· Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd...

86
Price Signals in Energy Optimization Claudia Sagastiz ´ abal (IMECC-UNICAMP, adjunct researcher) Workshop de Otimizac ¸˜ ao IM-UFRJ Rio, May 7, 2018

Transcript of  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd...

Page 1:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signalsin

Energy Optimization

Claudia Sagastizabal(IMECC-UNICAMP, adjunct researcher)

Workshop de Otimizacao

IM-UFRJRio, May 7, 2018

Page 2:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Joseph-Louis Lagrange (1736-1813)

source: Wikipedia

W.W.R. Ball, History of Mathematics (3rd Ed., 1901)

I regard as quite useless the reading of largetreatises of pure analysis: too large a number of

methods pass at once before the eyes. It is in theworks of application that one must study them; one

judges their utility there and appraises the manner ofmaking use of them.

Je considere comme completement inutile la lecture de grostraites d’analyse pure: un trop grand nombre de methodes

passent en meme temps devant les yeux. C’est dans les travauxd’application qu’on doit les etudier; c’est la qu’on juge leurs

capacites et qu’on apprend la maniere de les utiliser.

There are many great quotes by Lagrange . . .

Page 3:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Joseph-Louis Lagrange (1736-1813)

source: Wikipedia

W.W.R. Ball, History of Mathematics (3rd Ed., 1901)

I regard as quite useless the reading of largetreatises of pure analysis: too large a number of

methods pass at once before the eyes. It is in theworks of application that one must study them; one

judges their utility there and appraises the manner ofmaking use of them.

Je considere comme completement inutile la lecture de grostraites d’analyse pure: un trop grand nombre de methodes

passent en meme temps devant les yeux. C’est dans les travauxd’application qu’on doit les etudier; c’est la qu’on juge leurs

capacites et qu’on apprend la maniere de les utiliser.There are many great quotes by Lagrange . . .

Page 4:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Joseph-Louis Lagrange (1736-1813)

source: Wikipedia

Before we take to sea we walk on land.Before we create we must understand.

Following Lagrange’s leadin today’s talkwe shall walk on the landof energy optimizationand make a few interesting observations

Page 5:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Joseph-Louis Lagrange (1736-1813)

source: Wikipedia

Before we take to sea we walk on land.Before we create we must understand.

Following Lagrange’s leadin today’s talkwe shall walk on the landof energy optimizationand make a few interesting observations

Page 6:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Joseph-Louis Lagrange (1736-1813)

source: Wikipedia

Thanks to Lagrangewe know electricity prices

1st obse

rvat

ion

Page 7:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Walking on the application’s landManage at minimum cost the different technologies of aninterconnected electric power system, so that demand is satisfied

source: ONS

Page 8:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

Two power plants

pT ∈PT pH ∈PH

CT (pT ) CH(pH)

pT + pH = d (demand)

Page 9:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

Two power plants

pT ∈PT pH ∈PH

CT (pT ) CH(pH)

pT + pH = d (demand)

Page 10:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

Two power plants

pT ∈PT pH ∈PH

CT (pT ) CH(pH)

pT + pH = d (demand)

Page 11:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

Two power plants

pT ∈PT pH ∈PH

CT (pT ) CH(pH)

pT + pH = d (demand)

Page 12:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example min CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = dTwo power plants

pT ∈PT pH ∈PH

CT (pT ) CH(pH)

pT + pH = d (demand)

Page 13:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal costThe perturbation function

p( u ) :=

inf CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+uis the key to price electricity:what is the cost increaseif demand increases in one unit?

Answer:compute p(·)’s directional derivative at 0

(p(0) is the optimal value of our problem)

Page 14:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal costThe perturbation function

p( u ) :=

inf CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+uis the key to price electricity:what is the cost increaseif demand increases in one unit?

Answer:compute p(·)’s directional derivative at 0

(p(0) is the optimal value of our problem)

Page 15:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 16:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup??

of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 17:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 18:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u

← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 19:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x

+ ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 20:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 21:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 22:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u ← x + ↑

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

s.t. pT + pH = d + u

Page 23:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

supx

infp∈P

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

supx{F(x ,u) := 〈x ,u〉+ constant terms in u}

Page 24:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

supx

infp∈P

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

supx{F(x ,u) := 〈x ,u〉+ constant terms in u}

Page 25:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremFor max-functions, differentiating the argument provides the

directional derivative:

ϕ(u) = supx

F(x ,u) =⇒ ϕ′(u, ·) = 〈∇uF(x ,u), ·〉

p(u) = sup?? of the Lagrangian!

p(u) =

{inf

p∈PCT (pT ) + CH(pH)

s.t. pT + pH = d + u

= infp∈P

supx

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

supx

infp∈P

CT (pT ) + CH(pH) + 〈x ,d + u−pT −pH〉

supx{F(x ,u) := 〈x ,u〉+ constant terms in u}

Page 26:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremAt u = 0 Lagrange multiplier x gives directional derivative

p(0) =

min CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+0

= minp∈P

maxx

CT (pT ) + CH(pH) + 〈x ,d−pT −pH〉

L(p,x) = CT (pT ) + CH(pH) + 〈x ,d−pT −pH〉is the Lagrangian

p(0) = minp∈P

maxx

L(p,x)

maxx

minp∈P

L(p,x)

thanks to Lagrangewe know electricity prices

Page 27:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremAt u = 0 Lagrange multiplier x gives directional derivative

p(0) =

min CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+0= min

p∈Pmax

xCT (pT ) + CH(pH) + 〈x ,d−pT −pH〉

L(p,x) = CT (pT ) + CH(pH) + 〈x ,d−pT −pH〉is the Lagrangian

p(0) = minp∈P

maxx

L(p,x)

maxx

minp∈P

L(p,x)

thanks to Lagrangewe know electricity prices

Page 28:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremAt u = 0 Lagrange multiplier x gives directional derivative

p(0) =

min CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+0= min

p∈Pmax

xCT (pT ) + CH(pH) + 〈x ,d−pT −pH〉

L(p,x) = CT (pT ) + CH(pH) + 〈x ,d−pT −pH〉is the Lagrangian

p(0) = minp∈P

maxx

L(p,x)

maxx

minp∈P

L(p,x)

thanks to Lagrangewe know electricity prices

Page 29:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremAt u = 0 Lagrange multiplier x gives directional derivative

p(0) =

min CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+0= min

p∈Pmax

xCT (pT ) + CH(pH) + 〈x ,d−pT −pH〉

L(p,x) = CT (pT ) + CH(pH) + 〈x ,d−pT −pH〉is the Lagrangian

p(0) = minp∈P

maxx

L(p,x)

maxx

minp∈P

L(p,x)

thanks to Lagrangewe know electricity prices

Page 30:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: Danskin’s theoremAt u = 0 Lagrange multiplier x gives directional derivative

p(0) =

min CT (pT ) + CH(pH)s.t. pT ∈PT ,pH ∈PH

pT + pH = d+0= min

p∈Pmax

xCT (pT ) + CH(pH) + 〈x ,d−pT −pH〉

L(p,x) = CT (pT ) + CH(pH) + 〈x ,d−pT −pH〉is the Lagrangian

p(0) = minp∈P

maxx

L(p,x)

maxx

minp∈P

L(p,x)

thanks to Lagrangewe know electricity prices

Page 31:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost

At u = 0 Lagrange multiplier x gives directional derivative

p(0) = minp∈P

maxx

L(p,x)

maxx

minp∈P

L(p,x)

without convexity

Page 32:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: there is a catch!

At u = 0 Lagrange multiplier x gives directional derivative

p(0) = minp∈P

maxx

L(p,x)

p∗∗(0) = maxx

minp∈P

L(p,x)

without convexity p(0)≥ p∗∗(0)= dualPros and cons+ Separability ( 6= technologies )

+ Accurate marginal prices in short CPU timeif using a bundle method with on-demand accuracy

(subgradient/Uzawa-like methods do not get enough precision)– Loss of primal feasibility

Page 33:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: there is a catch!

At u = 0 Lagrange multiplier x gives directional derivative

p(0) = minp∈P

maxx

L(p,x)

p∗∗(0) = maxx

minp∈P

L(p,x)

without convexity p(0)≥ p∗∗(0)= dualPros and cons+ Separability ( 6= technologies )

+ Accurate marginal prices in short CPU timeif using a bundle method with on-demand accuracy

(subgradient/Uzawa-like methods do not get enough precision)– Loss of primal feasibility

Page 34:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: there is a catch!

At u = 0 Lagrange multiplier x gives directional derivative

p(0) = minp∈P

maxx

L(p,x)

p∗∗(0) = maxx

minp∈P

L(p,x)

without convexity p(0)≥ p∗∗(0)= dual

Pros and cons+ Separability ( 6= technologies )

+ Accurate marginal prices in short CPU timeif using a bundle method with on-demand accuracy

(subgradient/Uzawa-like methods do not get enough precision)– Loss of primal feasibility

Page 35:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Pricing via marginal cost: there is a catch!

At u = 0 Lagrange multiplier x gives directional derivative

p(0) = minp∈P

maxx

L(p,x)

p∗∗(0) = maxx

minp∈P

L(p,x)

without convexity p(0)≥ p∗∗(0)= dualPros and cons+ Separability ( 6= technologies )

+ Accurate marginal prices in short CPU timeif using a bundle method with on-demand accuracy

(subgradient/Uzawa-like methods do not get enough precision)– Loss of primal feasibility

Page 36:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variablesgenerates either 0 or P any power in [0,P]

PT = {0,P} PH = [0,P]expensive cheap

Suppose demand is d = 1.5P so that both power plants aredispatched: pT = P and pH = 1

2PThe dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 37:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variables

generates either 0 or P any power in [0,P]PT = {0,P} PH = [0,P]

expensive cheapSuppose demand is d = 1.5P so that both power plants aredispatched: pT = P and pH = 1

2PThe dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 38:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variablesgenerates either 0 or P any power in [0,P]

PT = {0,P} PH = [0,P]

expensive cheapSuppose demand is d = 1.5P so that both power plants aredispatched: pT = P and pH = 1

2PThe dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 39:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variablesgenerates either 0 or P any power in [0,P]

PT = {0,P} PH = [0,P]expensive cheap

Suppose demand is d = 1.5P so that both power plants aredispatched:

pT = P and pH = 12P

The dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 40:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variablesgenerates either 0 or P any power in [0,P]

PT = {0,P} PH = [0,P]expensive cheap

Suppose demand is d = 1.5P so that both power plants aredispatched: pT = P and pH = 1

2P

The dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 41:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variablesgenerates either 0 or P any power in [0,P]

PT = {0,P} PH = [0,P]expensive cheap

Suppose demand is d = 1.5P so that both power plants aredispatched: pT = P and pH = 1

2PThe dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]

Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 42:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple examplePower plants have same capacity (P) but different technology

0-1 variables continuous variablesgenerates either 0 or P any power in [0,P]

PT = {0,P} PH = [0,P]expensive cheap

Suppose demand is d = 1.5P so that both power plants aredispatched: pT = P and pH = 1

2PThe dual approach, that finds x via p∗∗(0), amounts to solvingan LP with PT = {0,P} replaced by conv(PT ) = [0,P]Because p∗∗ = conv(p) < p, the associated dispatch isinfeasible: pdual

T = 12P and pdual

H = P

Page 43:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

For our simple problem p(u) ={

inf CT (pT )+CH (pH )s.t. pT ∈PT ,pH ∈PH

pT +pH = d +u

is the minimum of two quadratic functions

There is a duality gap

when p is not convex(p(0) > p∗∗(0))

p(0)→← p∗∗(0)

Page 44:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

For our simple problem p(u) ={

inf CT (pT )+CH (pH )s.t. pT ∈PT ,pH ∈PH

pT +pH = d +u

is the minimum of two quadratic functions

There is a duality gap

when p is not convex

(p(0) > p∗∗(0))

p(0)→← p∗∗(0)

Page 45:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Illustration with a simple example

For our simple problem p(u) ={

inf CT (pT )+CH (pH )s.t. pT ∈PT ,pH ∈PH

pT +pH = d +u

is the minimum of two quadratic functions

There is a duality gap

when p is not convex(p(0) > p∗∗(0))

p(0)→← p∗∗(0)

Page 46:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks:

Price signals in Energy Optimization

I are always related to dual variables

I nonconvexity complicates calculations

I direct 0-1 solution not possible:there are no dual variables!

Page 47:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks:

Price signals in Energy Optimization

I are always related to dual variables

I nonconvexity complicates calculations

I direct 0-1 solution not possible:there are no dual variables!

Page 48:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks:

Price signals in Energy Optimization

I are always related to dual variables

I nonconvexity complicates calculations

I direct 0-1 solution not possible:there are no dual variables!

Page 49:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks:

Price signals in Energy Optimization

I are always related to dual variables

I nonconvexity complicates calculations

I direct 0-1 solution not possible:

there are no dual variables!

Page 50:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks:

Price signals in Energy Optimization

I are always related to dual variables

I nonconvexity complicates calculations

I direct 0-1 solution not possible:there are no dual variables!

Page 51:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Another star? in decomposition methods

source: Twitter

Page 52:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

A less simple example

Two power plants

pT ∈PT pH ∈PH

PT = [0,P] PH = [0,P]

x ∈ {0,1} and pT ≤ x P〈c,x〉+ CT (pT ) CH(pH)

pT + pH = d(demand)

Page 53:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

A less simple example

Two power plants

pT ∈PT pH ∈PH

PT = [0,P] PH = [0,P]x ∈ {0,1} and pT ≤ x P

〈c,x〉+ CT (pT ) CH(pH)

pT + pH = d(demand)

Page 54:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

A less simple example

Two power plants

pT ∈PT pH ∈PH

PT = [0,P] PH = [0,P]x ∈ {0,1} and pT ≤ x P〈c,x〉+ CT (pT ) CH(pH)

pT + pH = d(demand)

Page 55:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

A less simple example

Two power plants

pT ∈PT pH ∈PH

PT = [0,P] PH = [0,P]x ∈ {0,1} and pT ≤ x P〈c,x〉+ CT (pT ) CH(pH)

pT + pH = d(demand)

Page 56:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problemmin 〈c,x〉+CT (pT )+CH(pH)

s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]pT ≤ x PpT +pH = d

Page 57:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problemmin 〈c,x〉+CT (pT )+CH(pH)

s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]pT ≤ x PpT +pH = d

Page 58:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problem

min 〈c,x〉+ CT (pT ) + CH(pH)s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]

pT ≤ x PpT + pH = d

min 〈c,x〉+ Q(x)s.t. x ∈ {0,1}

Ax = bfor Q(x) :=

min CT (pT ) + CH(pH) ↔ C(p)s.t. pT ,pH ∈ [0,P] ↔ p ∈ [0,P]

pT ≤ x PpT + pH = d

}↔ Wp = h−Tx

Bender’s decomposition separates the optimizationproblem in two decision levels.

Page 59:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problem

min 〈c,x〉+ CT (pT ) + CH(pH)s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]

pT ≤ x PpT + pH = d

min 〈c,x〉+ Q(x)s.t. x ∈ {0,1}

Ax = bfor Q(x) :=

min CT (pT ) + CH(pH) ↔ C(p)s.t. pT ,pH ∈ [0,P] ↔ p ∈ [0,P]

pT ≤ x PpT + pH = d

}↔ Wp = h−Tx

Bender’s decomposition separates the optimizationproblem in two decision levels.

Page 60:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problem

min 〈c,x〉+ CT (pT ) + CH(pH)s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]

pT ≤ x PpT + pH = d

min 〈c,x〉+ Q(x)s.t. x ∈ {0,1}

Ax = bfor Q(x) :=

min CT (pT ) + CH(pH) ↔ C(p)s.t. pT ,pH ∈ [0,P] ↔ p ∈ [0,P]

pT ≤ x PpT + pH = d

}↔ Wp = h−Tx

Bender’s decomposition separates the optimizationproblem in two decision levels.

Page 61:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problem

min 〈c,x〉+ CT (pT ) + CH(pH)s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]

pT ≤ x PpT + pH = d

min 〈c,x〉+ Q(x)s.t. x ∈ {0,1}

Ax = bfor Q(x) :=

min CT (pT ) + CH(pH) ↔ C(p)s.t. pT ,pH ∈ [0,P] ↔ p ∈ [0,P]

pT ≤ x PpT + pH = d

}↔ Wp = h−Tx

Bender’s decomposition separates the optimizationproblem in two decision levels.

Page 62:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problem

min 〈c,x〉+ CT (pT ) + CH(pH)s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]

pT ≤ x PpT + pH = d

min 〈c,x〉+ Q(x)s.t. x ∈ {0,1}

Ax = bfor Q(x) :=

min CT (pT ) + CH(pH) ↔ C(p)s.t. pT ,pH ∈ [0,P] ↔ p ∈ [0,P]

pT ≤ x PpT + pH = d

}↔ Wp = h−Tx

Bender’s decomposition separates the optimizationproblem in two decision levels.

If d or C uncertain,this is a 2-stage stochastic programming problem

Page 63:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Optimization problem

min 〈c,x〉+ CT (pT ) + CH(pH)s.t. x ∈ {0,1} ,pT ,pH ∈ [0,P]

pT ≤ x PpT + pH = d

min 〈c,x〉+ Q(x)s.t. x ∈ {0,1}

Ax = bfor Q(x) :=

min CT (pT ) + CH(pH) ↔ C(p)s.t. pT ,pH ∈ [0,P] ↔ p ∈ [0,P]

pT ≤ x PpT + pH = d

}↔ Wp = h−Tx

Bender’s decomposition separates the optimizationproblem in two decision levels. If d or C uncertain,this is a 2-stage stochastic programming problem

Page 64:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx

← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x , ωs)

Page 65:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x , ωs)

Page 66:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}

⇒ E [Q(x ,ω)]≈ 1S ∑

sQ(x ,ωs) and E [π(x ,ω)]≈ 1

S ∑s

π(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x , ωs)

Page 67:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs)

and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x , ωs)

Page 68:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x , ωs)

Page 69:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}

⇒ E [Q(x ,ω)]≈ 1S ∑

sQ(x , ωs) and E [π(x ,ω)]≈ 1

S ∑s

π(x , ωs)

Page 70:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x , ωs)

Page 71:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x,ω)]≈ 1S ∑

sπ(x, ωs)

source: Clara Lage

How much prices change? 1

[1]R&D contract with , joint with C. Lage and M. Solodov

Page 72:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x,ω)]≈ 1S ∑

sπ(x, ωs)

source: Clara Lage

How much prices change? 1

[1]R&D contract with , joint with C. Lage and M. Solodov

Page 73:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Price Signals in 2-SLPmin 〈c,x〉+E [Q(x ,ω)]s.t. x ∈ {0,1}

Ax = bQ(x ,ω) =

min C(p)s.t. p ∈ [0,P]

Wp = h(ω)−Tx ← π(x ,ω)

SAA: ω ∈ {ω1, . . . ,ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x ,ωs) and E [π(x ,ω)]≈ 1S ∑

sπ(x ,ωs)

A different sample: ω ∈ {ω1, . . . , ωS}⇒ E [Q(x ,ω)]≈ 1

S ∑s

Q(x , ωs) and E [π(x,ω)]≈ 1S ∑

sπ(x, ωs)

source: Clara Lage

How much prices change? 1

[1]R&D contract with , joint with C. Lage and M. Solodov

Page 74:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Multiplier sensitivity to uncertainty representationA simple exercise

I Take a 2-SLP with uncertain right-hand side(h(ω))

I Given the sample ω ∈ {ω1m, . . . ,ω

Sm}:

I The SAAm has first-stage solution xmI The associated optimal mean price is

πm :=1S ∑

sπ(xm,ω

sm)

I Repeat for m = 1, . . . ,M and compute thevariance of π1, . . . , πM

Var [π] can be very large!

Page 75:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Multiplier sensitivity to uncertainty representationA simple exercise

I Take a 2-SLP with uncertain right-hand side(h(ω))

I Given the sample ω ∈ {ω1m, . . . ,ω

Sm}:

I The SAAm has first-stage solution xmI The associated optimal mean price is

πm :=1S ∑

sπ(xm,ω

sm)

I Repeat for m = 1, . . . ,M and compute thevariance of π1, . . . , πM

Var [π] can be very large!

Page 76:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variablesTo mitigate wild oscillations, for β > 0 consider

Qβ (x ,ω) :=

min C(p) + 1

2β‖Tx + Wp−h(ω)‖2

s.t. p ∈ [0,P]

Wp = h(ω)−Tx ↗

I corresponds to a quadratic term β

2‖π‖2 in the

dualI if pβ = pβ (x ,ω) solves the regularized problem,

then

πβ (x ,ω) :=

(Tx + Wpβ −h(ω)

)I xβ now solves

{min 〈c,x〉+E

[Qβ (x ,ω)

]s.t. Ax = b ,x ≥ 0

Page 77:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variablesTo mitigate wild oscillations, for β > 0 consider

Qβ (x ,ω) :=

min C(p) + 1

2β‖Tx + Wp−h(ω)‖2

s.t. p ∈ [0,P]

Wp = h(ω)−Tx ↗

I corresponds to a quadratic term β

2‖π‖2 in the

dual

I if pβ = pβ (x ,ω) solves the regularized problem,then

πβ (x ,ω) :=

(Tx + Wpβ −h(ω)

)I xβ now solves

{min 〈c,x〉+E

[Qβ (x ,ω)

]s.t. Ax = b ,x ≥ 0

Page 78:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variablesTo mitigate wild oscillations, for β > 0 consider

Qβ (x ,ω) :=

min C(p) + 1

2β‖Tx + Wp−h(ω)‖2

s.t. p ∈ [0,P]

Wp = h(ω)−Tx ↗

I corresponds to a quadratic term β

2‖π‖2 in the

dualI if pβ = pβ (x ,ω) solves the regularized problem,

then

πβ (x ,ω) :=

(Tx + Wpβ −h(ω)

)

I xβ now solves

{min 〈c,x〉+E

[Qβ (x ,ω)

]s.t. Ax = b ,x ≥ 0

Page 79:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variablesTo mitigate wild oscillations, for β > 0 consider

Qβ (x ,ω) :=

min C(p) + 1

2β‖Tx + Wp−h(ω)‖2

s.t. p ∈ [0,P]

Wp = h(ω)−Tx ↗

I corresponds to a quadratic term β

2‖π‖2 in the

dualI if pβ = pβ (x ,ω) solves the regularized problem,

then

πβ (x ,ω) :=

(Tx + Wpβ −h(ω)

)I xβ now solves

{min 〈c,x〉+E

[Qβ (x ,ω)

]s.t. Ax = b ,x ≥ 0

Page 80:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variablesTo mitigate wild oscillations, for βk → 0 consider

Qβk (x ,ω) :=

{min C(p) + 1

2βk‖Tx + Wp−h(ω)‖2

s.t. p ∈ [0,P]

I corresponds to a quadratic term βk2 ‖π‖

2 in thedual

I if pk = pβk (x ,ω) solves the regularized problem,then

πk(x ,ω) :=

1βk

(Tx + Wpk −h(ω)

)I xk now solves

{min 〈c,x〉+E

[Qβk (x ,ω)

]s.t. Ax = b ,x ≥ 0

Page 81:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variables: theoremUnder suitable CQ and 2nd order conditions,as βk → 0 ,

I the sequence {xk} is bounded,I each accumulation point x∞ solves the original

problemI the sequence of mean regularized prices{

πk :=

1S ∑

k(xk ,ωs)

}converges to the

multiplier π with minimum norm .

Page 82:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Stabilizing dual variablesAs βk decreases the variance Var

[πk(xk ,ω)

]increases

=⇒ a suitable value for βk must be found

Page 83:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks (suite):

Price signals in Energy Optimization

I are always related to dual variables

I uncertainty complicates calculations

I quadratic regularization in the dual stabilizes theoutcome

I nonconvexity complicates calculations

I regarding nonconvexity . . .

Page 84:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks (suite):

Price signals in Energy Optimization

I are always related to dual variables

I uncertainty complicates calculations

I quadratic regularization in the dual stabilizes theoutcome

I nonconvexity complicates calculations

I regarding nonconvexity . . .

Page 85:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Remarks (suite):

Price signals in Energy Optimization

I are always related to dual variables

I uncertainty complicates calculations

I quadratic regularization in the dual stabilizes theoutcome

I nonconvexity complicates calculations

I regarding nonconvexity . . .

Page 86:  · Joseph-Louis Lagrange (1736-1813) source: Wikipedia W.W.R. Ball, History of Mathematics (3rd Ed., 1901) I regard as quite useless the reading of large treatises of pure analysis:

Solucoes Matematicas para Problemas Industriais

I Modelling schoolI WorkshopI 2018 topic:

Formacao de precosno Despacho Hidrotermico

de Curto PrazoI July 9 to 20th, 2018I Sao Carlos, SP

http://www.cemeai.icmc.usp.br/3WSMPI/