Development of an algorithm for solving mixed integer and nonconvex problems arising in electrical supply networks

Transcription

1 Development of an algorithm for solving mixed integer and nonconvex problems arising in electrical supply networks E. Wanufelle 1 S. Leyffer 2 A. Sartenaer 1 Ph. Toint 1 1 FUNDP, University of Namur 2 Argonne National Laboratory One-day Symposium on Optimization and Engineering, CORE, May 24th 2006

2 Outline 1 Motivations 2 Piecewise linear approximations 3 and numerical results 4

3 The considered problem Studied application Unsuitable available methods The problem of tertiary voltage control (TVC) In alternating current: power is a complex number real part = real power imaginary part = reactive power reactive power transmission causes voltage drops and losses need a regulation of the reactive power produced by each generator of an electrical network under some physical laws problem and model provided by Tractebel Engineering

4 Modelling of the problem Studied application Unsuitable available methods 8 >< >: min P k N G w k (Q k obj k ) 2 s.t. P i P ic X P ik X P ik X P ik X P ik = 0, i N ik S s i ik S e i ik T s i ik T e i Q i Q ic + a i ν 2 i Q i0 X Q ik X Q ik X Q ik X Q ik = 0, i N X ik S s i ik S e i ik T s i Q ik = K ik B ν mini ν i ν maxi, a i binary i N P mini P i P maxi, Q mini Q i Q maxi i N G r minik r ik r maxik, r ik E disc discrete ik T θ mini θ i θ maxi, i N ik T e i

5 Studied application Unsuitable available methods Modelling of the problem (continued) where P ik = ν 2 i (y ik cos(ζ ik ) + g ik ) ν i ν k y ik cos(ζ ik + θ i θ k ), Q ik = ν 2 i (y ik sin(ζ ik ) h ik ) ν i ν k y ik sin(ζ ik + θ i θ k ), P ik = ν 2 i r 2 ik y ik cos(ζ ik ) ν i ν k r ik y ik cos(ζ ik + θ i θ k ), Q ik = ν 2 i r 2 ik y ik sin(ζ ik ) ν i ν k r ik y ik sin(ζ ik + θ i θ k ), P ki = ν 2 k (y ik cos(ζ ik ) + g ik ) ν i ν k y ik cos(ζ ik + θ k θ i ), Q ki = ν 2 k (y ik sin(ζ ik ) h ik ) ν i ν k y ik sin(ζ ik + θ k θ i ), P ki = ν 2 k (y ik cos(ζ ik ) + y 0 ik cos(ζ 0ik)) ν i ν k r ik y ik cos(ζ ik + θ k θ i ), Q ki = ν 2 k (y ik sin(ζ ik ) + y 0 ik sin(ζ 0ik )) ν iν k r ik y ik sin(ζ ik + θ k θ i ), ik Si e ik Si e ik Ti e ik Ti e ki Si s ki Si s ki Ti s ki Ti s highly nonlinear, nonconvex

6 Use of discrete variables Studied application Unsuitable available methods a i : binary (i N) variables on/off r ik E disc : discrete (ik T ) e.g.: E disc = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10} the transformer ratio can only be equal to some fixed values Mixed Integer NonConvex Programming problem

7 Motivation Motivations Studied application Unsuitable available methods Current approach: heuristics: Successive solutions of relaxed nonlinear problems wish to work with more reliable/robust methods Idea: use an appropriate linear approximation of the problem

8 Linear suitable function SOS approximation Decomposition of the problem How can we approximate a nonlinear component by a linear function? e.g.: sin

9 Linear suitable function SOS approximation Decomposition of the problem How can we approximate a nonlinear component by a linear function? e.g.: sin not accurate

10 Linear suitable function SOS approximation Decomposition of the problem How can we approximate a nonlinear component by a linear function? e.g.: sin piecewise linear approximation

11 Linear suitable function SOS approximation Decomposition of the problem Approximation by special ordered sets To approximate f (x) by f (x), we use f (x) f (x) = n λ i f (x i ) i=1 where x i are breakpoints, i = 1, n n x = λ i x i i=1 n λ i = 1, λ i 0, i=1 Refs: Beale, Tomlin, Martin i = 1, n

12 SOS condition: motivation Linear suitable function SOS approximation Decomposition of the problem If λ 1 0, λ 5 0 λ i = 0, i = 2,.., 4

13 SOS formulation (1 dimension) Linear suitable function SOS approximation Decomposition of the problem f (x) f (x) = n λ i f (x i ) i=1 where x i are breakpoints, i = 1,.., n n x = λ i x i i=1 n λ i = 1, λ i 0, i=1 SOS type 2 condition: At most 2 λ i can be nonzero. Moreover, these λ i must be adjacent. i = 1,..., n

14 SOS formulation (1 dimension) Linear suitable function SOS approximation Decomposition of the problem f (x) f (x) = n λ i f (x i ) i=1 where x i are breakpoints, i = 1,.., n n x = λ i x i i=1 n λ i = 1, λ i 0, i=1 i = 1,..., n At most 2 λ i can be nonzero. Moreover, these λ i must be adjacent. (LP) (LP) + branching

15 SOS formulation (2 dimensions) Linear suitable function SOS approximation Decomposition of the problem f (x, y) f (x, y) = n i=1 j=1 m λ ij f (x i, y j ) where (x i, y j ) are breakpoints, i = 1,.., n, j = 1,.., m n m x = λ ij x i y = n i=1 j=1 n i=1 j=1 i=1 j=1 m λ ij y j m λ ij = 1, λ ij 0, i = 1,..., n, j = 1,.., m At most 3 λ ij can be nonzero. Moreover, these λ ij must be adjacent on a triangle.

16 Illustration: xy Motivations Linear suitable function SOS approximation Decomposition of the problem On [ 2 : 2] [ 2 : 2] :

17 Illustration: xy Motivations Linear suitable function SOS approximation Decomposition of the problem Approximation by SOS: 3 breakpoints are used in each dimension

18 Illustration: xy Motivations Linear suitable function SOS approximation Decomposition of the problem Dividing the feasible domain into triangles

19 Linear suitable function SOS approximation Decomposition of the problem Approximation by special ordered sets (3 dimensions and more) the same reasoning could be used BUT introduction of a lot of variables into the problem: for k breakpoints in each dimension: 1 dim : k var λ 2 dim : k 2 var λ 3 dim : k 3 var λ... n dim : k n var λ Idea: decompose problem into components of 1 or 2 variables

20 Decomposition of the problem Linear suitable function SOS approximation Decomposition of the problem Computational graph for c = 4x 1 x x 2 x 4 sin(x 3 ) Decomposition of the problem into nonlinear components of 1 or 2 variables Approximation of each of these nonlinear components by new variables Computational graph not unique

21 Main components of the problem Linear suitable function SOS approximation Decomposition of the problem 3 main kinds of nonlinear components: square functions: x 2 trigonometric functions: sin(x), cos(x) bilinear functions: xy

22 Insufficient approximation Linear suitable function SOS approximation Decomposition of the problem Building of a linear approximation problem subject to SOS conditions There exists an efficient method (Martin) solution of an approximation problem the solution of that problem has little chance to be feasible for our our problem physical constraints must be absolutely satisfied use outer approximations to guarantee solution

23 Insufficient approximation Linear suitable function SOS approximation Decomposition of the problem Building of a linear approximation problem subject to SOS conditions There exists an efficient method (Martin) solution of an approximation problem the solution of that problem has little chance to be feasible for our our problem physical constraints must be absolutely satisfied use outer approximations to guarantee solution

24 Insufficient approximation Linear suitable function SOS approximation Decomposition of the problem Building of a linear approximation problem subject to SOS conditions There exists an efficient method (Martin) solution of an approximation problem the solution of that problem has little chance to be feasible for our our problem physical constraints must be absolutely satisfied use outer approximations to guarantee solution

25 Outer approximations Outer approximations Refinement of approximations Algorithm Numerical results Idea: replace each nonlinear component f by a linear domain which includes the nonlinear function.

26 Outer approximations Outer approximations Refinement of approximations Algorithm Numerical results Idea: replace each nonlinear component f by a linear domain which includes the nonlinear function. Idea recently used (Gatzke) Difference: use of linear big M approximations instead of SOS approximations

27 Determination of an outer domain Outer approximations Refinement of approximations Algorithm Numerical results For each component f, compute the approximation errors ɛ L (x i, x i+1 ) = max x [xi,x i+1 ]( f (x) f (x), 0) ɛ U (x i, x i+1 ) = max x [xi,x i+1 ](f (x) f (x), 0) and replace f (x) f (x) by f (x) ɛl (x i, x i+1 ) f (x) f (x) + ɛ U (x i, x i+1 ), x i x x i+1 on each piece

28 Too coarse approximations Outer approximations Refinement of approximations Algorithm Numerical results Results in an outer approximation BUT its solution can be very far from the true solution Need to refine the approximations

29 Refinement of approximations Outer approximations Refinement of approximations Algorithm Numerical results Use of a branch-and-bound tree: reduce the approximation interval, refine the mesh better approximations

30 Refinement of approximations Outer approximations Refinement of approximations Algorithm Numerical results Use of a branch-and-bound tree: reduce the approximation interval, refine the mesh better approximations

31 Refinement of approximations Outer approximations Refinement of approximations Algorithm Numerical results Use of a branch-and-bound tree: reduce the approximation interval, refine the mesh better approximations ideal framework to treat discrete variables 2 types of division guaranteed convergence to the global optimum in the end

32 Outer approximations Refinement of approximations Algorithm Numerical results Choices associated to the branch-and-bound process Choice of the node to refine: depth-first search Choice of the variable to divide: the variable of the starting problem leading to the largest error of approximation not on the SOS variables λ... inefficient Upper bound: the solution of the 1st linear problem is employed as starting point for the NLP problem to generate an upper bound

33 Algorithm Motivations Outer approximations Refinement of approximations Algorithm Numerical results 1 Build an outer approximation problem, (LP 0 ), for (P) k := 0 2 Propagate bounds through the computational graph and compute the approximation errors. Update (LP k ) 3 Solve (LP k ) ( x, f ) If f U the node can be cut, else if x is feasible for (P) and f ( x) < U U = f ( x), x = x and the node can be cut else choose a variable j and divide the pbm (LP k ) into 2 new subproblems 4 If the tree is completely explored: STOP else k := k + 1 choose a node which has not been examined yet and go to 2.

34 Numerical results Motivations Outer approximations Refinement of approximations Algorithm Numerical results Toy problem: (P) no discrete variables min w 1 sin w 4 s.t. 4w 1 w w 2 w 4 sin w 3 1 w 2 0.5w 2 w 4 cos w w w w 3 2π 0 w 4 2π 5 breakpoints for the trigonometric components, 3 for the others approximation problem: 69 variables and 46 constraints

35 Numerical results (continued) Outer approximations Refinement of approximations Algorithm Numerical results Nonlinear local optimization solvers available on NEOS: KO for 87.5% of the solvers Nonlinear global optimization solver, ACRS: OK but random BARON: not applicable due to sin(x), cos(x) Our method: global solution obtained (and proved) after the solution of 103 LP and 2 NLP (ɛ = 10E-6).

36 Future work More tests problems Increase the speed of convergence by improving presolve developing better rules to choose the variable to divide and the place to divide testing finer approximations (quadratic, inequalities of McCormick,...) adding cuts to the problem of approximation Introduction of discrete variables into the problem

37 Conclusion Motivations Promising approach Able to ensure convergence to the global optimum But convergence can be slow Solution of linear problems only But needs the introduction of new variables and constraints into the approximation problem