Download Dissertation Progress Report

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Polyhedral Risk
Measures
Vadym Omelchenko,
Institute of Information Theory and Automation,
Academy of Sciences of the Czech Republic.
The presentationโ€™s
structure
1. Definition of polyhedral risk measures (Two-stage)
2. Definition of polyhedral risk measures (Multi-stage)
3. Applications in the energy sector (CHP)
Definition of Polyhedral Risk
Measures (Two-Stage)
Polyhedral Risk Measures
โ€ข ฮณ = ๐ฟ๐‘ (๐น, โ„) be the usual Banach space of real
random variables on some probability space
(ฮฉ, F, P) for some p โˆˆ [1,โˆž).
Polyhedral Probability
Functionals
โ€ข Definition. A probability functional ๐‘…: ๐›พ โ†’ โ„ is called
๐‘๐‘œ๐‘™๐‘ฆโ„Ž๐‘’๐‘‘๐‘Ÿ๐‘Ž๐‘™ if there exist ๐‘˜0 , ๐‘˜1 โˆˆ โ„•, ๐‘ค๐‘– , ๐‘๐‘– โˆˆ โ„๐‘˜๐‘– , ๐‘– โˆˆ
0,1 , and non-empty polyhedral sets ๐‘‰๐‘– โŠ† โ„๐‘˜๐‘– , ๐‘– โˆˆ 0,1 ,
such that
๏ƒฌ
๏ƒผ
v1 ๏ƒŽ L p F , R k1 ,
๏ƒฏ๏ƒฏ
๏ƒฏ๏ƒฏ
R (Y ) ๏€ฝ sup ๏ƒญ c0 , v0 ๏€ซ E ๏€จ c1 , v1 ๏€ฉ vi ๏ƒŽ Vi , i ๏ƒŽ 0, 1 , ๏ƒฝ
๏ƒฏ
๏ƒฏ
w
,
v
๏€ซ
w
,
v
๏€ฝ
Y
๏ƒฏ๏ƒฎ
๏ƒฏ๏ƒพ
0
0
1 1
๏€จ
๏ป
๏€ฉ
๏ฝ
โ€ข for every Y โˆˆ ๐›พ . Here โ‹…,โ‹… denote scalar products on โ„๐‘˜0
and โ„๐‘˜1 .
โ€ข ๐‘‰1 โŠ† โ„๐‘˜1 has to be understood in the sense a.s.
Linear Reformulation
โ€ข Definition. A probability functional ๐‘…: ๐›พ โ†’ โ„ is called
๐‘๐‘œ๐‘™๐‘ฆโ„Ž๐‘’๐‘‘๐‘Ÿ๐‘Ž๐‘™ if there exist ๐‘˜0 , ๐‘˜1 โˆˆ โ„•, ๐‘๐‘– , ๐‘ค๐‘– โˆˆ โ„๐‘˜๐‘– ,
matrices ๐ด๐‘– , and vectors ๐‘๐‘– โˆˆ โ„๐‘˜๐‘– , ๐‘– โˆˆ 0,1 , such that
๏ƒฌ
๏ƒฏ
๏ƒฏ
R (Y ) ๏€ฝ sup ๏ƒญ E cT0 ๏ƒ— v0 ๏€ซ c1T ๏ƒ— v1
๏ƒฏ
๏ƒฏ
๏ƒฎ
๏€จ
๏€จ
๏€ฉ
๏€ฉ
v1 ๏ƒŽ L p F , R k1 , ๏ƒผ
๏ƒฏ
๏ƒฏ
A1v1 ๏‚ฃ b1 , a.s.,
๏ƒฝ
A0 v0 ๏‚ฃ b0 ,
๏ƒฏ
w0 ๏ƒ— v0 ๏€ซ w1 ๏ƒ— v1 ๏€ฝ Y ๏ƒฏ๏ƒพ
Example
โ€ข We consider the functional R Y = E u Y
โ€ข on ฮณ where u: โ„ โ†’ โ„ is of the form
๐‘˜
โ€ข ๐‘ข ๐‘ฅ = sup ๐‘, ๐‘ฃ : ๐‘ฃ โˆˆ โ„ ๐‘˜+๐‘Ÿ
+ , ๐‘–=1 ๐‘ฃ๐‘– = 1, ๐‘ค, ๐‘ฃ = ๐‘ฅ with
some ๐‘, ๐‘ฃ โˆˆ โ„ ๐‘˜+๐‘Ÿ , ๐‘˜, ๐‘Ÿ โˆˆ โ„• and hence it is concave
and polyhedral in kinks ๐‘ค๐‘– , ๐‘๐‘– โˆˆ โ„ 2, ๐‘– = 1, โ€ฆ , ๐‘˜.
โ€ข According to Rockafellar and Wets (1998), Theorem
14.60, we can reverse the order of sup and E.
Example
โ€ข We consider the functional R Y = E u Y
โ€ข on ฮณ where u: โ„ โ†’ โ„ is of the form
๐‘˜
โ€ข ๐‘ข ๐‘ฅ = sup ๐‘, ๐‘ฃ : ๐‘ฃ โˆˆ โ„ ๐‘˜+๐‘Ÿ
,
+
๐‘–=1 ๐‘ฃ๐‘– = 1, ๐‘ค, ๐‘ฃ = ๐‘ฅ with
some ๐‘, ๐‘ฃ โˆˆ โ„ ๐‘˜+๐‘Ÿ , ๐‘˜, ๐‘Ÿ โˆˆ โ„• and hence ๐‘(โ‹…) is concave
and polyhedral in kinks ๐‘ค๐‘– , ๐‘๐‘– โˆˆ โ„ 2, ๐‘– = 1, โ€ฆ , ๐‘˜.
โ€ข See Rockafellar and Wets (1998), Theorem 14.60.
Theorem Rockafellar and Wets
Popular examples
โ€ข CV@R is a polyhedral risk measure.
โ€ข Every linear combination of CV@Rs are polyhedral
risk measures
โ€ข V@R is not polyhedral.
Properties of Polyhedral
Functionals
โ€ข Let R be a functional of the form:
๏ƒฌ
๏ƒฏ๏ƒฏ
R(Y ) ๏€ฝ sup ๏ƒญ c0 , v0 ๏€ซ E ๏€จ c1 , v1
๏ƒฏ
๏ƒฏ๏ƒฎ
๏€ฉ
๏€จ
๏€ฉ
๏ƒผ
๏ƒฏ๏ƒฏ
vi ๏ƒŽ Vi , i ๏ƒŽ ๏ป 0, 1 ๏ฝ, ๏ƒฝ
๏ƒฏ
w0 , v0 ๏€ซ w1 , v1 ๏€ฝ Y ๏ƒฏ๏ƒพ
v1 ๏ƒŽ L p F , R k1 ,
โ€ข Let ๐‘‰๐‘– โŠ† โ„๐‘˜๐‘– , ๐‘– โˆˆ 0,1 be polyhedral cones and assume:
1.
2.
๐‘ค1 , ๐‘‰1 = โ„ (complete recourse),
๐ท โ‰” ๐‘ข โˆˆ โ„: ๐‘0 โˆ’ ๐‘ข๐‘ค0 โˆˆ ๐‘‰ 0โˆ—, ๐‘1 โˆ’ ๐‘ข๐‘ค1 โˆˆ ๐‘‰ 1โˆ— โ‰  โˆ… (dual feasibility.
Then R is finite, concave, and continuous on ฮณ.
Properties of Polyhedral
Functionals
โ€ข
Let R be a functional of the form:
๏ƒฌ
๏ƒฏ๏ƒฏ
R(Y ) ๏€ฝ sup ๏ƒญ c0 , v0 ๏€ซ E ๏€จ c1 , v1
๏ƒฏ
๏ƒฏ๏ƒฎ
๏€ฉ
๏€จ
๏€ฉ
๏ƒผ
๏ƒฏ๏ƒฏ
vi ๏ƒŽ Vi , i ๏ƒŽ ๏ป 0, 1 ๏ฝ, ๏ƒฝ
๏ƒฏ
w0 , v0 ๏€ซ w1 , v1 ๏€ฝ Y ๏ƒฏ๏ƒพ
v1 ๏ƒŽ L p F , R k1 ,
โ€ข Let ๐‘‰๐‘– โŠ† โ„๐‘˜๐‘– , ๐‘– โˆˆ 0,1 be polyhedral cones and assume:
1.
๐‘ค1 , ๐‘‰1 = โ„ (complete recourse),
2. ๐ท โ‰” ๐‘ข โˆˆ โ„: ๐‘0 โˆ’ ๐‘ข๐‘ค0 โˆˆ ๐‘‰ 0โˆ—, ๐‘1 โˆ’ ๐‘ข๐‘ค1 โˆˆ ๐‘‰ 1โˆ— โ‰  โˆ… (dual feasibility.
3.
1
1
๐‘ž โˆˆ 1, โˆž given by ๐‘ + ๐‘ž = 1
Then R admits the dual representation
๐‘น ๐’€ = ๐’Š๐’๐’‡ ๐‘ฌ ๐’€ ๐’ : ๐’ โˆˆ โ„ค โˆ—
Where โ„ค โˆ— is a subset of โ„ค = ๐ฟ๐‘ž (๐น, โ„) given by
โ„ค โˆ—= ๐‘ โˆˆ โ„ค: ๐‘0 โˆ’ ๐‘ค0 ๐ธ ๐‘ โˆˆ ๐‘‰ 0โˆ—, ๐‘1 โˆ’ ๐‘ค1 ๐‘ โˆˆ ๐‘‰ 1โˆ— .
Definition of Polyhedral Risk
Measures (Multi-Stage)
Polyhedral Multi-Period
Acceptability Functionals
โ€ข Let us denote ๐›พ =×๐‘‡๐‘ก=1 ๐ฟ๐‘ ๐น๐‘ก , ๐‘ โˆˆ [1, โˆž).
โ€ข Definition. A probability functional ๐‘…: ๐›พ โ†’ โ„ is called
๐‘๐‘œ๐‘™๐‘ฆโ„Ž๐‘’๐‘‘๐‘Ÿ๐‘Ž๐‘™ if there are ๐‘˜๐‘ก โˆˆ โ„•, ๐‘๐‘ก โˆˆ โ„๐‘˜๐‘ก , and nonempty polyhedral sets ๐‘‰๐‘ก โŠ† โ„๐‘˜๐‘ก , ๐‘ก = 0, . . ๐‘‡, ๐‘ค๐‘ก,๐œ โˆˆ
โ„๐‘˜๐‘กโˆ’๐œ , ๐œ = 0, . . , ๐‘ก, ๐‘ก = 0, . . ๐‘‡ such that
๏ƒฌ
๏ƒฏT
R (Y ) ๏€ฝ sup ๏ƒญ๏ƒฅ E ๏€จ ct , vt
๏ƒฏ t ๏€ฝ0
๏ƒฎ
๏€ฉ
๏€จ
๏€ฉ
๏€ฉ
vt ๏ƒŽ L p F , R k1 , vt ๏ƒŽ Vt , t ๏€ฝ 0,.., T ๏ƒผ
๏ƒฏ
t
๏ƒฝ
E wt ,๏ด , vt ๏€ญ๏ด ๏€ฝ Yt , t ๏€ฝ 1,.., T ๏ƒฏ
๏ƒฅ
๏ด ๏€ฝ0
๏ƒพ
๏€จ
โ€ข holds for every Y โˆˆ ๐›พ. Here โ‹…,โ‹… denotes scalar
products on every โ„๐‘˜๐‘ก .
Conditions for Supremal
Values
1.
2.
โ€ข
๐‘‰๐‘ก is a polyhedral cone for ๐‘ก = 0, . . , ๐‘‡ and ๐‘ค๐‘ก,0 , ๐‘‰๐‘ก = โ„ holds for every
๐‘ก = 1, . . , ๐‘‡ (complete recourse).
There exists ๐‘ข โˆˆ โ„๐‘‡ such that ๐‘0 โˆ’ ๐‘‡๐œ=1 ๐‘ค๐œ,๐œโˆ’๐‘ก ๐‘ข๐œ โˆˆ ๐‘‰0โˆ— , ๐‘๐‘ก โˆ’
๐‘‡
โˆ—
โˆ—
๐œ=๐‘ก ๐‘ค๐œ,๐œโˆ’๐‘ก ๐‘ข๐œ โˆˆ ๐‘‰๐‘ก , ๐‘ก = 1, . . , ๐‘‡, hold, where sets ๐‘‰๐‘ก are the polar cones to ๐‘‰๐‘ก .
(dual feasibility)
If 1. and 2. and the polyhedral function is defined by:
๏ƒฌ
๏ƒฏT
R(Y ) ๏€ฝ sup ๏ƒญ๏ƒฅ E ๏€จ ct , vt
๏ƒฏ t ๏€ฝ0
๏ƒฎ
โ€ข
๏€ฉ
๏€จ
๏€ฉ
๏€ฉ
vt ๏ƒŽ L p F , R k1 , vt ๏ƒŽVt , t ๏€ฝ 0,.., T ๏ƒผ
๏ƒฏ
t
๏ƒฝ
E wt ,๏ด , vt ๏€ญ๏ด ๏€ฝ Yt , t ๏€ฝ 1,.., T ๏ƒฏ
๏ƒฅ
๏ด ๏€ฝ0
๏ƒพ
๏€จ
R is finite, positively homogeneous, concave, and continuous on ๐›พ
Note on Multi-Stage
๏€จ
๏€ฉ
๏€ฉ
k1
๏ƒฌ
v
๏ƒŽ
L
F
,
R
, vt ๏ƒŽ Vt , t ๏€ฝ 0,.., T ๏ƒผ
t
p
T
๏ƒฏ
๏ƒฏ
R (Y ) ๏€ฝ sup ๏ƒญ๏ƒฅ E ๏€จ ct , vt ๏€ฉ t
๏ƒฝ
E wt ,๏ด , vt ๏€ญ๏ด ๏€ฝ Yt , t ๏€ฝ 1,.., T ๏ƒฏ
๏ƒฅ
๏ƒฏ t ๏€ฝ0
๏ด ๏€ฝ0
๏ƒฎ
๏ƒพ
โ€ข The dual solutions that correspond to the constraint
is the slope of the R.
๏€จ
E๏€จ w ๏ด , v ๏ด ๏€ฉ ๏€ฝ Y
๏ƒฅ
๏ด
t
๏€ฝ0
t,
t๏€ญ
t
โ€ข This problem is solved by means of cost-to-go
functions and bellmanโ€™s equation.
Note on Multi-Stage
๏€จ
๏€ฉ
๏€ฉ
k1
๏ƒฌ
v
๏ƒŽ
L
F
,
R
, vt ๏ƒŽ Vt , t ๏€ฝ 0,.., T ๏ƒผ
t
p
T
๏ƒฏ
๏ƒฏ
R (Y ) ๏€ฝ sup ๏ƒญ๏ƒฅ E ๏€จ ct , vt ๏€ฉ t
๏ƒฝ
E wt ,๏ด , vt ๏€ญ๏ด ๏€ฝ Yt , t ๏€ฝ 1,.., T ๏ƒฏ
๏ƒฅ
๏ƒฏ t ๏€ฝ0
๏ด ๏€ฝ0
๏ƒฎ
๏ƒพ
โ€ข The dual solutions that correspond to the constraint
is the slope of the R.
๏€จ
E๏€จ w ๏ด , v ๏ด ๏€ฉ ๏€ฝ Y
๏ƒฅ
๏ด
t
๏€ฝ0
t,
t๏€ญ
t
โ€ข This problem is solved by means of cost-to-go
functions and bellmanโ€™s equation.
Vt (Y ) ๏€ฝ max๏ปCt (Y , x) ๏€ซ E ๏€จVt ๏€ซ1 (Yt ๏€ซ1 (Yt , x)) | Y ๏€ฉ๏ฝ
x๏ƒŽ๏‘
Note on V@R
โ€ข If we use V@R, many problems will cease to be
linear and convex. However, replacing V@R with
CV@R enables us to preserve the convexity of the
underlying problem because this measure is
polyhedral.
Applications in the
Energy Sector (CHP)
Liberalization/Deregulation
of the Energy Markets
โ€ข The deregulation of energy markets has lead to an
increased awareness of the need for profit
maximization with simultaneous consideration of
financial risk, adapted to individual risk aversion
policies of market participants.
โ€ข More requirements on Risk management.
Liberalization/Deregulation
of the Energy Markets
โ€ข Mathematical modeling of such optimization
problems with uncertain input data results in mixedinteger large-scale stochastic programming models
with a risk measure in the objective.
โ€ข Often Multi-Stage problems are solved in the
framework of either dynamic or stochastic
programming.
โ€ข Simultaneous optimization of profits and risks.
Applications of
polyhedral Risk Measures
The problem of finding a strategy that yields the
optimal (or near optimal) profit under taking into
account technical constraint and risks.
min โ†’
1 โˆ’ ๐›พ โˆ— ๐‘…๐‘–๐‘ ๐‘˜ โˆ’ ๐›พ โˆ— ๐‘ƒ๐‘Ÿ๐‘œ๐‘“๐‘–๐‘ก
Specification of the
Problem
โ€ข The multi-stage stochastic optimization models are
tailored to the requirements of a typical German
municipal power utility, which has to serve an
electricity demand and a heat demand of
customers in a city and its vicinity.
โ€ข The power utility owns a combined heat and power
(CHP) facility that can serve the heat demand
completely and the electricity demand partly.
Stochasticity of the Model
Sources:
1. Electricity spot prices
2. Electricity forward prices
3. Electricity demand (load)
4. Heat demand.
Stochasticity of the Model
Multiple layers of seasonality
1. Electricity spot prices (daily, weekly, monthly)
2. Electricity demand (daily, weekly, monthly)
3. Heat demand (daily, weekly, monthly)
The seasonality is captured by the deterministic part.
Interdependency between the
Data (prices&demands)
โ€ข Prices depend on demands and vice versa
โ€ข Tri-variate ARMA models (demand for heat&electricity
and spot prices).
โ€ข Spot prices AR-GARCH.
โ€ข The futures prices are calculated aposteriori from the
spot prices in the scenario tree. (month average)
Parameters
Decision Variables
Objective
Objective โ€“ Cash Values
โ€ข Cash values are what we earn from producing heat
and electricity. We of course take into account
technical constraints.
Objective
Simulation Results
โ€ข The best strategy is to not use any contracts.
โ€ข Minimizing without a risk measure causes high spread for
the distribution of the overall revenue.
โ€ข The incorporation of the (one-period) CV@R applied to
z(T) reduces this spread considerably for the price of high
spread and very low values for z(t) at time t<T.
Simulation Results
Simulation Results
Simulation Results
Simulation Results
Simulation Results
Simulation Results
Conclusion
โ€ข Polyhedral risk measures enable us to incorporate
more realistic features of the problem and to
preserve its convexity and linearity.
โ€ข Hence, they enable the tractability of many
problems.
โ€ข V@R is a less sophisticated risk measure, but many
problems cannot be solved by using V@R unlike
CV@R.
Bibliography
โ€ข A. Philpott, A. Dallagi, E. Gallet. On Cutting Plane Algorithms
and Dynamic Programming for Hydroelectricity Generation.
Handbook of Risk Management in Energy Production and
Trading International Series in Operations Research &
Management Science , Volume 199, 2013, pp 105-127.
โ€ข A. Shapiro, W. Tekaya, J.P. da Costa, and M.P. Soares. Risk
neutral and risk averse Stochastic Dual Dynamic Programming
method. 2013.
โ€ข G. C Pflug, W. Roemisch. Modeling, Measuring and Managing
Risk. 2010.
โ€ข A. Eichhorn, W. Römisch, Mean-risk optimization of electricity
portfolios using multiperiod polyhedral risk measures. 2005