Feed Foward Non Real Iz Able

Advances in Feedforward Control
for Measurable Disturbances
José Luis Guzmán Sánchez
Department of Informatics
Engineering and Systems Area
University of Almerı́a (Spain)
joseluis.guzman@ual.es
Madrid, Febrero 2019
1/119 José Luis Guzmán Sánchez Advances in Feedforward Control for Measurable Disturbances
Outline
1 Introduction
2 Feedforward control problem
3 Nominal feedforward tuning rules

Non-realizable delay
4 Performance indices for feedforward control
5 Conclusions
Outline
1 Introduction

5 Conclusions
Introduction
What are load disturbances?

Typically low frequency input signals which affect the output of
processes but that cannot be manipulated
Disturbances
Manipulated
Process Outputs
Inputs
Introduction
Most industrial processes are subject to disturbances and the

nature and origin of disturbances may vary depending on the
process and the operational environments.
Effective disturbance effect reduction is a key topic in process
control. In fact, disturbances together with process uncertainty,
are one of the reasons for feedback control.
Introduction
Real plants at the Automatic Control research group in Almerı́a
Introduction
Motivation: feedback controller
Pd
r u y
Σ Cfb Pu Σ
−1
Introduction
Motivation: feedback controller

process output
0.2
0.15
d 0.1
0.05
0
Pd −0.05
0 2 4 6 8 10
time
r u y
Σ Cf b Pu Σ 0
control eﬀort
−0.1
−1 −0.2
−0.3
−0.4
−0.5
0 2 4 6 8 10
time
No reaction until there are discrepancies!
Introduction
Motivation: feedforward compensator
−C f f Pd
r u y
Σ Cf b Σ Pu Σ
−1
Pd
Cf f =
Pu
Y = ( Pd − Pu C f f ) D
Introduction
0.25
0.2
d
process output
0.15
−C f f Pd 0.1
r u y 0.05
Σ Cf b Σ Pu Σ
0
−1 −0.05
0 2 4 6 8 10
time
Pd
Ideal compensation: C f f = = Pd Pu−1
Pu
Introduction
0.25
0.2
d
process output
0.15
−C f f Pd 0.1
r u y 0.05
Σ Cf b Σ Pu Σ
0
−1 −0.05
0 2 4 6 8 10
time
Pd
Ideal compensation: C f f = = Pd Pu−1
Pu
Feedforward control problem
Perfect compensation is seldom realizable:
Non-realizable delay inversion.

Right-half plan zeros.
Integrating poles.
Improper transfer functions.
Classical solution
Ignore the non-realizable part of the compensator and implement the
realizable one. In practice, static gain feedfoward compensators are
quite common.
Introduction
Motivation: non-ideal feedforward compensator

0.6
0.5
0.4
process output
0.3
0.2
0.1
−0.1
0 1 2 3 4 5 6 7 8 9 10 11
time
Introduction
Motivation: non-ideal feedforward compensator

0.6
0.5
0.4
process output
0.3
0.2
0.1
−0.1
0 1 2 3 4 5 6 7 8 9 10 11
time
Introduction
Motivation: residual term

d
−C f f Pd
r u y
Σ Cf b Σ Pu Σ
−1
Pd
Cf f =
Pu
Y = ( Pd − Pu C f f ) D
Introduction
Motivation
(a) Open−loop response (b) Closed−loop response

0.14 0.3
Static
0.12
Without Feedforward
Lead−Lag 0.25
Static
0.1 Lead−Lag
0.2
0.08
0.15
0.06
y
0.1
0.04
0.05
0.02
0 0
−0.02 −0.05
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
0 Static 0
Without Feedforward
Lead−Lag Static
−0.1 −0.1
Lead−Lag
−0.2 −0.2
−0.3 −0.3
u
−0.4 −0.4
−0.5 −0.5
−0.6 −0.6
−0.7 −0.7
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
Introduction
Motivation
http://aer.ual.es/ilm/ http://fichas-interactivas.pearson.es/
Introduction
Motivation
An interaction between feedforward and feedback controllers arises
Pd − C f f Pu Pd − C f f Pu
y= d= d
1+L 1 + C f b Pu
Other design strategies are required!
Introduction
Motivation
An interaction between feedforward and feedback controllers arises
y= d= d
1+L 1 + C f b Pu
Other design strategies are required!
Introduction
Motivation
Surprisingly there are very few studies in literature (we starting the
project in 2010):
D. Seborg, T. Edgar, D. Mellichamp, Process Dynamics and Control,

Wiley, New York, 1989.
F. G. Shinskey, Process Control Systems. Application Design
Adjustment, McGraw-Hill, New York, 1996.
C. Brosilow, B. Joseph, Techniques of Model-Based Control,
Prentice-Hall, New Jersey, 2002.
A. Isaksson, M. Molander, P. Modn, T. Matsko, K. Starr, Low-Order
Feedforward Design Optimizing the Closed-Loop Response, Preprints,
Control Systems, 2008, Vancouver, Canada.
Introduction
Objectives
1 Study and development of a control methodology to improve

disturbance compensation in industrial processes
2 Definition of nominal simple optimal tuning rules for designing
feedforward compensators
3 Development of a robust methodology to cope with both
reference tracking and disturbance rejection, using feedforward
control structures
4 Integration of the obtained nominal and robust feedforward tuning
rules into a general dead-time compensation solution
5 Propose performance indices for feedforward control
Outline
1 Introduction

5 Conclusions
Feedforward control is an old topic in process control. In fact, its

first application dates from 1925, where a feedforward
compensator was used for drum level control of tanks connected
in series.
Many of the other early applications dealt with control of
distillation columns.
Since then, feedforward control has become a fundamental
control technique for the compensation of measurable
disturbances.
Nowadays, this mechanism is implemented in most distributed
control systems to improve the control performance.
The idea behind feedforward control from disturbances is to supply

control actions before the disturbance affects the process output:
−C f f Pd
r u y
Σ Cf b Σ Pu Σ
−1
Pd
Cf f =
Pu
In industry, PID control is commonly used as feedback controller and

four structures of the feedforward compensator are widely considered:
! "
1
Cfb = κfb 1+ + sτd
sτi
Static: Cf f = κf f
Static with delay: C f f = κ f f e−sL f f

1 + sβ f f
Lead-lag: Cf f = κf f
1 + sτ f f
1 + sβ f f −sL
Lead-lag with delay: Cf f = κf f e ff
1 + sτ f f
Then, if we consider that process transfer functions are modeled as

first-order systems with time delay, i.e.
κu −sλu κd
Pu = e , Pd = e−sλd
1 + τu 1 + sτd
The following feedforward compensator can be considered:
κd
Static: Cf f =
κu
κ
Static with delay: Cf f = d e− s ( λd − λu )
κu
κ 1 + sτu
Lead-lag: Cf f = d
κu 1 + sτd
κ 1 + sτu −s(λd −λu )
Lead-lag with delay: Cf f = d e
κu 1 + sτd
Lets consider the following example:
1 −s 1
Pu (s) = e , Pd (s) = e−2s
s+1 2s + 1
Static: Cf f = 1
Static with delay: C f f = e− s

1+s
Lead-lag: Cf f =
1 + 2s
1 + s −s
Lead-lag with delay: Cf f = e
1 + 2s
C f b is a PI controller tuned using the AMIGO rule, κ f b = 0.25 and
τi = 2.0.
Open−loop response Closed−loop response

0.02 0.25
Static Without Feedforward

0 0.2
Static−delay Static
−0.02 Lead−Lag 0.15 Static−delay
Lead−Lag−delay Lead−Lag
−0.04 0.1 Lead−Lag−delay
−0.06
y
0.05
y
−0.08 0
−0.1 −0.05
−0.12 −0.1
−0.14 −0.15
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
0.05 0.05
0
Static 0
Without Feedforward
Static−delay Static
−0.05 Lead−Lag −0.05 Static−delay
−0.1
Lead−Lag−delay −0.1
Lead−Lag
Lead−Lag−delay
−0.15 −0.15
u
−0.2 −0.2
−0.25 −0.25
−0.3 −0.3
−0.35 −0.35
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
Motivation
Then, lets consider a delay inversion problem, i.e., λd < λu . Then, the
resulting feedforward compensators are given by:
κd
Cf f = Kf f =
κu
κd τu s + 1
Cf f =
κu τd s + 1
Motivation
Example:
1 1 −s
Pu (s) = e−2s , Pd (s) = e
2s + 1 s+1
2s + 1
C f f = 1, Cf f =
s+1
The feedback controller is tuned using the AMIGO rule, which gives
the parameters κ f b = 0.32 and τi = 2.85.
Motivation
(a) Open−loop response (b) Closed−loop response

0.14 0.3
Static
0.12
Without Feedforward
Lead−Lag 0.25
Static
0.1 Lead−Lag
0.2
0.08
0.15
0.06
y
y
0.1
0.04
0.05
0.02
0 0
−0.02 −0.05
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
0 Static 0
Without Feedforward
Lead−Lag Static
−0.1 −0.1
Lead−Lag
−0.2 −0.2
−0.3 −0.3
u
−0.4 −0.4
−0.5 −0.5
−0.6 −0.6
−0.7 −0.7
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
−C f f Pd
r u y
Σ Cf b Σ Pu Σ
−1
y= d= d
1+L 1 + C f b Pu
−C f f Pd
r u y
Σ Cf b Σ Pu Σ
−1
r r + Pd∗ (e−λu s − e−λd s )d

e= , e= , Pd = Pd∗ e−λd
1 + Pu C f b 1 + Pu C f b
Outline
1 Introduction

5 Conclusions
Feedforward tuning rules
Cases to be evaluated in this research:

Non-realizable delay inversion.
Right-half plan zeros.
Integrating poles.
Outline
1 Introduction

5 Conclusions
Nominal feedforward design: non-realizable delay
Objective
To improve the final disturbance response of the closed-loop system
when delay inversion is not realizable (λu > λd )
Methodology
Adapt the open-loop tuning rules to closed-loop design
Obtain optimal open-loop tuning rules
Design a switching controller to improve the results
Two approaches:
d d
Cf f Pd H Cf f Pd
r u y r u y
Σ Cf b Σ Pu Σ Σ Cf b Σ Pu Σ
−1 −1
κk
Pk (s) = e− λk s k ∈ [u, d] λu > λd
τk s + 1
τi s + 1 βf fs + 1
C f b ( s) = κ f b C f f ( s) = κ f f
τi s τf f s + 1
Two approaches:
d d
Cf f Pd H Cf f Pd
r u y r u y
−1 −1
κk
τk s + 1
τi s + 1 βf fs + 1
C f b ( s) = κ f b C f f ( s) = κ f f
τi s τf f s + 1
Two approaches:
d d
Cf f Pd H Cf f Pd
r u y r u y
−1 −1
κk
τk s + 1
τi s + 1 βf fs + 1
C f b ( s) = κ f b C f f ( s) = κ f f
τi s τf f s + 1
replacements
Delay inversion: open-loop compensation
d
Pd
u y
−C f f Pu Σ
# $ κd τu s + 1
y = Pf f = Pd − C f f Pu d Cf f = ·
κu τd s + 1
replacements
1
process output
0.8
0.6
0.4
0.2
0
0 5 10 15 20 25
time
0.8
process output
0.6
0.4
0.2
0
0 5 10 15 20 25
time
# $ κd τu s + 1
y = Pf f = Pd − C f f Pu d Cf f = ·
κu τd s + 1
Pd C f fP u O v e rsh o ot e rror I n i t i al e rror
0.3
p r o c e ss ou t p u t 0.25
0.2
0.15
0.1
0.05
0
−0.05
0 5 10 15 20 25
tim e
P d − C f fP u O v e r sh o ot e r r or I n i t i al e rror
0.2
p r o c e ss ou t p u t
0.15
0.1
0.05
−0.05
0 5 10 15 20 25
tim e
# $
y = Pf f = Pd − C f f Pu d + u f b Pu
First approach
To deal with the non-realizable delay case, the first approach was to
work with the following:
Use the classical feedforward control scheme.

Remove the overshoot observed in the response.
Proposed a tuning rule to minimize Integral Absolute Error (IAE).
The rules should be simple and based on the feedback and
model parameters.
To remove the overshoot, the feedback control action is taken into

account to calculate the feedforward gain, κ f f .
κfb % κfb
∆u = edt = IE · d
τi τi
So, in the new rule, the goal is to take the control signal to the correct
stationary level −∆u in order to take the feedback control signal into
account and reduce the overshoot. The gain is therefore reduced to
kd κfb
κf f = − IE
ku τi
Closed-loop design
To remove the overshoot, the feedback control action is taken into

account to calculate the feedforward gain, κ f f .
κfb % κfb
∆u = edt = IE · d
τi τi
So, in the new rule, the goal is to take the control signal to the correct
stationary level −∆u in order to take the feedback control signal into
account and reduce the overshoot. The gain is therefore reduced to
kd κfb
κf f = − IE
ku τi
Closed-loop design
IE estimation:
Y = ( Pd − Pu C f f ) D = Pd D − Pu C f f D
t +
⎧ *
−
⎨ kd 1 − e d d
⎪ τ
0 ≤ t ≤ λb
y(t) − ysp = ! *
− τt
+ ! t −λ
− T b
""
⎩ kd
⎪ 1−e d − 1−e b d λb < t
λb = max(0, λu − λd ), Tb = τu + τ f f − β f f
IE estimation:
Y = ( Pd − Pu C f f ) D = Pd D − Pu C f f D
− τt
⎧ * +
⎨ kd 1 − e d d
⎪ 0 ≤ t ≤ λb
y(t) − ysp = ! *
− τt
+ ! t −λ
− T b
""
⎩ kd
⎪ 1−e d − 1−e b d λb < t
λb = max(0, λu − λd ), Tb = τu + τ f f − β f f
IE estimation:
% ∞
IE · d = (y(t) − ysp )dt
0
% λb * % ∞! t −λ
"
− τt − τt − T b
+
= kd 1 − e d d dt + kd −e d + e b d dt
0 λb
/∞
− τt λb
t −λ
.
− τt − T b
, -
= kd t + τd e d d + kd τd e d − Tb e b d
0 λb
! λ λ
"
− τb − τb
= kd λb + τd e d − τd − τd e d + Tb d
= kd (λb − τd + Tb ) d
IE estimation:
0
kd (τu − τd + τ f f − β f f ) λd ≥ λu
IE =
kd (λu − λd + τu − τd + τ f f − β f f ) λd < λu
kd κfb
κf f = − IE
ku τi
Lets consider the same previous example:
1 1 −s
Pu (s) = e−2s , Pd (s) = e
2s + 1 s+1
2s + 1
C f f = 1, Cf f =
s+1
(a) Static feedforward (b) Lead/Lag feedforward

0.3 0.3
Without Feedforward Without Feedforward

0.25
0.25
No K reduction No Kff reduction
ff
0.2 K reduction 0.2 K reduction
ff ff
0.15 0.15
y
y
0.1 0.1
0.05 0.05
0 0
−0.05 −0.05
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
0.05
0
0 Without Feedforward Without Feedforward
No K reduction No K reduction
−0.05 ff −0.1 ff
K reduction K reduction
−0.1 ff −0.2
ff
−0.15
−0.3
u
−0.2
−0.4
−0.25
−0.5
−0.3
−0.35 −0.6
−0.4 −0.7
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
The feedforward gain κ f f has been reduced from 1 to 0.778 for the
static feedforward and from 1 to 0.889 for the lead-lag filter.
Once the overshoot is reduced, the second goal is to design β f f and

τ f f to minimize the IAE value. In this way, we keep β f f = τu to cancel
the pole of Pu and fix the pole of the compensator:
% ∞ % t0 % ∞
I AE = |y(t)|dt = y(t)dt − y(t)dt
0 0 t0
where t0 is the time when y crosses the setpoint, with ysp = 0 and
d = 1.
− τt
⎧ * +
⎨ kd 1 − e d d
⎪ 0 ≤ t ≤ λb
y(t) − ysp = ! *
− τt
+ ! t −λ
− T b
""
⎩ kd
⎪ 1−e d − 1−e b d λb < t
% ∞ % t0 % ∞
I AE = |y(t)|dt = y(t)dt − y(t)dt
0 0 t0
t0 t0 − λ b τλ τd
= → t0 = d b = λ
τd Tb τd − Tb τu − τ f f b
Tb = τu + τ f f − β f f
% λ ! " % t ! t−λb " % ∞! t−λb "

b − τt 0 − τt − Tb − τt − Tb
I AE = 1−e d dt + −e d +e dt − −e d +e dt
0 λb t0
. /λb . t−λb /t 0 . t−λb /∞
− τt − τt − Tb − τt − Tb
= t + τd e d + τd e d − Tb e − τd e d − Tb e
0 λb t0
t t −λ
− τ0 − 0T b
= λb − τd + Tb + 2τd e d − 2Tb e b
λ λ
− τ −bT − τ −bT
= λb − τd + Tb + 2τd e d b − 2Tb e d b
! "
λ
− τb
= λb − τ 1 − 2e
with τ = τd − τ f f .
d λb λ λb
I AE = −1 + 2e− τ + 2 b e− τ = −1 + 2(1 + x)e− x = 0
dτ τ
where x = λb /τ . A numerical solution of this equation gives x ≈ 1.7,
which gives
λb
τ f f = Tb − τd + τu = τd − τ ≈ τd −
1.7
λu − λd ≤ 0
⎧
τ
⎨ d
⎪
λu − λd
τf f = τd − 0 < λu − λd < 1.7τd
⎪ 1.7
λu − λd > 1.7τd
⎩
0
Gain and τ f f reduction rule:
0.3
Without Feedforward
0.25
Open−loop rule
K reduction, T =T
0.2 ff p 3
K and T reduction
ff p
0.15
y
0.1
0.05
−0.05
0 5 10 15 20 25 30 35 40 45 50
t
−0.5
u
Without Feedforward
−1 Open−loop rule
K reduction, T =T
ff p 3
K and T reduction
ff p
−1.5
0 5 10 15 20 25 30 35 40 45 50
t
Gain and τ f f reduction rule:

u
y
0.3
Without Feedforward
0.25
Open−loop rule
0.2
Kff reduction, Tp=T3
Kff and Tp reduction
0.15
y
0.1
0.05
−0.05
0 5 10 15 20 25 30 35 40 45 50
t
−0.5
u
Without Feedforward
−1 Open−loop rule
Kff reduction, Tp=T3
Kff and Tp reduction
−1.5
0 5 10 15 20 25 30 35 40 45 50
t
No FF Open-loop rule κ f f reduction κ f f &τ f f reduction

IAE 9.03 1.76 1.37 0.59
First approach: Guideline summary
1 Set β f f = τu and calculate τ f f as:
λu − λd ≤ 0
⎧
τ
⎨ d
⎪
λu − λd
τf f = τd − 0 < λu − λd < 1.7τd
⎪ 1.7
λu − λd > 1.7τd
⎩
0
2 Calculate the compensator gain, κ f f , as
kd κfb
κf f = − IE
ku τi
0
kd (τ f f − τd ) λd ≥ λu
IE =
kd (λu − λd − τd + τ f f ) λd < λu
Second approach
To deal with the non-realizable delay case, the second approach was
to work with the following:
Use the non-interacting feedforward control scheme (feedback

effect removed).
Obtain a generalized tuning rule for τ f f for moderate, aggressive
and conservative responses.
The rules should be simple and based on the feedback and
model parameters.
Second approach: non-interacting structure
H Cf f Pd
r u y
Σ Cfb Σ Pu Σ
−1
Pf f + LH # $
y= d = Pf f ϵ + Hη d H = Pf f = Pd − C f f Pu
1+L
C. Brosilow and B. Joseph. Techniques of model-based control. Prentice Hall,
New Jersey, 2012.
H Cf f Pd
r u y
Σ Cfb Σ Pu Σ
−1
Pf f + LH # $
y= d = Pf f ϵ + Hη d H = Pf f = Pd − C f f Pu
1+L
C. Brosilow and B. Joseph. Techniques of model-based control. Prentice Hall,
New Jersey, 2012.
H Cf f Pd
r u y
Σ Cfb Σ Pu Σ
−1
r + ( H − Pd + Pu C f f )d
e= , H = Pf f = Pd − Pu C f f
1 + Pu C f b
Second approach
The main idea of this second approach relies on analyzing the residual
term appearing when perfect cancelation is not possible:
y
= Pd − Pu C f f = Pd − Pf f , Pf f = Pu C f f
d
y kd kd
= e− λd s − e− λu s
d τd s + 1 τf f s + 1
0.6
P3 output
0.5 Pff output
0.4 Initial Error
Overshoot Error
output
0.3
0.2
0.1
0 5 10 15 20 25
time
Process output
0.3 Initial Error
Overshoot Error
0.25
0.2
output
0.15
0.1
0.05
−0.05
0 5 10 15 20 25
time
y kd kd
d τd s + 1 τf f s + 1
From the previous analysis, it can be concluded that in order to totally

remove the overshoot for the disturbance rejection problem by using a
lead-lag filter, the settling times of both transfer functions must be the
same:
y kd kd
d τd s + 1 τf f s + 1
4τd + λd − λu λ
τf f = = τd − b , λb = λd − λu
4 4
From the previous analysis, it can be concluded that in order to totally

remove the overshoot for the disturbance rejection problem by using a
lead-lag filter, the settling times of both transfer functions must be the
same:
y kd kd
d τd s + 1 τf f s + 1
τf f = = τd − b , λb = λd − λu
4 4
Notice that the new rule for τ f f implies a natural limit on performance.
If parameter τ f f is chosen larger, performance will only get worse
because of a late compensation. The only reasons why τ f f should be
even larger is to decrease the control signal peak:
λb
τ f f = τd −
4
0.6
P output
3
0.5 P output
ff
0.4 Initial Error
Overshoot Error
output
0.3
0.2
0.1
0 5 10 15 20 25
time
Process output
0.3 Initial Error
Overshoot Error
0.25
0.2
output
0.15
0.1
0.05
−0.05
0 5 10 15 20 25
time
So, considering the IAE rule obtained for the first approach, two tuning
rules are available:
τf f = = τd − b
4 4
λu − λd λ
τ f f = τd − = τd − b
1.7 1.7
And a third one (a more agreessive rule) can be calculated to minimize
Integral Squared Error (ISE) instead of IAE such as proposed in the
first approach.
ISE minimization:
% ∞
1 ( t − λ b)
22
− τf f − τt
ISE = e −e d dt
λb
% ∞
1 2( t − λ b) τd ( t − λ b)+ τ f f t 2
− τf f − τd τ f f − τ2t
= e − 2e +e d dt
λb
3 2 ( t − λ b)
4∞ 3 τd ( t − λ b)+ τ f f t 4∞
τf f τd τ f f τd − τ2t ∞
. /
− τf f − τd τ f f
=− e +2 e − e d
2 τd + τ f f 2 λb
λb λb
τf f τf f λ
− b τ − 2λb
= − 2τd e τd + d e τd
2 τd + τ f f 2
ISE minimization:
1 2
d ISE 1 λ
− b 1 −τf f 1 2τd2 λ
− τb
= − 2τd e τd + = − e d = 0
d τf f 2 τd + τ f f ( τd + τ f f )2 2 ( τd + τ f f )2
λb
−
τ 2f f + 2τd τ f f + τd2 (1 − 4e τd )=0
5
λb
−
−2τd + 4τd2 − 4τd2 (1 − 4e τd )
1 6 2
λ
− τb
τf f = = τd 2 e d −1
2
Thus, three tuning rules are available:
λb
τ f f = τd −
4
λ
τf f = τd − b
1 1.7
6 2
λ
− τb
τ f f = τd 2 e d −1
which can be generalized as:
λb
τ f f = τd −
α
Second approach: Guideline summary
1 Set β f f = τu , κ f f = k d /k u and calculate τ f f as:

⎧
⎨ τd λb ≤ 0
λb
τf f = τ − 0 < λb < 4τd
⎩ d α
0 λb ≥ 4τd
2 Determine τ f f with λb /τd < α < ∞ using:
λb
⎧
⎪ ! √ " aggressive (ISE minimization)
− λ b /τd
2τd 1−
⎪
⎪
⎪ e
⎨
α= 1.7 moderate (IAE minimization)
⎪
⎪
⎪
⎪
4
⎩
conservative (Overshoot removal)
Example:
0.5 −2.25s 1
Pu (s) = e , Pd (s) = e−0.75s
5s + 1 2s + 1
0.35
Basic
0.3 Hast and Hägglund
ISE Minimization
0.25 IAE Minimization
Overshoot Removal
0.2
0.15
output
0.1
0.05
−0.05
−0.1
0 5 10 15 20 25
time
0
Basic
−5 Hast and Hägglund
ISE Minimization
−10 IAE Minimization
Overshoot Removal
−15
control signal
−20
−25
−30
−35
−40
0 5 10 15 20 25
time
ISE IAE uinit J1 J2
Hast and Hägglund 0.0739 0.6423 38.7800 2.5710 0.8979

ISE Minimization 0.0896 0.6021 8.0090 0.9993 0.8615
IAE Minimization 0.0975 0.5641 5.3680 0.9113 0.8315
Overshoot Removal 0.1277 0.6833 3.6920 0.9323 0.8870
% ∞
ISE( F ) ISC( F )
! "
1
J1 ( F, B) = + , ISC = u(t)2 dt
2 ISE( B ) ISC( B ) 0
% ∞
IAE( F ) IAC( F )
! "
1
J2 ( F, B) = + , IAC = |u(t)| dt
2 IAE( B ) IAC( B ) 0
Second approach: A switching solution
It is clear that if the compensation is made too fast, the output will
suffer a bigger overshoot error, while if it is too slow, the compensator
will take too much time to reject the disturbance and it will have a
bigger residual error. Therefore, a switching rule can be proposed in
such a way that the feedforward compensator reacts fast before the
outputs cross in order to decrease the residual error, and slower after
this time to avoid the overshoot because of the residual error.
0.6
P3 output
0.5 Pff output
0.4 Initial Error
Overshoot Error
output
0.3
0.2
0.1
0 5 10 15 20 25
time
Process output
0.3 Initial Error
Overshoot Error
0.25
0.2
output
0.15
0.1
0.05
−0.05
0 5 10 15 20 25
time
Pd C f fP u O v e rsh o ot e rror I n i t i al e rror
0.3
p r o c e ss ou t p u t 0.25
0.2
0.15
0.1
0.05
0
−0.05
0 5 10 15 20 25
tim e
P d − C f fP u O v e r sh o ot e r r or I n i t i al e rror
0.2
p r o c e ss ou t p u t
0.15
0.1
0.05
−0.05
0 5 10 15 20 25
tim e
# $
y = Pf f = Pd − C f f Pu d
0.6 P3 output
0.5 Pff output
Initial Error
0.4
output
0.3
0.2
0.1
0 t_change 4 t_cross 8 t_restore 12 16 20 24

time
Process output
0.3 Initial Error
0.25
0.2
output
0.15
0.1
0.05
−0.05
time
The idea is to set τ f f to a small value until the time when the
responses of both transfer functions cross. After this time, the new
value of τ f f will be τd . Once the load disturbance is rejected, τ f f will
be set again to the small initial value in order to be ready for new
coming disturbances.
Thus, the first step is to calculate the time it takes since a step change
in d appears at time instant td until the outputs of both transfer
functions cross. This time, tcross , corresponds to the point when the
step responses of Pf f and Pd are equal:
1 2
−( t cross − t d − λ d ) −( t cross− t d − λ u )
τf f
κd d e τd
−e =0
where it is straightforward to see that:
τd λu − τ f f λd
tcross = + td
τd − τ f f
On the other hand, notice that the time event of the switching rule is
really given at tchange = tcross − λu .
Once the disturbance has been rejected, the feedforward switching
controller should return to its original value in order to be ready for
possible new coming load disturbances. This change must be done at
a time instant, tr , which can be proposed as the settling time of
process Pd such as follows:
tr = 4τd + λd + td
Thus, τ f f should be equal to τd when td + tcross − λu ≤ t ≤ td + tr

and it must be tuned for a faster response otherwise, specially for
t < tchange .

t < tchange .

t < tchange .
0.6 P output
3
0.5 P output
ff
Initial Error
0.4
output
0.3
0.2
0.1

time
Process output
0.3 Initial Error
0.25
0.2
output
0.15
0.1
0.05
−0.05
time
τd λu − τ f f λd
tcross = + td tchange = tcross − λu
τd − τ f f
Second approach: the switching solution guideline
1 Set τ f f to a value as close to 0 as possible (tradeoff with the

control signal peak).
2 Wait until a step load disturbance is detected at time instant td .
Define tcross and trestore . Set tchange = tcross − λu .
3 Using a non-interacting scheme, set C f f and H as follows:
κd 1 + τu s
⎧
⎪
⎪ tchange ≤ t ≤ tr
⎨ κu 1 + τd s
⎪
⎪
C f f ( s) =
⎪
⎪ κd 1 + τu s
⎪
⎪
⎩ κ 1+τ s otherwise
u ff
4 Go to step 2.
0.35
ISE Minimization
0.3 IAE Minimization
Switching
0.25
0.2
0.15
output
0.1
0.05
−0.05
−0.1
0 5 10 15 20 25
time
0
−1
−2
−3
control signal
−4
−5
−6
−7
ISE Minimization
−8 IAE Minimization
Switching
−9
0 5 10 15 20 25
time
ISE IAE uinit J1 J2
ISE Minimization 0.0896 0.6021 8.0090 0.9993 0.8615

IAE Minimization 0.0975 0.5641 5.3680 0.9113 0.8315
Switching 0.0889 0.4252 6.2160 0.9062 0.7527
0.35
Hast and Hägglund
0.3 Switching
0.25
0.2
0.15
output
0.1
0.05
−0.05
−0.1
0 5 10 15 20 25
time
0
−5
−10
−15
control signal
−20
−25
−30
−35 Hast and Hägglund

Switching
−40
0 5 10 15 20 25
time
ISE IAE uinit J1 J2
Hast and Hägglund 0.0739 0.6423 38.78 2.5710 0.8979

Switching 0.0630 0.2878 38.78 2.6650 0.7149
Outline
1 Introduction

5 Conclusions
Performance indices for feedforward control
There exist metrics to evaluate feedback controllers for load

disturbance rejection problem based on the controller parameters. For
instance:
Pu (s) C f b (s) Pu (s) 1

Gy/d = = ω ↓↓
1 + C f b (s) Pu (s) 1 + C f b (s) Pu (s) C f b (s)
1 s
Gy/d ≈ ≈
C f b ( s) κi
0
10
Gy/d (ω )
−1
10
−2
10
−2 −1 0 1
10 10 10 10
ω
14
Ki=0.33
12 Ki=2
10
8
Salida
0
0 10 20 30 40 50 60 70 80 90 100
Tiempo (s)
0
Ki=0.33
−10
Ki=2
−20
−30
Señal de control
−40
−50
−60
−70
−80
−90
−100
0 10 20 30 40 50 60 70 80 90 100
Tiempo (s)
Bode Diagram
K_i=0.33
K_i=2
−10
−20
−30
Magnitude (dB)
−40
−50
−60
−70
−3 −2 −1 0 1 2
10 10 10 Frequency (rad/sec) 10 10 10
Objective
To proposed indices such that the advantage of using a feedforward
compensator with respect to the use of a feedback controller only can
be quantified.
Methodology
Propose different indices
Calculate the indices based on the process parameters
The two feedforward schemes are considered:
d d
Cf f Pd H Cf f Pd
r u y r u y
−1 −1
Assumptions:
κu −sλu κd
Pu = e , Pd = e−sλd
1 + τu 1 + sτd
Only, the non-inversion delay problem is analyzed:

κd 1 + sτu
Lead-lag: Cf f =
κu 1 + sτd
Assumptions:
! "
1
Cfb = κfb 1+
sτi
The lambda tuning rule is considered:
τi
κfb = , τi = τu
κu (λu + τbc )
where τbc is the closed-loop time constant.
The following index structure is proposed
I AEFF
IFF/FB = 1 − ,
I AEFB
where I AE FB is the integrated absolute value of the control error
obtained when only feedback is used, and I AE FF is the corresponding
I AE value obtained when feedforward is added to the loop.
As long as the feedforward improves control, i.e. I AE FF < I AE FB ,
the index is in the region 0 < IFF/FB < 1.
Calculation of I AE f b
In the feedback only case, the transfer function between disturbance d

and process output y is
e−sλd
κd
Pd (s) 1 + sτd
Gy/d (s) = =
1 + Pu (s)C f b (s) e−sλu 1 + sτi
1 + κu κfb
1 + sτu sτi
Assuming that r = 0 and d is a step with magnitude Ad and using the
final value theorem, the Integrated Error ( IE) value becomes (note that
e = −y, with r = 0)
% ∞
1 A τκ
IE FB = e(t)dt = lim s · E(s) = lim − Gy/d (s) d = − i d Ad
0 s →0 s s →0 s κu κ f b
Calculation of I AE f b
The magnitude of the IE value can be set equal to the I AE value

provided that the controller is tuned so that there are no oscillations:
τi κd
IAEFB = Ad
κu κ f b
Finally, considering the lambda tuning rule, it becomes
I AEFB = κd Ad (λu + τbc )
Calculation of I AEFF for classical FF scheme
In this case, the transfer function from the disturbance to the error is
e−sλd e−sλu
Pd (s) + Pu (s)C f f (s) κd − κd
1 + sτd 1 + sτd
Gy/d (s) = − = − sλ
1 + Pu (s)C f b (s) e u 1 + sτi
1 + κu κfb
1 + sτu sτi
Considering the lambda tuning rule and that the delays are
approximated as
e− λu s ∼
= 1 − λu s, e− λd s ∼
= 1 − λd s
It results in:
κd (λu + τbc )(λu − λd )s2
Gy/d (s) = −
(1 + τd s)(1 + τbc s)
The IE value for this case becomes

% ∞
1 Ad
IE FF = e(t)dt = lim s · Gy/d (s) = 0.
0 s →0 s s
which demonstrates that zero steady-state error can be achieved by
using feedforward control.
0.2
0.15
0.1
y
0.05 A1
0
A2
−0.05
0 10 20 30 40 50 60
time
Now, it is worth determining the expression of the error in the time

domain when a step signal of amplitude Ad is applied as a
disturbance. We have
⎧
κd Ad (λu + τbc )(λu − λd ) # −t/τ
bc − τ e − t/τd
⎪ $
⎪
⎪ τd e bc τbc ̸= τd
τbc τd (τbc − τd )
⎨
e( t) =
κ A ( λ + τ )( λ − λ )
! "
d d u bc u d t
1− e−t/τd τbc = τd
⎪
⎪
2
⎪
⎩ τd τ d
0.2
0.15
0.1
y
0.05 A1
0
A2
−0.05
0 10 20 30 40 50 60
time
0
τbc τd * +
τ
log τbc τbc ̸ = τd
t0 = τbc − τd d
τd τbc = τd
We can therefore calculate the area of the first part of the transient as
"− τbc
⎧
κ A
!
⎨ d d (λ + τ )(λ − λ ) τbc τbc − τd τ ̸= τ
% t0
⎪
⎪
u bc u d bc d
A1 = e(t)dt = τd τd
0
⎩ κd Ad (λu + τbc )(λu − λd )e−1
⎪
τbc = τd
⎪
τd
According to
% ∞
1 Ad
IE FF = e(t)dt = lim s · Gy/d (s) = 0.
0 s →0 s s
the area | A2 | in the previous figure is equal to | A1 |, and the I AE

value can finally be determined as
"− τbc
⎧
κ A
!
τ
⎪
⎨2 d d (λ + τ )(λ − λ ) bc τbc − τd τ ̸= τ
⎪
u bc u d bc d
IAE FF = 2| A 1 | = τd τd
⎩2 κd Ad (λu + τbc )(λu − λd )e−1
⎪
τbc = τd
⎪
τd
Calculation of I AE FF for non-interacting FF scheme
In this case, the I AE FF estimation can be obtained in a straightforward

manner, as the effect from the feedback controller is removed.
The IAE result obtained in the non-invertible delay case can be
reformulated as
λu −λ
! ! ""
− d
I AEFF = κd A d ( λu − λd ) − ( τd − τu − τu + τu ) 1 − 2e τd − τu − τu + τu
λu −λ
τd − τu − τu + τu
! ! ""
− τ − τu − τud+ τu
= κd Ad 1 − 1 − 2e d (λu − λd )
λu − λd
! "
1 2 −a
= κd Ad 1 − + e (λu − λd )
a a
= κd Ad α(λu − λd )
where
1 2 −a λu − λd
α = 1− + e , a=
a a τd − τu − τu + τu
Analysis and discussion on the indices
Feedback control without feedforward:

I AEFB = κd Ad (λu + τbc )
Feedforward with classical control scheme and classical tuning:
κd Ad
IAEFF = 2 (λu + τbc )(λu − λd ) f (τbc /τd ) (1)
τ
where
τbc
⎧
⎨ τbc − τ − τ
⎪
⎪ ! "
bc d τ ̸= τ
f (τbc /τd ) = τ bc d (2)
⎪ d
⎩ e −1
⎪
τ =τ bc d
Feedforward with non-interacting control scheme:
I AEFF = ακd Ad (λu − λd )
where α can vary based on the τ f f value.
Analysis and discussion on the indices
The ratio between the I AE value of the classical scheme and the
noninteracting scheme is
I AEclassical 2(λu + τbc ) f (τbc /τd )

=
I AEnoninteracting τd α
Therefore, the classical scheme gives a smaller I AE value when
2(λu + τbc ) f (τbc /τd )

τd >
α
Index interpretation
For the classical feedforward control case, the index becomes
I AEFF 2( λ u − λ d )
IFF/FB = 1 − = 1− f (τbc /τd )
I AEFB τd
For the noninteracting feedforward control scheme, the index is given

by
I AEFF α(λu − λd )
IFF/FB = 1 − = 1−
I AEFB λu + τbc
Example 1
e−2s e− s
Pu (s) = Pd (s) =
10s + 1 5s + 1
Using lambda tuning with τbc = τu = 10 gives the PI controller
parameters κ f b = 0.83 and τi = 10.
The feedforward compensators are defined as
10s + 1
C f f ( s) =
5s + 1
for the classical feedforward control scheme and as
10s + 1
Cf f =
4.4s + 1
for the non-interacting feedforward control scheme (to minimize IAE).
Example 1
Control scheme I AEr I AEe IFF/FB

Feedback 11.99 12 –
Classical FF 1.21 1.2 0.9
Non-interacting FF 0.63 0.63 0.95
Example 1
0.8
Feedback
Classical FF
0.6 Non−interacting FF
0.4
y
0.2
−0.2
0 10 20 30 40 50 60 70
0
Feedback
Classical FF
−0.5
Non−interacting FF
−1
u
−1.5
−2
−2.5
0 10 20 30 40 50 60 70
time
Example 2
The differences between the pure feedback scheme and the

feedforward schemes can be reduced by retuning the PI controller to
obtain a more aggressive response. Lets retune the PI controller only
for the case when pure feedback is used, by using τbc = 0.25τu .
Example 2
Control scheme I AEr I AEe IFF/FB

Feedback 4.5 4.5 –
Classical FF 1.21 1.2 0.73
Non-interacting FF 0.63 0.63 0.86
Example 2
0.6
Feedback
Classical FF
y
0.2
−0.2
0 10 20 30 40 50 60
0
Feedback
Classical FF
−0.5
−1
u
−1.5
−2
−2.5
0 10 20 30 40 50 60
time
Example 3
Assume that τbc = τu = λu . It means that we have a process model

Pu (s) where the delay is equal to the time constant and that the
lambda tuning rule is used with τbc = τu . Two different values of the
time constant τd = ηλu , where η = 1 or 10.
Example 3
τd = λu τd = 10λu
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
1 − λd /λu 1 − λd /λu
Example 3
τd Control scheme I AEr I AEe r

IFF/FB e
IFF/FB
λu Feedback 2.04 2.0
Classical FF 1.43 1.47 0.30 0.26
Non-interacting FF 0.63 0.63 0.69 0.69
10λu Feedback 2.00 2.0
Classical FF 0.34 0.31 0.83 0.85
Non-interacting FF 0.63 0.63 0.69 0.69
Example 3
τ=L
Feedback
1 Classical FF
0.5
y
−0.5
0 1 2 3 4 5 6 7 8 9 10
0.5
Feedback
0 Classical FF
−0.5
u
−1
−1.5
−2
−2.5
0 1 2 3 4 5 6 7 8 9 10
time
Example 3
0.2 τ = 10L
Feedback
Classical FF
0.1
y
0.05
−0.05
0 5 10 15 20 25 30 35 40
0 Feedback
Classical FF
−0.2 Non−interacting FF
−0.4
u
−0.6
−0.8
−1
0 5 10 15 20 25 30 35 40
time
Outline
1 Introduction

5 Conclusions
Conclusions
The motivation for feedforward tuning rules was introduced.

The feedback effect on the feedforward design was analyzed.
The different non-realizable situations were studied.
The two available feedforward control schemes were used.
Simple tuning rules based on the process and feedback
controllers parameters were derived.
Performance indices for feedforward control were proposed.
Conclusions
Future research
What else can be done?

Nominal tuning. Unified methodology for low-order feedforward
controllers tuning
Robust tuning. Scale up to other feedforward structures
DTC with feedforward action. Extension to MIMO processes
Experimental results. Validate the theoretically claimed benefits
Distributed parameter systems. Feedforward tuning rules to
deal with resonance dynamics
Conclusions
Bibliography
1 C. Brosilow, B. Joseph, Techniques of Model-Based Control, Prentice-Hall, New Jersey,

2002.
2 J.L. Guzmán, T. Hägglund. Simple Tuning rules for feedforward compensators. Journal of
Process Control. 21(1), 92-102, 2011.
3 J.L. Guzmán, T. Häggund, K. Aström, S. Dormido, M. Berenguel, Y. Piguet. Feedforward
control concepts through interactive tools. 18th IFAC World Congress, Milano, Italy, 2011.
4 J.L. Guzmán, T. Hägglund, A. Visioli. Feedforward Compensation for PID Control Loops.
In PID Control in the Third Millennium, Springer, 2012, pp. 207-234. ISBN
978-1-4471-2424-5.
5 J.L. Guzmán, T. Hägglund, M. Veronesi, A. Visioli. Performance indices for feedforward
control. Journal of Process Control, 2014 (Under review).
6 M. Hast, T. Hägglund, Design of optimal low-order feedforward controllers, in: IFAC
Conference on Advances in PID Control, Brescia, 2013.
7 A. Isaksson, M. Molander, P. Moden, T. Matsko, K. Starr, Low-Order Feedforward Design
Optimizing the Closed-Loop Response, Preprints, Control Systems, 2008, Vancouver,
Canada.
Conclusions
Bibliography
8 A. Pawlowski, J.L. Guzmán, J.E. Normey-Rico, M. Berenguel. Improving feedforward

disturbance compensation capabilities in generalized predictive control. Journal of
Process Control, 22(3), 527-539, 2012.
9 C. Rodrı́guez, J. L. Guzmán, M. Berenguel, T. Hägglund. Generalized feedforward tuning
rules for non-realizable delay inversion. Journal of Process Control, 23(9), 1241-1250,
2013.
10 C. Rodrı́guez, J. L. Guzmán, M. Berenguel, T. Hägglund. Optimal feedforward
compensators for systems with right-half plane zeros. Journal of Process Control, 24(4),
368-374, 2014.
11 C. Rodrı́guez, J. E. Normey-Rico, J. L. Guzmán, and M. Berenguel. A robust design
methodology for simultaneous feedforward and feedback tuning. Submitted to ISA
Transaction.
12 C. Rodrı́guez, J. L. Guzmán, M. Berenguel, and J. E. Normey-Rico. On the filtered Smith
predictor with feedforward compensation. Submitted to IEEE Transactions on Control
Systems Technology
Conclusions
Bibliography
13 C. Rodrı́guez, J. L. Guzmán, M. Berenguel, and J. E. Normey-Rico. Optimal feedforward

compensators for integrating plants. To appear in 19th IFAC World Congress. Cape Town,
South Africa, August 2014.
14 D. Seborg, T. Edgar, D. Mellichamp, Process Dynamics and Control, Wiley, New York,
1989.
15 F. G. Shinskey, Process Control Systems. Application Design Adjustment, McGraw-Hill,
New York, 1996.
16 R. Vilanova, O. Arrieta, P. Ponsa, IMC based feedforward controller framework for
disturbance attenuation on uncertain systems, ISA Transactions 48, 439-448, 2009.
17 S. Skogestad. Tuning for Smooth PID Control with Acceptable Disturbance Rejection.
Industrial Engineering and Chemical Research, 45, 7817-7822, 2006.
Collaborators
1 Carlos Rodrı́guez (UAL, Spain)

2 Tore Hägllund (Lund, Sweden)
3 Manuel Berenguel (UAL, Spain)
4 Julio Normey (Florianopolis, Brazil)
5 Antonio Visioli (Brescia, Italy)
6 Max Veronesi (Milano, Italy)
End of the presentation
Thank you for your attention

Feed Foward Non Real Iz Able

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Feed Foward Non Real Iz Able

Uploaded by

Copyright:

Available Formats

Advances in Feedforward Control

for Measurable Disturbances

José Luis Guzmán Sánchez

Madrid, Febrero 2019

2 Feedforward control problem

3 Nominal feedforward tuning rules

4 Performance indices for feedforward control

2 Feedforward control problem

3 Nominal feedforward tuning rules

4 Performance indices for feedforward control

What are load disturbances?

Most industrial processes are subject to disturbances and the

Real plants at the Automatic Control research group in Almerı́a

Motivation: feedback controller

Motivation: feedback controller

No reaction until there are discrepancies!

Motivation: feedforward compensator

Motivation: feedforward compensator

Motivation: feedforward compensator

Feedforward control problem

Perfect compensation is seldom realizable:

Non-realizable delay inversion.

Motivation: non-ideal feedforward compensator

Motivation: non-ideal feedforward compensator

Motivation: residual term

(a) Open−loop response (b) Closed−loop response

An interaction between feedforward and feedback controllers arises

Other design strategies are required!

An interaction between feedforward and feedback controllers arises

Other design strategies are required!

D. Seborg, T. Edgar, D. Mellichamp, Process Dynamics and Control,

1 Study and development of a control methodology to improve

2 Feedforward control problem

3 Nominal feedforward tuning rules

4 Performance indices for feedforward control

Feedforward control is an old topic in process control. In fact, its

Feedforward control problem

The idea behind feedforward control from disturbances is to supply

In industry, PID control is commonly used as feedback controller and

Static with delay: C f f = κ f f e−sL f f

Feedforward control problem

Then, if we consider that process transfer functions are modeled as

Lets consider the following example:

Static with delay: C f f = e− s

Feedforward control problem

Open−loop response Closed−loop response

Static Without Feedforward

Feedforward control problem

(a) Open−loop response (b) Closed−loop response

Feedforward control problem

r r + Pd∗ (e−λu s − e−λd s )d

2 Feedforward control problem

3 Nominal feedforward tuning rules

4 Performance indices for feedforward control

Cases to be evaluated in this research:

2 Feedforward control problem

3 Nominal feedforward tuning rules

4 Performance indices for feedforward control

Nominal feedforward design: non-realizable delay

Nominal feedforward design: non-realizable delay

Nominal feedforward design: non-realizable delay

Delay inversion: open-loop compensation

Pd C f fP u O v e rsh o ot e rror I n i t i al e rror