You are on page 1of 6

1

A New Probabilistic Load Flow Method


Chun-Lien Su, Member, IEEE
direct approach based on the principle of statistical least square estimation to compute the effects of uncertainties in input data on all output quantities and to obtain the expected value and variance of the solution of a load flow problem. A discrete frequency domain convolution technique by applying fast Fourier transformations and linearized power flow equation was proposed in [3] to increase the computation accuracy and to find the PDF of all output quantities. Reference [4] used a DC load flow model combined the concept of Cumulants and Gram-Charlier expansion theory to consider the bus injection uncertainties and to achieve enough accuracy with less computation effort to compute the approximate PDF and cumulative distribution function of network branch flows. A new boundary load flow method based on fuzzy/interval numbers was proposed in [5] to find the accurate boundary load flow solutions. [6] proposed a new PLF algorithm based on linearized models to account for the nonlinear network equations and to compute the distributions of output quantities. [7] used interval arithmetic to consider the uncertainty of nodal data and to find all possible solutions included within the bounds given by interval arithmetic. A new PLF method based on a description of bus power injections as random variables was proposed in [8] to consider the bus power injection uncertainties and operating practice of power systems. The main advantage of the analytical methods mentioned above is to avoid the computer simulations, but more assumptions and complex mathematical algorithms are required for these methods. To simplify computation process and maintain the computation accuracy, a MCS technique based on linear power flow equations combined with analytical convolution is used in [9]. Reference [10] proposed a new PLF algorithm combining MCS technique and multilinearized load flow equations to consider the greater level of uncertainty of nodal data. To consider network topology and nodal data uncertainties, a new PLF algorithm is proposed in [11]. In the method, the distributions of the load flow solution are evaluated conditioned by each possible network configuration. The final solution is obtained from weighted sum of density distributions by using the probability associated with each configuration. To enhance computation efficiency while maintaining sufficient accuracy, this paper proposes a new probabilistic load flow algorithm to account for input data uncertainties in the load flow calculation. This paper assumes the uncertainty of the line parameters and bus injections can be estimated or measured and shows how to estimate the true distributions of the solution of a load flow problem for networks with constant configuration. The statistical moments of the output quantities

AbstractWith the advent of open access transmission, the industry has faced many uncertainties related to the system planning. This paper proposes a new probabilistic load flow algorithm to account for the uncertainties involved in computing load flow solution. The method allows probabilistic modeling of generation power injections, load data, and line parameters and computes the statistical moments of load flow solution. The moments are then used in the probability distribution curve fitting. The computation of the proposed method is more efficient than Monte Carlo simulations. Performance of the proposed method is tested and verified on IEEE six bus test system. Index TermsProbabilistic load flow, Open access transmission, Point estimate method, and Power system planning.

I. INTRODUCTION ITH the advent of deregulated open access environments, additional sources of uncertainty are emerging. The industry has experienced greater uncertainty related to system planning. Traditional load flow approach is the deterministic load flow where the system condition represents a snapshot in time or a set of deterministic values chosen by the analyst for each input variable. From a system planning point of view, it has been shown worthwhile to approach the load flow problem as a probabilistic one. The probabilistic load flow study could take into account the uncertainty factor in the computation and compute the load flow results with probability of occurrence. Many probabilistic load flow (PLF) methods have been proposed to study load flow uncertainty problem [1-11]. These methods can be classified as simulation method, analytical method, or by a combination of both. The simplest evaluation of the PLF problem is through Monte Carlo simulation (MCS). This method requires that the data involved to be assigned a probability distribution that characterizes the possible variation in the parameters. The random values from these distributions are selected and used to arrive at an estimate of load flow solution. A large computation effort is required and for some cases the iterative process used to solve the nonlinear load flow equations may not successfully converge. In order to reduce the computational effort, the analytical methods were used [1-8]. In [1], a DC load flow model was proposed to take node data uncertainty into account and to find the distributions of branch flows. Reference [2] used a
This work was supported by the National Science Council of Taiwan under Grants NSC 92-2213-E-022-004. C. L. Su is with the Department of Marine Engineering, National Kaohsiung Marine University, Kaohsiung, 805 TAIWAN (e-mail: cls@mail.nkmu.edu.tw).

are obtained using an efficient point estimate method that is a numerical-based approach with simpler arithmetic process for statistical inference. The moments are then used in the probability distribution fitting. The proposed method can be used directly with any existing deterministic load flow program. The algorithm used is described and test results on IEEE 6 bus test system are reported. II. PROBLEM FORMULATION The load flow problem can be mathematically described by two sets of nonlinear equations. For a given network configuration, the load flow equations are:
Y = g ( X , L) Z = h( X , L )

III. APPLICATION OF TWO-POINT ESTIMATE METHOD TO PROBABILISTIC LOAD FLOW Several approximate methods have been used for the analysis of engineering systems under uncertainty [12-16]. Examples of these methods include the first order second moment (FOSM) method [12], the discretization method [13], and the point estimate method [14-15]. The FOSM uses a truncated Taylor series to expand the function F about the mean of each parameter. One of the disadvantages of the FOSM method is that it requires evaluation of the derivatives of F(Y,L) with respect to each parameter pi . The discretization method uses discrete probability distributions to replace continuous probability distributions. This method is easy to use and does not require approximation to F(Y,L) [17]. The two-point estimate method [14,15] can be used to calculate the moments of a random quantity that is a function of one or several random variables. Let pl denote a random variable with probability density function
f pl . When f pl by

(1) (2)

where Y is the input bus power injection vector L is the line parameter vector X is the state vector Z is the output line flow vector g, h are load flow equations Once the bus power injections and line parameters are specified, the state variable vector can be evaluated, and the output branch flow quantities expressed by Z are determined. Equation (1) can be rewritten by
X = g 1 (Y , L) = G (Y , L)

A=F(Y,L) is a function of pl , the two-point estimate method uses two probability concentrations to replace matching the first three moments of f pl . The concept of the scheme is depicted in Fig. 1, by a two-dimension representation [17].
F(Y) Ai (l ,1) = F ( p1 , p2 , K , p l ,1 , K , p Estimated
m 1

,p )
m

(3)
wl ,1
Ai (l ,1)

A=F(Y,L)

Substituting (3) into (2), then (2) is rearranged by


Z = h( g 1 (Y , L), L) = F (Y , L)

(4)
f pl

The F function is a nonlinear function. The branch flow solution set can then be expressed as follows Branch Flow Solution Set A = {z1 , z 2 ,..., z l }
= F (Y , L)

wl , 2
Ai (l ,2)
0
pl

pl , 2 wl ,2

pl ,1 wl ,1
m 1

= F ( p1, p 2 ,K, p m )

(5)

Ai (l ,2) = F ( p1 , p2 , K , pl , 2 , K , p

,p )
m

Fig. 1. Conceptual representation of the two-point estimate method [17]

where A is a set of branch flow solution, zi is the line flow of branch i, and pi is bus power injection or line parameter. Uncertainty in the parameter pi causes variations in the load flow solution. The variation of each load flow result quantity can be expressed as the expected value and standard deviation to present the possible ranges of result quantities corresponding to the known distributions of input data parameters. In order to achieve that, a two-point estimate method is proposed to compute the statistic moments of all unknown system states and network branch flows. In the whole process of computation, the nonlinear power flow equations are used.

An efficient two-point estimate method similar to that shown in [16,17] is adopted in this paper to estimate load flow uncertainty. The information about the two of pl variants, pl,1 and pl,2, are transferred to produce the two estimates of Ai variants, Ai(l,1) and Ai(l,2), through the functional relationship between A and F(Y,L). Two weightings wl,1 and wl,2 then scale these estimates and the standard deviations of Ai are computed through the first and second moments of Ai obtained from the two-point estimate method. Two concentrations of each variable are used to take into account the skewness of the probability density function of Ai. The standard deviation of Ai is then calculating using

Ai = var( Ai ) = E ( Ai2 ) [ E ( Ai )]2


m 2 m 2

(6)

E(Ai)=0; E(Ai2)=0 Select the uncertain parameter pl Compute the skewness coefficient l,3 k=1 Compute the parameter pl,k

where E ( Ai j ) wl , k [ Ai (l , k )] j = wl , k [ F ( p , p , 1 2
l =1k =1 l =1k =1

..., p l ,k ,..., pm 1 , pm )] is the jth moments of Ai. wl,k is the

location of the concentration, and


pl ,k = pl + l ,k pl

(7)

Compute the concentration wl,k Compute Ai(l,k) at ( p1 , p2 ,..., pl , k ,..., pm 1 , pm )

pl and pl are the mean and standard deviation of f pl ,


and l ,k = l ,3 / 2 + (1)
3k

n + (l ,3 / 2)

, k=1,2. l ,3 s
no

E(Ai)= E(Ai)+wl,kAi(l,k) E(Ai2)= E(Ai2)+wl,k[Ai(l,k)]2

are defined in (5). Each wl ,k denotes the weighting of the concentration located at ( p1 , p2 ,..., pl ,k ,..., pm 1 , pm ) and its value ranges from 0 to 1. The sum of all wl ,k ' s is unity. From the Appendix A, wl ,k can be expressed as
1 (1) k l ,3 k / l n
no

k=k+1
k>2?

yes
Have all uncertain parameters been taken care of ? yes Ai = E(Ai) Compute Ai

wl ,k =

(8)

where

l = 2 n + (l ,3 / 2) 2

and

l ,3

denotes

the

Plot accumulative probability density function


Fig. 2. Flow diagram of the proposed method [18]

coefficient of skewness of pl and can be obtained by


E[( pl pl ) ] ( pl )
3 3

IV. TEST RESULTS AND DISCUSSIONS

l ,3 =

(9)

where E[( p l pl ) 3 ] = ( p l ,t pl ) 3 Prob(pl,t). N is the


t =1

The performance of the proposed method was tested on IEEE 6 bus test system. The test was conducted on a Pentium 4 2.4GHz PC. Fig. 3 shows the tested system. The coefficient of variation (CV), defined as the standard deviation and mean value ratio, is used to indicate the dispersion of the random variables [17]. The uncertainties shown in Tables I and II are assumed.
Bus 3 Bus 2

number of observations of pl and Prob(pl,t) is the probability of each observation pl ,t . The distribution of each parameter
pl does not have to be normally distributed, but if it is indeed

normally or uniformly distributed, then l ,3 equals to 0 and


wl ,k equals to 1/(2m) [17].

Bus 6

The computational procedure of the proposed method is shown in Fig. 2. In the proposed method the nonlinear power flow equations of (1) and (2) are used. The proposed method estimates the statistic moments of system states and branch flows. The moments are then used in the computation of distributions of the load flow solution and in the cumulative probability density function fitting. It can be applied with the currently available load flow program. For a system with m uncertain system parameters, 2m load flow calculations are required. The number of load flow calculation is much lower than that required in the Monte Carlo style approaches.

Bus 1

Bus 5

Bus 4

Fig. 3 IEEE 6 bus test system

4 TABLE I DISTRIBUTIONS OF UNCERTAIN NODAL DATA (A) NORMAL DISTRIBUTION Bus Distributions of Bus Power Injection Voltage Active Power Reactive Power (p.u.) No. Type (p.u.) (%) (p.u.) (%) 1 Slack 1.05 2 Gen. 1.05 0.5 2.5 0.0 0.0 3 Gen. 1.07 0.6 3.0 0.0 0.0 4 Load -0.7 3.5 -0.7 3.5 5 Load -0.7 3.5 -0.7 3.5 6 Load 0.0 0.0 -0.7 3.5 Bus No. Type (B) DISCRETE DISTRIBUTION Distributions of Bus Power Injection Active Power Power (p.u.) -0.7 -0.6 -0.5 Probability 0.2 0.4 0.4

For the accuracy check, results from 5,000 repeated computations of load flow are also used. Table III shows the test results of some of load flow solution quantities. It can be seen from the results that the proposed method is about 25 times faster than the Monte Carlo simulation with 509 trails and has a good result in the estimate of result quantities distributions. Assuming that Monte Carlo approach can catch the stochastic behavior of the problem studied and provide true result of the load flow solution distributions [17], from Tables III it can be seen that the proposed method gives a fairly good modeling of the relationship between nonlinear power flow model and nodal data. Its computation requires a lower execution time than those required in the MCS method.
TABLE III COMPARISONS OF LOAD FLOW RESULT QUANTITIES Methods MCS Method MCS Method Proposed Load (509 trails) (5,000 trails) Method Flow Results
6 (degrees) V6 (p.u.) Q1-4 (p.u.) P1-5 (p.u.) P6-5 (p.u.) Execution Time (sec.) -4.7581 0.8559 1.0066 0.0025 0.2105 0.0126 0.3188 0.0322 0.0144 0.0196 5.606 -4.7512 0.8318 1.0067 0.0026 0.2104 0.0128 0.3162 0.0313 0.0131 0.0187 55.156 -4.7501 0.8320 1.0067 0.0026 0.2100 0.0127 0.3165 0.0314 0.0131 0.0188 0.219

Load

Using the results obtained from MCS method as the basis, three performance indices, denoted as , v, flow, are used to ascertain the performance of the proposed method for the load flow analysis with uncertain parameters. is defined as that the average relative errors, measured in
PEM MCS MCS ( ) / 100% i i i MCS MCS and (PEM ) /
i i i

100% ,

are

taken among 5 angles considering parameters mean () and standard deviation (). The performance index v is taken among 3 voltages corresponding to the system load bus and flow is taken among 44 line flows.
TABLE II DISTRIBUTIONS OF UNCERTAIN LINE PARAMETERS
Line Data R(p.u.) X(p.u.) From bus To bus Uniform Distribution 1 2 =0.1 =0.2 1 4 =0.05 =0.2 1 5 =0.08 =0.3 2 3 =0.05 =0.25 2 4 =0.05 =0.1 2 5 =0.1 =0.3 2 6 =0.07 =0.2 3 5 =0.12 =0.26 3 6 =0.02 =0.1 4 5 =0.2 =0.4 5 6 =0.1 =0.3 BCAP(p.u.) Binary Distribution prob{B12=0.02}=0.95, prob{B12=0}=0.05 prob{B12=0.02}=0.95; prob{B12=0}=0.05 prob{B12=0.03}=0.95; prob{B12=0}=0.05 prob{B12=0.03}=0.95; prob{B12=0}=0.05 prob{B12=0.01}=0.95; prob{B12=0}=0.05 prob{B12=0.02}=0.95; prob{B12=0}=0.05 prob{B12=0.025}=0.95; prob{B12=0}=0.05 prob{B12=0.025}=0.95; prob{B12=0}=0.05 prob{B12=0.01}=0.95; prob{B12=0}=0.05 prob{B12=0.04}=0.95; prob{B12=0}=0.05 prob{B12=0.03}=0.95; prob{B12=0}=0.05

First consider the bus power injection uncertainties. For the test system shown in Fig. 3, bus power injections are assumed to have a normal distribution with 5% of CV and a discrete distribution. An ac power flow model is used. Using the procedure shown in Fig. 2, the first and second moments (E(Ai), E( Ai2 )) of each component of the load flow solution are calculated, and the means and standard deviations of output quantities are estimated. The results are compared to those from the Monte Carlo simulation method. Two Monte Carlo simulations results are used for comparison of computation efficiency and accuracy [17]. In one simulation the termination criterion is set as the maximum differences of means and standard deviations of the calculated line flows in three successive trails are less than 10-4. Using this criterion, Monte Carlo simulation converged in 509 trails.

The probability distributions of active power on lines 1-5 and 6-5, by the Monte Carlo simulations with 5,000 trails, are reported in Fig. 4 and 5. As can be seen, the line flows on different branches can have different distributions. The mean and standard deviation of active power flow at line 1-5 obtained by the proposed method are 0.3165 and 0.0314 respectively, as compared with 0.3162 and 0.0313 for the Monte Carlo simulations. For line 6-5, they are 0.0131 and 0.0185 by the proposed method and 0.0131 and 0.0187 from the MCS method. Fig. 6 and 7 show the cumulative density functions of Fig. 4 and 5 respectively, obtained from the proposed method and the MCS method. In the proposed method, the first order moment of E(Ai) is recorded at each iteration computation and used for cumulative density function fitting. When 10% and 90% confidence levels of line flow are interested and used to roughly indicate the desired capacity of the branch, the estimates of Fig. 6 at 10% and 90% confidence levels are 0.273 and 0.3634 for the proposed method as compared to 0.279 and 0.3611 from the MCS method. In Fig. 7, they are -0.0180 and 0.0395 obtained from the proposed method as compared to -0.0157 and 0.0372 from the Monte Carlo simulations. From Fig. 4, 5, 6, and 7, it can be seen that the proposed method could accurately estimate load flow computation uncertainty even when it is nonGaussian type distribution. Test results also indicate that cumulative density functions of line flows can be accurately estimated by the proposed method and used to compute different confidence levels of the power flow on the branch to determine security levels of the branch with their probability of occurrence.

trails are used for accuracy comparisons. Fig. 8 shows the results obtained from the proposed method. As can be seen from Fig. 8, when there is an extremely high uncertainty in nodal injections, which is CV=100% means standard deviation of uncertain parameter equals to 100% of its expected value, the line flow average errors in mean and standard deviation are 0.312% and 0.660%, respectively. The distributions of the load flow solution obtained by the proposed method are close to those from the MCS method.

Fig. 4 Probability distribution of active power on line 1-5

Fig. 7 Cumulative density function of active power on line 6-5

Fig. 5 Probability distribution of active power on line 6-5

Fig. 8 Effects of variation of bus injections on line flow errors

Fig. 6 Cumulative density function of active power on line 1-5

Sensitivity analysis is performed to ascertain the performance of the proposed method with different numbers of CV of bus power injections. The line flow errors are used in the analysis. Fig. 8 shows the performance index flow for the proposed method with respect to the different CVs. The results obtained from Monte Carlo simulation with 5,000

The line parameter uncertainty can also be handled by the proposed method. Using the data shown in Tables I and II as basis, Table IV shows that the performance indices of the proposed method for the cases with different CVs of line parameters. For the results shown in Table IV, the results obtained from Monte Carlo simulation with 5,000 trails are used as a reference. As can be seen that, the errors in estimating output quantities distributions obtained from the proposed method slightly increase as the CVs of line parameter increase. From Fig. 8 and Table IV, test results have indicated that even when the uncertainty level in the input data is very high, the proposed method could have a good performance in estimating true distributions of the load flow solution.

6 TABLE IV SENSITIVITY OF PERFORMANCE OF THE PROPOSED METHOD TO COEFFICIENT OF VARIATION OF LINE PARAMETERS
CVs of Line Parameters (%) Performance Indices (%) 0 20 40 60 80 100

0.0371 0.4565 0.0072 0.3728 0.2354 0.5817

0.0410 0.4603 0.0111 0.3764 0.2582 0.6046

0.0456 0.4642 0.0154 0.3807 0.2703 0.6270

0.0498 0.4696 0.0192 0.3859 0.2756 0.6326

0.0537 0.4748 0.0236 0.3897 0.2988 0.6717

flow

0.058 1 0.478 9 0.027 8 0.394 5 0.326 9 0.693 9

[11] A. M. Leite da Silva, R. N. Allan, and V. L. Arienti, Probabilistic load flow considering network outages, IEE Proc. C: Generation, Transmission, Distribution, Vol. 132, No. 3, May 1985, pp. 139-145. [12] G. J. Hahn and S. S. Shapiro, Statistical Models in Engineering, John Wiley, New York, 1967. [13] R. E. Kurth and D. C. Cox, Discrete probability distribution for probabilistic fracture mechanics, Risk Analysis, Vol. 5, No. 1, 1985, pp. 235-240. [14] E. Rosenblueth, Point estimation for probability moments, Proc. Nat. Acad. Sci. U.S.A., Vol. 72, No. 10, 1975, pp. 3812-3814. [15] E. Rosenblueth, Two-point estimates in probability, Appl. Math. Modelling, Vol. 5, 1981, pp. 329-335. [16] H. P. Hong, An efficient point estimate method for probabilistic analysis, Reliability Engineering and System Safety, Vol. 59, 1998, pp.261-267. [17] C. L. Su and C. N. Lu, Two-point estimate method for quantifying transfer capability uncertainty, Accepted for Publication in the IEEE Trans. on Power Systems, 2004.

V. CONCLUSIONS In the open access transmission, many uncertainties have been introduced in the load flow study. In this paper, a new probabilistic load flow method is proposed to consider bus power injection and line parameter uncertainties. Test results have indicated that when the distributions of nodal data and line parameters are measured or estimated, the variations of load flow solution can be efficiently computed and the cumulative density function can be accurately estimated by the proposed method. The information obtained would provide system planners a better view of future system conditions in analyzing the effectiveness of alternative plans for system expansion. VI. REFERENCES
B. Borkowska, Probabilistic load flow, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-93, No. 3, May/June 1974, pp. 752759. [2] J. F. Dopazo, O. A. Klitin, and A. M. Sasson, Stochastic load flows, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-94, No. 2, March/April 1975, pp. 299-309. [3] R. N. Allan, A. M. L. Da Silva, and R. C. Burchett, Evaluation methods and accuracy in probabilistic load flow solutions, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-100, No. 5, May 1981, pp. 2539-2546. [4] P. Zhang and S. T. Lee, Probabilistic load flow computation using the method of combined Comulants and Gram-Charlier expansion, IEEE Trans. on Power Systems, Vol. 19, No. 1, February 2004, pp. 676-682. [5] A. Dimitrovski and K. Tomsovic, Boundary load flow solutions, IEEE Trans. on Power Systems, Vol. 19, No. 1, February 2004, pp. 348-355. [6] R. N. Allan and A. M. Liete da Silva, Probabilistic load flow using multilinearisations, IEE Proc. C: Generation, Transmission, Distribution, Vol. 128, No. 5, 1981, pp. 280-287. [7] Z. Wang and F. L. Alvarado, Interval arithmetic power flow analysis, IEEE Trans. on Power Systems, Vol. 7, No. 3, August 1992, pp. 13411349. [8] A. P. Sakis Meliopoulos, G. J. Cokkinides, and X. Y. Chao, A new probabilistic power flow analysis method, IEEE Trans. on Power Systems, Vol. 5, No. 1, February 1990, pp. 182-190. [9] A. M. Leite da Silva, V. L. Arienti, and R. N. Allan, Probabilistic load flow considering dependence between input nodal powers, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-103, No. 6, June 1984, pp. 1524-1530. [10] A. M. Liete da Silva and V. L. Arienti, Probabilistic load flow by a multilinear simulation algorithm, IEE Proc. C: Generation, Transmission, Distribution, Vol. 137, No. 4, July 1990, pp. 276-282. [1]

VII. BIOGRAPHIES
Chun-Lien Su (S97-M01) received the Ph.D. degree in Electrical Engineering from the National Sun Yat-Sen University, Kaohsiung, Taiwan in 2001. He is now an assistant professor in the Marine Engineering Department, National Kaohsiung Marine University. His main areas of interest are computer applications to power systems and shipboard electric power system.

You might also like