Renewable Energy xxx (2011) 1e8 Contents lists available at ScienceDirect Renewable Energy journal homepage: www.elsevier.com/locate/renene A fast algorithm based on the submodular property for optimization of wind turbine positioning Zhang Changshui a, Hou Guangdong a, *, Wang Jun b a b Department of Automation, Tsinghua University, FIT Building 3-120, 100084 Beijing, China Department of Control Science and Engineering, Tongji University, 201804 Shanghai, China a r t i c l e i n f o a b s t r a c t Article history: Received 29 December 2010 Accepted 31 March 2011 Available online xxx In the design of a wind farm, the placement of turbines is an important factor that affects the efficiency and profit, but automatic placement of turbines is still a challenging problem. This study reveals the “submodular” property of the wind turbine positioning problem based on Jensen wake model. Based on this property, a “lazy greedy” algorithm is used to optimize the placement. This method can obtain solutions with theoretical guarantee of quality. It can also estimate the lower bound of the optimal value of the objective function. This method is tested on three types of wind scenarios. Compared to previous research, this algorithm takes much less time, and always gains a better solution. To enlarge the application scope, the wake model is extended to the large scale complex terrain in this study. The present algorithm and some other algorithms are tested in the simulation of the complex terrain. The experimental results demonstrate the present method’s superior performance. 2011 Elsevier Ltd. All rights reserved. Keywords: Wind turbine positioning Wake model Submodular Optimization Genetic algorithm 1. Introduction Due to the depletion of fossil fuel resources and the impact of environmental concerns, attention has increasingly focussed on renewable energies, among which wind is one of the fastest developing energies. In order to extract more power while using wind energy, many factors should be taken into account, such as the type of wind energy conversion systems, the location of the farm, the local wind condition. When these factors are determined, the individual position of turbines in the wind farm becomes significant, mainly because the turbines will change wind condition and influence each other. That also makes wind turbine positioning a challenging task in wind farm design. There have been several studies on the optimization of wind turbine positioning [1e4]. Mosetti et al. designed a genetic algorithm for optimizing placement of turbines [1]. The objective function contained the investment cost and the total power produced by the wind farm. And their solutions were compared with configurations of randomly placed turbines. Though the solutions given by Mosetti et al. were proved better than random placement, they did not outperform the empirical placement. Grady et al. also used a genetic algorithm to demonstrate its effectiveness in turbine positioning [2]. * Corresponding author. Tel.: þ86 10 62796872. E-mail addresses: zcs@mail.tsinghua.edu.cn (Z. Changshui), hougd05@gmail. com (H. Guangdong), junwang@tongji.edu.cn (W. Jun). They run the algorithm with more individuals for sufficient generations and obtained improved solutions. Those works mentioned above are based on the genetic algorithm. Generic algorithms always take a lot of time to get a solution, and this will be an obstacle when the wind condition is complicated. Besides, there is no guarantee about the quality of the solution. That means there is no knowledge about how good the solution is. And in those works, only the flat terrain is considered. While in many real-world instances, wind farms are built in mountains or some other types of terrain, not always in flat area. In this work, a completely different method based on the conception named “submodular” is used. It takes much less time and always gains a better solution. And both the wake model and this turbine positioning approach are extended to the complex terrain. 2. Wake and cost model In both Mosetti and S.A. Grady’s work [1,2], the wake model used for wind farm calculation is similar to the model developed by Jensen et al. [5e7]. The model will be introduced in this section. It is based on the assumption that the momentum is conserved in the wake. For one single turbine, the downstream wake area will be regarded as a trapezoid. It means that the radius of the area increases at a rate proportional to the distance from the turbine. As shown in Fig. 1: 0960-1481/$ e see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.renene.2011.03.045 Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 2 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 3. The algorithm for flat terrain The mean wind speed obeys the formulation as follows: 2 3 6 u ¼ u0 6 41 7 2a 7 x2 2 5 1þa 3.1. Introduction of submodular property (1) r where a is the axial induction factor, x is the distance from the turbine, a is the entrainment constant, r is the downstream rotor radius. The relationship between r and the turbine radius rr is shown by equation (2): rffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1a r ¼ rr 1 2a (2) The axial induction factor can be calculated by the turbine thrust coefficient CT using the relationship CT ¼ 4a(1 a). The entrainment constant a, is empirically given as: a ¼ 0:5 z ln z0 (3) where z is the hub height of the wind turbine, and z0 is the surface roughness. Assuming that in multiple wakes, the kinetic energy loss is equal to the sum of the energy deficits. So for the downstream of N turbines, the velocity downstream can be expressed by the following equation: vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3 u N uX u 25 1 ui ¼ u0 41 t u0 Submodular is a useful property of some set functions. And it can be proved that when using Jensen’s model, the function which describes the extracted power P from certain placement of turbines is submodular. So the algorithm to find the maximum of a submodular function can be designed for turbine positioning. Definition. Assume that V ¼ {1,2,.,N} is a finite set, F: 2V/R is a function defined on the set V. If F meets the following condition: for all A; B4V, FðAÞ þ FðBÞ FðAWBÞ þ FðAXBÞ, then F is called submodular. And there is an equal definition: A set function F on V is called submodular if for all A4B4V; s;B, it holds that FðAWfsgÞ FðAÞ FðBWfsgÞ FðBÞ. The following is a simple example of a submodular function: V is the set of positions where circles can be placed. A, B are two subsets of V, F(A) is the total area covered by all the circles in A, as shown by Fig. 2. A, B are subsets of V, and A4B, it is obvious that for all s;B, it holds that FðAWfsgÞ FðAÞ FðBWfsgÞ FðBÞ, so F is submodular. In next subsection, the proof is presented that the turbine positioning problem based on current wake model is in nature a problem of maximizing a submodular function. 2 (4) 3.2. The submodular property of the turbine positioning problem i¼1 The total power P extracted from the wind is a function of local wind speed, as shown in the following expression: P ¼ N X 0:3 u3i (5) i The investment cost of a wind farm is assumed to be determined only by the number of turbines. The total cost per year for the entire wind park can be expressed as follows [1]: cost ¼ N 2 1 0:00174 N2 þ e 3 3 (6) PðAWfsgÞ PðAÞ ¼ The following objective function will be optimized: Objective ¼ cost Ptotal In the turbine positioning problem, the available terrain can be subdivided into cells. To keep necessary distance between two adjacent turbines, the size of a cell is suitably chosen and every turbine is installed only on the center of a cell. So the available positions are finite, V denotes the set of all these positions, thus one possible solution A is a subset of V. F(A) is the total power of A extracted for a period of time. There is the proof that F(A) is submodular. Assume A and B are two solutions, B contains all the positions in A. s is an available position which neither in A nor B. Adding s to A, the total power increment is: jAWfsgj X i¼1 (7) Minimizing this objective function leads to a solution with lowest cost of per unit of wind energy production. ¼ Ps0 in Pi0 A þ jAj X Pi i¼1 jAj X i¼1 Pi0 in A Pi in (8) A Adding s to solution B, the total power increment is: PðBWfsgÞ PðBÞ ¼ jBWfsgj X i¼1 þ jAj X i¼1 Ps0 in A Pi0 Pi0 jBj X i¼1 in B Pi ¼ Ps0 in Pi B in B þ jBj X i ¼ jAjþ1 Pi0 Pi ð9Þ Ps0 in B where and are respectively the power extracted by s PjAj after added in A and B. It can be proved that i ¼ 1 ðPi0 in A Pi in B Þ PjAj 0 ðP Pi in B Þ, so the following inequality is obtained: i ¼ 1 i in A PðBWfsgÞ PðBÞ ½PðAWfsgÞ PðAÞ jBj X i ¼ jAjþ1 Fig. 1. The wake model. Pi0 Pi þ Ps0 in B Ps0 in A (10) After a new turbine s placed, the total power of an original turbine decreases, if this turbine is in an area influenced by the Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 3 While using this algorithm to solve turbine positioning problem in present study, it always leads to much better solutions than 63% P (Aoptimal), and even better than genetic algorithm solutions. Fig. 2. A simple example of a submodular function. wake of s, otherwise its extracted power does not change. So Pi0 Pi , PjBj and i ¼ jAjþ1 ðPi0 Pi Þ 0 Let uA,uB be the wind speeds of s in A,B solutions, u0 is the wind speed of the same position before s added. According to the multiple wakes model, for A: uA u0 1 2 ¼ jAj X 1 i¼1 ui u0 2 IðiÞ ¼ SA (11) for B: 1 uB u0 and 2 IðiÞ ¼ ¼ jBj X j¼1 1 uj u0 2 IðjÞ ¼ SA þ jBj X j ¼ jAjþ1 1 uj u0 2 IðjÞ (12) 1 ; s is in the wake area of i 0 ; s is out of the wake area of i PjBj For ð1 uj =u0 Þ2 IðjÞ 0 and P ¼ 0.3u3, so uA uB j ¼ jAjþ1 and Ps0 in A Ps0 in B , PðAWfsgÞ PðAÞ PðBWfsgÞ PðBÞ, it means that the total extracted power P is a submodular function defined on the set of turbine positions V. 3.3. Algorithms for the turbine positioning 3.3.1. A basic greedy algorithm for turbine positioning In last subsection, it is proved that P(A) is submodular, so the turbine positioning problem can be considered as the maximization of a submodular function. The maximization of a submodular function is NP-hard unfortunately, however there are efficient algorithms to get an approximate solution which has a guarantee of quality. A very simple greedy algorithm (Algorithm 1) is designed to solve this optimization problem. This algorithm can give a placement for turbines with specified number k to (locally) maximize the total extracted power. Algorithm 1. Greedy Algorithm For Turbine Positioning 1: 2: 3: 4: 5: 6: 3.3.3. The bound of the best solution Using submodular property, the bound of the optimal placement’s total power can be got. A rough bound is contained in equation 13, while a better bound can be gained by a data dependent method developed by Minoux [9]. This method starts with a random solution A, for cs˛V=A, gets ds ¼ PðAWfsgÞ PðAÞ. The results are sorted, in the order d1 d2 . dn, then Minoux P proved that PðAÞ þ ki¼ 1 di is a limit of the P (Aoptimal). 4. Experiment and results Initialize: A0 ¼ B; for all i such that 1 i k do si ¼ arg maxfPðAi1 WsÞ PðAi1 Þg s Ai ¼ Ai1 Wsi end for return Ak 4.1. Numerical procedure First, a square area is used as the farm and subdivided into 100 cells. In the center of one cell a turbine would be placed. The width of each cell is 200 m, which is five times the length of the rotor diameter For such a simple algorithm, Nemhauser et al. proved that it can provide a solution with a guarantee of quality [8]. They proved that for a monotone submodular function PðAÞ and PðBÞ ¼ 0, the greedy solution Agreedy with k elements has the property [8]: 1 maxjAjk PðAÞz63% P Aoptimal P Agreedy 1 e 3.3.2. The lazy greedy algorithm for turbine positioning In the above algorithm, i < j0Ai 4Aj . And according to submodular property, Ai 4Aj 0ds ðAi Þ ds ðAj Þ. This property can be used to speed up the basic greedy algorithm. Interestingly, Minoux developed an accelerated greedy algorithm which speeds up the greedy algorithm substantially [9], when the submodular property holds. Compared to the basic greedy algorithm, Minoux’s algorithm can find a solution with a guarantee of quality in much less time. The algorithm used in present study with name “lazy greedy” is derived from Minoux’s method with some slight modifications. Algorithm 2 shows its whole procedure. This “lazy greedy” algorithm starts with solution A0 ¼ B, adds elements in this solution one by one, like Algorithm 1 does. The difference is: Algorithm 2 sorts the increments ds ¼ PðAWsÞ PðAÞ of all elements s in descending order to create a list at first, then updates this list in every iteration and according to this list, selects elements to add to the solution. This procedure helps to save a lot of running time. Using submodular property, the values in this list do not increase during the whole procedure, so it is easy to update this list. In each iteration, the values of ds are tested from the top of the list, if it decreases, its position will be updated in the list, otherwise the corresponding element will be removed to the solution set then one iteration is completed. Most times, only several values need to be updated in each iteration, while the basic greedy algorithm actually calculates all the values in this list in one iteration. That makes the “lazy greedy” algorithm much faster than basic greedy algorithm, especially when the turbines number is large. It is noted that in “lazy greedy” algorithm, the power increment brought by a newly added turbine decreases as the turbines number increases. When the increment in one iteration is less than 60% of that the fist added one provided, the algorithm stops. Then it gets a series of solutions: Ai, i ¼ 1,.,j. Among these solutions, the one with the max value of cost ðiÞ=PðiÞ will be chosen as the final solution. (13) Table 1 Wind turbine properties. Wind turbine properties Value Hub height (z) Rotor radius (rr) Thrust coefficient (CT) 60m 40m 0.88 Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 4 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 Case c): Multiple directions and variable speeds of 8, 12, and 17 m/s. The distribution is shown in Fig. 3 [2]. Wind Distribution in Case c) 7 17 m/s 12 m/s 8 m/s 6 Algorithm 2. “lazy greedy” Algorithm For Turbine Positioning Wind Frequency (%) 5 4 3 2 1 0 0 50 100 150 200 250 300 350 Angle(o) Fig. 3. The wind distribution in Case c) [2]. D in this numerical experiment. Other parameters of the turbines are shown in Table 1. The values are the same as those in [1] and [2]. Three types of wind scenarios will be considered. Case a): Uniform wind direction and a constant wind speed of 12 m/s; Case b): A constant mean wind speed of 12 m/s and 36 directions (every 10 ); Initialize: A0 ¼ B; 2: for all i such that 1 i N do di ¼ PðfigÞ 4: end for Sort fdi ; 1 i Ng in descending order to create a list L 6: Poriginal ¼ d1 j ¼ 0; * 8: while ds > 60% Poriginal do * Pick the first s in list L, calculate ds ¼ arg maxfPðAj WsÞ PðAj Þg * 10: if ds < ds in the list then ds ¼ d*s ; 12: Update the position of ds to keep the descending order of the list else 14: Ai ¼ Ai1 Ws j¼jþ1 16: Remove ds from the list. end if 18: end while return Agreedy ¼ arg maxfPðAk Þjk ¼ 1; .; jg Ak 4.2. Results and analysis Case a). Since “lazy greedy” algorithm picks the positions one by one, the total power and the objective (cost/power) value of every stage will be gained. The less objective value means the less cost Fig. 4. Optimal configurations and objective curve in Case a). Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 Table 2 Comparison of solutions in Case a). 5 Table 3 Comparison of solutions in Case b). Placement Turbine number Total power Efficiency Objective Placement Turbine number Total power Efficiency Objective Mosetti Grady Present Study 26 30 30 12352 14310 14310 91.645 92.015 92.015 16197 15436 15436 Mosetti Grady Present study Present study Present study 19 39 19 39 40 9245 17220 9354 17611 17991 93.86 85.17 94.97 87.11 86.76 17371 15666 17169 15318 15280 and the more extracted power. The “lazy greedy” algorithm takes the minimum point on the curve as the result. It is shown in Fig. 4d. Fig. 4a and b illustrate Mosetti et al. and Grady et al. solutions. Both of them use genetic algorithms. Fig. 4c is the optimal solution given by present study. Table 2 gives the comparison of solutions numerically. In this simple wind scenario, the turbines in one column cannot influence the ones in other columns. Grady pointed out that each 10-cell column can be viewed as a single independent optimized unit in Case a). The optimization for such a column can be extended to the whole considered domain. So the solution for the whole domain can be simply verified on a column by testing all the 210 ¼ 1024 placements. That shows Grady’s method and the present method both give the optimal solution. But the present method uses much less time. In this scenario, the turbines could be influenced only by the one in the same column. As the number of turbines in one column increases, the average power and the average cost of each turbine decrease. From Fig. 4d, it can be seen that when each column has 3 turbines, the objective function is optimal. Mosetti et al. did not find the best solution. Compared to the work by Grady et al., that may be caused by not sufficient individuals or generations. Case b). In this case, any two turbines influence each other. It makes hard to place the turbines by experience. Fig. 5 shows the solutions given by Mosetti et al., Grady et al. and present study. Table 3 is the comparison of the numerical results of these methods. When using the same number of turbines, the present work increases the total power by 1.19% than Mosetti et al. and 2.27% than Grady et al. Using 40 turbines, the objective can be further improved. Case c). In this case, the present study also gets good solutions. Fig. 6 shows the solutions given by Mosetti et al., Grady et al. and the present study. Table 4 is the comparison of the numerical results of these methods. The present method increases the total power by 6.39%, when using the same number of turbines with Mosetti et al. It also increases the total power by 4.73%, when using the same number 39 with Mosetti et al. The method chooses 40 turbines at final to get a better objective value. In these solutions, the density of placed turbines is high in the edge and low in the inner. This character is more obvious than Case b). Fig. 6f also illustrates the comparison of time cost between basic greedy and “lazy greedy” algorithm in Case c). It can be found that the latter performs much better in time cost. And generally speaking, even the basic greedy algorithm always costs less time than genetic algorithm in the same case. This is shown in next section. 5. Complex terrain 5.1. Modified model For many real-world instances, wind farms are not always built in flat area. Mountains or some other types of terrain are often used Fig. 5. Optimal configurations and objective curve in Case b). Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 6 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 Fig. 6. Optimal configurations and objective curve in Case b). for the rich wind resource. Some differences must be taken into account when placing turbines in these complex terrains. Firstly, even before the turbines placed, the wind conditions in different parts of a complex terrain are usually different, unlike in flat areas. In a complex terrain area, wind conditions may have opposite polarities in different parts. Some parts, like a ridge or a valley between two mountains, may have strong wind, while some may be opposite, like the lee side of a hill. Thus, the wind distribution for the whole terrain needs to be simulated before the evaluation of the algorithms. Secondly, in flat terrain, the same type of turbines has the same hub heights. So all the hubs of the turbines are in one plane. The influence of the turbine wake can be calculated simply in this plane, and the wake area of one single turbine is a trapezoidal in this plane as shown in Fig. 1. In complex terrain, the different positions make the hub heights different, and the shape of the influenced area is like a cone behind the turbine shown in Fig. 7. Besides this, the axis of the wake area (L0 in Figs. 1 and 7) is horizontal in flat area, while this property does not always hold in complex terrain. These differences require us to extend the Jensen wake model to the stereoscopic space. For this task, an assumption needs to be made at first. In large scale complex terrain, when the local slope does not change drastically, its tangent plane can be taken as the local terrain with very little loss of accuracy. Based on this assumption, the modified wake model shown in Fig. 7 is used. The wake area of one turbine will be regarded as a cone. The sectional radius increases at a rate proportional to the distance from the turbine. The wind speed decreases according to equation (3). The axis of this area has the same direction with the average of the wind vectors in the windward of the turbine. In Fig. 7, V0, V1 are the average vector of the wind in the windward of the turbine. L2 is the distance vector of these two turbines and L1 is the length of the projection perpendicularity to L0. It can be determined that whether the influence exits by comparing L1 with r1. 5.2. Result and analysis The modified wake model and the “lazy greedy” positioning algorithm are tested in simulated complex terrain, and compared with the genetic algorithm under the same situation. The terrain is shown in Fig. 8a and b. The wind distribution is illustrated by Fig. 8c. The simulation of the wind distribution is not very precise. That does not influence the research much, since the purpose here is evaluating the optimal algorithms but not getting accurate simulation result. Table 4 Comparison of solutions in Case c). Placement Turbine number Total power Objective Mosetti Grady Present study Present study Present study 15 39 15 39 40 13460 32038 14320 33553 34271 994.05 803.14 934.38 802.36 802.15 Fig. 7. The modified wake model. Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 7 Fig. 8. Solutions of “lazy greedy” and genetic algorithms. As shown in Fig. 8, two normal functions are used to simulate two mountains, with heights about 45 m and 15 m respectively. The area is 4 km 4 km, and is subdivided into 400 cells. The top is north and the input wind is 15 m/s Fig. 8b shows the direction of the input wind. Although the distribution of the wind power is more uneven in complex terrain, in this scenario, the total power is also submodular. Fig. 8d shows the result gained by the “lazy greedy” algorithm, the total power is 79584.8 kW, objective value is 3382. In this solution, turbines mainly distribute on the windward side of the ridge and the valley between two hills. Fig. 8e is the result of the genetic algorithm in the same situation. The two solutions are similar, while Table 5 gives the details of them which prove that the “lazy greedy”’s result is better. In Table 5, the genetic algorithm uses 40 individuals. When it stops, the total generation is about 2100. It is noted that the lazy Table 5 Comparison of characters of algorithms on complex terrain. Algorithm Greedy Lazy Greedy Genetic Algorithm Time Cost Total Power Cost/Total Power (Objective) 2 m 32 s 79584.8 3383 About 10 s 79584.8 3383 1 h 40 m 78850 3414 greedy algorithm has outstanding performance in time cost. Besides this the total power and objective are both better than traditional genetic algorithm. 6. Conclusion The present study reveals the submodular property of the turbine positioning problem based on Jenson model. Taking advantage of this property, a “lazy greedy” algorithm is used to handle the turbine positioning problem in much less time than some other popular methods. The solution gained by this algorithm has a theoretical guarantee. In fact, the experiments in present work on three scenarios show that this algorithm always finds a better solution than genetic algorithms. An effort is made to extend the wake model to complex terrain, which makes it possible to handle the more general cases. And for complex terrains, the more uneven wind distribution makes the greedy based algorithms more effective. Acknowledgments This work was supported by 973 Program (2009CB320602) and NSFC (Grant No. 61021063 and No. 61075064). Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045 8 Z. Changshui et al. / Renewable Energy xxx (2011) 1e8 References [1] Mosetti G, Poloni C, Diviacco B. Optimization of wind turbine positioning in large wind farms by means of a genetic algorithm. Journal of Wind Engineering and Industrial Aerodynamics 1994;51(1):105e16. [2] Grady SA, Hussaini MY, Abdullah MM. Placement of wind turbines using genetic algorithms. Renewable Energy (Elsevier) 2005;30:259e70. [3] Mora José Castro, Calero Barón JM, Riquelme Santos Jesús M, Payán Manuel Burgos. An evolutive algorithm for wind farm optimal design. Neurocomputing (Elsevier) 2007;70:2651e8. [4] Marmidis Grigorios, Lazarou Stavros, Pyrgioti Eleftheria. Optimal placement of wind turbines in a wind park using Monte Carlo simulation. Renewable Energy 2008;33:1455e60. [5] Jensen NO. A note of wind generator interaction. Roskilde, Denmark: RisØ National Laboratory; 1993. [6] Frandsen S. On the wind speed reduction in the center of large clusters of wind turbines. Amsterdam, The Netherlands: EWEC’91; 1992. pp. 375e380. [7] Katic I, Hojstrup J, Jensen NO. A simple model for cluster efficiency. In: Proceedings of the European wind energy association conference and exhibition; 1986. p. 407e10. [8] Nemhauser G, Wolsey L, Fisher M. An analysis of the approximations for maximizing submodular set functions. Mathematical Programming 1978;14: 265e94. [9] Minoux M. Accelerated greedy algorithms for maximizing submodular set functions. In: Stoer J, editor. Actes congres IFIP. Berlin: Springer Verlag; 1977. p. 234e43. Please cite this article in press as: Changshui Z, et al., A fast algorithm based on the submodular property for optimization of wind turbine positioning, Renewable Energy (2011), doi:10.1016/j.renene.2011.03.045