Vous êtes sur la page 1sur 31

Applications of optimization methods in electrical

power systems

Florin Capitanescu

Interdisciplinary Centre for Security, Reliability, and Trust (SnT)


University of Luxembourg

Convex optimization course (instructor Gan Zheng)


SnT Luxembourg
6 December 2012

Florin Capitanescu Optimization methods in power systems (1 / 31)


Outline of the presentation

Day-ahead decision making in power systems


security-constrained optimal power flow (SCOPF)
SCOPF problem decomposition
Iterative approaches to reduce the very large size of the
problem
Handling of discrete variables
SCOPF and decision making in real-time
Conclusions and future works
Interior-point method for nonlinear programming

Florin Capitanescu Optimization methods in power systems (2 / 31)


Electric power system picture

picture taken from www.altenergymag.com


Florin Capitanescu Optimization methods in power systems (3 / 31)
Decision making in power systems : stages, time-scales,
and requirements for optimization tools

grid planning (years ahead of operation)


very accurate optimization tools with no special solution time
constraints
grid maintenance planning (years to months ahead of
operation)
very accurate optimization tools with no special solution time
constraints
...
operational planning (day-ahead of operation)
accurate optimization tools with stringent solution time
constraints (few minutes to one hour)
real-time operation
very fast optimization tools using reasonable approximate
models (solution desired between few seconds and 15 minutes)

Florin Capitanescu Optimization methods in power systems (4 / 31)


Day-ahead operational planning
contingency : the outage of any equipment of the system (e.g.
line, generator)
input data : the load forecast of the 24 states of the system
the next day
aim : for each anticipated state of the next day the system
must operate at minimum cost while being able to withstand
the loss of any single equipment (N-1 reliability criterion)
ensure a stable transition towards a viable equilibrium point
very complex optimization problem :
multi-period optimization (solution coupled over 24 hours
including usually 24 states)
very large scale (consider a large number of contingencies)
nonlinear algebraic and differential equations (model the
system behaviour for postulated contingencies)
with a large number of variables (binary, discrete, and
continuous)
with stringent solution time requirements (e.g. less than 1
hour) !
Florin Capitanescu Optimization methods in power systems (5 / 31)
Day-ahead operational planning
problem decomposition in sequential sub-problems
(economics vs. security) :
(market-based) unit commitment : determines the status
on/off of generators for each period of time and the
generators active power according to their bids
very large MILP problem (neglects nonlinear network equations
or uses their linearization)
SCOPF : determines optimal preventive/corrective control
actions to satisfy security constraints (overloads, abnormal
voltages) for the 24/48 anticipated operation states of the
power system for the next day
very large MINLP problem
time-domain (dynamic) simulation : check system stability
for the postulated contingencies
numerical integration of dynamic phenomena with different
time scales (e.g. miliseconds to minutes)

Florin Capitanescu Optimization methods in power systems (6 / 31)


Conventional Optimal Power Flow (OPF) formulation

min f (x, uc , ud ) ← generation cost, power losses


x,uc ,ud

s.t. g(x, uc , ud ) = 0 ← AC power flow equations


h(x, uc , ud ) ≤ 0 ← operational limits on I, P, S, V
uc ≤ uc ≤ uc ← bounds on control variables
T
ud = [ud1 . . . udi . . . udnd ]
1 j p(i)
udi ∈ {udi , . . . , udi , . . . , udi } ← discrete variables values

x - state/dependent variables :
magnitude V and angle θ of complex voltage at all buses
uc - continuous control variables :
generator active power, generator terminal voltage, etc.
ud - discrete/binary control variables :
transformer ratio, phase shifter angle, shunt
capacitors/reactors reactive power, network topology, etc.
Florin Capitanescu Optimization methods in power systems (7 / 31)
Conventional SCOPF formulation
SCOPF is OPF extension which considers constraints for :
base case and (mainly “N-1”) post-contingency states
min f (x0 , u0 )
x0 ,...,xc ,u0 ,...,uc
s.t. g0 (x0 , u0 ) = 0 ← base case constraints
h0 (x0 , u0 ) ≤ 0 ← base case constraints
gk (xk , uk ) = 0 k = 1, . . . , c ← contingency k constraints
hk (xk , uk ) ≤ 0 k = 1, . . . , c ← contingency k constraints
|uk − u0 | ≤ ∆umax
k k = 1, . . . , c ← “coupling” constraints

the “coupling” constraints avoid unrealistic control variations


limited by ∆umax
k = Tk du/dt, where :
Tk is the time horizon assumed to remove violated limits
du/dt is the controls ramp
post-contingency constraints may be duplicated to satisfy
several security thresholds
Florin Capitanescu Optimization methods in power systems (8 / 31)
Features of the conventional SCOPF problem
nonlinear : includes power flow equations and other nonlinear
inequality constraints
non-convex : includes power flow equations and bounds on
other nonlinear inequality constraints
as opposed to convex optimization cannot guarantee global
optimality but only local optimality
global optimum may even be undesired from operational point
of view
if human operators are not used with those grid operating
conditions the solution is more difficult to manage and
understand
if the gain wrt local optimum is not significant
if it requires large amounts of control means redispatch
with continuous, discrete, and binary variables
static : refers to a single operating point in time
important drawback as the system operating conditions are
dynamically (and even significantly) changing during 30
minutes/1 hour
Florin Capitanescu Optimization methods in power systems (9 / 31)
Features of the conventional SCOPF problem

large scale : the SCOPF problem for a 1000-bus system and


999 contingencies contains :
around 2000 x 1000 = 2.000.000 equality constraints
around 6000 x 1000 = 6.000.000 inequality constraints
around 500 x 1000 = 500.000 control variables
intractable on a normal computer due to memory limitation
the problem can be only partially paralelized
stringent solution time requirements
fortunately not all constraints are binding
identification of binding constraints at the optimum is essential

Florin Capitanescu Optimization methods in power systems (10 / 31)


OPF/SCOPF problem decomposition according to the
type of control variables

OPF/SCOPF belongs to the class of optimization problems


called Mixed-Integer Nonlinear Programming (MINLP)
nowadays there is no conventional MINLP solver able to solve
large scale SCOPF problems (prohibitive CPU times and
memory requirements due to large combinatorial nature) !
hence one has to rely on heuristics
conventional approach :
handle separately continuous and discrete variables
continuous variables settings computed by NLP solvers
discrete variables settings determined by heuristic approaches :
(at worst) rounding-off, problem linearization and MILP, etc.

Florin Capitanescu Optimization methods in power systems (11 / 31)


Approaches for solving the “continuous” SCOPF

If discrete variables are fixed or assumed continuous then


the OPF becomes a nonlinear programming (NLP) problem
direct approach
iterative (active contingency set) approach using the full
contingency model
iterative approaches using approximate contingency models
Benders decomposition
linearization of relevant post-contingency constraints
network compression of post-contingency states
use of DC problem approximation

Florin Capitanescu Optimization methods in power systems (12 / 31)


OPF/SCOPF as a NLP : solution methods background

1968 : gradient method (H. Dommel and W. Tinney)


1973 : sequential quadratic programming (G. Reid and L.
Hasdorf)
1974 : sequential linear programming (O. Alsac and B. Stott)
1984 : Newton Method (Sun et al.)
1987 : Benders decomposition (A. Monticelli, M. Pereira, S.
Granville) for the corrective SCOPF
1994 : Interior-point method (Y. Wu et al., and S. Granville)
2007 : Trust regions (H. Wang et al., A. Sousa and G.L
Torres)
2008 : (Extended) conic programming (R. Jabr)
2012 : Semi-definite programming (J. Lavaei and S. Low)

Florin Capitanescu Optimization methods in power systems (13 / 31)


SCOPF approaches : iterative approaches

motivation : in real-life applications most contingencies do


not constrain the optimum
aim : to quickly identify the set of binding contingencies and
to solve the SCOPF by including only these contingencies
efficient contingency filtering techniques are essential !
definition : the set of binding contingencies is the smallest
subset of the full contingency set which provides the same
optimal objective value as the full set
binding contingencies generally lead to active post-contingency
constraints on operational limits (e.g. branch currents, voltage
magnitudes) at the SCOPF optimum
if the computer memory does not allow to include in SCOPF
at least all binding contingencies one has to rely on
approximate models for post-contingency states !

Florin Capitanescu Optimization methods in power systems (14 / 31)


Challenges to the SCOPF problem solution

Handling of discrete variables


Variables with small discrete steps :
transformer ratio
phase shifter angle
shunt reactors/capacitors reactive power
Variables with large discrete steps and binary variables :
network switching
unit start-up/shut down
shunt reactors/capacitors reactive power
Binary variables stemming from :
realistic modelling of TSO’ operational rules
problem reformulation to enable the use of mature solvers

Florin Capitanescu Optimization methods in power systems (15 / 31)


Handling of discrete variables

Variables with small discrete steps :


many approaches proposed that provide acceptable results
unless if the proposed discrete variables settings are infeasible
need further improvement to enable convergence to feasible
solutions
Variables with large discrete steps : very challenging due to
highly nonlinear impact on the network state !
separate vs. simultaneous handling of these two classes of
variables

Florin Capitanescu Optimization methods in power systems (16 / 31)


Handling of discrete variables
MINLP classical methods are not yet sufficiently mature to
cope with very large size problems
handling of discrete variables is a trade-off between :
degree of sub-optimality
reliability (ability to deal with infeasible discrete variables
settings)
computational speed
proposed approaches :
simple heuristics : round-off, progressive round-off
penalty functions within NLP or LP solvers
ordinal optimization
mixed-integer linear programming (MILP)
interior point cutting plane
mathematical programs with equilibrium constraints (MPEC)
developed by RTE within PEGASE project
sensitivity of discrete variables wrt objective/constraints and
merit functions (ULg)
evolutionary optimization methods (e.g. GA, SA, TS)
Florin Capitanescu Optimization methods in power systems (17 / 31)
SCOPF and decision-making in real-time

ultimate (utopian ?) OPF application goal : use OPF in


real-time in an automatic closed-loop control fashion
optimum is a continuously moving target
in real-time the OPF is rarely used and only in open-loop i.e.
the system operator is responsible for implementing its
solution
secondary voltage control is the single optimization-based
automatic closed-loop controller present in few systems
it is a kind of model predictive control

Florin Capitanescu Optimization methods in power systems (18 / 31)


SCOPF application in real-time

gaps to bridge for an automatic closed-loop OPF use in real-time :


OPF must be very reliable and very fast
nowadays NLP field has reached maturity
solvers are very reliable and fast (e.g. generic solvers such as
KNITRO, LOQO, IPOPT, etc.)
OPF/SCOPF should have warm start capability
interior-point method do not generally warm start well
system operators do not need optimal solutions from an
algorithmic perspective (e.g. NLP) but near-optimal solutions
from operational perspective
in real life systems some constraints are discontinuous (bidding
curves) and non-differentiable (valve effect)
system operators need a limited number of control actions
system operators need a sequence of control actions

Florin Capitanescu Optimization methods in power systems (19 / 31)


SCOPF application in real-time
gaps to bridge for an automatic closed-loop OPF use in real-time :
accept that multiple local optima may exist due to problem
non-convexity
OPF should have the ability to detect/handle infeasibility
provide a solution with least violation of soft constraints
progressively removing constraints or adding control variables
implement constraints/controls priority rules
cope with data errors : in real-time OPF relies on the output
of the state estimator which is not always “clean”
ability to filter out cases where the power system snapshot is
taken during fast dynamics and identify cyber-attacks
study interaction with Automatic Generation Control (AGC)
for frequency regulation
hybrid control system as an intermediate step :
OPF-based control with the system operator in the loop

Florin Capitanescu Optimization methods in power systems (20 / 31)


Conclusions and future works
the current process of optimal decision making in power
systems is still challenging, especially in real time
various optimization fields are not sufficiently mature (e.g.
MINLP, BP, MPEC)
this process becomes more complex in order to meet the
objectives of “smart grids”
need robust closed loop controllers paying attention to
cybersecurity aspects
many research paths are open
SCOPF under uncertainty (e.g. multiple base case SCOPF)
multi-period SCOPF to optimize objectives over longer
horizons
minimize power losses over the next day
there is a need for look-ahead prediction-based
decision-making tools
dynamic programming
model predictive control

Florin Capitanescu Optimization methods in power systems (21 / 31)


Interior-Point Method (IPM) - Introduction

emerged in the middle 50’s (Frisch)


largely developed in the late 60’s (Fiacco and McCormick)
very popular since 1984 : IPM clearly outperforms the classical
simplex method for linear-programming problems (Karmarkar)
IPM successfully applied to nonlinear programming problems

Florin Capitanescu Optimization methods in power systems (22 / 31)


Obtaining optimality conditions :
Interior-Point Method for Nonlinear Programming

Denote y = [x, uc , ud ]T and assume ud fixed or continuous


General form of the nonlinear programming problem :

min f (y)
y

s.t. g(y) = 0
h(y) ≤ 0

f (y), gi (y) and hi (y) are twice continuously differentiable


and at least one of these functions is nonlinear

Florin Capitanescu Optimization methods in power systems (23 / 31)


IPM - Obtaining optimality conditions (step 1)

One adds non-negative slack variables to inequality constraints,


transforming them into equality constraints :

h(y) ≤ 0 ⇔ −h(y) − s = 0, s ≥ 0

The problem becomes :

min f (y)
y,s

s.t. g(y) = 0
− h(y) − s = 0
s≥0 ← non-negativity conditions

y and s are called primal variables.

Florin Capitanescu Optimization methods in power systems (24 / 31)


IPM - Obtaining optimality conditions (step 2)

Non-negativity conditions are implicitly handled by appending


them to the objective function as logarithmic barrier terms :
p
X
min f (y) − µ ln si
y,s
i=1
s.t. g(y) = 0
− h(y) − s = 0

µ is a positive scalar called barrier parameter


and p = dim h
The inequality constrained optimization problem has been
transformed into an equality constrained one !

Florin Capitanescu Optimization methods in power systems (25 / 31)


IPM - Obtaining optimality conditions (step 3)

One transforms the equality constrained optimization problem into


an unconstrained optimization one, by defining the Lagrangian :
p
X
Lµ (z) = f (y) − µ ln si − λT g(y) − π T [−h(y) − s]
i=1

λ, π are called dual variables (or Lagrange multipliers)


and let z = [s π λ y]T

Florin Capitanescu Optimization methods in power systems (26 / 31)


IPM - Obtaining optimality conditions (step 4)

The first order perturbed Karush-Kuhn-Tucker (KKT) optimality


conditions are obtained by setting to zero the Lagrangian’s first
derivatives wrt all unknowns :
   
∇s Lµ (z) −µ1 + Sπ
 ∇π Lµ (z)   h(y) + s 
 = =0
 ∇ Lµ (z)   −g(y)
λ

∇y Lµ (z) T T
∇f (y) − Jg (y) λ + Jh (y) π

∂g(y) ∂h(y)
where Jg (y) = and Jh (y) =
∂y ∂y
the exact KKT conditions correspond to µ = 0
Fiacco & McCormick theorem :
as µ → 0, y(µ) → y? , a local optimum of the original problem

Florin Capitanescu Optimization methods in power systems (27 / 31)


Pure Primal-Dual (PD) Algorithm

Solve for the Newton direction the linear system :


   
∆s µ1 − Sπ
 ∆π   −h(y) − s 
H(z)  ∆λ  = 
  
g(y) 
∆y −∇f (y) + Jg (y)T λ − Jh (y)T π

∂ 2 Lµ (z)
H(z) = is the second order derivatives Hessian matrix
∂z2
IPM idea : solve this system iteratively for decreasing values of µ

Florin Capitanescu Optimization methods in power systems (28 / 31)


Outline of the pure Primal-Dual (PD) algorithm

1 Set k = 0. Initialize z0 , such that s0 , π 0 > 0. Chose µ0 > 0.


2 Solve the linear system of equations for the Newton’s direction ∆zk :

H(zk ) ∆zk = −∇Lµ (zk )

3 Determine the largest step αk ∈ (0, 1] along the direction ∆zk such that
(sk+1 , π k+1 ) > 0. Update solution :

zk+1 = zk + αk ∆zk

4 Check convergence. A (locally) optimal solution is found when :


primal feasibility, dual feasibility, complementarity gap and objective function
variation from an iteration to the next fall below some tolerances.
5 Compute the barrier parameter for the next iteration :

ρk
µk+1 = σ k ,where usually σ k = 0.2
p

where ρk = (sk )T π k is called complementarity gap. Set k = k + 1. Go to step 2.

Florin Capitanescu Optimization methods in power systems (29 / 31)


Interior-Point Method (IPM) - concluding remarks

advantages of the IPM :


fast convergence
adequate handling of inequality constraints by
logarithmic barrier functions
a strictly feasible initial point is not required
drawbacks of the IPM :
the heuristic of decreasing the barrier parameter
slack variables and their corresponding dual variables
must remain positive at every iteration
does not warm start well

Florin Capitanescu Optimization methods in power systems (30 / 31)


For further reading : works that inspired this presentation
B. Stott, O. Alsac, A.J. Monticelli, “Security analysis and
optimization” (Invited Paper), Proc. of IEEE. 75, 1987.
A. Papalexopoulos, “Challenges to On-Line OPF Implementation”,
IEEE/PES Winter Meeting, New York (USA), 1995.
F. Capitanescu, J.L. Martinez Ramos, P. Panciatici, D. Kirschen, A.
Marano Marcolini, L. Platbrood, and L. Wehenkel, “State-of-the-art,
challenges, and future trends in security constrained optimal power
flow”, EPSR Journal, vol. 81, no. 8, 2011, pp 1731-1741.
P. Panciatici, S. Fliscounakis, Y. Hassaine, J.L. Martinez-Ramos, M.
Ortega-Vazquez, L. Platbrood, L. Wehenkel, “Security management
under uncertainty : from day-ahead planning to intraday operation”,
IREP Symposium, Buzios (Brazil), 2010.
M. Ilic, “AC OPF and smart grids”, presentation at FERC
Conference on Enhanced Optimal Power Flow Models, Washington,
USA, 2010.
J. Bialek, P. Varaiya, F. F. Wu, J. Zhong, “Smart Operation of
Smart Grid”, Proc. of the IEEE, 2010.
Florin Capitanescu Optimization methods in power systems (31 / 31)

Vous aimerez peut-être aussi