* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download ASU_OptimalPowerFlow
Survey
Document related concepts
Transcript
1 A Relaxed AC Optimal Power Flow Model Based on a Taylor Series Hui Zhang, Student Member, IEEE, Vijay Vittal, Fellow, IEEE, Gerald T. Heydt, Life Fellow, IEEE, and Jaime Quintero, Member, IEEE School of Electrical, Computer and Energy Engineering, Arizona State University, Tempe, AZ 85281, USA {hui.zhang}, {vijay.vittal}, {heydt}, {jaime.quintero.1}@asu.edu Abstract—Model accuracy plays an important role in optimal power flow (OPF) studies. However, the traditional DC model used in OPF is potentially inaccurate in many cases. This paper presents a relaxed ACOPF model based on a Taylor series. The proposed model includes reactive power and off-nominal bus voltages. A mixed-integer linear programming (MILP) model is developed for modeling network losses and the possible relaxations of the MILP model are discussed. Conditions when the relaxations are exact are given and proved. Simulation results show that the proposed model is computationally effective and provides a better approximation for the original AC network. Index Terms— AC optimal power flow, convexification, loss modeling, piecewise linearization, relaxation, Taylor series. NOMENCLATURE ag bg bk bk0 cg ck gk k(l) L Pk PDd PGg PGgmax PGgmin PLk Qk QDd QGg QGgmax QGgmin QLk Skmax Vi ΔVi ΔVmax ΔVmin uk(l) δk θk θmax Quadratic cost coefficient of generator g Linear cost coefficient of generator g Series admittance of branch k, a negative value Shunt admittance of branch k, a positive value Fixed cost coefficient of generator g Investment cost of the branch k Conductance of branch k, a positive value The slope of the lth piecewise linear block Number of linear blocks Active power flow on branch k Active power demand of load d Active power generated by generator g Maximum active power output of generator g Minimum active power output of generator g Active power loss on branch k Reactive power flow on branch k Reactive power demand of load d Reactive power generated by generator g Maximum reactive power output of generator g Minimum reactive power output of generator g Reactive power loss on branch k MVA rating of branch k Bus voltage magnitude in p.u. at bus i Voltage magnitude deviation from 1 p.u. at bus i Upper bound on the voltage magnitude deviation Lower bound on the voltage magnitude deviation Binary variable for the lth linear block Binary variable for modeling |θk| Phase angle difference across branch k Maximum angle difference across a branch θk+, θk– Δθk(l) Nonnegative slack variables used to replace θk The lth linear block of θk I. INTRODUCTION The objective of the optimal power flow (OPF) problem is to minimize the total energy cost in the power grid subject to the system and resource constraints. These constraints include bus voltages magnitudes, angles, line flows as well as generator outputs. The AC model of the OPF problem (ACOPF) uses the nonlinear AC power flow equations in the constraints. Among the solution techniques for the ACOPF problems, Newton’s method, due to its fast convergence near the solution, was widely adopted in early literature [1]-[2]. However, Newton’s method has difficulties in handling inequality constraints, and the performance of Newton’s method also depends largely on the starting point. As algorithms developed, the interior point method based algorithms [3]-[5] have become the mainstream algorithms for solving ACOPF problems. Despite the development of nonlinear algorithms, obtaining a robust solution for large-scale ACOPF problems efficiently is still challenging. Therefore, for problems such as real time economic dispatch where the speed is a primary concern, a DC model is often used. The DC model of the OPF problem (DCOPF) is a linearized version of the ACOPF model [6]. It assumes fixed voltage magnitude at every bus as well as negligible reactive power and network losses. Thus, the original nonconvex ACOPF model can be relaxed to a quadratic programming (QP) model (assuming that generators have quadratic cost curves), which is convex and easier to solve. Due to the approximations made, the accuracy of the DC model may be poor in some cases [7]. Endeavors have been made recently to search for better approximations to the AC model. Some related work is summarized as follows. The core concepts of the work being done are convexification and relaxation. The purpose of formulating a convex model is to obtain the global optimal solution. During the convexification process, if certain conditions do not hold, then relaxations may be needed. A linear programing approximation to the AC power flow equations was presented in [8], where the cosine term in the power flow equations was piecewise linearized and other nonlinear terms are approximated by the Taylor series. In [9], the ACOPF problem was reformulated using a semi-definite programming (SDP) model and solved by the interior point method. A zero-duality SDP model based on the Lagrange dual of the ACOPF problem was proposed in [10]. For the zero-duality to hold, small 2 modifications to the original systems may be needed. Studies including branch flow model [11] and branch-and-bound algorithms [12]-[13] have been conducted recently to further explore the zero-duality feature. Compared to the DC model, the models presented in the above work provide better approximations to the ACOPF model. However, these models are still incomplete and require further investigations. For example, the zero-duality may not hold when certain constraints, such as line flow or lower bounds on reactive power generation, are enforced [13]-[14]. In addition, these models are complicated and may not be easily extended to other applications such as transmission expansion planning. This paper develops a relaxed OPF model based on a Taylor series. The proposed model provides a better approximation to the AC network by retaining reactive power, off-nominal bus voltage magnitudes as well as network losses. Contributions of this work mainly include the following: A relaxed ACOPF model with a mixed-integer linear programing (MILP) based loss model. Relaxations of the MILP-based loss model and proofs of the conditions when the relaxations are exact. The rest of this paper is organized as follows: Section II presents the relaxed ACOPF model. Section III investigates the loss model and its relaxations. Simulation results are presented in Section IV. Conclusions are drawn in Section V. II. THE RELAXED ACOPF MODEL The proposed relaxed ACOPF model is derived in this section. The derivation is based on the following assumptions: Similar to the standard ACOPF model, the proposed model takes the following form: min f (x) (1a) h(x) c (1b) g(x) b (1c) T (1d) . In the above model, bold face refers to vectors. The objective function (1a) is the summation of the quadratic cost functions of each generator and has the following form: :i (3a) QG Q 0.5QL QD d :i (3b) gi g bus i k ki k ki k ki Pkij Cg PGg ag PGg2 bg PGg cg . d i k ki d i Pkji rk + jxk bus j Qkji Qkij 0.5(PLk + jQLk) 0.5(PLk + jQLk) Figure 1. Modeling of network losses as bus fictitious demand The inequality constraints (1c) represent the apparent power flow limit on each branch: Pk2 Qk2 (Skmax )2 . (4) Notice that (4) is a set of second order cone constraints. This type of constraint is still convex and can be handled by linear solvers such as Gurobi [19]. However, if a solver requires the constraints to be strictly linear, a piecewise linearized version for (4) can also be derived. Neglecting the effects of off-nominal transformer turns ratios and phase shifters yield the full AC power flow through branch k as follows, Pk Vi 2 gk VV i j g k cos k bk sin k (2) The equality constraints (1b), which are further elaborated in (3), represent the active and reactive power balance equations at every bus. In the proposed model, the terms corresponding to network losses PLk and QLk are explicitly added to the nodal balance equations. Define bus i and bus j to be the “from” bus and the “to” bus of branch k. As shown in Fig. 1, the losses on the branch are split in half and attached to the (5a) (5b) Rewriting the bus voltage magnitude as, Vi 1 Vi . (6) Based on the assumptions, ΔVi is expected to be small. Substituting (6) into (5) and neglecting higher order terms, Pk 1 2Vi gk 1 Vi V j subject to d g gi 2 The angle difference across a branch is small so that sin(θk) ≈ θk and cos(θk) ≈ 1 can be applied. x PGg , QGg ,Vi ,i PG P 0.5PL PD Qk Vi bk bk 0 VV i j bk cos k g k sin k . All bus voltage magnitudes are close to 1.0 p.u. xmin x xmax two terminal buses. The expressions of PLk and QLk will be elaborated further in Section III. g k bkk (7a) Qk 1 2Vi bk bk 0 1 Vi V j bk gkk . (7b) Notice that (7) still contains nonlinearities. However, since ΔVi, ΔVj and θk are expected to be small, the products ΔViθk and ΔVjθk can be treated as second order terms and hence negligible. Therefore, the linearized power flow through branch k metered at bus i are obtained as follows, Pk Vi V j gk bkk Qk 1 2Vi bk 0 Vi V j bk gkk . (8a) (8b) The bounds on variables (1d) include upper and lower limits on bus voltage magnitudes and angles as well as the generator active and reactive outputs: Vi min Vi Vi max max k max (9a) (9b) 3 PGgmin PGg PGgmax QGg QG min g max g QG (9c) . θk 2 (9d) The complete relaxed ACOPF model is described by (2)(4) and (8)-(9). III. k(L) NETWORK LOSSES MODELING As shown in (3), the network losses can be included in the proposed model. This section derives the network losses PLk and QLk and investigates the possible relaxations. Applying the second order approximation of cosθk and neglecting high order terms, the active and reactive network losses can be approximated as, PLk gkk2 (10a) QLk bk . (10b) 2 k Notice that (10a) and (10b) are nonconvex. The following two approaches are developed to render them convex. A. Piecewise Linearized Relaxation The piecewise linearized model is to approximate θk2 by a series of linear blocks. Since (10a) and (10b) are nonconvex, certain line losses in the resultant linear model may fail to converge to the correct value and cause the “fictitious loss” problem as pointed out in [15] and [16]. The following MILP model introduces a series of binary variables to prevent the presence of the fictitious losses. L PLk g k k (l ) k (l ) θkmax k(1) Δθk(1) |θk| … Δθk(L) Figure 2. Piecewise linearization of θk2 The above MILP model eliminates the fictitious losses by adding the binary variables. These binary variables, however, could prevent the resultant model from being solved efficiently. In fact, the MILP model can be relaxed to a more efficient linear programming (LP) model by discarding the binary variables δk from (11h) and (11j) and retaining only (11a)-(11j). When certain conditions are met, this relaxation is exact. The condition is investigated in the following. First, let ℒ(x) be the Lagrangian function. (x) f (x) φT c h(x) μT b g(x) where, φ and μ represent the dual variables associated with the equality and inequality constraints respectively. Bounds on variables are converted to inequality constraints and included in g(x). The dual variables needed are given in (11a) to (11j). According to Lagrangian duality theory, the dual variables are non-positive for “≤” constraints, and are free of sign restriction for equality constraints. All terms are moved to the right-hand side of the constraints. The optimality condition requires the following constraints hold simultaneously, : k (11a) QLk bk k (l ) k (l ) :k (11b) PLk 0.5(i j ) k 0 k k : k (11c) QLk 0.5(i j ) k 0 l 1 L l 1 k L l 1 k (l ) k k : k k (l ) 0, l 1,..., L k (l ) max : L , l 1,..., L 0 k (11d) l k (11e) : l k (11f) :k (11g) (11h) k kkmax :k (11i) k 0 : k k (1 k )kmax : k (11j) k (l ) k (l 1), l 2,..., L (11k) max max L k (l 1) uk (l 1) L , l 2,..., L (11l) max k (l ) 1 uk (l 1) L l 2,..., L (11m) k (l ) (2l 1) max L . + – In (11c), θk is replaced by slack variables θk and θk . Binary variable δi together with (11g) to (11j) ensure the right-hand side of (11d) equals the absolute value of θk, i.e., at most one of θk+ and θk– can be nonzero. Constraints (11k)-(11m) guarantee that the linear block on the left will always be filled up first to eliminate fictitious losses. The piecewise linearization is illustrated in Fig. 2. (12a) (12b) (12c) k (l ) ( gk k bk k )k (l ) k 0 (12d) l k l k k k k k k 0 (12e) k k k k 0 (12f) k Theorem 1: If (gkγk – bkωk) > 0, then the MILP model and the LP model are equivalent. Proof: Without losing generality, consider Δθk(l) to be the lth linear block for θk where 1 < l < L. If 0 < Δθk(l) < θmax/L, then by complementary slackness (CS), (12d) becomes: ( gk k bk k )k (l ) k (13a) Let Δθk(l – m) and Δθk(l + n) be any linear block before and after Δθk(l) respectively, where 1 ≤ m < l and 1 ≤ n ≤ (L – l), then for these two linear blocks, (12d) gives: ( gk k bk k )k (l m) k kl m kl m ( gk k bk k )k (l n) k l n k l n k (13b) (13c) Provided the OPF model is feasible, (13a)-(13c) must have a solution. Since (gkγk – bkωk) > 0 and k(L) > k(l) > 0, by CS, kl m kl n 0 . This indicates that any linear block before 4 Corollary 1: For an OPF model that only considers active power constraints, if the sum of the locational marginal price (LMP) at two terminal buses of a branch is positive, then the LP relaxation is exact. 250 200 k k For the piecewise linearized model, Theorem 1 provides an approach to identify the branches where the fictitious losses may be created. Binary variables are needed only for these branches instead of all the branches in the system. Based on Theorem 1, for an OPF model that only considers the active power constraints, the following corollary can be derived. 150 kk Equation (13a) shows that βk > 0. By CS, it can be observed from (12e) and (12f) that if either θk+ or θk– is nonzero, then the other must be at its lower bound, i.e., zero. This indicates that θk is either equal to θk+ or –θk– depending on the sign of θk. Hence, Theorem 1 is proved. ■ can observed from the figure, (gkγk – bkωk) values of all the branches are positive except for branch 16, which is zero. According to Theorem 1, the LP relaxation should be exact and no fictitious losses should be created except for branch 16. The active power loss for each branch is plotted in Fig. 4 using (11a) and (10a) respectively. In the figure, fictitious losses are observed at branch 16. This is because with (g16γ16 – b16ω16) equals to zero, the piecewise linear model fails to select the correct linear blocks for branch 16 and causes excessive losses which will not be observed in reality. For other branches, since (gkγk – bkωk) > 0, the LP relaxation is exact and therefore no fictitious loss is created. 100 g r -b w Δθk(l) must be at its upper bound, i.e., Δθk(l – m) = θmax/L and any linear block after Δθk(l) must be at its lower bound, i.e., Δθk(l + n) = 0. Referring to Fig. 2, the linear blocks will be filled continuously starting with the leftmost one. Proof: LMPs are the dual variables associated with the active power nodal balance constraints, i.e., λi. If reactive power is neglected, then the term –bkωk will drop out from (12d). From (12b), it is clear that γk is positive. If gk > 0, then gkγk > 0 and according to Theorem 1, the LP relaxation is exact. If gk = 0, then the active power loss for that branch is always zero. Hence, Corollary 1 is proved. ■ 50 0 (MW) 4 (14) This inequality relaxation is exact when (14) is binding. The following condition need to be satisfied. (15) In (15), it is easy to show that if (λi + λj) > 0, then γk > 0. By CS, (14) must be binding. ■ IV. SIMULATION RESULTS The proposed model and its relaxations are evaluated in this section. All models are programmed in AMPL [18]. The solvers used are Gurobi 5.0.2 [19]. The computer used for simulation has an Intel E8500 CPU with 3.2 GB of RAM. First, a case is presented to verify Theorem 1. The test system is constructed based on the IEEE 24-bus RTS system [23] with modified generator cost data. The system has 38 branches. The OPF is solved using the relaxed LP model presented in Section III. The LMP is positive at every bus. The (gkγk – bkωk) value of each branch is plotted in Fig. 3. As one 30 35 Ploss by (10a) Fictitious losses 3 2 0 0 5 Proof: Let ℒ be the Lagrangian function and consider (14), then (12b) can be rewritten as: PLk 0.5(i j ) k 0 15 20 25 Line number 1 Theorem 2: If the sum of the LMPs at the two terminal buses of a branch is positive, then (14) is binding. 10 Ploss by (11a) loss : k 5 5 P PLk gk 0 Figure 3. (gkγk – bkωk) value of each branch B. Quadratic Inequality Relaxation If reactive power losses are neglected, i.e., (10b) is removed, then (10a) can be relaxed using the following inequality constraint, 2 k g16r16 - b16w16 = 0 10 15 20 25 Line number 30 35 Figure 4. Active power loss of each branch As shown in the Table I, the LP model creates 8.9 MW fictitious losses at branch 16, whereas the full MILP model and the reduced MILP model are free of fictitious losses. In terms of simulation time, the LP model is the fastest among the three models, which is solved in less than 0.1 s. The full MILP model is solved in 2.3 s and this time is reduced by 35% if the reduced MILP model is used. TABLE I. Loss model1 Full MILP Reduced MILP LP 1. 2. COMPARSION OF DIFFENENT LOSS MODELS Problem Type2 MISOCP MISOCP SOCP Fictitious losses (MW) 0 0 8.9 (at branch 16) Simulation time (s) 2.3 1.5 < 0.1 Full MILP model: binary variables are added to every branches. Reduced MILP model: binary variables are only added to branches with (gkγk – bkωk) = 0 MISOCP: Mixed-integer second order cone programming. SCOP: Second order cone programming 5 The proposed model and its relaxations are also applied to multiple test cases [20] and the results are reported in Table II. The number of linear blocks used is 40. It can be observed that for all test cases, the optimal solutions obtained from the proposed model provide good approximations to the full ACOPF solutions. The proposed model is also computationally efficient. All test cases are solved within 0.6 s. TABLE II. Test cases IEEE 14 IEEE 24 IEEE 39 IEEE 57 IEEE 118 [5] [6] [7] [8] ACCURACY OF THE PROPOSED MODELS Piecewise linearized loss model Gap† Time Objective (%) (s) 8067 0.2 0.06 63397 0.07 0.11 41876 0.03 0.17 41497 0.6 0.27 129612 0.04 0.55 Quadratic inequality loss model Gap† Time Objective (%) (s) 8094 0.3 0.06 63352 < 10–6 0.09 41861 < 0.01 0.16 41448 0.7 0.22 129471 0.1 0.52 † Gap is the percentage mismatch between the objective values given by the proposed models and the full ACOPFs solution using MATPOWER [20]. V. [9] [10] [11] [12] [13] CONCLUSIONS This paper develops a relaxed OPF model based on a Taylor series. The proposed model retains the reactive power, off-nominal bus voltage magnitudes as well as network losses. A MILP-based loss model is developed to eliminate the fictitious losses. Relaxations of the MILP model are investigated. It is proved that the relaxations are exact if the system meets certain specified conditions. Based on the results in this paper, the following conclusions are drawn: [14] [15] [16] In the piecewise linear model, the branches that may create fictitious losses can be identified. Binary variables are only needed for these branches instead of all the branches in the system. [17] In the piecewise linearization model, even if the all the LMPs are positive, the fictitious losses may still be present if the reactive power losses are considered. [19] If reactive power losses are neglected, then the quadratic inequality relaxation is exact for active power losses. Inclusion of reactive power and the off-nominal bus voltage magnitudes improves the model accuracy. The proposed model is computationally efficient and provides a better approximation to the full ACOPF model. REFERENCES [1] [2] [3] [4] H. W. Dommel, and W. F. Tinney, “Optimal power flow solutions,” IEEE Trans. Power App. Syst., vol. PAS-87, no. 10, pp. 1866–1876, Oct. 1968. D. I. Sun, B. Ashley, B. Brewer, A. Hughes, and W. F. Tinney, “Optimal power flow by Newton approach,” IEEE Trans. Power App. Syst., vol. PAS-103, no. 10, pp. 2864-2880, Oct. 1984. H. Wei, H. Sasaki, J. Kubokawa, and R. Yokoyama, “An interior point nonlinear programming for optimal power flow problems with a novel data structure,” IEEE Trans. Power Syst., vol. 13, no. 3, Aug. 1998. W. Yan, J. Yu, D. C. Yu, and K. Bhattarai, "A new optimal reactive power flow model in rectangular form and its solution by predictor corrector primal dual interior point method," IEEE Trans. Power Syst., vol. 21, no. 1, Feb. 2006. [18] [20] [21] [22] [23] H. Wang, C. E. Murillo-Sánchez, R. D. Zimmerman, and R. J. Thomas, “Computational issues of market-based optimal power flow,” IEEE Trans. Power Syst., vol. 22, no. 3, pp. 1185-1193, Aug. 2007. O. Alsac, J. Bright, M. Prais, and B. Stott, “Further developments in LP-based optimal power flow,” IEEE Trans. Power Syst., vol. 5, no. 3, pp. 697-711, Aug. 1990. B. Stott, J. Jardim, and O. Alsac, “DC power flow revisited,” IEEE Trans. Power Syst., vol. 24, no. 3, pp. 1290–1300, Aug. 2009. C. Coffrin, and P. V. Hentenryck, “A Linear-programming approximation of AC power flows,” June 2012, [Online]. Available: http://arxiv.org/abs/1206.3614. X. Bai, H. Wei, K. Fujisawa, and Y. Wang, “Semidefinite programming for optimal power flow problems,” Intl J. of Electrical Power & Energy Systems, vol. 30, no. 6-7, pp. 383–392, July-Sept. 2008. J. Lavaei, and S. Low, “Zero duality gap in optimal power flow,” IEEE Trans. Power Syst., vol. 27, no. 1, pp. 92–107, Feb. 2012. M. Farivar, and S. Low, “Branch flow model: relaxations and convexification,” Nov. 2012, [Online]. Available: http://arxiv.org/pdf/1204.4865.pdf. D. T. Phan, “Lagrangian duality and branch-and-bound algorithms for optimal power flow,” Operations Research, vol. 60, no. 2, pp. 275–285, March-April 2012. A. Gopalakrishnan, A. U. Raghunathan, D. Nikovski, and L. T. Biegler, "Global optimization of optimal power flow using a branch & bound algorithm,” Nov. 2012, [Online]. Available: http://numero.cheme.cmu.edu/uploads/globalOptBBOPF_v6(1).pdf. B. Lesieutre, D. Molzahn, A. Borden, and C. L. DeMarco, “Examining the limits of the application of semidefinite programming to power flow problems,” in Proc. of the Allerton Conf. on Comm., Ctrl. and Computing, Sept. 2011. P. Sánchez-Martín, and A. Ramos, “Modeling transmission ohmic losses in a stochastic bulk production cost model,” Oct. 1997, [Online]. Available: http://www.iit.upcomillas.es/aramos/papers/losses.pdf. R. Palma-Benhke, A. Philpott, A. Jofré, and M. Cortés-Carmona, “Modeling network constrained economic dispatch problems,” Optimization and Engineering, Springer, Oct. 2012. Z. Hu, H. Cheng, Z. Yan, and F. Li, “An iterative LMP calculation method considering loss distributions,” IEEE Trans. Power Syst., vol. 25, no. 3, pp. 1469-1477, Aug. 2010. R. Fourer, D. M. Gay, and B. W. Kernighan, AMPL: A Modeling Language for Mathematical Programming, 2nd edition, Florence, KY: Duxbury Press, 2002. Gurobi Optimization, “Gurobi optimizer reference manual,” [Online]. Available: http://www.gurobi.com. R. D. Zimmerman, C. E. Murillo-Sánchez, and R. J. Thomas, “Matpower: steady-state operations, planning and analysis tools for power systems research and education,” IEEE Trans. Power Syst., vol. 26, no. 1, pp. 12-19, Feb. 2011. D. G. Luenberger, and Y. Ye, Linear and Nonlinear Programming (3rd ed.), vol. 116, Springer, 2008. G. Andersson, Modeling and Analysis of Electric Power Systems, ETH Zurich, Sept. 2008. IEEE RTS Task Force, “The IEEE reliability test system-1996,” IEEE Trans. Power Syst., vol. 14, no. 3, pp. 1010-1020, Aug. 1999. Hui Zhang (S’09) received the B.E. degree from Hohai University, Nanjing, China, in 2008 and the M.S. degree from Arizona State University, Tempe, AZ, in 2010. He is currently pursuing the Ph.D. degree at Arizona State University, Tempe. Vijay Vittal (S’78–F’97) received the Ph.D. degree from Iowa State University, Ames, IA, in 1982. He is currently the Director of the Power Systems Engineering Research Center (PSERC). Gerald T. Heydt (S’62–M’64–SM’80–F’91–LF’08) received the Ph.D. degree in electrical engineering from Purdue University, West Lafayette, IN, in 1970. He is a Regents’ Professor at Arizona State University, Tempe. Jaime Quintero (M’06) received the Ph.D. degree in electrical engineering from Washington State University, Pullman, WA, in 2005. Currently, he is a postdoctoral researcher at Arizona State University, Tempe.