23.01.2014 Views

Hedging Strategy and Electricity Contract Engineering - IFOR

Hedging Strategy and Electricity Contract Engineering - IFOR

Hedging Strategy and Electricity Contract Engineering - IFOR

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

R<br />

R<br />

M<br />

R<br />

R<br />

64 Risk management<br />

where moreover<br />

argmin<br />

argmin<br />

x† X<br />

H xg j@L † X`‰‡<br />

argmin<br />

S xŠ G?R Š V<br />

Fd S x GaRbV<br />

Fd S xŠ G?RbV<br />

xŠ<br />

U d S xV?G‹R Š<br />

†m‡ j<br />

A proof of Theorem 3.4 <strong>and</strong> 3.5 is given in [RU00]. lS If G x V Y is convex with<br />

respect to x, then (3.2) <strong>and</strong> (3.3) can with help Fd S of GaRbV x be solved with<br />

convex programming, since both the objective function <strong>and</strong> the constraint are<br />

then convex.<br />

The integral Fd S in G?RKV x can be approximated by sampling from the probability<br />

distribution of Y . 9 If we sample a collection of vectors y P?P@P<br />

G 1G y J then<br />

S x GaRbV can be approximated by<br />

Fd<br />

S 1<br />

1<br />

c V J<br />

J<br />

j… 1<br />

M P<br />

Fd S x GaRKV<br />

lS x G y jV<br />

If the loss lS function G x V Y is linear in x, the optimization problem involving<br />

CVaR can be solved with linear programming. This is a very nice feature of<br />

CVaR, since linear programming can h<strong>and</strong>le very large problems efficiently.<br />

The lS terms G x y jV Fd S in GaRKV x are not linear, but piecewise linear. This<br />

can however be resolved by replacing these terms by the auxiliary variables z j ,<br />

<strong>and</strong> imposing the constraints z j x G y jV R , z lS j 0, 1G<br />

P?P@P<br />

G j J. The<br />

optimization problem (3.2) can then be reduced to the linear program<br />

min<br />

1<br />

n g z†ˆ‡<br />

J g j †m‡ H 1I d L J j… 1 z j<br />

x†m‡<br />

s.t. z j x G y jV RŒG z lS j j 1G 0G<br />

1<br />

J<br />

J<br />

J<br />

j… 1 lS x G y jV R,<br />

P?P@P<br />

G J<br />

(3.4)<br />

<strong>and</strong> (3.3) similarly to<br />

9 There is a whole theory on how to sample <strong>and</strong> generate r<strong>and</strong>om numbers. We have<br />

in this work used the Monte Carlo method, but there are a vast number of variancereduction<br />

techniques, such as Latin hypercube sampling. For more on sampling see, for<br />

example, [Nie92].

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!