EE5712 Power System Reliability

:: Resource Planning Problem

Panida Jirutitijaroen
Problem statement
Proposed solution methods
Global decomposition with Dynamic programming
Stochastic programming

OUTLINE
Objective and challenges
System modeling
Reliability evaluation tools

PROBLEM STATEMENT
 Problem Statement
• Multi-area power systems
• Long term planning 1 2 8
• Demand growth in the
future.
• Make decision on where
to locate capacity at how 6 3 9
much that will yield
minimum cost and
acceptable reliability. 5 11 7
– Expected Unserved Energy
(EUE): Expected energy not
supplied to load.
– Loss of Load Probability. 10 4 12
 Objective
• To investigate an effect of additional capacity
in the system to system reliability.
• To quantify in terms of improved reliability
that additional capacity can contribute to
system reliability.
• To propose systematic approaches of deciding
the best placement of additional capacity in
the system.
 Challenges
• No explicit reliability equation in terms of
decision variables of additional capacity
• Randomness in generation, tie-line capacity
and load and in additional capacity (its
availability)
• Reliability evaluation is computationally
expensive
• Decision variables are binary
 System Modeling
gi   li  
tij  

…….

…….
s Power system network t
(only tie line capacity)

• Capacity flow network

• Network random variables: area generation
capacity, load, tie line capacity between areas
• Discrete probability distribution functions of all
random variables are constructed.
 Reliability Evaluation Tools
• Monte Carlo simulation
– Create history of components by using their
probability distributions.
– Reliability indices estimated by statistical
inference
• Decomposition-analytical method
– To categorize states into acceptable, unacceptable
sets and compute for reliability indices.
Sensitivity analysis
Scenario analysis
Dynamic programming
Hybrid method
Stochastic programming
Stochastic programming with sample-average approximation

PROPOSED SOLUTION METHODS

 Sensitivity Analysis
• Monte Carlo simulation
• Increase generation in
-3
x 10 Loss of Load Probability with Additional Generation in Different Areas
5

each area 4.5

Area 1,2,3,6,7,9,10,11
Area 4
4

• Find loss of load

Area 5,12
Area 8
Area 13

Loss of Load Probability

3.5

probability with respect 3

to additional generation 2.5

in each area 2

1.5
0 10 20 30 40 50 60 70 80 90 100
Generation Addition with Incremental of 10 MW
 Scenario Analysis
• Use global decomposition to evaluate system
reliability
• Analyze sets of scenarios chosen by the
experts
• Smart comparison along with reliability
evaluation
• Require less memory usage
• No need to perform complete decomposition
 Decomposition Stages
State space
A set (Unclassified Set) L set
“U” set

First stage of decomposition

A set U set L set … … A set U set L set

A set U set L set

Second stage of decomposition

A set U set L set … A set U set L set A set U set L set … A set U set L set A set U set L set … A set U set L set
 Comparison Algorithm
• Initialization
– System state space is the first U set.
– Possible solution list ← every combination i.
• Decomposition and Evaluation
– Partition U sets into A and L sets
– Compute and update PAi and PLi for every combination i.
– Find PAmax  max
i
PAi
• Comparison
– If PLi  1  PAmax  , delete combination i from possible solution list.
• Stopping criterion
– Stop if number of combination is 1 or for all i PUi  1  PAi  PLi  
– Otherwise, create U sets and go to step 1.
 Scenario Analysis Results
Decomposition Number of Decomposition Number of omitted
Stage omitted scenarios Stage scenarios
1-14 0 1-41 0
15 36 43 78
16 84 44 0
18 36 45 12
19 8

• Less memory usage

• Reduce number of possible solution at each
stage of decomposition
 Dynamic Programming
• Use global decomposition to evaluate system
reliability.
– Decomposition depends on state capacity NOT
probability
• Reliability equation is derived and
approximated in a form of separable
structure.
• Dynamic programming is applied to solve the
approximated problem.
 Stochastic Programming
• General framework for capacity expansion
problem with explicit reliability consideration.
• Formulate as two-stage recourse model and use
expected unserved energy as reliability index.
• Variety in network modeling; for example,
network capacity flow, DC flow.
• Solve by well known L-shaped method.
• However, a large number of scenarios affects
computational time.
 Sample Average Approximation
• Stochastic programming with two-stage
recourse model
• Application of sampling techniques to reduce
number of scenarios to evaluated in the
solution procedures.
• Effectively estimate the optimal solution with
a small number of samples.
• Reduce computation time.
Reliability equation
Generation Capacity Probability Table
Dynamic programming application to generation expansion planning
Three-area system example
Hybrid method

GLOBAL DECOMPOSITION WITH

DYNAMIC PROGRAMMING
 Global Decomposition
• State space includes all possible generation and is
valid for all possible generation combination.
• Generation addition in the candidate locations
are artificially included in the model.
• Decomposition depends on capacity NOT
probability!
• After decomposition, reliability can be calculated
for any area generation combination by assigning
zero probability to the omitted states.
 Probability Calculation
• Given a set, X, can assume any value between its minimum
and maximum state,
M1X  M nX 
X  X X 
 1
m  m n 
• Probability is found from, N
Pr X    p xk
k 1 mkX  x k  M kX
px k= probability of state xk of arc k
• Using cumulative probabilities,
 
N
Pr  X    pM X  pm X 1
k k
k 1
px k = cumulative probability of state xk of arc k
 Generation Capacity Probability Table

Origi With 1 unit With 2 units

MW -nal
Gn 2 - - PG0n PU1 PU2
Gn 1 - PG0n PU1 PG0n1 PD1 PD2  PG0n PD1 PU2  PG0n PU1 PD2  PG0n1 PU1 PU2
Gn PG0n PG0n PD1  PG0n1 PU1 PG0n PD1 PD2  PG0n1 PD1 PU2  PG0n1 PU1 PD2  PG0n2 PU1 PU2
… … … …
G2 PG02 PG02 PD1  PG01 PU1 PG02 PD1 PD2  PG01 PD1 PU2  PG01 PU1 PD2
G1 PG01 PG01 PD1 PG01 PD1 PD2
 Generation Capacity Probability Equations

• Assume that the addition units are identical in each area, i. e.

PU1  PU2    PUm  PU , and, PD1  PD2    PDm  PD
i
where PD is probability of i th unit addition being down
PUi is probability of i th unit addition being up
• For m unit addition, we will have the following equation,
 m 0  m
P  P PD   P  PD  PU  PGi2  PD  PU     PG0im PU 
m 0 m 0 m 1 m2 2 m
Gi Gi Gi 1
1 2
m
m
Or, P   P
m
Gi
0
Gi k  PD m  k PU k where PG0  0 if i  k
k
i k
k 0

Note that, this probability is in terms of number of unit addition only

 Probability Equation for a Set

Pr  X    g j m j 
n

j 1 ,
m j  M jA  m j  
g j m j      PG j  FOR j  j 1  FOR j  
0 m  k k

k  0  i  m Aj i k 

  k 
where
m j = Number of unit addition in area j
FORj = Forced outage rate or failure probability of a unit addition in
area j
PG0 j = Probability of generation at level i-k before unit addition in area j
ik
 Reliability Equation
• Generation capacity probability equations for
each area are derived in terms of number of
unit additions.
• Probability of any set is computed from
multiplication.
• Then, loss of load probability equation is
derived as a function of unit addition in each
area by summing up all the L sets.
 Approximation
• The first decomposition produces 1 A set and
n L sets.
• Probability equation of one (either A or L) set
has special form: it is a separable function!
• Moreover, probability equation of the first L
sets also has separable structure.
• The separable structure allows Dynamic
programming to solve the problem.
 Dynamic Programming
• The function needs to be in separable form.
• Problem can be divided into subproblems.
• In each subproblem, there will be a certain
number of possible solutions in it.
• A recursive formula can be written in terms of
previously computed solution from
proceeding subproblem.
The First A Set Probability Equation

Pr  A   h j m j 
n

j 1

 0  m j  
h j m j     PG j  PG j  FORj  1  FORj  
mj
0 m j k k
 A k 1 
k 
k 0 
 jM A k m j   
where
m j = Number of unit addition in area j
PG0j = Cumulative probability of generation at level i-k before unit
i k
addition in area j

This equation is probability of system success after the first

decomposition.
 The First L Sets Probability Equation

       
n
Pr L    Pr Li   g11 m1   1  g11 m1  g 22 m2   1  g 22 m2   g nn11 mn1   1  g nn11 mn1  g nn mn 
i 1

 m  k
g m j     PG0j  j FORj  j 1  FORj  
mj
j m k
j
k 0 
vi 1k
 k  
where0
PG j
= Cumulative probability of capacity arc generation area j
vi 1 k
before
unit addition

This equation is LOLP (probability of loss of load) after the first

decomposition
 Optimization Problem
• Let the budget constraint be of the following
form,
a1m1  a2m2   an mn  R
where
a j = Cost per one unit addition to area j
R = Budget
• Maximizing reliability subject to cost
– the first A set optimization: maximize system availability.
– the first L sets optimization: minimize the LOLP.
– LOLP is widely used and practical reliability index, emphasize on the
optimization for the first L sets.
 Dynamic Programming
• Problem can be divided into subproblems
called stages.
• In each stage, there will be a certain number
of states (possible solutions) in it.
• A decision is made at each stage which will
transform the state at current stage to the
state at the following stage.
• A recursive formula can be written in terms of
previously computed solution from
preceeding stage.
 Application to GEP
• Reliability equation is of separable function,
i.e. multiplication single variable function.
• DP can be applied to separable problem (A)
max Pr  A   g j m j 
n

j 1

s.t. a1m1  a2 m2    an mn  R
mj  I 
where a j is cost per unit addition to area j
R is total budget available
m j is number of unit addition in area j
• Definition of stages and states
– Stages are area
– States are possible value of budget available at each stages

S2 = 0,1,…,R Sn = 0,1,…,R
S1 = R
Stage 1 Stage 2 …. Stage n
m1 m2 mn

where S j is possible states at stage j

• At each stage, a decision is made on how much to
spend on each area generation
• Note that the problem is divided into subproblems.
Each subproblem consists of smaller number of
areas.
m1, m2 ,, mn1, mn , m1, m2 ,, mn2 , mn1,, mn1, mn , mn 

S1 = R S2 = 0,1,…,R Sn-1 = 0,1,…,R Sn = 0,1,…,R

Stage 1 Stage 2 …. Stage n-1, Stage n

m1 m2 mn-1 mn

an1mn1  an mn  Sn1 an mn  Sn
a2 m2    an1mn1  an mn  S2
a1m1  a2m2   an1mn1  an mn  S1  R
• At the last stage, the problem is to
max g n mn 
s.t. an mn  S n
mn  I 
where S n is possible states at stage n, takes value from 0 to R

• At any stage, the problem is to

n
max  g k mk 
k j

s.t. a j m j  a j 1m j 1    an mn  S j
mk  I 
where S j is possible states at stage n, takes value from 0 to R
 Recursive Function
• Dynamic programming solves the smallest
subproblem first (smallest number of variables
in it), for this problem it is the last stage.
• The optimal solution to the current
subproblem (current stage) is calculated using
the optimal solution from the previously
smaller subproblem (following stage).
• Let f j S j  be the optimal objective function at
stage j with available budget S j
• At the last stage, the problem is
max g n mn 
s.t. an mn  S n
mn  I 
where S n is possible states at stage n, takes value from 0 to R

• The optimal solution is

 Sn 
f n Sn   gn mn  where 0  mn   
  Sn    an 
or, f n S n   g n    
  an  
where S nis possible states at stage n, takes value from 0 to R
• At any stage, the problem is to
n
max  g m  k k
k j

s.t. a j m j  a j 1m j 1    an mn  S j
mk  I 
where S j is possible states at stage j, takes value from 0 to R

• The optimal solution is

f j S j   maxg j m j  f j 1 S j  m j 
Sj 
where 0  mj   
 a j 
S j is possible states at stage j, takes value from 0 to R
• The same function is also applied to the first
stage where S1  R .
• The recursive function is in the following,
 S j 
 g j 
aj  , jn
f j S j   
  
 max Sj 
g j m j  f j 1 S j  m j  , j  1,, n  1
0 m j   a j 
  
 Example
Area Area
Let budget = 10 m$
1 2
State Area 1 Area 2 Area 3
of
Cap. Cap Prob Cap Prob Cap Prob
Arc (MW) (MW) (MW)
Area
7 600 0.262144
3
6 500 0.32768 500 0.393216 500 0.32768

Area j $ Load 5 400 0.40960 400 0.245760 400 0.40960

millions (MW) FORj 4 300 0.20480 300 0.081920 300 0.20480
aj 3 200 0.05120 200 0.015360 200 0.05120
1 3 500 0.15
2 100 0.00640 100 0.001536 100 0.00640
2 5 600 0.05
1 0 0.00032 0 0.000064 0 0.00032
3 4 500 0.1
 First Global Decomposition
• From the budget constraint, we can add at most 3 units in area 1, 2 units
in area 2, and 2 units in area 3.
• These units are included before performing decomposition. State of
capacity arc will start from 1 to 9.
• After the first decomposition, the first A set is

9 9 8 2 2 2
A 
• The first L sets are 6 7 6 1 1 1 

3 9 8 2 2 2 9 4 8 2 2 2  9 9 3 2 2 2 
L1    L2    L3   
1 1 1 1 1 1 4 1 1 1 1 1  4 5 1 1 1 1 
The First A Set Optimization: Problem Formulation

max Pr  A  h1 m1   h2 m2   h3 m3 

s.t. 3m1  5m2  4m3  10
mj  I 

m1
h1 m1    P
k 0
0
G19k
P 0
G15k
 m1 
 0.15m1  k 0.85k
k 

h m    P
2 2
m2

k 0
0
G 92k
 PG02
6 k

 m2 
 0.05m2  k 0.95k
 k 

h m    P
3 3
m3

k 0
0
G83k
 PG03
5k
 k 0.1
m 
3 m3  k
0.9k
The First L Sets Optimization: Problem Formulation

    
min Pr L   g11 m1   1  g11 m1  g 22 m2   1  g 22 m2  g 33 m3 
s.t. 3m1  5m2  4m3  10
mj  I 
m1
 0  m1  k
g m1     PG1  0.15 0.85 
1 m1 k
1
k 0 
3 k
k 
m2
 0  m2  k
g 2 m2     PG2  0.05 0.95 
2 m2 k

k 0 
4 k
k  
m3
 0  m3  k
g m3     PG3  0.1 0.9 
3 m3 k
3
k 0 
3 k
k  
Solution with Load (300, 400,300)

Possible Cost LOLP of the first L set LOLP

combination obtained from
m1 m2 m3 global decomposition

0 0 2 8 0.0019 0.0115
0 1 1 9 0.0005 0.0057
0 2 0 10 0.0006 0.0074
1 0 1 7 0.0017 0.0092
1 1 0 8 0.0005 0.0061
2 0 1 10 0.0016 0.0070
3 0 0 9 0.0019 0.0105

Optimal solution = Solution from the first L sets optimization

Solution with Load (400, 500,400)
Solution from the first L set optimization

Possible Cost LOLP of the first L set LOLP

combination obtained from
global decomposition
m1 m2 m3
0 0 2 8 0.0237 0.1485
0 1 1 9 0.0100 0.1157
0 2 0 10 0.0136 0.1247
1 0 1 7 0.0192 0.1354
1 1 0 8 0.0103 0.1219
2 0 1 10 0.0181 0.0837
3 0 0 9 0.0236 0.1174

Optimal solution
 Advantage of DP
• Systematic approach VS. Exhaustive search
• Less number of mathematical operation, in 3 area
system
Operation Exhaustive Search Dynamic Programming
Multiplication 4×3×3×2 = 72 22
Comparison 4×3×3-1 = 35 10
• The difference will be larger for large system
• The problem constraint can be in any nforms of
summationn of one variable function, h j m j  For
example,  a j m 2j  b j m j  c j j 1

j 1
This will give variations in modeling budget
constraint.
• It is also possible to apply DP to multiple constraints
problem; however, the algorithm may not be
efficient.
 Limitation of DP
• Can applied only to one set (A, L)
• Real application, Loss of load probability
computed from many L sets.
• Approximation not valid for L sets, cause ‘bias
information’.
• Multiple sets destroy ‘separable’ structure
 Hybrid Method
• A combination of heuristic
technique and classical 8.5
x 10
-3

optimization. 8
Solution from DP
Random Sampling 1

• Dynamic programming is 7.5

Random Sampling 2
Random Sampling 3

applied to produce good 7

starting solution for Tabu

LOLP
search. 6.5

• Neighborhood function is a 6

random sampling of 5.5

locations to add and drop 5

one generator. 4.5

0 1 2 3 4 5 6 7 8 9 10
• Objective function is Iteration

calculated from global

decomposition technique.
A two-stage recourse model
Solution algorithm
Sample-average approximation

STOCHASTIC PROGRAMMING
 Problem Formulation
• Two-stage recourse model
• Mixed-integer stochastic programming.
• Use expected unserved energy as reliability index.
• Decision variables:
– First stage
xig : number of additional generators at area i, integer
– Second stage
yij  : flow from arc i to j
• Objective function: Minimize expansion cost in the first stage
and expected unserved energy in the second stage.
 Stochastic Programming Model
minimize z   cig xig  E~  f  x, ~ 
iI
s.t.
x
iI
i
g
N g
Depends on duration
x 0
i
g of state ω
Expected
where f  x,    Min  cil  li    yit   Unserved
iI Energy
ysi    gi    M ig xig
s.t.
Maximum i  I Cost
Capacity y ji    yij    tij   i, j  I , i  j
Flow in
Network yit    li   i  I
Network Flow ysi     y ji     yij    yit   i  I
Conservation j I jI
j i j i

yij  , ysj  , yit    0 i, j  I

 Loss of Load Cost Coefficient
• Find state duration in hours ( D ) from,
24
D  ml

 g  g  t  t  l


iI
 
i

 

iI
i

 

i , jI
ij
i , jI
ij
k 1
k

i j i j
x  Equivalent transition rate of component x from a capacity of state ω to
higher capacity (per day)
x  Equivalent transition rate of component x from a capacity of state ω to lower
capacity (per day)

• Calculate this cost (dollar per kW-h) from

customer damage function.
c l D   e 6.48005  0.38489 D 0.02248 D
2
 
 Solution Procedure
• L-shaped algorithm
• 1 Master problem – many sub-problems
• Each iteration,
– Approximate second stage objective function by
generating piecewise linear function.
– Append this linear function to master problem
• Solve master problem for candidate solution
• Use this candidate solution in the sub-problems
• Repeat until converged (Upper bound = Lower
bound).
Illustration of L-Shaped Algorithm

Second Stage Objective Function

Approximation at each iteration

First stage decision variable

Optimality Gap
(generation location in this case)
(UB – LB)
 L-Shaped Algorithm
• Step 0: Initialization
• Step 1:
– Solve all sub-problems.
– Obtain dual solutions and generate cut.
– Update UB.
• Step 2:
– Solve master problem with cuts.
– Update LB.
• Step 3:Convergence
 Example System
• Three-area power system
• Problem stats:
– 3 integer variables
Area 1 Area 2
– 12 positive continuous
variables
– 16 constraints
Area 3 – 10080 states
 Result
• Converged in 6 iterations. 202

• Optimal solution is to 201

locate one generator in

200
area 1 and one generator

UB and LB
in area 2 199

• EUE cost of 1.48 million 198

dollar. 197

• Expansion cost of 200 196

1 2 3 4 5 6

million dollar. Iteration

 Unit Availability Consideration Model
minimize z    cig,k xig,k  E~  f  x, ~ 
iI kK
s.t.
x
Expected Unserved
g
i ,kN g
Energy Cost
Depends on duration
iI kK g
x 0
i ,k
of state ω

where f  x,    Min  cil  li    yit  

iI
ysi    gi    M ig,k  xig,k
s.t.
Maximum i  I
Capacity y ji    yij    tij   i, j  I , i  j
Flow in Unit Availability
Network yit    li   Representation i  I
Network Flow ysi     y ji     yij    yit   i  I
Conservation j I jI
j i j i

yij  , ysj  , yit    0 i, j  I

 Example System
• Three-area power system
• Problem stats
– 6 binary variables
Area 1 Area 2
– 12 positive continuous
variables
– 16 constraints
Area 3 – 64512 scenarios
 Result
• Converged in 2 iterations.
• Optimal solution is to locate one generator in area 1 and one
generator in area 2
• EUE cost of 1.7 million dollar.
• Expansion cost of 200 million dollar.

Iteration Lower Bound Upper Bound % Optimality Gap

0 198.02 211.36 6.312

1 201.62 201.70 0.040
2 201.70 201.70 0.000
 12-Area Test System
• 137 generating units
1 2 8 • 169 t-lines connection
• 4 additional generators,
6 3 9 200 MW each.
• Maximum of 5
additional transmission
5 11 7 lines, 200 MVA each.
• 1.62 × 10⁴⁹ states
10 4 12
 Stochastic Programming Model
Minimize z   cig xig   cijt xijt  E~  f  x, ~ 
iI iI jI
s.t. g i j
x  N g Expected Unserved
i Energy Cost
iI

 ij
x t

iI jI
N t

i j Depends on duration
x  0, x  0, integer
i
g t
ij of state ω

where f  x,    Min  cil  li    yit  

iI
Maximum s.t. ysi    gi    M ig xig i  I
Capacity y ji    yij    tij    M ijt xijt i, j  I , i  j
Flow in yit    li   i  I
Network ysi     y ji     yij    yit   i  I
j I jI
Flow Conservation j i j i

yij  , ysj  , yit    0 i, j  I

 Sample Average Approximation
• Reduce computation burden.
• Approximate second stage objective value by
sampling.
• Exterior sampling method.

Objective Function:
N
ˆf x   1
N
N
 f  x,  
r 1
r

where N is a sample size.

 SAA Problem
1 N
Min zˆ N   c x   c x   cil k  ylik k 
g g
i i
t t
ij ij
iI iI jI N k 1 iI
i j

 i
x
iI
g
 N g

 ij
x t

iI jI
N t

y k   gi k   M ig xig ; i  I , k  1,2,, N 

i j k
gi

y kji k   yijk k   tij k   M ijt xijt ; i, j  I , i  j , k  1,2,  , N 
yitk k   li k ; i  I , k  1,2,, N 
ysik k    y kji k    yijk k   yitk k ; i  I , k  1,2,, N 
jI jI
j i j i
x  0, xijt  0, integer
i
g

y gik k , ylik k   0,ik k unrestricted ; i  I , k  1,2,, N 

 Sample-based Solution
• Solution quality depends on sampling
technique and sample size, n.
• Optimality is attained with sufficiently large
sample size.
• Repeat the experiment to make statistical
inference of Upper Bound and Lower Bound of
the objective value
 Bounds of Optimal Objective Value
• Let x*N and zˆ *N be the optimal solution and objective value to
SAA problem. If x * and z *are the optimal solution and
objective value to the original minimization problem,
   
z *  zˆ*N and zˆ*N  zˆN x*N  zˆN x*
• Take expectation on both sides,

    
E zˆ*N  E zˆN x*  z *

Bounds of Objective Function:

 
E zN  z  zN
ˆ * *
ˆ *
 Lower Bound Estimate
  
• Estimate E zˆ*N x*N
• Generate N L batches, each with M L samples
• Lower bound estimate, 1 M L *, s
LN L , M L  
M L s 1
zˆ N L
• By central limit theorem,
LN L ,M L  N  L ,  L2 
• Two-Sided 100(1 − β)% Confidence Interval

 z / 2 sL z / 2 sL 
LN L ,M L  ,LN L ,M L  
 ML M L 
where
 L  E zˆ*N 
sL 
2 1 L M L *, s
M L  1 s 1 L
ˆ
z 
 N N L ,M L
 L  2

Pr z / 2  N 0,1  z / 2   1  
 Upper Bound Estimate
*
• Given a sample-based solution, xN
• Estimate zˆ*N x*N 
• Generate NU batches, each with MU samples
• Upper bound estimate,
U NU , M U x    NU N 

M
1 U
*
N ˆ
z s
x *

MU s 1
• By central limit theorem,
 
U NU ,MU x*N  N U ,  U2  
• Two-Sided 100(1 − β)% Confidence Interval


z
 
s
 /2 U x *


z   s
 /2 U x *

   
 NU ,MU N 
* N * N
U x ,U NU , M U x N
 M U MU 
U  Ezˆ N x*N 
where U

s x  M zˆ x   U x 
M
1

U
2 * *, s 2 * *

1
U N NU N NU , M U N
U s 1

Pr z / 2  N 0,1  z   1  /2

 Sampling Techniques
• Monte Carlo simulation
– Random sampling based on the failure and repair
rate of each components in the system
• Latin Hypercube simulation
– Stratified sampling technique
– Sampling based on the distribution
– Reduce variance of the samples
 Approximate Solution
• Found when a unique solution is obtained
from solving several SAA problems with
different samples of a given size, N.
• Optimality should be attained with sufficiently
large N.
• Possible to have different solutions for small
sample size.
 Computational Results
• Two planning problems
– Generation expansion problem (GEP)
– Transmission expansion problem (TEP)
• Conduct three tests
– Lower bound estimates
– Upper bound estimates
– Optimal solution approximation
• Two sampling techniques
– Monte Carlo sampling
– Latin hypercube sampling
• Sample size 200, 500, 1000, 5000, and 10000
 GEP Lower Bound Estimates
• Monte Carlo sampling • Latin Hypercube
sampling
Sample Lower Bound Sample Lower Bound
size ( $m, 95% size ( $m, 95%
confidence interval ) confidence interval )

200 1971.29 ± 95.34 200 1904.48 ± 51.35

500 1874.14 ± 120.17 500 1902.03 ± 64.83
1000 1941.27 ± 92.08 1000 1835.58 ± 31.40
2000 1844.44 ± 16.54 2000 1833.12 ± 18.24
5000 1886.82 ± 31.81 5000 1838.68 ± 8.20
 GEP Approximate Solution from MC
Sampling
Sample Batch # of Additional Generator at area Upper Bound ( $m, 95%
Size confidence interval )
1 2 9 10
1000 1 0 2 1 1 1886.41 ± 20.09
2 0 2 1 1 1883.47 ± 18.40
3 0 4 0 0 1878.17 ± 30.83
4 0 2 1 1 1886.52 ± 19.05
5 0 2 1 1 1886.35 ± 11.02
5000 1 0 3 0 1 1880.09 ± 15.59
2 0 3 0 1 1866.34 ± 12.08
3 0 3 0 1 1887.42 ± 16.71
4 0 3 0 1 1879.49 ± 25.38
5 0 3 0 1 1878.91 ± 25.16
10000 1 0 3 0 1 1862.52 ± 21.02
2 0 3 0 1 1866.07 ± 13.08
3 0 3 0 1 1896.01 ± 18.42
4 0 3 0 1 1880.29 ± 17.67
5 0 3 0 1 1889.96 ± 22.27
 GEP Approximate Solution from LHS
Sampling
Sample Batch # of Additional Generator at area Upper Bound ( $m, 95%
Size confidence interval )
1 2 9 10
1000 1 0 4 0 0 1839.79 ± 8.55
2 0 2 0 2 1847.51 ± 10.65
3 0 2 1 2 1844.94 ± 8.69
4 0 3 0 1 1836.73 ± 6.80
5 0 2 1 1 1848.27 ± 6.24
5000 1 0 3 0 1 1841.75 ± 8.38
2 0 3 0 1 1845.62 ± 9.69
3 0 3 0 1 1835.45 ± 9.13
4 0 3 0 1 1835.86 ± 8.13
5 0 3 0 1 1841.31 ± 11.87
10000 1 0 3 0 1 1843.41 ± 7.96
2 0 3 0 1 1853.04 ± 9.57
3 0 3 0 1 1848.59 ± 8.86
4 0 3 0 1 1840.28 ± 11.16
5 0 3 0 1 1842.60 ± 10.70
 Bounds of GEP
• Monte Carlo sampling • Latin Hypercube
sampling
9 9
x 10 x 10
2.2 2.2
Lower Bound Lower Bound
Upper Bound Upper Bound
2.1 2.1

2 2
Objective value

Objective value
1.9 1.9

1.8 1.8

1.7 1.7

1.6 1.6
200 500 1000 5000 10000 200 500 1000 5000 10000
Number of sampled state Number of sampled state
 TEP Lower Bound Estimates
• Monte Carlo sampling • Latin Hypercube
sampling
Sample Lower Bound Sample Lower Bound
size ( $m, 95% size ( $m, 95%
confidence interval ) confidence interval )

200 1089.67 ± 95.07 200 1020.03 ± 45.86

500 990.16 ± 125.11 500 1017.50 ± 63.29
1000 1058.94 ± 94.24 1000 951.69 ± 29.50
2000 960.17 ± 16.87 2000 947.89 ± 17.21
5000 1004.24 ± 33.21 5000 955.72 ± 8.44
 TEP Approximate Solution from MC
Sampling
Sample Batch # of Add. Tie-lines bet. areas Upper Bound ( $m, 95%
Size 1-6 2-8 4-10 9-11 10-12 confidence interval )
1000 1 1 1 0 0 3 994.61 ± 20.95
2 1 1 0 0 3 992.31 ± 25.44
3 1 1 0 0 3 1016.12 ± 25.88
4 1 1 0 0 3 993.39 ± 25.40
5 1 1 0 0 3 1004.71 ± 12.26
2000 1 1 1 0 0 3 991.57 ± 19.00
2 1 1 0 0 3 995.88 ± 16.35
3 1 1 0 0 3 1005.41 ± 34.68
4 1 1 0 0 3 1000.54 ± 20.75
5 1 1 0 0 3 990.69 ± 21.69
5000 1 1 1 0 0 3 982.25 ± 14.30
2 1 1 0 0 3 992.23 ± 22.49
3 0 1 0 0 3 993.75 ± 25.93
4 1 1 0 0 3 1007.58 ± 14.78
5 0 1 0 0 3 1004.28 ± 10.19
 TEP Approximate Solution from LHS
Sampling
Sample Batch # of Add. Tie-lines bet. areas Upper Bound ( $m, 95%
Size 1-6 2-8 4-10 9-11 10-12 confidence interval )
1000 1 1 1 0 0 3 949.07 ± 10.61
2 1 1 0 0 3 963.54 ± 9.35
3 1 1 0 0 3 952.85 ± 11.67
4 1 1 0 0 3 957.23 ± 11.52
5 0 1 0 0 3 962.39 ± 14.10
2000 1 1 1 0 0 3 965.12 ± 9.41
2 1 1 0 0 3 961.22 ± 7.70
3 1 1 0 0 3 964.60 ± 9.10
4 0 1 0 0 3 960.57 ± 10.33
5 1 1 0 0 3 960.82 ± 8.10
5000 1 0 1 0 0 3 958.37 ± 6.46
2 0 1 0 0 3 958.99 ± 9.07
3 0 1 0 0 3 959.41 ± 9.10
4 1 1 0 0 3 965.68 ± 10.35
5 1 1 0 0 3 956.73 ± 11.29
 Bounds of TEP
• Monte Carlo sampling • Latin Hypercube
sampling
9 9
x 10 x 10
1.2 1.2
Lower Bound Lower Bound
1.15 Upper Bound 1.15 Upper Bound

1.1 1.1

1.05 1.05
Objective value

Objective value
1 1

0.95 0.95

0.9 0.9

0.85 0.85

0.8 0.8
200 500 1000 5000 10000 200 500 1000 5000 10000
Number of sampled state Number of sampled state
 Results
• System-wide reliability optimization.
• Original problem has very large number of system
states (1.62 × 10⁴⁹ states), SAA technique requires
workable sample of 10000.
• Offer a methodology to solve a reliability constrained
planning problem
• Applicable to the multi-stage planning problem.
• Applicable to security and operation problem
• Flexible modeling, techniques applicable to DC flow
formulation
 IEEE RTS 24-bus system
• 24 buses.
• 10 generation buses, 3045
MW installed capacity.
• 17 load buses, 2850 MW
peak load.
• Five buses chosen as
candidate locations, which
are bus 1, 2, 7, 15, and 22.
• Budget of 100 million.

May 25-28, 2008 10th PMAPS, Rincon, Puerto Rico 81

 SAA Problem
1 N 
Min zˆ N   c x  P    yil  r    
i
g
i
g r

iI  N r 1 iI 
s.t.   cig,k xig,k  Budget
iI kKi

y gir r   g i r    M g

i ,k r xig,k ; i  I , r  1,2,, N 
kKi
ylir r   li r ; i  I , r  1,2,, N 
bij  ir  r    jr  r   t ij  r ; i, j  I , i  j , r  1,2, , N 
 B 
ij j
r
 r   y r
gi  r   y li r   li r ; i  I , r  1,2, , N 
r

y gi r , ylir r   0, ir r unrestricted ; i  I , r  1,2,, N 

jI r

xig,k , binary

May 25-28, 2008 10th PMAPS, Rincon, Puerto Rico 82

 Objective Function Estimate

Sample Objective Function Lower Bound

size ( 95% confidence interval )
Expansion Cost ($m) Expected Power Loss (MW)
1000 70 ± 37.53 0.0249 ± 0.0455
2000 100 0.1083 ± 0.0301
8000 100 0.0812 ± 0.0373
12000 100 0.0720 ± 0.0169

May 25-28, 2008 10th PMAPS, Rincon, Puerto Rico 83

 References
• Panida Jirutitijaroen and Chanan Singh, “Reliability Constrained Multi-Area Adequacy Planning
Using Stochastic Programming with Sample-Average Approximations”, IEEE transactions on Power
Systems, Vol. 23, No. 2, May 2008, presented in the 2008 Power Engineering Society General
Meeting, Pittsburg, Pennsylvania, July 2008.
• Panida Jirutitijaroen and Chanan Singh, “Composite-System Generation Adequacy Planning Using
Stochastic Programming with Sample-Average Approximation”, in Proceedings of the 16th Power
Systems Computation Conference, Glasgow, Scotland, July 2008.
• Panida Jirutitijaroen and Chanan Singh, “A Hybrid Method for Multi- Area Generation Expansion
using Tabu-search and Dynamic Programming”, in Proceedings of the 2006 International
Conference on Power System Technology , Chongqing, China, October 2006.
• Panida Jirutitijaroen and Chanan Singh, “Reliability and Cost Trade- Off in Multi-Area Power System
Generation Expansion Using Dynamic Programming and Global Decomposition”, IEEE Transactions
on Power Systems, Vol. 21, No. 3, August 2006.
• Panida Jirutitijaroen and Chanan Singh, “A Global Decomposition Algorithm for Reliability
Constrained Generation Planning and Placement”, in Proceedings of the 9th International
Conference on Probabilistic Method Applied to Power Systems, Stockholm, Sweden, June 2006.