Sunteți pe pagina 1din 51

Multi-Objective Dynamic

Optimization
using
Evolutionary Algorithms
by
Udaya Bhaskara Rao N.
under the guidance
of
Dr. Kalyanmoy Deb
Professor
Department of Mechanical Engineering

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Birds view

Introduction to DMO.

Test problems in DMO.

NSGA-II application in DMO.

Introduction to hydrothermal scheduling problem.

NSGA-II application on hydrothermal scheduling problem.

Hydrothermal scheduling problem formulation as DMO.

Modifications in the proposed algorithm.

Results and discussion.

Conclusions.

Future scope of work.

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Introduction to DMO
Dynamic optimization is optimization in dynamic environment.
i.e. either objective function or constraints are time dependent.

The dynamic multi-objective optimization (DMO) is multiobjective optimization in dynamic environment.


Classification in DMOs :

POS
POF

No change

Change

No change

Type IV

Type I

Kanpur Genetic Algorithms


Laboratory Type III
Change
IIT Kanpur
25, July 2006 (11:00 AM)

Type II

Introduction to DMO
It is better to go for DMO, whenever the problem is time
dependent.
Advantages in using DMO:
1. By relating time with generation number, number of
variables reduce i.e. the dimension of problem reduces.
2. Whenever problem changes, the new problem adopts
the old solution, which helps in faster convergence.
3. Results for all the problems can be found in one run.
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Test problems in DMO

The following test problems are formulated by Farina et. al.


(2004).
FDA 1 : Constant convex Pareto-optimal front in objective space and linear
change in solution space.

FDA 2 : Pareto-optimal front changes from convex to non convex and no


change in solution space.

FDA 3 : Change in Pareto-optimal front but all convex and linear change in
solution space.

FDA 4 : Constant non convex Pareto-optimal front and linear change in


solution space which is three Dimensional space.

FDA 5 : Change in Pareto-optimal front but all non convex and linear
in solution
space which is three dimensional space.
Kanpur Geneticchange
Algorithms
Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Test problems in DMO


FDA 1:

Type I
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Test problems in DMO


FDA 2:

Type III
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Test problems in DMO


FDA 3:

Type II
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Test problems in DMO


FDA 4:

Type I
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Test problems in DMO


FDA 5:

Type II
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

NSGA-II application in DMO


Present algorithm is developed based on NSGA-II.
NSGA-ll algorithm can not be applied straightaway on DMO
problems.
Elitism, restricts the upward movement of Pareto-optimal
front in NSGA-II and hence removed.
The term time is correlated with generation number.

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Modified NSGA-II algorithm-I

Elitism removed

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Modified NSGA-II algorithm-II

Elitism
introduced
interactively

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

FDA 2 simulation

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

FDA 3 simulation

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

FDA 5 simulation

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Introduction to hydrothermal
scheduling problem
In hydrothermal systems both hydroelectric and thermal
generating units are to be utilized together to meet the total
power demand.
The hydrothermal problem here consists of Ns number of
thermal and Nh number of hydroelectric generating units
sharing the total power demand.
Minimizing both fuel cost and emission of nitrogen oxides from
the thermal generating units.
The static problem formulation is taken from the work done by
M. Basu (2005). (Weighted sum approach using simulated
annealing)

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Introduction to hydrothermal
scheduling problem
In this present work the problem is formulated for two
hydraulic units and four thermal units.
Problem is defined for four timeslots of each 12 hours.
So the total number of variables are 24.
The demand values for these four time slots are as follows:

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Objective functions
Economy:

Emission:

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Constraints
Power balance constraints:

Water availability constraints:

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Constraint handling
Step 1 : The procedure is to be started with the two water
available constraints, as they are independent of variables
related to thermal units.

Step 2 : Constraint equation can be written as,

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Constraint handling
Step 3 : Start with h = 1, m = 1
For finding Phm value from constraint equation, first rewrite
the equation in terms of Phm by taking all four Ph values of the
present hydro unit from GA solution and finding out the
ratios with respect to Phm.
The obtained quadratic equation in terms of Phm is solved
algebraically to get Phm value. Subsequently the positive
value is chosen, so that the lower limit is satisfied
automatically. If it is also satisfied the upper limit go to Step
5, else go to Step 4.
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Constraint handling
Step 4 : m = m + 1, and if m 4 repeat Step 3 else go to Step
6.

Step 5 : Change all four Ph value by using previously calculated


ratios and h = h + 1, if h 2 repeat Step 3, else Exit.

Step 6 : The constraint is not satisfied, so for the present


variable values, the fitness function is to be penalized with
the constraint violation.

If both water availability constraints are satisfied through the


above process, similar analysis is to be done on power
balance constraints
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

NSGA-II application on hydrothermal


scheduling problem
Input parameters :
Population size = 240
Number of generations = 1000
Crossover probability = 0.9
Mutation probability = 0.04
Distribution index for crossover =
20
Distribution index for mutation =
50

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Analysis of obtained results

Ph1 vs. F1
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Ph2 vs. F1

Analysis of obtained results

Ps1 vs. F1
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Ps2 vs. F1

Analysis of obtained results

Ps3 vs. F1
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Ps4 vs. F1

Hydrothermal scheduling problem


reformulation as DMO

As the problem parameters change with time, this


problem comes under DMO.

Few modifications required in problem formulation, they


are as follows:
1. Term time should be removed as a variable.
2. All time variable parameters should be directly related with
generation number.

The dimension of problem has reduced from 24 to 6.

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Modifications in the proposed


algorithm

The proposed algorithm can handle the changes occurring


after every 12 hours.

The algorithm is further modified to handle frequent


changes.

The main modifications are as follows:


1. Introducing new solutions at change by generating random
solutions.

2. Introducing new solutions at change by mutating old solutions .


Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

First modification

Introducing
new
random
solutions
at
change

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Second modification

Introducing
new
mutated
solutions
at
change

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Interpolation

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Results and discussion


Comparison among three modified algorithms :

4
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

8
timeslots

Results and discussion


Comparison among three modified algorithms :

16
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

48
timeslots

Results and discussion


Comparison among three modified algorithms :

96
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

192
timeslots

Results and discussion


Percentage of random new solutions verses performance
index :

4
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

8
timeslots

Results and discussion


Percentage of random new solutions verses performance
index :

16
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

48
timeslots

Results and discussion


Percentage of random new solutions verses performance
index :

96
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

192
timeslots

Results and discussion


Percentage of mutated new solutions verses
performance index :

4
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

8
timeslots

Results and discussion


Percentage of mutated new solutions verses
performance index :

16
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

48
timeslots

Results and discussion


Percentage of mutated new solutions verses
performance index :

96
timeslotsLaboratory
Kanpur Genetic Algorithms
IIT Kanpur
25, July 2006 (11:00 AM)

192
timeslots

Conclusions
1. Static to dynamic conversion of the problem, increases its
convergence rate and simultaneously there also exists a
possibility for dimensionality reduction.
2. Modified NSGA-II algorithms, has yielded better results for
all test problems.
3. The reformulated hydrothermal scheduling problem has
been solved efficiently.
4. The static analysis of hydrothermal scheduling problem
with modified NSGA-II produced better results compared to
previous works.
Kanpur Genetic Algorithms Laboratory
IIT Kanpur
25, July 2006 (11:00 AM)

Conclusions
5.

Best results are produced when the hydrothermal


scheduling problem is formulated into DMO problem,
with considerable reduction in computational time over
static problems.

6.

The final proposed algorithm has increased the


possibility in achieving Pareto-optimal front within short
time period and performs best up to one hour time slot.

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Future scope of work


1.

Generalization of the proposed algorithm, would make it


user friendly.

2.

The hydrothermal scheduling problem defined for


individual hydro units, can be extended for cascaded
hydro units.

3.

The present algorithm is used to search for Pareto-optimal


front, this algorithm can be slightly modified to get
reliable and robust Pareto-optimal solutions also.

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Back up
slides

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Parameter Analysis

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Nomenclature

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Input parameters

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Input parameters

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

Pareto front
Non-dominated front is Pareto-optimal front.
Trade-off of optimal solutions on F1 vs F2 plot.

Kanpur Genetic Algorithms Laboratory


IIT Kanpur
25, July 2006 (11:00 AM)

S-ar putea să vă placă și