Homogeneous Interior Point Method For Constrained Power Scheduling
Homogeneous Interior Point Method For Constrained Power Scheduling
R.A.Jabr and A.H.Coonick Abstract: A method for solving the active power economic dispatch problem is presented. The problem is formulated a s a convex program which allows [he inclusion o f a second-ordcr network model, phasc shifters and operational constraints. An accurate representation o l the line losses is achieved through the second-order model. The problem is solved using a homogeneous interior point (HIP) algorithm. The HIP algorithm yields either an approximate global optimum or detects the possible infeasibility or unboundedness of the problem. Moreover, this algorithm does not require any skill for setting a propcr starting point. The algorithm is tested on standard IEEE networks and on a practical network. Computational experience shows that the method is efficient.
Introduction
The objective of active power dispatching is to satisfy the consumer demand at minimum cost, without violating any of the physical or technical constraints of the system. The dispatch schedule at the true minimum cost requires correct modelling of network losses. However, iterative procedures, such a s sequential linear programming (SLP), are usually needed to reflect accuratcly transmission losses in the generator output [I]. These algorithm require a heuristic procedure to adjust the stcp-size limit and guide convergence towards optimality. However, it cannot be proved mathematically that this heuristic method leads to the optimum solution. Classical one-step methods usually satisfy an cstimated transmission loss [2], or it global loss term obtained through the B-matrix loss formula [3]. The hoiiiogencous interior point (HIP) algorithm is a one-step method that takes into account local transmission losses without any need of recursive programming. This algorithm is an extension of the simplified homogeneous and self-dual linear programming model [4] for convex optimisation. Theoretically, it is similar to the algorithm presented in [5],but the computational procedure is different. First, the algorithm is presented for symmetric convex programming, which is different from the standard form in [5].Secondly, the search direction is obtained from a positive definite matrix through Cholesky decomposition, and thirdly, our implementation allows different primal and dual step lengths as opposed to a cominoii one. Recently, interior point (IP) methods have found applications in power system optimisation. Yan and Quintana [GI proposed the Mehrotra predictor corrector (MPC) IP method to solve the loss-less constrained economic dispatch and the optimal reactive power dispatch problem by SLP.
0IEE, 2000 IEE Iroc,eedings onlinc no. 20000437 DOL 10.1049/ipgtd:20000437
Papel- first rmived 2nd Novembcr 1999 and in rcviscd form 1st March 2000 Thc authors are with the Department of Electrical and Elwtronic Engineering, Impcrial College, London, SW7 2BT, UK
Vargas ct d.[7] used the dual affine-scaling (DAS) IP method to solve the economic dispatch with network constraints and transmission losses, also through SLP. The SLICcess of the recursive methods in [6, 71 d i e d upoii heuristically setting a step-size limit. Indeed, using different values for this limit resulted in sub-optimal results [7].Wci et d. proposed a primal-dual (PD) IP method and the [8] MPC IP method to solve the OPF, with the AC-load flow equations included directly in the optiniisation constraints. However, since the OPF is not convex, convergence is not guaranteed. The proposcd approach has two new features that are not present in the IP nicthods surveyed above. First, it is less sensitive than the other inethods to the selection of the initial point. Numerical experience, together with theoretical evidencc [c)], shows that we can always start the algorithm successfUy from a simple well centred initial poinl. This eliminates the nced for special heuristics to get a good starting solution in PD/MPC IP inethods and big-M sclieines in affine-scaling algorithms. Secondly, it offers an elegant and reliable approach to detect if the solution is infeasible or unbounded. Solution infeasibility may arise owing to a variety of reasons, such as a change in network topology following a line contingency or generation shortfall. Moreover, since the HIP solves a convex program, convergence to the global optimum, if it exists, is guaranteed. Compared to the PDiMPC algorithms, the only disadvantage of this method is that it requircs one additional step, with thc factorisation of the reduced Newton equations matrix in each iteration. Unlike SLP, the proposed nicthod does not suffer from the inappropriate setting of the trust region radius. Also in SLP, if an LP sub-problem is infeasible, it is usually taken as an indication that the non-linear program itself is infeasible, although this cannot be guaranteed.
Problem formulation
The constrained economic dispatch (CED) problem considered consists of minimising the toVal dispatch cost, subject to satisfying the power-balance constraints, network constraints and generator dispatch limits. The second-order network model accounts for line losses locally [I]:
219
where PGj = powcr generation at node i P D j = powcr demand at node i dj = voltage angle at iiodc i y j j = line admittance = gii + d(-I) . / I , k(i) = set of nodcs coiinccted to node i uji = v , , , = V,V$;j i # , j VVh;
- 1 6,s - fibf L (r 4 . (4) where (4, @)are thc lower and uppcr phase-shift limits. Moreover, we ciisurc in the optimisation process that thc pliase shiftcr powcr SlJS extracted from M is injected into S.
V, = voltage magnitude at nodc i The average power flow on line (i,j)is limitcd to tlie hie capacity Ti;
-T,ij
5 T,j
= -aij(O,
h3) 5
T,;,, (2)
Section 2. I describes tlic modclling of tlie phase sliiftcr, which is coiisidcred one of tlic classical devices uscd in power-system-security control calculations. The objcctive function of the economic dispatch is tlie total gciieration dispatch cost. Following the power pool operation of England and Walcs, the gcnerator cost ratc function is assumed to be convex pieccwise linear. It can be cxpresscd using scparable programming [ 101 as:
k=2
CX,,<l; 2 0 ; k = 2 , . . . , q x,r,
L=2
(3)
where pi/,
CIA
generalion axis
= cost rate of generation of generator i at
the kth break point C,(PGi) = cost rate of generation of gencrator i at level P(,/
4 ,
M
t,o d ( z ) 2 0 alld x 2 0 (5) where x E !)IL. firnctioiif: !Ji The is convex wlicreas thc component functions dj : !W + 91, i = I ... /n are concave. Moreover, f and rl are twicc continuously difrerentiablc. In the casc wherc,faiid d are affine, cqn. 5 rcduces to the wcll known symmetric linear program. Thereforc eqn. 5 is denoted a s the synimctric form of the convex program. We arc intercstcd in tlic standard form of eqn. 5, because it yields it sparsc symmclric positivc definite matrix for the solution of tlic search direction in the normal system approach. Symmetric positivc definitencss is the Iiighcst state a matrix can aspirc to be [12], since it permits economy in storagc and numerical stability in tlie solution of linear systems. The CED problem can be casily put into tlic symtnctric form of ecyn. 5. First, by choosing the reference anglc at tlie slack nodc to be sufficiently greater than zero (in practice between 1 and 10 rad), all tlic angles at thc other buses can be constrained to be positive. Sccondly, in accordancc with eqn. 2, the phase shiftcr power is constrained to be grcater than tlie negative of thc rating of line MN, see Fig. 2. This constraint is further transformed into a positivity constraint using a variable sliift. Moreover, in the CED problcm, tlie active power balance, eqn. I , is usually expressed as an equality. I T wc replacc it by thc inequality form, eqn. 6, tlie dispatch will not lead to oversatisfaction of the load, provided that lhc problcm is feasible and the cost curves are positive and convcx [2].
rninirnisc! f ( z )siihj
d, =
-
riDt
[-fl,7,7(&
- h,,)
jtk(2)
1 + -GL3(hi - hy)2 2
2. I Phase shifter model A phasc shiftcr (PS) is a devicc that is able to rcgulate voltage and current phase angles in ordcr to rc-route power
flow. Fig. I shows the oiic-line diagram of the phasc shifter where lc = lL# is a complex ratio. A direct consequence of this is that the bus admittance matrix ( Y bus) bccomes unsymmetrical. In order to make the Y bus symmetrical, Srinivasan et d.[I I] approximately simulated the phaseshifting elTect by installing additional injections at the buses. Our approach is dirferent. For a phase-shiftcr branch, we introduce a voltagc controllcd bus S (slave), as shown in Fig. 2. The control is done through thc master bus M according to the following constraint:
240
20 (6) Since (A; S)TV2dj(A; = .2jEk(j)~j,(dj 15) -~ 5 0 for every (A; 15)# 0, thc firiictions d, given by eyn. 6 are coiicave and thcreforc satisfy the requirement of the convex program. Hcre, and in tlic rollowing Sections, wc use the notation (A; 6) to denote a vector that is tlie concatenation of the vectors A and S. A bcneficial property or coiivex programming is that the Kuhii--Tucker conditions are sufficient for a point to be a global optiinum [13].
~
qz, ) = f(.) z
IEE Im-Geiio..
Ihii,si7i.
z%(z)
(7)
where z is a vector of Lagrangc multipliers. The KuhnTucker conditions for optiinality arc
y and
V : , L ( Z , %= s )
d(z) = v
rL) = s
rp = v
r Z 7d(x/r)
s7'x = 0;
/, %I = 2)
x ,s , U , z 2 0 (8) The conditions in eqn. 8 arc cquivaletit to a iiionotonc coinplirnentarity problem (M CP) [5]. Introd ticing tlic homogenisation variablc, we obtain the homogenised nionotone coinplimentarity pro blcm (H M C P): Miniiiiisc suhjcct to
(IC;v; ) " ' ( s ;z ; K ) T
(13) Thc capital letters, e.g. X , S, V, 2, denotc a diagonal matrix with the elements of the small lettcr vector, e.g. x, s, v, z, on the diagonal. If we use eqn. 11 to eliminate (b; AV; AK) in eqn. 10 wc obtain the following reduced system of equations:
D,
v:j
rV,L(x/r, z / r ) = s
r d ( x / T ) = 'U
-
-?]'r/)
,,
-
. 7
TLLC
Z ~ ~ V : , . L ( Zz Tr ) - z / d ( x / r ) = t; // .
(9) HMCP is (asymptotically) feasible and cvcry (asymptotically) feasible point is an (asymptotically) complimentary solution [9]. HMCP also has a maximal complimentary solution. Let (x*, v*, z*, .t*,K") be a maxiinal complimens', tary solution to the HMCP. Eqn. 9, then the following holds [5, 91: The MCP eqn. 8 has a solution if; and only if, 7' > 0. I n this case, (x"/lz*, s"ld, v*/+', z'ld) is aconiplimcntary solution for the MCP. The MCP cyn. 8 is infeasible if and only if K* > 0.
2 ,s, ,z , r , K, v
20
where
1 iirc:
~ l r -U+ p
- K,
Z-'y//.c
+ T-~yp
(14)
D , = H + X - ' S ; D? = 2 ' V ,
U , = H,q+rpl/i (15)
Under suitable assumptions 191, the HIP algorithm can be proved to generate a maximal complimentary feasible solution for the HMCP. Therefore we can use this maximal solution to get the inlbrniation needed about tlie MCP.
Note that the dimclision of thc Newton equations systcm solved by tlie HIP algorithm cqn. 14 is slightly larger than the corresponding system solved by the PDIMPC algorithm. In fact the dimension is increased by exactly one. The method can be implemented so that the same kictorisation as in tlie PDiMPC algorithm is computed in each iteration. However, the factorisation needs to be used in onc more solve to obtain the solution of the Ncwlon equations systcni. Details are given bclow. Let
4. I
Search direction
Assume an initial point (x, s, v , z, z, K) > 0 is given. Thcn the search direction employed by thc algorithm is defined by the linear cquations (eqns. 10 and I 1 arc known a s tlic Newton equations system): where ( ; E"), and E.(, arc defined by tlie right-hand side of F , eqn. 14. Under the assumption that K), is non-singular, we can define (pl;p 2 ) and ( q I ;(12) as the unique solutions to:
K,l(P, ;p.L)= (-vu1 ; VJ)
(18)
and K n (41 ; q'L) = ( + I > ; (19) then AT and (Ax; Az) can be deduced from cqn. 17 as: where and
Ar =
PG
( v 2 ;--'U#
(4,; ) q2
U,
('u2;
-W:dT(P, ; P a )
(20)
( A x ; a z )= (q,;4,) - ( P , ; P a ) A r (21) The rest of the search direction (As; AV; Ak) is obtained from eqn. I I .
4.2 Solving the Newton equations The solution of the Newton equations system is the major computational burdcn of the IP method. As seen in Section 4.1, this system reduces to eqns. 18 and 19. Our initial assumption of not incltiding equality constraints (D, is diagonal positive definite) allows u s to solve a sparse positive definite systcm for the search direction. We describe
24 I
our approach herc. Sincc the solution procedurc for eqns. 18 and 19 is idcntical, without loss of gencrality wc restrict our attention to e c p 19. From the second row in eqn. 19 we can obtain q2 as
(22)
and by substituting in the first row of eqn. 19:
-
( D , + s''V-1zB) q1 = P r ,
B ' V ZPp
4.4
'
(23) Since eqn. 5 is a convex program, H is symmetric positive semi-definite aiid K = (0, BTV '2%) is symmetric posi+ tive definite. Therefore 'Cholesky-infinity' decomposition [14] can be used to factor K . Minimum degree ordering is performed prior to decomposition in order to reduce fill in the Cholesky factor and speed up the process. ql is obtained by performing two triangular solves.
An important issue in a primal-dual interior point method is thc choice of the centring parameter y. Mehrotra [15] suggests varying y dynamically during the iterations. Owing to the success of this approach in linear programming, we use a variant of it in our solution program. We actually use the following rule: let
(zi
Np
pa =
' +
ND
Az )
I '
+m + 1
(30) where (AxL', AS", AV., Ai?, A f , AIP) is the affine-scaling direction obtained by setting y = 0 and 11 = 1 in eqns. 10 and 1 1. a: and aK+ step sizes of AZ" and AK" according are to eqri 28. Then choose
"/
/d2
rniri{aZ~,cv,,a,~a,}
(24)
where
q=1--/ (32) The above choice in eqn. 32 keeps y 5 0.1 and y~ 2 0.9, which leads to a large reduction in primal and dual infeasibility when a long step is taken. We also inake use of second-order correction for the complimentarity condition
L = T ; , u , z , S , 7 ,ti: (25) we then check if it is beneficial to take different step sizes according to the following rule: If ap > an. then if sTAx + z"Av 5 0, we take different step sizes; If a,, aD,then if'x7hs + vTAz 5 0, we take different step sizes. Otherwise, we use the same step size for the primal and dual updates. If we assume that diffcreiit step sizes are taken, then for the primal update: x+ = x a p a x
[U.
4.5 Stopping criteria
The prograin is terminated in one of two states: Problem (cqn. 31) is near infeasible or unbounded if
p/po <
(33) An optimal (approximate) solutioii (x*/T*,.?/+*:, v*/7*, Z"/ is obtained if the relative prinial infeasibility cafi.u;s), d u d infeasibility (&m) and optiiiiality (opt) are less than an exit tolerance (to[ = I x
i?)
U '
= U +npAu
+ apAr
(as)
(35)
The stopping criteria in eqn. 34 are equivalent to those used in conventional PD/MPC methods.
242
Bus
Generators 5 10 6 7 54 54 54 69
Lines 20
Phase
Cost segme'lts
Without PS variables 28 37 47 77 279 260 260 506 constraints 59 110 118 224 544 835 863 1191
With PS
~-
3
5 4 17 9
0
15 14 18 21 162 86 86 207
38
41 80 186 303 317 41 1
905 1344
14 51
288 608
= e , so = e , v" = e , z0 = e ,
7"
= 1 and
K"
=I
HOW
(36)
where e is a vector of ones of appropriate size.
I I I I
F F F
I
I
8
21 14 13 13
8
14
8
Numerical results
I
F F
-
The proposed method was prograninled in MATLAB running on a Pentium I1 400MHz PC with 128 Mbyte RAM. The testing was carried out on the IEEE 14, 24, 30, 57, 11 8, 300 bus test systems and on a 175 bus network. The problem statistics are summarised in Table 1. Using the HIP algorithm, each of the systems was dispdtched at ten different load lcvels giving a total of 150 tcst cases. Table 2 shows typical computational results (ITN = number of iterations, TIME = CPU time in seconds) with and without phase shifters (PS) installed in the network. The last column is the ratio of the dispatch cost using PS to that without tlicni. The fact that the cost ratio 5 1 validates the results, since the PS increase the feasible region of the problem [17]. However, it is important to note that the PS are not expected to have a significant impact on decreasing tlie dispatch cost. Their importance is for security control. The flow limits on some lines were reduced to make the dispatch effectively constrained by network limits. For instance, the 300-bus system has 14 flows binding in the base case and with phase shifters installed the number becomes 31. Table 2 also shows that the largest problem was solved within less than 3 seconds. Since MATLAB runs as an interpretcr (it compiles and then executes each command, one at a time), computational savings can be further anticipated by programming the code in C++. To simulate the presence of infeasibility, each of the systems
Table 2:CED computational results
~ _ _ _ _ _
15
-
1 '3
15
13 20
1.10 2.58
F
F
-~
With PS ITN 9 10 14 9 TIME 0.22 0.22 0.38 0.28 0.54 __ 1.15 2.70 HOW
Without PS
Bus
CG ITN TIME 0.16 0.16 0.33 0.27 0.55 0.99 0.88 1.64 HOW 1 4 1 5 54 6 6 56 7
I I I I I
-
8
13 9
I
I
I
8 13
12 14
I F F I
8
-
14 21
F F
ITN
7 8
15 6 13 13 13 18
15
-
1.04 2.80
0.9998 0.9963
13 22
was also dispatched with all single line contingencies that do not cause network splitting, and all single generator contingencies giving a total of 2745 cases. Tables 3 and 4 show the cffect of a selected sample of l i t e and generator contingencies on the dispatch solution. The selection of contingencies was done, such that some of the infeasible operating conditions are presented. In these Tables, CL represents the number of the contingent line and CG represents the number of the contingent generator. The solution state is shown undcr HOW, thus I indicates that the solution is infeasible, and F that the solution is feasible. The results demonstrate that the HIP algorithm is capable of detecting infeasible problem el'ficiently. Moreover, the validity of the optiniisation outcome is again shown by the ability of the phase shifters to restore solution feasibility following a line contingency, e.g. 57 and 300 bus test systems, or a generator contingency, e.g. 300 bus test system.
241
Conclusions
In this paper, wc have formulated the economic dispatch with a second-ordcr network model, network constraints and phasc shifters as a convex program. Thc HIP algorithm is proposcd to solve this problem. Compared to the PDiMPC algorithm, the HIP algorithm requires one additional solve per iteration in order to obtain a solution of the Newton systcin. At the expense of this, the HIP algorithm can always start from a simple starting solution, either to find an approximate global optimum solution or to detect if the problem is either infxsible or unbounded. I n addition, it docs not suffer froin the disadvantages of classical SLP inethods. Tcsting on the various systems shows that the method is efficient.
Acknowledgment
R.A. Jabr would like to thank the committee of vice-chancellors and principals of the univcrsities of the U K (CVCP) for thc overseas rcsearch sludcnt (ORS) award.
8
1
References
BERRY, P.E., and DUNNETT, R.M.: Contingency constrained cconomic dispatch algorithm for transmission planning, / L E Proc.. C, G(.r?er.T/.cu?.cm. /li.stri/i., 1989, 136, (4), pp. 238-224 2 IRVING, M.R., and STERLING, M.J.11.: Economic dispatch of active power by quadratic programming using a sparsc linear compleE/iqq,y Sjj.st., 1985, 7, (I), pp. 2-6 mentary algorithm, E M r . I(JIIW 3 AOICI, K., and SATOH, T.: Economic dispatch with nclwork scctirity constraints using parametric quadratic programming, T/O/?.S. PO\lJel A/)/)(/!.SjJJt., 1982, 101, (12), pp. 4548-4556 4 XU, X., HUNG, P., and YE, Y.: A simplified homogcncous and sclldual liticar programming algorithm and its iiiiplcmcntation, Awi. Ope/.. Res., 1996, 62, pp. 151-171
5 ANDERSEN, ED., and YE, Y.: A conipiilational study or the homogcncous algorithm for large scale coiivcx programming, C b m pur. Olprinz. A/>/)/.,1998, 10, pp. 243 269 6 YAN, X., and QUINTANA, V.H.: Improving a n interior-pointbased OPF by dynamic adjustments ol step sizes and tolerances, IEEE 7 m P ~ I ISysr.,I 1999, 14, ( ) pp. 709-716 ~ ~~ . 2, 7 VARGAS, L., QUINTANA, V., and VANNELLI, A.: A tutorial description of an interior point method and its applications to sccurity constrained ccoiioniic dispatch, IEIX 7i.rnr.s. P~JWY Sysr., 1993, 8, (3), p p 1315 1324 8 WEI, 13,. SASAKI, H., KUBAKAWA, J., and YOKOYAMA, It.: An interior point non-linear progrmitning for optimal power Ilow problcms with a novel data structure, I 1 TWt/.S.IOlLW sJ,Sr., 1998, 13, ( 3 ) , pp. 870-877 9 ANDEIZSEN, E.D., and YE, Y.: On a homogeneous algorithm Tor the monotone complimentarity problcm, M i i t / / . Progrcitn., 1999, 84, pp. 375-309 10 UAZARAA, M.S., and SHETIY, C.M.: Nonlinear programmingtheory and algorithms (J. Wiley & Son.;, Ncw York, 1979) I 1 SKINIVASAN, N., PRAKAESA, 1C.S.. a d VENKATA, S.S.: 011or phasc shilter distribution factors and line load 7i.cu/.s. I(JIWT Syst., 1985, 104, (7), pp. I656 Apptn.. I662 12 HLGHAM, N.J.: Computing 21 ncarcst synimclric positive scmidclinile matrix, Lirirrrr. Aljic/i/,. App/., 1988, 103, pp. 103-1 18 13 FLETCHEI1, I<.: Practical methods 01 optimisation (J. Wiley & Sons, New York, 1987) 14 Matlab iiscrs guidc (The Mathworks, Inc., Natick, Massachtiselts, 1996) 15 LUSTIG, 1.J.. MARSTEN, R.E., and SHANNO, D.F.: On implementing Melirotms prcdictor-corrector intcrior-point method Cor lincar programming, SIAM -1. Cotzfrd Optbn., 435 449, 1992, (2), pp, 3 16 GAY, D.M., OVEIZTON, M.L., and WRIGHT, M.H.: A primaldual interior method for ~ionconvcxnonlinear programming. Tcclincal Report 97-4-08, Computing Sciences Ilcscarcli Center, Bell Laboratories, Murray Hill, New Jersey 07974, 1997 17 GALIANA, F.D., ALMEIDA, K., TOUSSAINT, M., GRIITIN, J., ATANACK-OVIC, D., oor, B.T., MCGTLLIS,D.T.: ~ s s c s s mcnt and control of the impact of facts dcviccs on power system performancc, [EEL Trcim Power Sy.st., 1996, I I , (4), pp. 1931-1936
244