” in World Congress, S. Bittanti,
V. ACKNOWLEDGMENT A. Cenedese, and S. Zampieri, Eds. Milano, Italy: IFAC-PapersOnLine,
2011, vol. 18, pp. 2452–2457.
This work has been partially supported by the grants APVV- [13] M. Huba and P. Ťapák, “Experimenting with modified smith predictors
0343-12 Computer aided robust nonlinear control design and using b&r automation studio target for simulink,” in 11th IFAC/IEEE
International Conference on Programmable Devices and Embedded
VEGA 1/0937/14 Advanced methods for nonlinear modeling Systems, Brno, Czech Republic, 2012.
and control of mechatronic systems. [14] T. Huba, M. Huba, P. Ťapák, and P. Bisták, “New Thermo-Optical Plants
for Laboratory Experiments,” in IFAC World Congress, Cape Town,
R EFERENCES South Africa, 2014.
[15] P. Ťapák, M. Huba, and T. Huba, “Experimental comparison of a 2DOF
[1] M. Huba, “Performance measures, performance limits and optimal PI
PI and DO-FPI,” in Int. SSKI Conference Oscadnica, 2014.
control for the IPDT plant,” Journal of Process Control, vol. 23, 4, pp.
500–515, 2013. [16] J. Normey-Rico and E. Camacho, “Unified approach for robust dead-
time compensator design,” J. Process Control, vol. 19, 1, pp. 38–47,
[2] M. Huba, “Comparing 2DOF PI and Predictive Disturbance Observer
2009.
Based Filtered PI Control,” Journal of Process Control, vol. 23, 10, pp.
1379–1400, 2013.

You might also like