潮流不同排序方案的比較畢業(yè)論文文獻(xiàn)翻譯中英文對(duì)照

1、<p><b>　　中文4900字</b></p><p>　　A Comparison of Power Flow by Different Ordering Schemes</p><p>　　Abstract—Node ordering algorithms, aiming at keeping sparsity as far as possible,

2、 are widely used today. In such algorithms, their influence on the accuracy of the solution is neglected because it won’t make significant difference in normal systems. While, along with the development of modern power s

3、ystems, the problem will become more ill-conditioned and it is necessary to take the accuracy into count during node ordering. In this paper we intend to lay groundwork for the more rationality order</p><p>

4、　　Keywords—power flow calculation; node ordering; sparsity; accuracy; Newton-Raphson method ; linear equations</p><p>　　I. INTRODUCTION</p><p>　　Power flow is the most basic and important co

5、ncept in power system analysis and power flow calculation is the basis of power system planning, operation, scheduling and control [1].Mathematically speaking, power flow problem is to find a numerical solution of nonlin

6、ear equations. Newton method is the most commonly used to solve the problem and it involves repeated direct solutions of a system of linear equations. The solving efficiency and precision of the linear equations directly

7、 influences the</p><p>　　Jacobian matrix in power flow calculation, similar with the admittance matrix, has symmetrical structure and a high degree of sparsity. During the factorization procedure, nonzero en

8、tries can be generated in memory positions that correspond to zero entries in the starting Jacobian matrix. This action is referred to as fill-in. If the programming terms is used which processed and stores only nonzero

9、terms, the reduction of fill-in reflects a great reduction of memory requirement and the number of</p><p>　　After sparse matrix methods, sparse vector methods [5], which extend sparsity exploitation to vecto

10、rs, are useful for solving linear equations when the right-hand-side vector is sparse or a small number of elements in the unknown vector are wanted. To make full use of sparse vector methods advantage, it is necessary t

11、o enhance the sparsity of L-1by ordering nodes. This is equivalent to decreasing the length of the paths, but it might cause more fill-ins, greater complexity and expense. Counterin</p><p>　　Up to now, on th

12、e basis of the assumption that an arbitrary order of nodes does not adversely affect numerical accuracy, most node ordering algorithms take solving linear equations in a single iteration as research subject, aiming at th

13、e reduction of memory requirements and computing operations. Many matrices with a strong diagonal in network problems fulfill the above assumption, and ordering to conserve sparsity increased the accuracy of the solution

14、. Nevertheless, if there are junctions of ve</p><p>　　Based on the existing node ordering algorithm mentioned above, this paper focus attention on the contradiction between memory and accuracy during node or

15、dering, research how could node ordering algorithm affect the performance of power flow calculation, expecting to lay groundwork for the more rationality ordering algorithm. This paper is arranged as follows. The contrad

16、iction between memory and accuracy in node ordering algorithm is introduced in section II. Next a simple DC power flow is showe</p><p>　　CONTRADICTION BETWEEN MEMORY AND ACCURACY</p><p>　　IN

17、 NODE ORDERING ALGORITHM</p><p>　　According to numerical mathematics, complete pivoting is numerically preferable to partial pivoting for systems of liner algebraic equations by Gaussian Elimination Meth

18、od (GEM). Many mathematical papers [9-11] focus their attention on the discrimination between complete pivoting and partial pivoting in (GEM). Reference [9] shows how partial pivoting and complete pivoting affect the sen

19、sitivity of the LU factorization. Reference [10] proposes an effective and inexpensive test to recognize numeri</p><p>　　The node reordering algorithms guided by sparse matrix technology have wildly used in

20、power system calculation, aiming at minimizing memory requirement. In these algorithms, the nodes with fewer adjacent nodes tend to be numbered first. The result is that diagonal entries in node admittance matrix tend to

21、 be arranged from least to largest according to their module. Analogously, every diagonal submatrices relate to a node tend to be arranged from least to largest according to their determinants. </p><p>　　III

22、. DIFFERENCE PRECISION OF THE SOLUTION USING PARTICAL PIVOTING AND COMPLETE PIVOTING</p><p>　　It is said that complete pivoting is numerically preferable to partial pivoting for solving s

23、ystems of linear algebraic equations. When the system coefficients are varying widely, the accuracy of the solution would be affect by rounding errors hardly and it is necessary to take the influence of the ordering on t

24、he accuracy of the solution into consideration. </p><p>　　Fig.1 DC model of Sample 4-node network</p><p>　　As an example, consider the DC model of sample 4-node system shown in Figure 1. Node 1

25、 is the swing node having known voltage angle; nodes 2-4 are load nodes. Following the original node number, the DC power flow equation is:</p><p>　　To simulate computer numerical calculation operations, fou

26、r significant figures will be used to solve the problem. Executing GEM without pivoting on (1) yields the solution[ θ2,θ3,θ4]T=[-0.3036,-0.3239,-0.3249]T, whose components differ from that of the exact solution [θ2, θ3,θ

27、4]T=[-0.3,-0.32,-0.321]T. A more exact solution could be obtained by complete pivoting: [θ2,θ3, θ4]T=[-0.3007,-0.3207,-0.3217]T, and the order of the node after row and column interchanges is 3,2,4. So this is a more rea

28、so</p><p>　　IV. THE INFLUENCE OF NODE REODERING ON THE PERFORMANCE OF NEWTON-RAPHSON POWER FLOW METHOD</p><p>　　Fig.2 Sample 6-node network</p><p>　　On the basis of the

29、 above-mentioned analysis, the scheme for node reordering will not only affect memory requirement but also the accuracy of the solution in solving linear simultaneous equations. So performance of Newton-Raphson power flo

30、w method will be different with various node ordering. In this section three schemes of ordering for different purpose will be applied to a sample 6-node network shown in Fig 2 to compare the influence of them on the acc

31、uracy of the solution, the convergence ra</p><p>　　Puropse 1 Saving Memory as far as possible </p><p>　　At present, there are various schemes widely used for node numbering in near-optimal orde

32、r to reduce fill-ins and save memory. The only information needed by the schemes is a table describing the node-branch connection pattern of the networks. An order that would be optimal for the reduction of the admittanc

33、e matrix of the network is also optimal for the table of factors related Jacobian matrix. Different schemes reach different compromise between programming complexity and optimality. In this p</p><p><b>

34、;　　Scheme I </b></p><p>　　a) Number the node degree of which is one. If more than one node meet this criterion, number the node with the smallest original number. If there are not sucn nodes any more

35、, start with step b); </p><p>　　b) Number the node so that no equivalent branches will be introduced when this node is eliminated. If more than one node meets this criterion, number the one with the smalle

36、st original number. If we can not start with step a) or step b), turn to step c); </p><p>　　c) Number the node so that the fewest branches will be introduced when this node is eliminated. If not only node

37、could introduce fewest branches, number the one with the largest degree.</p><p>　　Once certain node is numbered in the step above, update the degree of relevant nodes and topological information. Until all t

38、he nodes are numbered, the process of node numbering ends up. </p><p>　　TABLE I. REORDERED NODES USING SCHEME ONE</p><p>　　Following the steps of scheme I, the sequence of the node numbere

39、d for the 6-node network is given in table I. No fill-in will be introduced during the procedure of solving the linear equation, so the table of factors and the Jacobian matrix will have completely identical structure. S

40、o the memory requirement for the table of factors is 0.256Kb, which is the same with that for the Jacobian matrix. Normally, an acceptable solution can be obtained in four or five iterations by Newton-Raphson method</

41、p><p>　　B. Puropse 2: Improving Accuracy Using Complete Pivoting </p><p>　　Considering that complete pivoting is numerically preferable to partial pivoting, in this section complete pivoting is

42、adopted to improve accuracy of the solution of the linear equations, aiming at reducing the number of iterations. Here nodes relate to large determinant of the diagonal submatrices intend to be arrange in front. To some

43、extern, the modulus of the entries on the main diagonal of the admittance matrix could indicate the magnitude of the determinant of the submatrices on the main d</p><p>　　Scheme II </p><p>　　a)

44、 Form the nodal admittance matrix; </p><p>　　b) Factorize the nodal admittance matrix with complete pivoting. Record the changes on the position of the nodes; </p><p>　　c) Determine the new

45、 number of the node according to the positong of node in the end of the factorization; </p><p>　　TABLE II. REORDERED NODES USING SCHEME TWO</p><p>　　Executing scheme II, complete pivoting

46、might automatic performed without row and column exchanges. The module of entries on main diagonal corresponding to such node may become larger by summing more branch parameter, as a result, the nodes, degree of which is

47、 larger, tend to be numbered first. So the results of such scheme may depart form the principle of node numbering guided by sparse matrix methods and many fill-ins might be introduced. The sequence of the node numbered f

48、or 6-node network is </p><p>　　C. Puropse 3: Improving Accuracy while preserving the sparsity </p><p>　　Only one small impedance branch exists in the system, so only four entries (submatrices)

49、 corresponding to node 4 and node 6 are very large in admittance matrix (Jacobin matrix). During the process of forward substitution, once node 4 or node 6 is elimination, the submatrix comprised of rest elements could k

50、eep good numerical stability and numbering of rest nodes would not make a difference to the accuracy of the solution. To take both accuracy and sparsity into account, we numbered node 4 first,</p><p>　　Since

51、 only one small impedance branch exists in the system and it connects to node 4, the degree of which is one. Scheme III will meet the request of purpose 1. So the number of fill-ins, memory requirements and operations ne

52、eded for factorization are all the same with scheme I. Only nine iterations will be needed to insure the convergence, result in a large save of calculation (only 2107 multiply operations). The reduction on the number of

53、iterations indicates that more exact solutions for the </p><p>　　? The diagonal element related to node 4 is just a little smaller than the one related to node 6, so eliminate node 4 first will not decrease

54、 accuracy. The scheme could meet complete pivoting approximately. </p><p>　　? Fewer operations in scheme III reduce the rounding error of calculator floating-point numbers. Especially, if eliminate node 6 f

55、irst, very small value might be added to diagonal element of node 2 and node 5, which would cause serious rounding error. While, if eliminate node 4 first, a sizable value will be added to diagonal element of node 6, pro

56、ducing a value in the normal range. </p><p>　　TABLE III. REORDERED NODES USING SCHEME THREE</p><p>　　TABLE IV. PERFORMACNE OF NEWTON POWER FLOW USING DIFFERENT SCHMEMS OF NODE OR

57、DERING</p><p>　　V. CONCLUSION</p><p>　　Theoretical analysis and the result of numerical calculating suggest that it is necessary to consider the influence of node ordering on the accuracy of t

58、he power flow calculation. If the node ordering algorithm takes both memory and accuracy into account reasonably, the performance of power flow calculation could be further improved. Elementary conclusions of this paper

59、are as follows:</p><p>　　For the well-conditioning power system, the influence of node ordering on the accuracy of power flow calculation could be neglect. It is more important to focus our attention on keep

60、ing the sparsity to save memory requirement and compute operations.</p><p>　　For the ill-conditioning power system, the accuracy must be considered in node ordering algorithm to speed up the convergence rate

61、. On this basis, if the sparsity is considered meanwhile, more accuracy might be obtained because of the reduction of float point computation.</p><p>　　VI. REFERENCES</p><p>　　[1] Allen J. Woo

62、d and Bruce F. Wollenberg, “Power Generation, Operation and Cotrol (Second Edition),” Tsinghuo University Press, 2003. </p><p>　　[2] W. F. Tinney and J. W. Walker. “Direct solutions of sparse network equati

63、ons by optimally ordered triangular factorization,” Proceedings of the IEEE, vol. 55, No.11, pp. 1801-1809, November 1967. </p><p>　　[3] K. M. Sambarapu and S. M. Halpin, “Sparse matrix techniques in power

64、systems,” Thirty-Ninth Southeastern Symposium on System Theory, March 2007. </p><p>　　[4] W. F. Tinney and C. E. Hart, “Power flow solution by Newton's Method,” IEEE Transactions on Power Apparatus and

65、Systems, Vol. PAS-86, No. 11, pp. 1449-1460, November 1967. </p><p>　　[5] W. F. Tinney, V. Brandwajn, and S. M. Chan, “Sparse vector methods,” IEEE Transactions on Power Apparatus and Systems, Vol. PAS-104,

66、 No.2, pp. 295-301, February 1985. </p><p>　　[6] R. Betancourt, “An efficient heuristic ordering algorithm for partial matrix refactorization,” IEEE Transactions on Power Systems, Vol. 3, No. 3, pp. 1181-11

67、87, August 1988. </p><p>　　[7] A. Gomez and L.G. Franquelo. “An efficient ordering algorithm to improve sparse vector methods,” IEEE Transactions on Power Systems, Vol. 3, No. 4, pp. 1538-1544, November 198

68、8. </p><p>　　[8] B. Stott, “Review of load-flow calculation methods,” Proceedings of the IEEE, Vol. 62, No. 7, pp. 916-929, July 1974. </p><p>　　[9] X. W. Chang and C. C. Paige, “On the sensit

69、ivity of the LU factorization,” BIT, Vol. 38, No. 3, pp. 486-501, 1998. </p><p>　　[10] P.A. Businger, “Monitoring the numerical stability of Gaussian elimination,” Numer. Math, Vol. 16, pp. 360-361, 1971.

70、</p><p>　　[11] Paola Favati, Mauro Leoncini, and Angeles Martinez, “On the robustness of gaussian elimination with partial pivoting,” BIT, Vol. 40, No.1, pp.062-073, 2000</p><p>　　潮流不同排序方案的比較

71、 </p><p>　　摘 要：今天被廣泛應(yīng)用的節(jié)點(diǎn)排序算法，旨在盡可能地保證電力系統(tǒng)的稀疏性。在這些算法中，因?yàn)樵谡５南到y(tǒng)中算法對(duì)每種解決方案的精確度不會(huì)有顯著的差異，所以它的影響通常被忽略。然而隨著現(xiàn)代電力系統(tǒng)的發(fā)展，這個(gè)問(wèn)題會(huì)變得更加嚴(yán)重，并且在節(jié)點(diǎn)排序過(guò)程中必須要把計(jì)數(shù)精度考慮在內(nèi)。在本文中，我們?cè)噲D為更多合理性排序算法奠定了基礎(chǔ)，這樣可以使內(nèi)存和準(zhǔn)確性之間進(jìn)行合理的比較。本文列舉出了三種不同目的的排序

72、方案，旨在比較潮流計(jì)算的形式，并且以一個(gè)六節(jié)點(diǎn)網(wǎng)絡(luò)為例進(jìn)行具體討論。</p><p>　　關(guān)鍵詞：潮流計(jì)算，節(jié)點(diǎn)排序，稀疏性，精確度，牛頓—拉夫遜算法，線性方程組</p><p><b>　　1引言</b></p><p>　　潮流是在電力系統(tǒng)的分析中最基本和最重要的概念，而潮流計(jì)算則是進(jìn)行電力系統(tǒng)規(guī)劃，運(yùn)行，調(diào)度和控制的基礎(chǔ)。從數(shù)學(xué)上來(lái)講，潮

73、流問(wèn)題是要找到一個(gè)非線性方程組的數(shù)值解。牛頓—拉夫遜算法是解決這個(gè)問(wèn)題最常用的方法，它涉及到一系列線性方程組重復(fù)的直接求解。線性方程組求解的效率和精度直接影響了牛頓 - 拉夫遜潮流算法的性能。在潮流計(jì)算中，電力系統(tǒng)的數(shù)值和物理特性，學(xué)者們通過(guò)重新安排節(jié)點(diǎn)的數(shù)目，致力于研究以便改善線性方程組的計(jì)算效率，并獲得了很大的成功從而為電力系統(tǒng)的分析奠定了堅(jiān)實(shí)的基礎(chǔ)。</p><p>　　在潮流計(jì)算中的雅可比矩陣，類似于導(dǎo)納

74、矩陣，有著對(duì)稱的結(jié)構(gòu)和高度的稀疏性。在分解過(guò)程中，內(nèi)存中的位置可以產(chǎn)生非零輸入，從而在原始的雅可比矩陣中產(chǎn)生零輸入。這一行動(dòng)被稱為最小填充。如果用只能處理和存儲(chǔ)非零輸入的編程術(shù)語(yǔ)，最小填充的減少反映了內(nèi)存需求和執(zhí)行分解所需的操作數(shù)量的大大減小。所以廣泛的研究與最小填充的極小值有關(guān)。雖然很難找到為確定絕對(duì)的最佳排序的有效的算法，但是有著接近最好效果的一些有效算法已經(jīng)得到了實(shí)際應(yīng)用。每種策略是在結(jié)果和執(zhí)行速度兩者之間的折中，并且它們都被大部

75、分工業(yè)所采納。上面提到的稀疏性的編程排序消除，在電力系統(tǒng)網(wǎng)絡(luò)計(jì)算中這是一個(gè)突破，這使得牛頓法的計(jì)算速度和存儲(chǔ)需求顯著提高。</p><p>　　在稀疏矩陣的方法之后，稀疏向量擴(kuò)展到向量的稀疏性探索的方法，當(dāng)右手邊的向量是稀疏的或在未知向量元素少數(shù)想用于求解線性方程組時(shí)，這種方法對(duì)求解線性方程是有用的。為了充分利用稀疏向量方法的優(yōu)點(diǎn)，通過(guò)節(jié)點(diǎn)排序加強(qiáng)L-1的稀疏性是十分必要的。這相當(dāng)于減少路徑的長(zhǎng)度，但它可能會(huì)導(dǎo)致

76、更多的最小填充，更大的復(fù)雜性和費(fèi)用。為了解決這個(gè)問(wèn)題，提出了一些節(jié)點(diǎn)排序算法，這種算法試圖通過(guò)減少路徑的長(zhǎng)度，同時(shí)保持矩陣的稀疏性來(lái)增強(qiáng)稀疏向量方法。</p><p>　　到目前為止，在任意一個(gè)節(jié)點(diǎn)的次序不會(huì)對(duì)數(shù)值精度產(chǎn)生負(fù)面影響的假設(shè)的基礎(chǔ)上，大多數(shù)節(jié)點(diǎn)排序算法通常會(huì)采取單一迭代解決線性方程組作為研究對(duì)象的方法，旨在減少內(nèi)存需求和計(jì)算操作。許多在網(wǎng)絡(luò)問(wèn)題中的強(qiáng)大對(duì)角線矩陣滿足上述假設(shè)，并且為了保證稀疏性的排序方

77、法增加了解決方案的準(zhǔn)確性。然而，如果在潮流系統(tǒng)模型中存在一系列非常高或低的阻抗，長(zhǎng)的超高壓線路，串聯(lián)和并聯(lián)補(bǔ)償?shù)葐?wèn)題，對(duì)角占優(yōu)將被削弱和假設(shè)可能并不總是站不住腳的。此外，隨著現(xiàn)代電力系統(tǒng)的發(fā)展，不同數(shù)量級(jí)參數(shù)下的新模型出現(xiàn)在潮流模型中。分布式發(fā)電的推廣也使我們堅(jiān)定地把分布網(wǎng)絡(luò)和傳輸系統(tǒng)融入到整個(gè)電力系統(tǒng)潮流計(jì)算中，當(dāng)然它會(huì)造成更嚴(yán)重的數(shù)值問(wèn)題。上面提到的所有這些事情會(huì)使問(wèn)題變得更加糟糕。因此，有必要討論節(jié)點(diǎn)編號(hào)對(duì)計(jì)算精度的影響。<

78、/p><p>　　基于上述提出的節(jié)點(diǎn)排序算法，本文重點(diǎn)關(guān)注這種節(jié)點(diǎn)排序在內(nèi)存和準(zhǔn)確性之間的矛盾，研究節(jié)點(diǎn)排序算法如何能影響的電力系統(tǒng)潮流計(jì)算的性能，從而為更理性的排序算法奠定基礎(chǔ)。本文安排如下：在第二部分介紹了節(jié)點(diǎn)排序算法的內(nèi)存和準(zhǔn)確性之間的矛盾。接下來(lái)的第三部分通過(guò)一個(gè)簡(jiǎn)單的直流潮流來(lái)說(shuō)明節(jié)點(diǎn)的順序可能會(huì)影響算法的精度。然后在第四部分以6個(gè)節(jié)點(diǎn)的網(wǎng)絡(luò)作為一個(gè)例子，對(duì)于節(jié)點(diǎn)排序?qū)Τ绷餍阅艿挠绊戇M(jìn)行了詳細(xì)分析。在第六部

79、分給出了結(jié)論。</p><p>　　2 節(jié)點(diǎn)排序算法中內(nèi)存和精確度之間的矛盾</p><p>　　根據(jù)計(jì)算數(shù)學(xué)，對(duì)于用高斯消元法求解的系統(tǒng)的線性代數(shù)方程組，完全消元法在數(shù)值上比部分消元法更可取。許多數(shù)學(xué)論文[9-11]都會(huì)關(guān)注高斯消元法的完全消元法與部分消元法的區(qū)別。參考文獻(xiàn)[ 9 ]表明部分消元法和完全消元法是如何影響LU分解的靈敏度。參考文獻(xiàn)[ 10 ]提出了一種有效而廉價(jià)的測(cè)試，從而

80、找到在部分消元法在使用時(shí)的數(shù)學(xué)難題。一旦不能滿足評(píng)估標(biāo)準(zhǔn)，就會(huì)采用完全消元法，以獲得更好的數(shù)值穩(wěn)定性。在潮流計(jì)算中，部分消元法可以再?zèng)]有任何行交匯的情況下自動(dòng)實(shí)現(xiàn)，因?yàn)樵诖蠖鄶?shù)情況下，雅可比矩陣的對(duì)角占優(yōu)的功能可以保證在浮點(diǎn)運(yùn)算的數(shù)值穩(wěn)定性的。雖然由于舍入誤差，部分消元法在有些極限點(diǎn)附近不能提供準(zhǔn)確的解決方法。如果采用完全消元法，上面執(zhí)行過(guò)程中的每一步，關(guān)鍵因素通常選擇最大的模塊元素。這相當(dāng)于調(diào)整潮流計(jì)算的節(jié)點(diǎn)排序。因此，與最大的模塊元

81、素有關(guān)的節(jié)點(diǎn)往往安排在前面以達(dá)到提高精度的目的。</p><p>　　以稀疏矩陣技術(shù)為導(dǎo)向的節(jié)點(diǎn)重新排序算法已廣泛應(yīng)用于電力系統(tǒng)計(jì)算中，旨在最大限度地減少內(nèi)存需求。在這些算法中，有著較少相鄰節(jié)點(diǎn)的節(jié)點(diǎn)往往首先被編號(hào)。其結(jié)果是在節(jié)點(diǎn)導(dǎo)納矩陣的對(duì)角線項(xiàng)往往根據(jù)自己的模塊被安排從最小到最大排列。類似地，每一個(gè)涉及到一個(gè)節(jié)點(diǎn)的對(duì)角線子矩陣，往往根據(jù)他們的行列式按照從最小到最大的順序進(jìn)行排列。這樣從這些算法形式中的獲得的結(jié)

82、果只會(huì)偏離形成的原則，但是后續(xù)的解決方案的精度將提高。這是我們所說(shuō)的按照內(nèi)存原則進(jìn)行節(jié)點(diǎn)排序和精確度之間是有矛盾的。</p><p>　　3 使用部分消元法和完全消元法所產(chǎn)生的精確度差異</p><p>　　對(duì)于解決系統(tǒng)的線性代數(shù)方程組，完全消元法在數(shù)值上比部分消元法更可取。當(dāng)系統(tǒng)系數(shù)變廣時(shí)，解的精度幾乎不可能受舍入誤差的影響，因此把排序?qū)τ诮鉀Q方案的準(zhǔn)確性的順序考慮在內(nèi)是必要的。<

83、/p><p>　　圖1有著四個(gè)節(jié)點(diǎn)的網(wǎng)絡(luò)樣本的直流模型</p><p>　　以圖1所示的有著四個(gè)節(jié)點(diǎn)的網(wǎng)絡(luò)樣本的直流模型為例。節(jié)點(diǎn)1是已知電壓相角擺動(dòng)節(jié)點(diǎn);節(jié)點(diǎn)2-4負(fù)荷節(jié)點(diǎn)。按照原來(lái)的節(jié)點(diǎn)數(shù)量，直流潮流方程是：</p><p>　　為了模擬計(jì)算機(jī)數(shù)值計(jì)算操作，我們用四個(gè)有效數(shù)字來(lái)解決這個(gè)問(wèn)題。沒(méi)有消元地對(duì)公式（1）執(zhí)行高斯消元得到的解為[ θ2,θ3,θ4]T=[-0

84、.3036,-0.3239,-0.3249]T，其與精確解[θ2, θ3,θ4]T=[-0.3,-0.32,-0.321]的部分元素不同，通過(guò)完全消元法可以得到一個(gè)更加精確的解：[θ2,θ3, θ4]T=[-0.3007,-0.3207,-0.3217]，并且行和列的交匯處的節(jié)點(diǎn)的排序是3,2,4 。所以這是一個(gè)為了獲得更高精確度的一個(gè)更加合理的方案。</p><p>　　4 節(jié)點(diǎn)排序?qū)εｎD-拉夫遜潮流計(jì)算方法

85、的表現(xiàn)形式的影響</p><p>　　圖2有著六個(gè)節(jié)點(diǎn)的網(wǎng)絡(luò)樣本的直流模型</p><p>　　在上述分析的基礎(chǔ)上，對(duì)節(jié)點(diǎn)重新排序的方案將不僅影響到內(nèi)存的要求，而且影響到求解線性方程組時(shí)解的精度。因此，牛頓 – 拉夫遜潮流方法的性能將隨著節(jié)點(diǎn)排序的變化而不同。在本節(jié)中將把三種不同的排序方案應(yīng)用到如圖2所示的6個(gè)節(jié)點(diǎn)的網(wǎng)絡(luò)，以便對(duì)它們對(duì)潮流計(jì)算中解的精度、收斂速度、計(jì)算量和內(nèi)存需求量進(jìn)行比較

86、。表四所示的是性能的細(xì)節(jié)。</p><p>　　A 目的一：盡可能地節(jié)省內(nèi)存</p><p>　　目前，以減少最小優(yōu)化和節(jié)省內(nèi)存節(jié)點(diǎn)，有各種各樣的方案應(yīng)用于近優(yōu)化的節(jié)點(diǎn)排序。這種方案所需要的唯一信息是描述網(wǎng)絡(luò)節(jié)點(diǎn)分支連接模式的一個(gè)表。對(duì)減少網(wǎng)絡(luò)的導(dǎo)納矩陣有著最佳效果的排序也是相關(guān)的雅可比矩陣表的最優(yōu)的因素。在編程的復(fù)雜性和最優(yōu)性之間不同的方案可以達(dá)成不同的妥協(xié)。在本文中，我們關(guān)注的編號(hào)的結(jié)

87、果是如何影響計(jì)算性能。編程效率是超出了目前的工作范圍。為了節(jié)省內(nèi)存，在這一部分中，提出了與[2]中提出的第三種方案類似一個(gè)動(dòng)態(tài)節(jié)點(diǎn)排序方案。該算法的執(zhí)行步驟如下。</p><p><b>　　方案一</b></p><p>　　a 定義其中一個(gè)節(jié)點(diǎn)度為一。如果一個(gè)以上的節(jié)點(diǎn)符合這個(gè)標(biāo)準(zhǔn)，選擇最原始的節(jié)點(diǎn)。如果沒(méi)有任何節(jié)點(diǎn)符合要求，啟動(dòng)步驟b ；</p>

88、<p>　　b 當(dāng)這個(gè)節(jié)點(diǎn)被淘汰，編號(hào)那些沒(méi)有等效的分支節(jié)點(diǎn)可以被引入的節(jié)點(diǎn)。如果一個(gè)以上的節(jié)點(diǎn)符合這個(gè)標(biāo)準(zhǔn)，選擇最原始的節(jié)點(diǎn)。如果我們不能啟動(dòng)步驟a和步驟b，打開(kāi)步驟c ；</p><p>　　c 當(dāng)這個(gè)節(jié)點(diǎn)被淘汰，編號(hào)那些有最少分支的節(jié)點(diǎn)。如果不止一個(gè)節(jié)點(diǎn)可以引入最少的分支節(jié)點(diǎn)，給那個(gè)最大節(jié)點(diǎn)度的節(jié)點(diǎn)編號(hào)。</p><p>　　一旦在上述步驟中某個(gè)節(jié)點(diǎn)被編號(hào)，更新相關(guān)節(jié)點(diǎn)

89、度和拓?fù)湫畔?。直到所有的?jié)點(diǎn)都編上號(hào)，節(jié)點(diǎn)編號(hào)就完成了。</p><p>　　表1 用方案一給節(jié)點(diǎn)再排序</p><p>　　緊跟著方案一之后，6節(jié)點(diǎn)網(wǎng)絡(luò)的節(jié)點(diǎn)編號(hào)次序如表1所示。在求解線性方程組的過(guò)程中，沒(méi)有引進(jìn)最小填充，所以表格的因素和雅可比矩陣將有完全一致的結(jié)構(gòu)。所以表格的因素的內(nèi)存需求是0.256Kb的，這個(gè)與該雅可比矩陣相同。通常情況下，通過(guò)四五次牛頓-拉夫遜迭代方法就可以得到解

90、?？墒沁@個(gè)例子所需的迭代次數(shù)是三十三次，因?yàn)樾∽杩狗种斐傻牟B(tài)性。在每次迭代期間前后替代的過(guò)程中需要123次乘法運(yùn)算，整個(gè)解答過(guò)程需要7456次乘法運(yùn)算。</p><p>　　B 目的二：用完全迭代法改善精確度</p><p>　　考慮到完全消元法在數(shù)值上比部分消元法更可取，在本節(jié)中，為了提高解決線性方程組的準(zhǔn)確性而采用完全消元法，旨在減少迭代次數(shù)。這里涉及到大量的對(duì)角線子矩陣行列式

91、的節(jié)點(diǎn)以便安排在前面。在某種程度上，導(dǎo)納矩陣的主對(duì)角線上的入口模數(shù)可以表明雅可比矩陣的主對(duì)角線上的子矩陣的行列式的幅度。為方便起見(jiàn)，我們利用導(dǎo)納矩陣確定數(shù)字的順序。</p><p><b>　　方案二</b></p><p>　　a 形成節(jié)點(diǎn)導(dǎo)納矩陣；</p><p>　　b 用完全消元法因式分解節(jié)點(diǎn)導(dǎo)納矩陣。記錄節(jié)點(diǎn)的當(dāng)前位置上的變化；&

92、lt;/p><p>　　c 根據(jù)因式分解的節(jié)點(diǎn)的最終位置確定節(jié)點(diǎn)的新編號(hào)；</p><p>　　表1 用方案二給節(jié)點(diǎn)再排序</p><p>　　執(zhí)行方案二，完整消元法可以再?zèng)]有行和列交匯的情況下自動(dòng)進(jìn)行。對(duì)應(yīng)這些節(jié)點(diǎn)的主對(duì)角線的入口模數(shù)通過(guò)總結(jié)更多的分支參數(shù)而變得更大，因此，節(jié)點(diǎn)度越大的往往首先被編號(hào)。因此，該方案的的結(jié)果可能與稀疏矩陣方法和許多引入的最小填充下形成的

93、節(jié)點(diǎn)編號(hào)的原則相異。6節(jié)點(diǎn)網(wǎng)絡(luò)的節(jié)點(diǎn)編號(hào)次序如表2所示。將產(chǎn)生6個(gè)最小填充，所以在一次迭代中前后替代過(guò)程中將花費(fèi)更多的內(nèi)存（ 0.488Kb ）和更多的操作（ 321個(gè)乘法運(yùn)算） ，所需的迭代總數(shù)減少到十三次，這表明線性方程組的計(jì)算精度通過(guò)完全消元法得以提高。最后，由于迭代次數(shù)的減少乘法運(yùn)算的次數(shù)減少到5573次。</p><p>　　C 目的三：保持稀疏性的同時(shí)提高精確度</p><p&g

94、t;　　在系統(tǒng)中只存在一個(gè)小的阻抗分支，所以相應(yīng)于節(jié)點(diǎn)4和節(jié)點(diǎn)6的只有四個(gè)條目（子矩陣）是非常大的導(dǎo)納矩陣（雅可比矩陣）。在提出替代的過(guò)程中，一旦節(jié)點(diǎn)4和節(jié)點(diǎn)6被消除，其余元素組成的子矩陣能保持良好的數(shù)值穩(wěn)定性，并且其余節(jié)點(diǎn)的編號(hào)不會(huì)對(duì)解決方案的精確度產(chǎn)生影響。把準(zhǔn)確性和稀疏性都考慮在內(nèi)，我們把4節(jié)點(diǎn)編為1號(hào)，然后按照目的一的方法給其他節(jié)點(diǎn)編號(hào)。這就是我們所說(shuō)的6個(gè)節(jié)點(diǎn)網(wǎng)絡(luò)的方案三。6節(jié)點(diǎn)網(wǎng)絡(luò)的節(jié)點(diǎn)編號(hào)次序如表3所示。</p>

95、;<p>　　由于在系統(tǒng)中只存在一個(gè)小的阻抗分支，并且它連接到節(jié)點(diǎn)度為1的節(jié)點(diǎn)4。方案三將符合目的一的要求。因此，最小填充的數(shù)目，內(nèi)存需求和分解所需的操作次數(shù)與方案一相同。為了保證收斂性，只需要9次迭代，導(dǎo)致計(jì)算量大大減少（僅2107次乘法運(yùn)算）。迭代次數(shù)的減少表明，使用方案三可以使線性方程組得到更加精確的求解。經(jīng)過(guò)分析和比較，原因如下：</p><p>　　與節(jié)點(diǎn)4有關(guān)的對(duì)角線元素比節(jié)點(diǎn)6小一點(diǎn)，

96、因此首先消除節(jié)點(diǎn)4不會(huì)降低精確度。該方案能夠滿足完全消元法。</p><p>　　方案三的更少的操作次數(shù)減少計(jì)算器浮點(diǎn)數(shù)的舍入誤差。尤其注意的是，如果首先消除節(jié)點(diǎn)6，非常小的值可能被添加到節(jié)點(diǎn)2和節(jié)點(diǎn)5的節(jié)點(diǎn)元素，這將導(dǎo)致嚴(yán)重的舍入誤差。然而，如果首先消除節(jié)點(diǎn)4，一個(gè)相當(dāng)大的值將被添加到節(jié)點(diǎn)6的對(duì)角線元素上，產(chǎn)生的新值在正常范圍內(nèi)。</p><p>　　表3用方案二給節(jié)點(diǎn)再排序</p

97、><p>　　表4用不同節(jié)點(diǎn)排序方案的進(jìn)行牛頓潮流計(jì)算的性能比較</p><p>　　a 這里“操作次數(shù)”只代表乘法次數(shù)因?yàn)槌朔ㄗ詈臅r(shí)。</p><p><b>　　5 結(jié)論</b></p><p>　　理論分析和數(shù)值計(jì)算結(jié)果表明：在潮流計(jì)算中考慮節(jié)點(diǎn)排序是十分必要的。如果節(jié)點(diǎn)排序算法合理地考慮內(nèi)存和精確度，潮流計(jì)算性能可以

98、進(jìn)一步改善。本文的基本結(jié)論如下：</p><p>　　對(duì)于良好的電力系統(tǒng)，節(jié)點(diǎn)排序?qū)Τ绷饔?jì)算精確度的影響是可以忽略的。更重要的是，我們的注意力應(yīng)該集中在保持稀疏性以以節(jié)省內(nèi)存需求和計(jì)算操作上。</p><p>　　對(duì)于病態(tài)的電力系統(tǒng)，為了加快收斂速度，在節(jié)點(diǎn)排序算法中必須考慮精確度。在此基礎(chǔ)上，如果同時(shí)要保證稀疏性，為了減少浮點(diǎn)計(jì)算精度的需求，我們應(yīng)該獲得更多的精確度。</p>

99、<p><b>　　6 參考文獻(xiàn)</b></p><p>　　[1] Allen J. Wood and Bruce F. Wollenberg, “Power Generation, Operation and Cotrol (Second Edition),” Tsinghuo University Press, 2003. </p><p>