ECE 5314: Power System Operation & Control

Lecture 5: Economic Dispatch

Vassilis Kekatos

R2 A. Gomez-Exposito, A. J. Conejo, C. Canizares, Electric Energy Systems: Analysis and

Operation, Chapter 5

Lecture 5 V. Kekatos 1
Problem statement

• N power generation units

P G1
~
• serving a given load D P G2
~ Load%
• Pi : power output of unit i [MW] PL

• Ci (Pi ): operation cost [$/h] P GN

~

Find the most economic dispatch (ED) of units

N
X
min Ci (Pi )
{Pi }N
i=1 i=1
N
X
s.to Pi = D
i=1

Pimin ≤ Pi ≤ Pimax , i = 1, . . . , N

Convex problem if Ci (Pi ) are convex functions

Lecture 5 V. Kekatos 2
Cost functions and energy markets

• Typical options for cost functions (linear, quadratic, piecewise linear)

• Who solves the economic dispatch?

• Fuel-cost curves or market bids

• Bilateral contracts and spot markets (2010)

• Day-ahead and real-time markets

Lecture 5 V. Kekatos 3
Dispatch with linear costs
N
X
min ai Pi
{Pi }N
i=1 i=1
N
X
s.to Pi = D
i=1

Pimin ≤ Pi ≤ Pimax , i = 1, . . . , N

• solved as an LP or by simply sorting the cost coefficients

• uniqueness issues (case when a2 = a3 of the example)

Pimin [MW] Pimax [MW] ai [$/MW]

G1 0 100 20
Example:
G2 0 400 25
D = 350MW G3 0 100 22
G4 0 200 18

Lecture 5 V. Kekatos 4
Dispatch with convex piecewise-linear costs

The range [0, Pimax ] is divided into Ki blocks

Cost Ci (Pi ) expressed as Ki pairs: {(pi,k , ai,k ) = (block size, price)}Ki
k=1

• Any power in [0, pi,1 ] has incremental cost ai,1 $/MW

• Any power in [pi,1 , pi,2 ] has incremental cost ai,2 $/MW ...

Ci (PGi )

slope = ai,1

pi,3 = PGmax
i

pi,1 pi,2 P Gi

Cost Ci (Pi ) is convex if prices are increasing (a1 < . . . < ai,Ki )

Lecture 5 V. Kekatos 5
Rewriting convex piecewise linear costs

Ci (PGi )

slope = ai,1

pi,3 = PGmax
i

pi,1 pi,2 P Gi

Express each cost function as the pointwise maximum of linear functions:

Ci (Pi ) = max {ai,k Pi + bi,k }

k∈{1,...,Ki }

Q.5.1 Show that bi,k = (ai,k−1 − ai,k )pi,k−1 + bi,k−1 .

Lecture 5 V. Kekatos 6
Using the epigraph form

dispatch with convex piecewise-linear costs dispatch in epigraph form

N N
X
X
min max {ai,k Pi + bi,k } min ti
{Pi } k∈{1,...,Ki } {Pi ,ti }
i=1 i=1
N
X s.to ai,k Pi + bi,k ≤ ti , ∀ i, k
s.to Pi = D
i=1 N
X
Pi = D
0 ≤ Pi ≤ Pimax , ∀ i i=1

0 ≤ Pi ≤ Pimax , ∀ i

Lecture 5 V. Kekatos 7
An alternative formulation

Ci (PGi )
• Introduce one variable per block

Ki
X
Pi = Pi,k slope = ai,1
k=1

• Let pi,0 = 0 and pi,Ki = Pimax pi,3 = PGmax

i
pi,1 pi,2
PGi,1 PGi,2 PGi,3 P Gi

Ki
N X
X
min ai,k Pi,k
{Pi,k }
i=1 k=1
Ki
N X
X
s.to Pi,k = D
i=1 k=1

0 ≤ Pi,k ≤ pi,k − pi,k−1 , ∀i, k

∗ ∗
• At the optimum: Pi,k = pi,k − pi,k−1 if Pi,k+1 >0

Lecture 5 V. Kekatos 8
Comparing the two formulations

Epigraph form Form with one variable per block

N
X Ki
N X
X
min ti min ai,k Pi,k
i=1 i=1 k=1
Ki
N X
s.to ai,k Pi + bi,k ≤ ti ∀ i, k X
s.to Pi,k = D
N
X i=1 k=1
Pi = D
i=1 0 ≤ Pi,k ≤ pi,k − pi,k−1 ∀ i, k

0 ≤ Pi ≤ Pi ∀ i
N
P
variables: 2N variables: Ki
i=1
equality constraints: 1 equality constraints: 1
N N
P
inequality constraints:
P
Ki + 2N inequality constraints: 2 Ki
i=1 i=1
solved by simply sorting prices

Lecture 5 V. Kekatos 9
Dispatch with convex quadratic costs

N
X
min (ai Pi + ci Pi2 )
{Pi }N
i=1 i=1
N
X
s.to Pi = D
i=1

Pimin ≤ Pi ≤ Pimax , i = 1, . . . , N

• convex quadratic costs ci > 0 [$/(MW)2 ]

• solved as a quadratic program (QP)

• no uniqueness issues

Lecture 5 V. Kekatos 10
Lagrangian function

For general (strictly) convex and differentiable Ci (Pi )

Introduce Lagrange multipliers:
N
X
min Ci (Pi )
{Pi }N
i=1 i=1
N
X
s.to Pi = D ←− λ
i=1

Pimin ≤ Pi ≤ Pimax , i = 1, . . . , N ←− {(µi , µi )}N

i=1

Lagrangian function
N N
!
X X
L= Ci (Pi ) − λ Pi − D
i=1 i=1
N
X N
X
+ µi (Pimin − Pi ) + µi (Pi − Pimax )
i=1 i=1

Lecture 5 V. Kekatos 11
Optimality conditions

For i = 1, . . . , N,
PN
1. i=1 Pi∗ = D; Pimin ≤ Pi∗ ≤ Pimax

2. µ∗i ≥ 0; µ∗i ≥ 0

3. Ci0 (Pi∗ ) = λ∗ − µ∗i + µ∗i

4. µ∗i (Pimin − Pi∗ ) = 0; µ∗i (Pi∗ − Pimax ) = 0

Conditions 3 and 4 can be equivalently written as


λ∗ , if Pimin < Pi∗ < Pimax





Ci0 (Pi∗ ) = λ∗ + µ∗ , if Pi∗ = Pimin

 i

λ∗ − µ∗i , if Pi∗ = Pimax



Lecture 5 V. Kekatos 12
Optimality conditions simplified

Because µ∗i , µ∗i ≥ 0, optimality conditions become

N
X
1. Pi∗ = D
i=1

0 ∗ ∗
Ci (Pi ) = λ , if Pimin < Pi∗ < Pimax





2.0 Ci0 (Pi∗ ) ≥ λ∗ , if Pi∗ = Pimin



 0 ∗
Ci (Pi ) ≤ λ∗ , if Pi∗ = Pimax



∗
• Any PG 1
, . . . , PG∗ N , λ∗ satisfying the above conditions will be optimal

• Ci0 (Pi ) is the marginal or incremental cost for unit i

• Optimal dispatch when all units not operating at their limits have the

same incremental cost

Lecture 5 V. Kekatos 13
Solving economic dispatch

• Strict convexity implies that Ci0 (Pi ) is increasing in Pi

• Define the increasing function fi = (Ci0 )−1

• Given a λ, the {Pi }N

i=1 satisfying condition 2’ are expressed as

fi (λ), if Ci0 (Pimin ) < λ < Ci0 (Pimax )






Pi (λ) = Pimin ,
 if λ ≤ Ci0 (Pimin )


if λ ≥ Ci0 (Pimax )
 max

Pi ,

• Goal: find λ∗ such that {Pi }N

i=1 satisfy condition 1 (power balance) too
PN
• Pi (λ) and Pi (λ) − D are increasing in λ
i=1

• Bisection on λ until | N
P
i=1 Pi (λ) − D| ≤  (tolerance)

Lecture 5 V. Kekatos 14
λ-iteration or bisection method

Given tolerance  > 0, start with λ and λ̄ so that λ ≤ λ∗ ≤ λ̄

for example: λ = mini Ci0 (Pimin ) and λ̄ = maxi Ci0 (Pimax ).

1. Set λ = (λ + λ̄)/2
2. Find Pi (λ) for all i as follows
2.1 If Ci0 (Pimin ) ≥ λ, set Pi (λ) = Pimin
2.2 If Ci0 (Pimax ) ≤ λ, set Pi (λ) = Pimax
2.3 Otherwise, set Pi (λ) as the solution to Ci0 (Pi ) = λ
PN
3. If i=1 Pi (λ) − D > , set λ̄ := λ and go to Step 1
PN
4. If i=1 Pi (λ) − D < −, set λ := λ and go to Step 1

5. Else {Pi (λ)}N

i=1 is the solution within the specified tolerance

Lecture 5 V. Kekatos 15
Graphical illustration

λ-iteration method for two generators with quadratic costs

Lecture 5 V. Kekatos 16
λ-iteration with convex piecewise linear costs

Convex piecewise linear Ci (Pi ) is non-differentiable

KKT conditions with constraints Pimin ≤ Pi ≤ Pimax kept implicit
N N
!
X X
Lagrangian function : L(P1 , . . . , PN , λ) = Ci (Pi ) − λ Pi − D
i=1 i=1

Optimality conditions:
N
X
1. Pi∗ = D
i=1

2. Pi∗ ∈ arg min {Ci (Pi ) − λ∗ Pi } ∀i

Pimin ≤Pi ≤Pimax

Step 2 of λ-iteration now becomes

min {Ci (Pi ) − λPi }

Pimin ≤Pi ≤Pimax

Lecture 5 V. Kekatos 17
Economic interpretation

Optimality conditions with general costs

N
X
1. Pi∗ = D
i=1

2. Pi∗ ∈ arg min {Ci (Pi ) − λ∗ Pi } ∀i

Pimin ≤Pi ≤Pimax

Interpret multiplier λ [$/MW] as the price at which unit i will be compensated

for producing Pi MW

Given λ, unit i chooses Pi so its net cost Ci (Pi ) − λPi is minimized

(net revenue maximized)

Optimal multiplier λ∗ maximizes the total net revenue

Lecture 5 V. Kekatos 18
Sensitivity interpretation

N
X
C(D) := min Ci (Pi )
{Pimin ≤Pi ≤Pimax }N
i=1 i=1
N
X
s.to Pi = D ←− λ
i=1

Cost of generation as a function of demand D is a perturbation function!

If the problem is convex, then function C(D) is convex

If C(D) is differentiable and strong duality holds, then C 0 (D) = λ∗ (D)

Marginal price λ∗ : extra price one would pay for slight increase in demand

Lecture 5 V. Kekatos 19
Participation factors

Consider an optimal dispatch P and a small load variation ∆D

Assume new dispatch does not alter which units are at Pimin or Pimax

current ED perturbed ED
X X
Pi = D ∆Pi = ∆D
i i

λ = Ci0 (Pi ) ∆λ = Ci00 (Pi )∆Pi

Participation factors: How much each unit contributes to serve the new load?

∆Pi ∆λ/Ci00 (Pi ) 1/Ci00 (Pi )

= P 00
= P 00
∈ (0, 1)
∆D k ∆λ/Ck (Pk ) k 1/Ck (Pk )

constant for quadratic costs

∆Pi
Used for fast generation in response to load variation Pinew = Pi + ∆D
∆D

Lecture 5 V. Kekatos 20
Elastic demand

Elastic demand is characterized by its utility function Uj (Dj ) [$/h]:

the benefit by consuming Dj MW for 1h

N
X M
X
min Ci (Pi ) − Uj (Dj )
{Pi }N M
i=1 ,{Dj }j=1 i=1 j=1
N
X M
X
s.to Pi = Dj
i=1 j=1

Pimin ≤ Pi ≤ Pimax , i = 1, . . . , N

Djmin ≤ Dj ≤ Djmax , j = 1, . . . , M

• demand bids Uj (Dj ) are submitted by utilities, load serving entities,

aggregators, or industrial costumers

• convex problem for concave Uj (Dj )’s (diminishing returns or buy in bulk)

Lecture 5 V. Kekatos 21
Transmission losses

Accounting for losses on transmission lines (3-5%):

N
X
min Ci (Pi )
{Pi }N
i=1 i=1
N
X
s.to Pi = Ploss (p) ←− λ
i=1

Pimin ≤ Pi ≤ Pimax , i = 1, . . . , N

where p := [P1 . . . PN ]> captures generation and (fixed or elastic) demand

• Losses Ploss (p) is a non-linear function of p (non-convex problem)

• In the past, modeled as quadratic functions: Ploss (p) = p> Bp + c> p

• Now typically calculated by (successive) linearization of power flow

equations or via the optimal power flow problem

Lecture 5 V. Kekatos 22
Penalty factors

Lagrangian function (non-convex problem; KKT conditions are only necessary)

N N
! N N
X X X X
L= Ci (Pi )−λ Pi − Ploss (p) + µi (Pimin −Pi )+ µi (Pi −Pimax )
i=1 i=1 i=1 i=1

Optimality conditions simplified:

N
X
1. Pi∗ = Ploss (p∗ )
i=1
  
0 ∗ ∗ ∂Ploss (p)
C (P ) = λ 1− , if Pimin < Pi∗ < Pimax

i i

 ∂Pi


 p=p∗
  
2.0 Ci0 (Pi∗ ) ≥ λ∗ 1 − ∂Ploss (p)
∂Pi
, if Pi∗ = Pimin

 p=p∗

  
∂Ploss (p)
Ci0 (Pi∗ ) ≤ λ∗ 1 − if Pi∗ = Pimax


∂Pi
,
p=p∗

 −1
∂Ploss (p)
Marginal costs: fi Ci0 (Pi∗ ) for penalty factors fi = 1− ∂Pi
p=p∗

Lecture 5 V. Kekatos 23
Solving economic dispatch with losses

Solve economic dispatch ignoring losses to get initial dispatch p0

At iteration t = 0, 1, . . ., do
 −1
∂Ploss (p)
1. Calculate Ploss (pt ) and fit := 1− ∂Pi
for i = 1, . . . , N
pt

PN
2. Find pt+1 via λ-iteration with costs {fit Ci }i and i=1 Pi = Ploss (pt )

3. If kpt+1 − pt k2 <  stop; otherwise go to Step 1

• heuristic manner to account for losses

• how to compute ∇Ploss (p)?

Lecture 5 V. Kekatos 24
Computing penalty factors

Active power injections are not independent (swing bus 1 makes up for losses)
 
p1 (p̃)
p=  ∈ RN
p̃

Active power losses: Ploss (p) = p> 1N = p1 (p̃) + p̃> 1N −1

∂Ploss (p)
Interest in finding ∇p Ploss (p) (assume ∂p1
= 0)

Gradient vector for losses: ∇p̃ Ploss (p̃) = 1 + ∇p̃ p1 (p̃)

Resort to power flow equations to find ∇p̃ p1 (p̃)

Lecture 5 V. Kekatos 25
Finding ∇p̃ p1 (p̃) from power flow equations

power flow equations (dependence on voltage magnitudes has been ignored)

p1 = f1 (θ)

p̃ = f2 (θ)

where θ := [θ2 · · · θN ]> is the vector of voltage phases (θ1 = 0)

∂ p̃(θ) ∂f2 (θ)

Jacobian matrix: ∂θ
= ∂θ
= J(θ) ∈ R(N −1)×(N −1)

Jacobian matrix of inverse function: under technical conditions

∂θ(p̃) ∂f2−1 (p̃)
∂ p̃
= ∂ p̃
= [J(θ(p̃))]−1 ∈ R(N −1)×(N −1)

Gradient vector:
 >
∂θ(p̃)
∇p̃ Ploss (p̃) = 1 + ∂ p̃
∇θ f1 (θ) = 1 + [J(p̃)]−> ∇θ f1 (θ)

Lecture 5 V. Kekatos 26