1184 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 2, MAY 2005 TABLE II COMPARISON OF FUNCTIONS BETWEEN LSM AND HUBER increased level of confidence in each of the estimates, in spite of structural imperfections in H . V. CONCLUSIONS TABLE III STATISTICAL TEST RESULTS FOR TWO CASE STUDIES The Huber function technique is demonstrated for a robust estimation application in power engineering. A novel statistical test (see Table I) is presented to examine structural imperfections in the process matrix. The results offer an improvement over least-squares estimation in cases where the process matrix is characterized by structural imperfections. REFERENCES TABLE IV LSM AND HUBER ESTIMATES FOR SYNCHRONOUS GENERATOR PARAMETERS synchronous generator parameters using actual data. Usually, available data for synchronous generators are the stator phase currents and voltages at the terminals of the machine as well as the field voltage and current. Real data may be contaminated due to meter and communication errors, incomplete metering, or inaccuracy of metering equipment. Further, model selection and specifications may introduce structural defects in the process matrix. Actual data from a synchronous generator are considered. The generator is located in the southwest USA and is rated 213.7 MVA and 18 kV. The linearized generator model derived in [4] is used in this example. The least-squares method has provided satisfactory estimates of the mutual inductances LAD and LAQ and the field resistance rF . However, when the above parameters and the stator resistance r are estimated simultaneously, unrealistic values of r are obtained due to multicollinearity in H . This is evidenced by the results of the statistical test described in Section II and shown in Table III. The results of the statistical test show a “not high” statistical confidence for both cases. An examination of the flags raised by the test indicates that there are problems associated with multicollinearity in H (VIF flag), the condition number of H (eigenanalysis flag), and change in the confidence interval (Willan–Watts test flag). Table IV depicts the least-squares and Huber estimates of the synchronous generator parameters for both cases considered in Table III. For the Huber method, a bisectional search was used to find the optimum value of t. In this case, t = 2:1 yields optimum results. Table IV shows that even though statistical confidence is “not high” in case 1, the estimated parameters are satisfactory. The Huber technique does not produce estimates that are significantly different from the least-squares technique. This result is coincidental; it is generally advised to use robust estimation methods when the statistical confidence in the estimates is not high. This becomes evident in case 2, where four parameters are estimated simultaneously; a significant decrease in the percent deviation from manufacturer data is observed when robust estimation is used. Perhaps the most important benefit of this method is that the estimated parameter r is no longer negative (unrealistic for this application). The robust estimation results offer an [1] D. C. Montgomery, E. A. Peck, and G. G. Vining, Introduction to Linear Regression Analysis. New York: Wiley, 2001. [2] L. Mili, G. Steeno, F. Dobraca, and D. French, “A robust estimation method for topology error identification,” IEEE Trans. Power Syst., vol. 14, no. 4, pp. 1469–1476, Nov. 1999. [3] S. Suryanarayanan, “Accommodation of loop flows in competitive electric power systems,” Ph.D. dissertation, Dept. Elect. Eng., Arizona State Univ., Tempe, 2004. [4] E. Kyriakides, G. T. Heydt, and V. Vittal, “On-line estimation of synchronous Generator parameters using a damper current observer and a graphic user interface,” IEEE Trans. Energy Convers., vol. 19, no. 3, pp. 499–507, Sep. 2004. Introduction to Cumulant-Based Probabilistic Optimal Power Flow (P-OPF) Antony Schellenberg, William Rosehart, and José Aguado Abstract—This letter introduces the cumulant method for the probabilistic optimal power flow (P-OPF) problem. The proposed adaptation of the cumulant method to the P-OPF problem is given. Index Terms—Cumulants, optimization. optimal power flow, probabilistic I. INTRODUCTION Optimal power flow (OPF), which is a mathematical programming extension of the power flow problem, is a tool that has been commonly used within the power systems industry for many years [1]. In addition, many papers have been published on the subjects of probabilistic power flow (P-PF) [2] and probabilistic optimal power flow (P-OPF) [3], [4]. The goal of the P-OPF problem is to determine the probability density functions (PDFs) for all unknown variables present in the problem. These PDFs are the distributions of the optimal solutions. This letter proposes the adaptation of the cumulant method used in P-PF studies [5]–[7] to the P-OPF problem based on a logarithmic barrier interior Manuscript received March 17, 2004; revised June 18, 2004. Paper no. PESL00032-2004. A. Schellenberg and W. Rosehart are with the University of Calgary, Calgary, AB T2N 1N4 Canada (e-mail: schellen@enel.ucalgary.ca; rosehart@enel.ucalgary.ca). J. Aguado is with the University of Malaga, 29071 Malaga, Spain (e-mail: jaguado@uma.es). Digital Object Identifier 10.1109/TPWRS.2005.846188 0885-8950/$20.00 © 2005 IEEE IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 2, MAY 2005 1185 point method (LBIPM) [8]-type solution when random bus loading is considered. II. PROBABILITY AND STATISTICS BACKGROUND The cumulant method is based on a statistical measure known as cumulants. Therefore, some background information in probability and statistics is presented to provide a basis for the cumulant method. The expected value of a random variable x is defined as E [x ] = 1 01 xfx (x)dx (1) where fx (x) is the PDF of x. The nth-order raw moment mn is defined in the following manner: m n = E [x n ] : the moment generating function, 8z (s), for the random variable z can be written as follows: 8z (s) = E [esz ] = E [es(a x = E [es(a x +a ) e x + s(a x 111+a ) x ...e ) ] s(a x (7a) ) ]: Since x1 ; x2 ; . . . ; xn are independent 8z (s) = E [es(a x ) ]E [es(a x ) ] . . . E [es(a x ) ] = 8x (a1 s)8x (a2 s) . . . 8x (an s): (3) The nth raw moment is computed from the moment generating function by taking the nth derivative with respect to s and evaluating at s = 0. For example, the third raw moment can be computed as follows: d 3 8 x (s ) ds3 3 sx = d Eds[e3 ] 3 sx =E ddse3 m3 = 9z (s) = ln(8z (s)) = ln(8x (a1 s)8x (a2 s) . . . 8x (an s)) = ln(8x (a1 s)) + ln(8x (a2 s)) + + ln(8x (an s)) = 9 x (a 1 s ) + 9 x (a 2 s ) + + 9 x (a n s ): 111 s=0 s=0 The cumulant generating function [9], denoted by 9x (s), can be written in terms of the moment generating function 8x (s) as follows: (5) The cumulant generating function is employed in the same manner as the moment generating function; successive derivatives are taken with respect to s and evaluated at s = 0. The nth cumulant is denoted as n . (9b) (9c) (9d) (10) 111 (11) 111 (12) Evaluating (12) at s (4) (9a) The zero-, first- and second-order cumulants for a random variable 9 z (s ) = 9 x (a 1 s ) + 9 x (a 2 s ) + + 9 x (a n s ) 9z0 (s) = a1 90x (a1 s) + a2 90x (a2 s) + + an 9x0 (an s) 00 9z (s) = a21 9x00 (a1 s) + a22 9x00 (a2 s) + + a2n 9x00 (an s): s=0 9x (s) = ln 8x (s): (8b) z are computed as 111 = E x3 esx s=0 = E [x 3 ]: (8a) The cumulants for the variable z can be computed using the cumulant generating function (5) in terms of the component variables as follows: 111 8x (s) = E [e ]: (7c) (2) It is possible to compute the raw moments through the use of the moment generating function 8x (s) [9]. Mathematically, this function is stated as sx (7b) = 0 gives 9z00 (0) = a21 9x00 (0) + a22 9x00 (0) + 111 + a2n 9x00 (0): (13) It is noted that there is no limitation to second-order cumulants. Therefore, third and higher order cumulants can be computed following the same procedure. In general, the nth-order cumulant for z, which is a linear combination of independent random variables, can be determined with the following equation: n = 9z(n) (0) = an1 9x(n) (0) + an2 9x(n) (0) + 1 1 1 + ann 9x(n) (0) (14a) (14b) where the exponent (n) denotes the nth derivative with respect to s. III. BASIC CUMULANT METHOD DERIVATION IV. ADAPTATION TO P-OPF Given a new random variable z, which is the linear combination of independent random variables, x1 ; x2 ; . . . ; xn The cumulant method is adapted from the basic derivation above to accommodate the P-OPF problem when an LBIPM-type solution is used. The Hessian of the Lagrangian is necessary for the computation of the Newton step in the LBIPM. The inverse of the Hessian, however, can be used as the coefficients for the linear combination of random z = a 1 x1 + a 2 x2 + 1 1 1 + an xn (6) 1186 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 2, MAY 2005 variables around the current operating point. The pure Newton step is computed at iteration k of the LBIPM using the following equation: H (yk )1yk = 0F (y ) (15) k ( ) 1 where y is the vector of system variables, H yk is the Hessian of the Lagrangian function evaluated at the point yk , yk is the pure Newton step, and F yk is the mismatch vector evaluated at yk . Replacing yk with yk+1 0 yk in (15) and rearranging yields ( ) 1 yk+1 = 0H 01 (yk )F (yk ) + yk : (16) The similarity between (16) and a general linear equation is noted. Therefore, the matrix H 01 yk , which is the inverse Hessian, in (16) contains the multipliers for a linear combination of PDFs. This can be used to map the PDFs from known random variables into unknown random variables. In the case of random bus loading, a change in the bus load maps directly into a change in the mismatch vector. Therefore, the linear mapping information contained in the inverse Hessian can be used to determine cumulants for other variables when bus loading is treated as a random variable. If the multipliers in row i of the inverse ; ai;n , then the v th cumulant for the ith variHessian are ai;1 ; ai;2 ; able in y is computed using the following equation: ( ) ... yi;v =a 1 x v i; ;v +a 2 x v i; ;v + +a 111 v i;n x ;v REFERENCES [1] M. Huneault and F. Galiana, “A survey of the optimal power flow literature,” IEEE Trans. Power Syst., vol. 6, no. 2, pp. 762–770, May 1991. [2] M. T. Schilling, A. L. da Silva, R. Billinton, and M. El-Kady, “Bibliography on power system probabilistic analysis (1962–98),” IEEE Trans. Power Syst., vol. 5, no. 1, pp. 1–11, Feb. 1990. [3] M. Madrigal, K. Ponnambalam, and V. H. Quintana, “Probabilistic optimal power flow,” in IEEE Canadian Conf. Elect. Comput. Eng., vol. 1, May 1998, pp. 385–388. [4] G. Viviani and G. Heydt, “Stochastic optimal energy dispatch,” IEEE Trans. Power App. Syst., vol. PAS-100, no. 7, pp. 3221–3228, Jul. 1981. [5] W. Tian, D. Sutanto, Y. Lee, and H. Outhred, “Cumulant based probabilistic power system simulation using Laguerre polynomials,” IEEE Trans. Energy Convers., vol. 4, no. 4, pp. 567–574, Dec. 1989. [6] J. Stremel, R. Jenkins, R. Babb, and W. Bayless, “Production costing using the cumulant method of representing the equivalent load curve,” IEEE Trans. Power App. Syst., vol. PAS-99, no. 5, pp. 1947–1956, Sep./Oct. 1980. [7] P. Zhang and S. T. Lee, “Probabilistic load flow computation using the method of combined cumulants and Gram–Charlier expansion,” IEEE Trans. Power Syst., vol. 19, no. 1, pp. 676–682, Feb. 2004. [8] G. Torres and V. Quintana, “An interior-point method for nonlinear optimal power flow using voltage rectangular coordinates,” IEEE Trans. Power Syst., vol. 13, no. 4, pp. 1211–1218, Nov. 1998. [9] A. Papoulis and S. Pillai, Probability, Random Variables, and Stochastic Processes, 4th ed. New York: McGraw-Hill, 2002. (17) where yi is the ith element in y , and x ;v is the v th cumulant for the j th component variable. During an iteration of the cumulant method for P-OPF, the cumulants for unknown random variables are computed from known random variables. Next, PDFs are reconstructed using the Gram–Charlier/Edgeworth Expansion theory [7]. A statistical step is computed and combined with the pure Newton step in the LBIPM. The statistical step is computed by examining the system variable’s PDFs to increase the probability of occurrence. V. CONCLUSIONS This letter introduces the cumulant method for the P-OPF problem. By noting that the inverse of the Hessian used in the logarithmic barrier interior point can be used to perform linear mapping, cumulants can be computed for unknown system variables. The distributions are reconstructed from the PDFs, and a statistical step is computed. The statistical step is combined with the pure Newton step to determine a new step that attempts to increase the likelihood of occurrence of a solution while retaining the convergence properties associated with the pure Newton step. Although results are not included in a tabular form due to space limitations, numerical results using this approach support its validity. Results from the proposed approach are compared against Monte Carlo simulations in a nine-bus system. The mean values computed are well within 1% of the Monte Carlo result, and the variances are within 6%. As proposed here, the method makes the assumption that bus loads are independent. However, results can be modified to include covariances between two random variables. The method proposed in this letter is different from existing methods since it uses cumulants rather than moments for random variable information in a P-OPF setting. In addition, it makes use of the Hessian of the Lagrangian directly without any additional sensitivity/derivative techniques. Finally, the proposed method makes additional use of statistical information in step determination without including random variable information directly in the system equations. A Novel Approach to Estimate Load Factor of Variable-Speed Wind Turbines D. D. Li and C. Chen, Senior Member, IEEE Abstract—This paper proposes an approach to estimate the load factor for variable-speed wind turbines, which is an index of high interest for economic evaluation of the wind power generation systems. The proposed method is based on the statistical discipline of the wind resource and shorttime performance of wind turbines. Calculation results are given for a case study. Index Terms—Load factor, spectral analysis, statistics, wind energy. I. INTRODUCTION Due to changeability of the wind energy, a wind power generation system does not always operate at its rated condition, leading to inconvenience for calculation of actual output of the system. As the energy product is an important index for economic evaluation of the wind power generation systems, efforts on this topic have been reported in the literature [1]–[4]. Efficiencies of the electrical part of different wind power generation systems are examined with a statistical method in [1]. The statistical method is also used to study the annual production of wind energy systems [2]. Energy capture between various types of systems is compared based on a particular short-time wind speed series in [3]. Calculation of capacity credit of wind energy conversion systems is performed with chronological and probability methods in [4]. This paper will focus on the load factor of variable-speed wind turbines in Manuscript received April 7, 2004. This work was supported in part by the Alliance for Global Sustainability. Paper no. PESL-00041-2004. The authors are with the Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China (e-mail: sjtuldd@yahoo.com.cn; chchen@online.sh.cn). Digital Object Identifier 10.1109/TPWRS.2004.841154 0885-8950/$20.00 © 2005 IEEE