Framing an Appropriate Research Question 6b9b26d93da94caf993c038d9efcdedb.pdf
Arman cdc11
1. Wholesale Electricity Market: Dynamic Modeling and
Stability
Arman Kiani and Anuradha Annaswamy
Institute of Automatic Control Engineering, Technische Universit¨t M¨nchen, Germany,
a u
Department of Mechanical Engineering, Massachussets Institute of Technology
arman.kiani@tum.de
January 23, 2012
50th IEEE Conference on Decision and Control 2011
2. Table of contents
1 Motivation
Next Generation Grid
Electricity Market
2 Dynamic Modeling
Dynamical Market
State Based Games
Market Model: Stability Analsys
Asymptotic Stability
3 Simulation Results
4 Summary and Ongoing work
3. Motivation Next Generation Grid
Next Generation Grid: What makes a grid smart?
4. Motivation Next Generation Grid
Next Generation Grid: What makes a grid smart?
Smart Resources: Renewable Energy Resources
5. Motivation Next Generation Grid
Next Generation Grid: What makes a grid smart?
Smart Resources: Renewable Energy Resources
Smart Participation: Real-Time Pricing + Demand Response
6. Motivation Next Generation Grid
Next Generation Grid: What makes a grid smart?
Smart Resources: Renewable Energy Resources
Smart Participation: Real-Time Pricing + Demand Response
Smart Sensors: Advanced metering infrastructure (Smart Meters)
7. Motivation Next Generation Grid
Next Generation Grid: What makes a grid smart?
Smart Resources: Renewable Energy Resources
Smart Participation: Real-Time Pricing + Demand Response
Smart Sensors: Advanced metering infrastructure (Smart Meters)
Smart Market Design: optimize assets, operate efficiently → utilize dynamic
information
8. Motivation Next Generation Grid
Next Generation Grid: What makes a grid smart?
Smart Resources: Renewable Energy Resources
Smart Participation: Real-Time Pricing + Demand Response
Smart Sensors: Advanced metering infrastructure (Smart Meters)
Smart Market Design: optimize assets, operate efficiently → utilize dynamic
information
9. Motivation Electricity Market
Auction Process in Electricity Market
Power generation scheduling is conducted through a market mechanism:
Use of an auction market - bids from Generating Companies (GenCo)
and Consumer Companies (ConCo).
Any uncertainties are managed through a contingency analysis.
11. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
12. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
In Smart Grid we have several active agents as self-interested decision
makers.
Game theory is beginning to emerge as a powerful tool for the design and
coordinate of multiagent systems.
Utilizing Game theory for this purpose requires two steps.
1 Modeling the agent as self-interested decision makers in a game
theoretic environment. Defining a set of choices and a local objective
function for each decision maker.
2 Specifying a distributed learning algorithm that enables the agents to
reach a desirable operating point, e.g., a Nash equilibrium of the
designed game.
Learning Dynamics in Games = Dynamics of Disequilibrium
What Does Disequilibrium Mean? A situation where internal and/or external
forces (Uncertainties and Perturbations) prevent market equilibrium from being
reached or cause the market to fall out of balance.
13. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
14. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
Bayesian Learning Dynamics
Update beliefs (about an underlying state or opponent strategies)
based on new information optimally (i.e., in a Bayesian manner)
15. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
Bayesian Learning Dynamics
Update beliefs (about an underlying state or opponent strategies)
based on new information optimally (i.e., in a Bayesian manner)
Adaptive Learning Dynamics
Adjusting adaptively the expectations, myopic [Example: Gradient
Play, Best Response Dynamics, Fictitious Play, ... ]
16. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
Bayesian Learning Dynamics
Update beliefs (about an underlying state or opponent strategies)
based on new information optimally (i.e., in a Bayesian manner)
Adaptive Learning Dynamics
Adjusting adaptively the expectations, myopic [Example: Gradient
Play, Best Response Dynamics, Fictitious Play, ... ]
Learning Dynamics in Games = Dynamics of Disequilibrium
What Does Disequilibrium Mean? A situation where internal and/or external
forces (Uncertainties and Perturbations) prevent market equilibrium from being
reached or cause the market to fall out of balance.
17. Dynamic Modeling Dynamical Market
Learning Dynamics in Games
Kenneth Arrow: ”The attainment of equilibrium requires a disequilibrium
process.”
Bayesian Learning Dynamics
Update beliefs (about an underlying state or opponent strategies)
based on new information optimally (i.e., in a Bayesian manner)
Adaptive Learning Dynamics
Adjusting adaptively the expectations, myopic [Example: Gradient
Play, Best Response Dynamics, Fictitious Play, ... ]
Learning Dynamics in Games = Dynamics of Disequilibrium
What Does Disequilibrium Mean? A situation where internal and/or external
forces (Uncertainties and Perturbations) prevent market equilibrium from being
reached or cause the market to fall out of balance.
We will use the terms dynamics and learning dynamics in games interchangeably.
18. Dynamic Modeling State Based Games
State Based Games
Using the notion of state based game, we consider an extension to
the framework of strategic form games and introduces an underlying
state space to the game theoretic framework.
In the proposed state based games we focus on myopic players and
static equilibrium concepts similar to that of pure Nash equilibrium.
The state can take on a variety of interpretations ranging from
1 Dynamics for equilibrium selection
2 Dummy players in a strategic form game that are preprogrammed to
behave due to the specific strategy
3 Disequilibrium process to attain the equilibrium
19. Dynamic Modeling State Based Games
State Based Games
Definition: A State Based game
A State Based game G characterized by the tuple G =
N, X , (Ai )i∈N , (Ji )i∈N , f , which consists of
Player set N
Underlying finite coordination state space X
State invariant action set Ai
State dependent cost function Ji : X × A → R
Coordinator mechanism function f : X × A → X
The sequence of actions a(0), a(1), .. and coordination states x(0), x(1), ...
is generated according to the disequilibrium process. At any time t0 , each
player i ∈ N myopically selects an action ai (t) ∈ Ai according to some
specified decision rule.
20. Dynamic Modeling State Based Games
Electricity Market: Our proposed Model Set Up
A dynamic model based on sub-gradients in a nonlinear optimization
problem stated below:
Minimize f (x)
s.t. gi (x) = 0, ∀i = 1, . . . , N
N
Rji hi (x) ≤ cj , ∀j = 1, . . . L
i=1
21. Dynamic Modeling State Based Games
Electricity Market: Our proposed Model Set Up
A dynamic model based on sub-gradients in a nonlinear optimization
problem stated below:
Minimize f (x)
s.t. gi (x) = 0, ∀i = 1, . . . , N
N
Rji hi (x) ≤ cj , ∀j = 1, . . . L
i=1
Lagrange function L(x, λ, µ)
L(x, λ, µ) is called Lagrange function of the above optimization problem with
Lagrange multipliers λ and µ as
N L
L(x, λ, µ) = f (x) + λi gi (x) + µj (Rji hi (x) − cj )
i=1 j=1
22. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Gradient play can be viewed as progressively adjusting x, λ and µ as
follows:
x(t + ε) = x(t) − kx x L(x, λ, µ)ε
λ(t + ε) = λ(t) + kλ λ L(x, λ, µ)ε
+
µ(t + ε) = µ(t) + kµ [ µ L(x, λ, µ)]µ ε
where kx , kλ and kµ are positive scaling parameters which control the
amount of change in the direction of the gradient.
23. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Gradient play can be viewed as progressively adjusting x, λ and µ as
follows:
x(t + ε) = x(t) − kx x L(x, λ, µ)ε
λ(t + ε) = λ(t) + kλ λ L(x, λ, µ)ε
+
µ(t + ε) = µ(t) + kµ [ µ L(x, λ, µ)]µ ε
where kx , kλ and kµ are positive scaling parameters which control the
amount of change in the direction of the gradient.
Nonnegative projection of congestion cost
+
[h(x, y ) y denotes the projection of h(x, y ) on euclidean projection on the
nonnegative orthant in R+ m
+ h(x, y ) if y > 0,
h(x, y ) y
=
max(0, h(x, y )) if y = 0.
24. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Given the following DC Economic Dispatch with quadratic cost functions
Maximize SW = UDj (PDj ) − CGi (PGi )
j∈Dq i∈Gf
s.t. − PGi + PDj + Bnm [δn − δm ] = 0; ρn
i∈θn j∈ϑn m∈Ωn
max
Bnm [δn − δm ] ≤ Pnm ; γnm , ∀n ∈ N; ∀m ∈ Ω
25. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Given the following DC Economic Dispatch with quadratic cost functions
Maximize SW = UDj (PDj ) − CGi (PGi )
j∈Dq i∈Gf
s.t. − PGi + PDj + Bnm [δn − δm ] = 0; ρn
i∈θn j∈ϑn m∈Ωn
max
Bnm [δn − δm ] ≤ Pnm ; γnm , ∀n ∈ N; ∀m ∈ Ω
Use the gradient play, we will have:
τGi P˙ = ρn(i) − cGi PGi − bGi
Gi → Dynamics for GenCoi
26. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Given the following DC Economic Dispatch with quadratic cost functions
Maximize SW = UDj (PDj ) − CGi (PGi )
j∈Dq i∈Gf
s.t. − PGi + PDj + Bnm [δn − δm ] = 0; ρn
i∈θn j∈ϑn m∈Ωn
max
Bnm [δn − δm ] ≤ Pnm ; γnm , ∀n ∈ N; ∀m ∈ Ω
Use the gradient play, we will have:
τGi P˙ = ρn(i) − cGi PGi − bGi
Gi → Dynamics for GenCoi
τD P˙ = cDj PDj + bDj − ρn(j)
j Dj → Dynamics for ConCo j
27. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Given the following DC Economic Dispatch with quadratic cost functions
Maximize SW = UDj (PDj ) − CGi (PGi )
j∈Dq i∈Gf
s.t. − PGi + PDj + Bnm [δn − δm ] = 0; ρn
i∈θn j∈ϑn m∈Ωn
max
Bnm [δn − δm ] ≤ Pnm ; γnm , ∀n ∈ N; ∀m ∈ Ω
Use the gradient play, we will have:
τGi P˙ = ρn(i) − cGi PGi − bGi
Gi → Dynamics for GenCoi
τD P˙ = cDj PDj + bDj − ρn(j)
j Dj → Dynamics for ConCo j
τδn δ˙n = − Bnm [ρn − ρm + γnm − γmn ] → Phase angles at bus n
m∈Ωn
28. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Given the following DC Economic Dispatch with quadratic cost functions
Maximize SW = UDj (PDj ) − CGi (PGi )
j∈Dq i∈Gf
s.t. − PGi + PDj + Bnm [δn − δm ] = 0; ρn
i∈θn j∈ϑn m∈Ωn
max
Bnm [δn − δm ] ≤ Pnm ; γnm , ∀n ∈ N; ∀m ∈ Ω
Use the gradient play, we will have:
τGi P˙ = ρn(i) − cGi PGi − bGi
Gi → Dynamics for GenCoi
τD P˙ = cDj PDj + bDj − ρn(j)
j Dj → Dynamics for ConCo j
τδn δ˙n = − Bnm [ρn − ρm + γnm − γmn ] → Phase angles at bus n
m∈Ωn
τρn ρ˙n = − PGi + PDj + Bnm [δn − δm ] → Real-Time Price at bus n
i∈θn j∈ϑn m∈Ωn
29. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
Given the following DC Economic Dispatch with quadratic cost functions
Maximize SW = UDj (PDj ) − CGi (PGi )
j∈Dq i∈Gf
s.t. − PGi + PDj + Bnm [δn − δm ] = 0; ρn
i∈θn j∈ϑn m∈Ωn
max
Bnm [δn − δm ] ≤ Pnm ; γnm , ∀n ∈ N; ∀m ∈ Ω
Use the gradient play, we will have:
τGi P˙ = ρn(i) − cGi PGi − bGi
Gi → Dynamics for GenCoi
τD P˙ = cDj PDj + bDj − ρn(j)
j Dj → Dynamics for ConCo j
τδn δ˙n = − Bnm [ρn − ρm + γnm − γmn ] → Phase angles at bus n
m∈Ωn
τρn ρ˙n = − PGi + PDj + Bnm [δn − δm ] → Real-Time Price at bus n
i∈θn j∈ϑn m∈Ωn
max +
τnm γnm = [Bnm [δn − δm ] − Pnm ]γnm
˙ → Congestion Price for line n − m
30. Dynamic Modeling State Based Games
Dynamic Model of Wholesale Market
τGi P˙ = ρn(i) − cGi PGi − bGi
Gi
τD P˙ = cDj PDj + bDj − ρn(j)
j Dj
τδn δ˙n = − Bnm [ρn − ρm + γnm − γmn ]
m∈Ωn (1)
τρn ρ˙n = − PGi + PDj + Bnm [δn − δm ]
i∈θn j∈ϑn m∈Ωn
max +
τnm γnm = [Bnm [δn − δm ] −
˙ Pnm ]γnm
Trajectory of (1):
Distinct from the equilibrium (solutions of KKT conditions)
Converges to the equilibrium if stable
Represents desired exchange of information between key players in the
market to arrive at the equilibrium
32. Dynamic Modeling State Based Games
A New Market Model
Use of feedback in converging to the equilibrium.
GenCos and ConCos adjust their power level using a recursive process.
Price is a Public Signal that guides all entities to adjust efficiently.
33. Dynamic Modeling Market Model: Stability Analsys
Market Model: Stability Analysis
A compact representation of the model:
x1 (t)
˙ A1 A2 x1 (t) b
= + (2)
x2 (t)
˙ 0 0 x2 (t) f2 (x1 , x2 )
where
x1 (t) = PG PD δ ρ T +N +2N−1)×1
(Ng d
T −1 T
A2 = 0 0 −Bline Ar τδ 0
x2 (t) = γ1 ... γNt T
Nt×1
−1 −1 T
−τg cg τg A T T −1 T −1
0 0 g b = bg τg bd τd 0
−1 −1
0 τd cd 0 −τd AT d
A1 =
−1
−τδ AT Bline A −1
0 0 0 r f2 (x1 , x2 ) = τγ [Bline Ar Rx1 − P max ]+
x
−1 −1 −1 2
−τρ Ag τρ Ad τρ AT Bline Ar 0
34. Dynamic Modeling Market Model: Stability Analsys
Market Model: Stability Analysis
Let x = [x1 x2 ]T , E = {(x1 , x2 )|A1 x1 + A2 x2 + b = 0 ∧ f2 (x1 , x2 ) = 0},
T T
and Ω(γ) := {x| ||x|| < γ}
Definition of Market Stability
∗ ∗
The equilibrium point (x1 , x2 ) ∈ E is stable if given > 0, ∃σ such that
x(t) ∈ Ω(σ) ∀x(0) ∈ Ω( )
There exist the feasible sequences of PGi , PDj , and δn such that solutions
starting ”close enough” to the equilibrium (x(0) ∈ Ω( )) remain ”close
enough” forever (x(t) ∈ Ω(σ)).
35. Dynamic Modeling Market Model: Stability Analsys
Market Model: Stability Analysis
Let x = [x1 x2 ]T , E = {(x1 , x2 )|A1 x1 + A2 x2 + b = 0 ∧ f2 (x1 , x2 ) = 0},
T T
and Ω(γ) := {x| ||x|| < γ}
Definition of Market Stability
∗ ∗
The equilibrium point (x1 , x2 ) ∈ E is stable if given > 0, ∃σ such that
x(t) ∈ Ω(σ) ∀x(0) ∈ Ω( )
There exist the feasible sequences of PGi , PDj , and δn such that solutions
starting ”close enough” to the equilibrium (x(0) ∈ Ω( )) remain ”close
enough” forever (x(t) ∈ Ω(σ)).
Is the market stable?
36. Dynamic Modeling Asymptotic Stability
Market Model: Stability Analysis
∗ ∗
Let y1 = x1 − x1 , y2 = x2 − x2 , a positive definite Lyapunov function
V (y1 , y2 ) = y1 P1 y1 + y2 P2 y2 , and d = 2λmin (P2 )ψmin2 min (Q)
T T
τγ β
λ
max
37. Dynamic Modeling Asymptotic Stability
Market Model: Stability Analysis
∗ ∗
Let y1 = x1 − x1 , y2 = x2 − x2 , a positive definite Lyapunov function
V (y1 , y2 ) = y1 P1 y1 + y2 P2 y2 , and d = 2λmin (P2 )ψmin2 min (Q)
T T
τγ β
λ
max
Theorem (Asymptotic Stability)
∗ ∗
Let A1 be Hurwitz. Then the equilibrium (x1 , x2 ) ∈ E is asymptotically
stable for all initial conditions in Ωcmax = {(y1 , y2 ) | V (y1 , y2 ) ≤ c} for a
cmax > 0 such that Ωcmax D = {(y1 , y2 ) | ||y2 || ≤ d}.
38. Dynamic Modeling Asymptotic Stability
Market Model: Stability Analysis
∗ ∗
Let y1 = x1 − x1 , y2 = x2 − x2 , a positive definite Lyapunov function
V (y1 , y2 ) = y1 P1 y1 + y2 P2 y2 , and d = 2λmin (P2 )ψmin2 min (Q)
T T
τγ β
λ
max
Theorem (Asymptotic Stability)
∗ ∗
Let A1 be Hurwitz. Then the equilibrium (x1 , x2 ) ∈ E is asymptotically
stable for all initial conditions in Ωcmax = {(y1 , y2 ) | V (y1 , y2 ) ≤ c} for a
cmax > 0 such that Ωcmax D = {(y1 , y2 ) | ||y2 || ≤ d}.
Remarks
The region of attraction Ωmax for which stability and asymptotic
stability hold places an implicit bound on the congestion price.
39. Dynamic Modeling Asymptotic Stability
Market Model: Stability Analysis
∗ ∗
Let y1 = x1 − x1 , y2 = x2 − x2 , a positive definite Lyapunov function
V (y1 , y2 ) = y1 P1 y1 + y2 P2 y2 , and d = 2λmin (P2 )ψmin2 min (Q)
T T
τγ β
λ
max
Theorem (Asymptotic Stability)
∗ ∗
Let A1 be Hurwitz. Then the equilibrium (x1 , x2 ) ∈ E is asymptotically
stable for all initial conditions in Ωcmax = {(y1 , y2 ) | V (y1 , y2 ) ≤ c} for a
cmax > 0 such that Ωcmax D = {(y1 , y2 ) | ||y2 || ≤ d}.
Remarks
The region of attraction Ωmax for which stability and asymptotic
stability hold places an implicit bound on the congestion price.
In particular, it implies that the congestion price needs to be smaller
than d, which is proportional to thermal limit of transmission lines.
41. Summary and Ongoing work
Summary
A New Market Model was proposed
Recursive, dynamic convergence to equilibrium
Enables stability analysis
Not globally stable
”Domain of attraction” result
Related to congestion rent
Advantages: Model allows us to
design a stable market
utilize uncertain renewable generation
incorporate elastic demands
42. Summary and Ongoing work
Ongoing work
Strong relation to state-based games
Dynamic model: A disequilibrium process
Kenneth Arrow: ”The attainment of equilibrium requires a
disequilibrium process.”
Effect of renewable sources uncertainty on stability of electricity
market
Uncertainty analysis
Equality constraints and local stability