11/17/2019

Ali M. Sahlodin
Department of Chemical Engineering
AmirKabir University of Technology
1397 S.H

 Parametric Dynamic Systems

 Parametric sensitivities
 Finite differences
 Forward sensitivities

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 2

1
 11/17/2019

Optimization algorithms need

sensitivities of objective/constraints
w.r.t parameters

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 3

 Dynamic optimization (DO) problem

ABC
Maximize final production of B

 Find optimal temperature

 very simplified DO! Only one parameter.
Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 4

2
 11/17/2019

 Form the system of forward sensitivities

 t ; p)  f (t , x(t; p), p)
x(

x p j (t ; p)  f x (t , x(t; p), p)x p j (t; p)  f p j (t , x(t; p), p)

x p j (t0 ; p)  h p j (p)

The system parameter is temperature θ

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 5

 Forward sensitivity equations

 dk 
(C A )    1 C A2  2k1C A (C A ) 
 d 
dk  dk 
(C B )  1 C A2  2k1C A (C A )   2 CB  k2 (CB ) 
d  d 
dki b b
 ai ( 2 ) exp(  )
d R R
at t  0 (C A )  (CB )  0

1. Integrate the above sensitivity system together

with the original ODEs
2. Evaluate sensitivity for the objective function
Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 6

3
 11/17/2019

•Finite differences: (np+1) integrations of nx eqs. Computationally expensive

•Forward sensitivity: 1 integration of (np+1)nx eqs. for even small plant models

What sensitivities do we actually need?

Sensitivities of these
are finally needed.
State sensitivities are
intermediate results and
not required per se.

Can we bypass state sensitivities and directly calculate sensitivities of

objective/constraints?
Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 7

Objective function J (p) :  (x(tf ; p), p)

dJ   dx
Sensitivity at p j : (p)  (x(tf ; p), p)  (x(tf ; p), p) (tf ; p)
dp j p j x dp j

1. Adjoin the differential equations to the objective function

tf
J (p) :  (x(tf ; p), p)   λ(t )T f (t , x(t; p), p)  x(
 t; p) dt
t0

2. Differentiate the augmented functional

dJ dJ
(p)  (p)   p j (x(tf ), p)  x (x(tf ), p)T x p j (tf )
dp j dp j
t
  λ(t )T f p j (t , x(t ), p)  f x (t , x(t ), p)x p j (t )  x p j (t ) dt
f

t0  

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 8

4
 11/17/2019

3. Apply integration by part

tf t tf
t λ(t )T x p j (t )dt   λ(t )T x p j (t )    λ(  t )T x (t )dt
f

0
  t0 t0 pj

4. Substitute

dJ dJ
(p)  (p)   p j (x(tf ), p)  λ(t0 )T h p j (p)
dp j dp j

 
T
 x (x(tf ), p)-λ(t f ) x p j (tf )
t T
 t )  x ) dt
  (λ(t )T f p j (t , x(t ), p)  f x (t , x(t ), p)T λ(t )  λ(
f

t0   pj

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 9

5. To eliminate state sensitivities, their multipliers must be zero.

λ * (t )  f x (t , x(t ), p)T λ* (t ) λ* (t f )  x (x(tf ), p)

5. Now, sensitivities of objective function is simplified as

Direct influence of
parameter on objective

dJ dJ
(p)  (p)   p j (x(tf ), p)  λ* (t0 )T h p j (p)
dp j dp j
tf
  (λ* (t )T f p j (t , x(t ), p)) dt Influence of initial conditions
t0

Influence of system
dynamics
Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 10

5
 11/17/2019

1. Integrate system dynamics (forward in time)

 t ; p)  f (t , x(t; p), p) x(t0 )  h(p)
x(
 Store state values at time points
t0  t1    t f
2. Integrate the adjoint system (backward in time)
λ * (t )  f x (t , x(t ), p)T λ* (t ) λ* (t f )  x (x(tf ), p)

q j =-λ(t )T f p j (t , x(t ), p) q j (t f )=0, j  1,, n p

3. Compute gradients
dJ
(p)   p j (x(tf ), p)  λ* (t0 )T h p j (p)  q j (t0 )
dp j

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 11

 Maximize J ( )  CB2  a 2

 Construct adjoint equations

λ * (t )  f x (t , x(t ), p)T λ* (t ) λ* (t f )  x (x(tf ), p)

q j =-λ(t )T f p j (t , x(t ), p) q j (t f )=0, j  1,, n p

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 12

6
 11/17/2019

 Adjoint equations
1* (t )  T
 2k1C A 0  1 (t ) 
* 1* (t f )  0 
      ,   
2* (t )   2k1C A  k2  2* (t )  2C (
2 (t f )   B f 

*

t )

T
 dk1 2 
  d C A  1* (t ) 
q =-     , q(t f )=0
 dk1 C 2  dk2 C  2* (t ) 
 d A d B 

dJ
( )  2a  λ* (t0 )T h p j (p)  q(t0 )
d  
0

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 13

Adjoint Sensitivity
 More efficient than forward sensitivity when there is a
large number of parameters and few functionals.
 Requires more memory for storing state values in the
forward integration.
 With variable step integrators, interpolation is needed to
evaluate state values for the adjoint system.
 Does NOT provide state sensitivities.
 Does NOT provide sensitivity at intermediate times.

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 14

7
 11/17/2019

 Partial derivatives should be obtained via

symbolic or automatic differentiation
 Avoid finite differences if you can.
 Generally, sensitivity results are local: only
valid at p .

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 15

Solver Website License Forward Adjoint

CVODES http://acts.nersc.gov/sundial Free Yes Yes
s/
IDAS http://acts.nersc.gov/sundial Free Yes Yes
s/
DASPK3.1 http://www.engineering.ucsb Free for acad. Yes
.edu/
~cse/software.html
DASPKadjoint http://www.engineering.ucsb Free for acad. Yes
.edu/
~cse/software.html

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech. 16