Electric Power Systems Research 104 (2013) 917

Contents lists available at SciVerse ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Robust WLS estimator using reweighting techniques for electric energy systems
Eduardo Caro a, , Roberto Mnguez b , Antonio J. Conejo c
a b c

Universidad Politcnica de Madrid, Madrid, Spain Universidad de Cantabria, Cantabria, Spain Universidad de Castilla-La Mancha, Ciudad Real, Spain

a r t i c l e

i n f o

a b s t r a c t
The state estimator is a key tool in the operation of any real-world electric energy system. In this paper, a state estimator based on a weighted least squares model is proposed which is robust against outliers. This algorithm presents two relevant features: robustness that is achieved by readjusting measurement weights, and accuracy that is attained by considering measurement dependencies. The proposed method is tested in the IEEE 57-bus and 118-bus systems and the obtained results are analyzed using Design of Experiments and ANOVA techniques. 2013 Elsevier B.V. All rights reserved.

Article history: Received 31 July 2012 Received in revised form 24 April 2013 Accepted 31 May 2013 Keywords: Weighted least squares Power system state estimation Outlier detection Dependent Gaussian measurements

1. Introduction 1.1. Motivation In any real-world electric energy system, the Control Center monitors and controls the functioning of the network in real-time, ensuring operational security. To accomplish this task the Control Center needs to know accurately the actual state of the system (node voltages, power ows, etc.) at any time. These values are estimated by the state estimator (SE). The state estimator is a mathematical algorithm which computes the most-likely state of the network, given a redundant set of measurements captured from the system. From the statistical point of view, the state estimation algorithm is a nonlinear multiple regression problem, whose parameters to be estimated are those which characterize the network state: node voltage magnitudes and angles. This estimated state is generally computed using the Maximum Likelihood Estimator, minimizing the weighted sum of the squared residuals (i.e., Weighted Least Squares approach). Once the most-likely state is obtained, the Control Center performs a bad data detection and identication procedure to detect and eliminate those measurements whose associated standardized errors are larger than a pre-established tolerance. The statistical tests commonly employed for these tasks are the 2 -test and the Largest

Normalized Residual test, and are well established in the technical literature [1]. Once outliers have been removed, the nonlinear multiple regression problem is solved again, and the nal state estimate is obtained. If outliers are not properly detected or eliminated, the nal estimate will be biased, and the Control Center will not have an accurate knowledge of the actual state of the system, leading occasionally to an insecure operation of the network. For this reason, the detection and identication of bad measurements have a notorious relevance in the estimation process. In fact, an adequate and secure control is only achieved in the case that the SE procedure is robust enough to detect and eliminate the presence of corrupt measurements. Traditionally, the outlier elimination problem is solved iteratively by detecting/removing suspected measurements and re-estimating the state disregarding the rejected data. These estimators are based on the weighted least squares, which shows a notable computational efciency; however the lack of robustness deteriorates signicantly their performance in the presence of bad measurements. Specically, the presence of multiple conforming bad measurements in the measurement set may provoke a masking effect: good measurements may be rejected whereas corrupted ones may not. This undesirable situation occurs when measurement dependencies are not properly modeled. 1.2. Aim The aim of this paper is to present a robust state estimator based on a weighted least squares regression, which carries out the estimation and the bad data detection/identication processes simultaneously by successively adjusting the weighting matrix and

Corresponding author at: C/ Jos Gutirrez Abascal, 2,28006 Madrid, Spain. Tel.: +34 913363149. E-mail address: Eduardo.Caro@upm.es (E. Caro). 0378-7796/$ see front matter 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.epsr.2013.05.021

10

E. Caro et al. / Electric Power Systems Research 104 (2013) 917

considering the effect of measurement dependencies. The obtained estimate does not require further bad measurement processing algorithms.

2. Dependent state estimation model Any state estimator can be formulated as a nonlinear multiple regression problem, where the unknown parameters are the node voltage magnitude and angle of every node, represented by Vi and i , respectively. These two sets of variables form the state vector x = [V T ] . There are n state variables. The unknown true state is represented by xtrue . The unknown parameters are estimated using the information provided by observations {z1 , . . ., zm }. These observations are captured from the system using measuring devices, and are related with x by means of a multifunctional vector h( x). Depending on the measurement type, the functions hi ( x) differ. Expressions of functions hi ( x) are well-established in the technical literature [1]. The error terms used throughout this paper are dened below: Measurement residual: difference between the measurement zi , and the function hi ( x) evaluated at the optimal state x Residuali = zi hi (x)|x=x x) = ri . = zi hi ( (1)
T T

1.3. Literature review The technical literature is rich in references concerning the state estimation problem, for instance, [2] or [1]; and there is a signicant number of references on outlier detection: [311]. The previous works are focused mainly on the area of least squares linear regression. Other statistical models and estimation methods, such as reweighed techniques [1215], non-linear methods [16], variancevarying models [17], or some robust estimators [1820] have received comparatively less attention. Nevertheless, [21] report successful results from the application of the reweighted least deviation method developed by [14], to detect data related to hurricanes and typhoon on wave hindcast databases. However, no so many works address the power system WLS estimator using adjusted measurement weights. The pioneering work reported in [22] proposes a method for readjusting the measurement variances based on the residuals of previous estimations. Ref. [23] develops this approach, improving the computational efciency and ensuring mathematical convergence. Work [24] propose an iterative reweighted least-squares estimator that is based on Givens Rotations and improves the robustness against outliers. In [25], the weights of the WLS estimator are articially manipulated, leading to a more robust estimator with the properties of the weighted least absolute value approach. Recently, in [26], the WLS regression is addressed using estimated weights based on the measurement variances. All the aforementioned works consider that the measurement covariance matrix is diagonal. However, recent works [27,28] show that this matrix is generally non-diagonal. Thus, the reweighting techniques previously proposed in the technical literature can be improved, since such techniques cannot deal with measurement dependencies.

Measurement error: difference between the measurement zi and the function hi ( x) evaluated at any state x, Errori = zi hi (x)|x=x = zi hi (x) = ei . (2)

Metering error: difference between the measured value and the unknown true value, Metering error i = zi hi (x)|x=xtrue = zi hi (xtrue ) = zi zitrue . (3) Note that the term measurement residual is solely used in the case of comparing measurement value zi with the function hi () . Similarly, the term measurement evaluated at the optimal state x error is solely employed for comparing the measured value zi with the function hi (), evaluated at any state x. Measurement errors have been traditionally modeled as an independent unbiased Gaussian-distributed random variable. The factual metering infrastructure within substations results in significant statistical correlations between measurement errors. Works [27] and [28] numerically show that these correlations are significant, and its consideration may improve the quality of the nal estimate. Therefore, hereafter measurement errors are assumed to be dependent Gaussian-distributed unbiased random variables. The dependence structure is modeled by means of positive-denite non-diagonal variance-covariance matrix C z , which can be easily computed using the Point Estimate method [28]. As it is customary in the technical literature, all measurements are considered synchronous. This is reasonable since in steady state power system magnitudes change very slowly with respect to the time needed to transfer measurements to the EMS from Remote Terminal Units (RTUs) or Phasor Measurement Units (PMUs). 2.1. State estimation Given the previous assumptions, the estimation of the state variables are obtained by minimizing the weighted sum of squared measurement errors of the multiple nonlinear regression model, leading to a nonlinear optimization problem:
1 minimize J = [z h(x)]T C z [z h(x )] x

1.4. Contribution The contribution of this paper is threefold:

First, it provides a mathematic procedure that allows applying a reweighting estimation technique (originally designed for diagonal covariance matrices) to a non-diagonal estimation problem. Second, an iterative state estimator is proposed, showing both robustness against outliers and computational efciency. Specifically, it requires signicantly less time that similar methods proposed in the technical literature [29]. Finally, Design of Experiments and ANOVA procedures are used to compare the performance of the proposed method with statistical rigor.

1.5. Paper organization The rest of this paper is organized as follows. Section 2 develops and formulates the Reweighted Least Squares Estimator considering measurement dependencies. Section 3 applies the Design of Experiments and ANOVA procedures to the considered estimation problem. Section 4 provides and analyzes results from four realistic case studies. Finally, Section 5 provides some relevant conclusions.

(4a)

subject to c (x) = 0 g (x) 0 (4b)

where the scalar J is the objective function and c ( x) and g ( x) are the equality and inequality constraints modeling zero-injections nodes and physical operating limits, respectively. Note that matrix C z is

E. Caro et al. / Electric Power Systems Research 104 (2013) 917

11

not diagonal, since the considered model takes into account the statistical correlations between measurement errors, but positivedenite as it is a variance-covariance matrix.

The reweighting method is based on an iterative algorithm which updates the weighting factors in every iteration, according to the Tuckeys biweight formula:

2.2. Reweighted least squares formulation The Weighted Least Squares model (4a) accounts for the error measurements with different variances and their dependency. However, an alternative weighting matrix can be used accounting for (i) different variances, (ii) dependency, (iii) and degree of condence related to each measurement. In order to derive this weighting matrix let us consider the vector of measurement errors e = z h( x) and optimization problem (4). Using the Cholesky decomposition of matrix C z , the objective function in (4a) becomes
1 J = [e]T C z [e ]

wii =

yi 1 6
0

2 2

if

|yi | 6 (8)

if |yi | > 6

where yi = ui / * is the standardized residual related to uncorrelated vector u, and * is the scaled median absolute deviation estimator *:

mediani |ui | mediani |ui | 0.6745 (3/4)

(9)

= eT (LLT )
T

e = eT (L1 ) L1 e

(5)

= (L1 e) L1 e = uT u where u is a vector of standard independent normal random variables, with covariance matrix equal to the identity matrix I , and L is the lower-triangular Cholesky factor of matrix C z . Note that the derivation above is based on the assumption that the estimated is sufciently close to the true state xtrue , which is a common state x assumption in the state estimation Literature. The aim of most outlier detection methods is to determine whether or not a measurement should be considered as an outlier, without allowing for intermediate situations. In contrast, the method proposed in this paper, originally developed by [14], aims at empirically determining a diagonal matrix W to be included in model (5), i.e. JR = uT W u (6)

Reweighting techniques constitute an effective and computationally attractive alternative to solve M-estimators. A signicant number of M-estimators have been developed in the technical literature, such as least-absolute value, Huber, Cauchy, Welsh, Tuckey, etc., and each type has an associated updating formula for weights, such as (9) for Tuckeys case. Work [30] provide a useful discussion about the choice of the appropriate function for practical cases. Specically, Tuckeys biweight do not guarantee uniqueness of the solution, but the inuence of gross errors is reduced considerable or even eliminated [31]. Thus, the RWLS problem formulation is: minimize J = [z h(x)]T W R [z h(x)]
x

(10a)

subject to c (x) = 0 is: Initial non-dependent estimation. An initial WLS estimation is performed to estimate the measurement variance-covariance 0) 0) , matrix C z is matrix. Using the initial estimation obtained x computed via the Point-Estimate method in [28]. Parameter initialization. Weights wii are set to one: w0) = 1, i ii {1, m}. The iteration counter is set to 1, = 1. Dependent state estimation. The state estimation problem (10) is solved considering measurement dependencies and the 1) reweighted matrix W R . The obtained estimates are denoted . as x ) Convergence checking. Once the estimates x are available, if ) 1) x ||x || > the estimation process continues in 5). ) and the algorithm Otherwise, a solution with a tolerance is x concludes. Update weighting matrix W R . Once the estimates of the state vari) x ), the weighting matrix W R ) is updated ables are available ( using (8). Set + 1 and go to step (3). The computational efciency of this algorithm can be improved 1) by using x as the initial values of the estimation in step (3). The convergence of the iterative process does not only depend on the reweighted technique employed, but also on the structure of the considered optimization problem (i.e., convexity or linearity of the objective function, presence of equality/inequality constraints, types of variables, etc.). Thus, convergence rules are not general. Specically, work [30] discusses the choice of an appropriate function, whereas [32] shows convergence for certain weighting functions. For the optimization problem considered in the paper,
)

g (x) 0

(10b)

The algorithm of the RWLS estimator considering dependencies

where wii is a weight for every observation ranging continuously from 0, for observations that are completely unreliable, up to 1, for observations that are completely reliable. Considering (6), the objective function (5) becomes: JR = eT (L1 ) W L1 e = eT W R e. From (5) and (7), the following observations are in order: The measurement error vector e (dependent normal random variables) is transformed into a vector of independent standardized normal variables u [29]. The objective function J can be expressed as the sum of a set of squared independent standardized normal random variables. The objective function JR is computed as the weighted sum of squared independent standardized normal random variables. is multiplied by the weighting factor wii [0, 1]. Each factor u2 i If the ith weighting factor is null (wii = 0), then the component u2 is not considered in the objective function JR . If, on the other i is fully considered in JR . hand, wii = 1, the component u2 i The underlying idea of the RWLS method is to adjust empirically the weighting factors, based on the degree of condence of each measurement. The coefcients for those measurements completely unreliable are adjusted to zero and, similarly, the weighting factors for those measurements completely reliable are adjusted to one.
T

(7)

12

E. Caro et al. / Electric Power Systems Research 104 (2013) 917 Table 1 ANOVA model: factors and levels. Names Levels WLS DWLS RWLS 1,. . ., n

fast convergence is achieved in all cases, regardless the measurement scenario, presence of bad data, or network used. 3. Design of Experiments Section 2 above presents a novel algorithm to estimate the state of a power system in a robust manner. In this section, the statistical procedures Design of Experiments and ANOVA are briey described as they are used (i) to analyze the performance of the proposed method, and (ii) to compare it with other existing methodologies. The methods considered in this paper are listed below: WLS. The common Weighted Least Squares estimator is used as basic benchmark, using the 2 -test and the Largest Normalized Residual (LNR) test to detect and identify bad measurements, respectively. These algorithms are well-established in the technical literature [1]. The WLS results are the nal estimates once the 2 -test and LNR test have been successfully passed. DWLS. The estimation, detection, and identication algorithms considering dependencies, proposed in [27] and [29], are employed to estimate the state and to detect bad measurements including the statistical correlation between measurements. Thus, the DWLS results correspond to the nal estimate once the bad measurements have been removed. 3.1. Performance assessment In order to rigorously determine which is the method with best performance, a Design of Experiments is carried out. This statistical procedure allows determining if the proposed algorithm is significantly better than the rest of the approaches with a pre-specied condence level, and taking into consideration the dissimilarities among the considered measurement scenarios. The performance of each method is assessed by means of the following metrics: Metric V , dened as the average absolute error of the voltage MET, magnitude estimate for the th measurement scenario, considering the method MET, i.e.,
V MET,

Factors

Method (MET) Scenario ()

Table 2 ANOVA table structure. Source Method Scenario Residual Squared Sum SSM SSD SSerror Deg. of freedom 31 n 1 (n 1)(3 1) Mean-sq
2 M s 2 D s 2 error s

F-stat
2 2 M error s /s 2 2 D error s /s

Fcrit
Method Fcrit Scenario Fcrit

where uMET, N (0,

iid n

2)

and:

MET = 0;
MET =1

=0

n |V MET i=1 i,

true | Vi,

(11)

Note that the previous metric is measured in p.u. Metric , dened as the average absolute error of the voltage MET, angle estimate for the th measurement scenario, considering the method MET, i.e.,
MET,

where is the global effect, i.e., the average value of the considered metric yMET, . Parameter MET is the main effect of the estimation method, and measures the increase/decrease of the average response of the factor Method (MET) with respect to the average level. Likewise, parameter is the main effect of the factor Scenario (), and it measures the increase/decrease of the average response for all the methods with respect to the average level at the th measurement scenario. Finally, the random effect uMET, includes the effects of all other causes not modeled. Taking into consideration that the particularities of each measurement scenario may have inuence on the methods performance, the effect Scenario is included in model (13). The factors considered in this ANOVA analysis and the levels corresponding to each factor are provided in Table 1. Parameter n stands for the number of measurement scenarios considered. The background hypotheses of this model are: (i) normality, (ii) constant variance, and (iii) independence. To ensure this statistical properties, an appropriate diagnosis procedure is performed after the residual computation. Since the aim of this study is to nd the most accurate estimation method and to check if it is signicantly different from the other methods, the following tests are performed: H0 : MET = 0, MET H1 : MET|MET = / 0 H0 :

= 0, = 1, . . . , n

H1 : |

= / 0

n | MET i=2 i,

true | i,

n1

(12)

The previous metric is measured in radians. Note that the considered reference angle is located at node 1, i.e., 1 = 0 rad, for all the considered scenarios. Metric CPUMET , dened as the required CPU time to obtain the nal estimate considering the method MET for the th measurement scenario. Note that this metric is measured in seconds. 3.2. ANOVA model The model employed in this Design of Experiments procedure comprises the factors Method and Scenario. This model is described below, yMET, = + MET +

+ uMET,

(13)

The null hypothesis for the rst test corresponds to the no statistically-signicant inuence of the method on the average performance. The alternative hypothesis establishes that it exists at least one method performing different from the average. The second test is analogous but related to the factor Scenario. To perform the above two statistical hypothesis testing, the ANOVA table is computed and analyzed. Table 2 provides the general structure for this table, particularized for the problem under consideration. The computation of the elements of the table is well established in the technical literature [34]. Appendix A provides further details concerning the ANOVA table structure and the computation of its terms. Once the ANOVA table is obtained, the value of the F-statistic for the factor Method allows deciding whether or not the performances of the different methods are statistically different. If methods performances are statistically similar, F-stat will be low. Otherwise, if there are signicant dissimilarities among methods

E. Caro et al. / Electric Power Systems Research 104 (2013) 917 Table 3 118-Bus system case study: ranges. Vm 118 m [2; 4] Pm [10; 103] Qm [10; 103] PFm [208; 231] QFm [208; 231] Im [5; 20] Buses 118

13

Branches 179

performances, the value of F-stat will be high. In order to check if Method : the resulting F-stat is low or high, it is compared with Fcrit If F-statistic > F Method , then the related null hypothesis is rejected crit and it is concluded that the response variable is signicatively affected by the method employed. Then, the average values for each method and its condence intervals are plotted to determine which is the method with best performance. If F-statistic < F Method , then it is concluded that the performances crit of all methods are similar.
Method Within a certain condence level (1 ), the parameter Fcrit is dened as the maximum admissible value for a random variable which follows an F-Snedeccor distribution with the corresponding Method can be degrees of freedom. Alternatively, the parameter Fcrit dened as the limit value for the marginal right-side region with an area equal to for the F probability density function (see Fig. 6). Method is computed considering a condence level of The value of Fcrit 95%, leading to false alarm probability lower than 5%.

The numbers of measurements are well-balanced across measurement types. Measurements are geographically allocated in an uniform manner throughout the considered system. The random allocation of metering devices is realistic since the quantity and proportion of the allocated measurements are based on reasonable ranges. Table 3 provides the ranges used for this random allocation of measurements. Columns Vm , m , Pm , Qm , PFm , QFm , and Im correspond to the voltage meters, PMU devices, active/reactive power injection meters, active/reactive power ow meters and current meters, respectively. PMU measurements are modeled as in [35,36] The number of buses and branches of the considered system are provided in the last two columns. The computational analyses have been performed using a Windows-based personal computer with a 64-bits four-core thirdgeneration i5 processor at 1.73 GHz and 4 Gb of RAM. Sections 4.1 to 4.4 consider the IEEE 118-bus system, whereas Section 4.5 considers the IEEE 57-bus system. 4.1. First case study: no gross errors

4. Case studies In this section, four case studies are analyzed to check the estimation accuracy and computational efciency of the proposed state estimator. The networks under study are the IEEE 57-bus and the 118bus systems.1 To obtain statistically sound conclusions: (i) a set of one hundred randomly-generated measurement scenarios is considered, and (ii) an ANOVA procedure is performed to analyze the obtained results. Each scenario involves: (i) a random active/reactive power consumption level, (ii) random locations of voltage and active/reactive power meters (ensuring observability of the whole system), (iii) a random redundancy level, and (iv) Gaussian-distributed random errors in all measurements, (standard deviations of 0.01 pu and 0.02 pu for voltage and power measurements, respectively). Note that each scenario involves random errors in all measurements. These errors are modeled as Gaussian-distributed unbiased random variables, taking into consideration the actual metering infrastructure of each substation. Ref. [27] provides more details concerning the random measurement generation. The measurement set is generated assuming perfect synchrony throughout all substations. Future work will consider generation of measurements exhibiting accidental asynchrony and its impact on the performance of the proposed method. The allocation of meters is based on a random algorithm. This algorithm has been specically designed for allocating meters in a realistic way, providing scenarios that satisfy the following requirements: The resulting measurement conguration provides whole observability of the system. The measurement conguration provides a realistic redundancy ratio, ranged from 2.9 to 3.3. In this case, the measurement vector z is free of gross errors. For each measurement scenario, the estimates for methods WLS, DWLS, and RWLS are computed and an ANOVA analysis is performed. Table 4 provides the ANOVA analysis for metrics V , , and MET MET CPUMET , and Fig. 1 depicts the average value for these metrics and the condence intervals for a 95% condence level. From Table 4 and Fig. 1, the following observations are in order: From the ANOVA analysis of metric V in Table 4, it is observed MET Method = 3.03. This indicates that the accuthat F-stat = 39.2 > Fcrit racy of the methods for estimating voltage magnitudes are not the same. Thus, in order to determine graphically which is the method with higher accuracy, the mean and condence interval of metric V are plotted in Fig. 1. MET Similarly, the F-statistics for metrics and CPUMET are higher MET than the critical values. Then, the mean and condence interval of these metric are plotted in Fig. 1 to allow determining which is the method with better accuracy estimating the voltage angle, and which is the method computationally more efcient. Regarding estimation accuracy (left and center subplots), the WLS estimator is the less accurate, whereas the DWLS and RWLS
Table 4 ANOVA table for the rst case study. Source Metric Method Scenario Residual Metric MET Method Scenario Residual Metric CPUMET Method Scenario Residual
V MET

Squared Sum 2.4 106 4.1 105 6.0 106 2.1 105 4.0 105 1.9 105 6.0 101 4.3 102 5.0 102

DoF 2 99 198 2 99 198 2 99 198

Mean-sq 1.2 106 4.1 107 3.0 108 1.1 105 4.0 107 9.6 108 30 4.3 2.5

F-stat 39.232 13.490

Fcrit 3.03 1.30

111.094 4.196

3.03 1.30

Power Systems Test Case Archive. Available at: http://www.ee.washington. edu/research/pstca/.

11.975 1.721

3.03 1.30

14

E. Caro et al. / Electric Power Systems Research 104 (2013) 917

Fig. 1. Results for the case study with no gross errors: performance comparison.

procedures provide the most accurate results for a condence level of 95%. There is no signicant difference between the accuracy provided by the DWLS and RWLS methods. Regarding the required CPU time, methods WLS and RWLS are the most efcient. 4.2. Second case study: three gross errors In this case, the measurement vector contains bad data. The corresponding scenarios are generated as follows [27]: 1. Vector z true is derived from a power ow solution. Then, random Gaussian-distributed noise is added to all measurements in z true to derive measurement vector z . 2. Three substations are randomly selected, and for each of them: A metering transformer is randomly selected. Both phase and transformer type (current or voltage) are also randomly chosen. Then, a gross error is included in the selected transformer signal. All measurements from this substation are computed taking into account the erroneous transformer signal. Thus, two or more processed measurements in this substation are distorted as a result of an interacting and conforming error [29]. Again, Table 5 provides the ANOVA analysis for metrics V , MET and CPUMET , and Fig. 2 depicts the average value for these metrics and the condence intervals for a 95% condence level. From Table 5 and Fig. 2, the following observations are in order:
, MET

computational efciency. Thus, the condence interval plots (Fig. 2) are studied to determine which method provides the best performance. Regarding estimation accuracy (left and center subplots), the WLS estimator is the less accurate, whereas the DWLS and RWLS procedures provide the most accurate results for a condence level of 95%. There is no signicant difference between the accuracy provided by the DWLS and RWLS methods. Regarding the required CPU time, the RWLS estimator is the most efcient. The efciency provided by the WLS and DWLS algorithms are statistically similar. 4.3. Third case study: six gross errors In this case, six substations are randomly chosen, and a set of multiple gross errors is located in each substation. Thus, the measurement vector is corrupted by six sets of multiple bad data. Table 6 provides the ANOVA analysis for metrics V , , and MET MET CPUMET , and Fig. 3 depicts the average value for these metrics and the condence intervals for a 95% condence level. From Table 6 and Fig. 3, the following observations are in order: Results from Table 6 allow withdrawing the same conclusions as in the previous section: there are signicant differences among methods accuracy and efciency. Again, the WLS estimator is the least accurate, whereas the DWLS and RWLS procedures provide the most accurate results. Regarding the required CPU time, the RWLS estimator is again the most efcient, and the DWLS algorithm is the less efcient one. The required CPU time for the RWLS estimator is approximately 75% and 72% smaller than the CPU times required by the DWLS and WLS methods, respectively.

From Table 5, note that the three F-stat values corresponding to the factor Method are higher than 3.03, denoting that there are statistically signicant differences regarding the performance of the considered methods, for numerical accuracy and
Table 5 ANOVA table for the second case study. Source Metric Method Scenario Residual Metric MET Method Scenario Residual Metric CPUMET Method Scenario Residual
V MET

Table 6 ANOVA table for the third case study. Mean-sq 2.6 106 6.3 107 9.9 108 1.7 105 4.6 107 1.8 107 2.3 102 8.5 4.3 F-stat 26.339 6.406 Fcrit 3.03 1.30 Source Metric Method Scenario Residual Metric MET Method Scenario Residual Metric CPUMET Method Scenario Residual
V MET

Squared Sum 5.2 106 6.3 105 2.0 105 3.4 105 4.6 105 3.5 105 4.7 102 8.5 102 8.4 102

DoF 2 99 198 2 99 198 2 99 198

Squared Sum 2.6 105 9.1 105 4.5 105 3.0 105 6.6 105 3.0 105 2.2 103 2.1 103 1.6 103

DoF 2 99 198 2 99 198 2 99 198

Mean-sq 1.3 105 9.1 107 2.3 107 1.5 105 6.7 107 1.5 107 1.1 103 2.1 101 8.1 100

F-stat 57.645 4.035

Fcrit 3.03 1.30

95.202 2.597

3.03 1.30

100.764 4.475

3.03 1.30

54.899 2.010

3.03 1.30

133.953 2.573

3.03 1.30

E. Caro et al. / Electric Power Systems Research 104 (2013) 917

15

Fig. 2. Results for the case study with three gross errors: performance comparison.

Fig. 3. Results for the case study with six gross errors: performance comparison.

4.4. IEEE 118-bus system: results comparison Analyzing jointly the results obtained for the three case studies above, the following conclusions can be withdrawn: 1. The estimation accuracy of the proposed RWLS algorithm is signicantly better than the one provided by the traditional WLS procedure, for a 95% condence level and considering scenarios with zero, three, and six sets of multiple gross errors. 2. For a 95% condence level, there is no signicant difference between the estimation accuracy degree of methods RWLS and DWLS. That is, the obtained estimate for both methods have the same accuracy.

3. The computational efciency of the RWLS remains unaltered with independence of the number of multiple gross errors corrupting the measurement set. This consideration can be graphically illustrated by plotting the average required CPU time for each method as a function of the number of multiple bad measurement sets (see Fig. 4). 4. Note that the most accurate estimates does not correspond to the WLS results. This is so because WLS estimation does not consider measurement dependencies. 4.5. Fourth case study: IEEE 57-bus system Finally, methods WLS, DWLS and RWLS are applied to another system: the IEEE 57-bus network.2 Results are graphically shown in Fig. 5, using the same format as in Figs. 13. From Fig. 5, the following observations are in order: 1. The estimation accuracy provided by algorithms DWLS and RWLS are similar. 2. The accuracy level of method WLS is signicantly lower than the levels provided by approaches RWLS and DWLS. 3. The required CPU time of the method RWLS is signicantly lower than the CPU times required by approaches WLS and DWLS. The saving is up to 80%.

Average required CPU time (sec.)

8 6 4 2 0

WLS DWLS RWLS

6
2 Power Systems Test Case Archive. Available at: http://www.ee.washington. edu/research/pstca/.

Number of errors
Fig. 4. Required CPU time comparison.

16

E. Caro et al. / Electric Power Systems Research 104 (2013) 917

Fig. 5. Performance of the methods: IEEE 57-bus system.

5. Conclusion The method proposed in this paper outperforms alternative algorithms reported in the technical literature in the considered cases, especially if the number of gross errors or outliers in the measurement set is large. If the proposed RWLS approach is implemented in real-world systems, a similar behavior is expected, in terms of both computational efciency and numerical accuracy. The proposed RWLS estimator takes into consideration measurement dependencies to improve accuracy and its weights are automatically readjusted to increase robustness. In summary, it provides accurate estimates, is robust against outliers, and is computationally efcient. The methods performance is tested using several realistic case study with single and multiple gross errors, and considering a high number of scenarios. These results are compared in detail using ANOVA techniques, which allows proving the outperformance of the proposed method from a statistical point of view. Since most of the state estimation algorithms used in practice are based on WLS techniques, such algorithms can be easily adapted or modied to include RWLS features.

Appendix A. ANOVA table description In this Appendix, the ANOVA table structure presented in Section 3.2 is described in detail. The rst column provides the sources of variability according to model (13): factor Method, factor Scenario and residuals. The corresponding variabilities are quantied by the Squared Sum terms, provided in the second column. These terms are computed using the following expressions:
3

SS M = n
M=1

) M y (y

(A.1)

SS D = 3
=1

y ) (y

(A.2)

SS error =
=1 M=1

M yM y + y ) (y

(A.3)

Acknowledgements Eduardo Caro would like to thank partial nancial support by Project DPI2011-23500, Ministerio de Ciencia e Innovacin, Spain.

M is the average response for the METth method considwhere y is the average response for the th scenario ering all scenarios; y is the average response considering all considering all methods; y scenarios and methods.

E. Caro et al. / Electric Power Systems Research 104 (2013) 917

17

Fig. 6. F-Snedeccor distribution.

These Squared Sum terms are divided by their corresponding degrees of freedom3 (third column), obtaining the variances for each factor, denoted as SSM , SSD , and SSerror . The aim of the ANOVA analysis is to determine whether or not the factors affect the response variable. This can be checked by comparing the above variances. The way of comparing them is by 2 /s 2 /s 2 2 M error D error computing the ratios s and s . If the performance is not inuenced by the method employed, 2 /s 2 M error the ratio s is low. On the other hand, if the performance varies depending on the method used, this ratio is high. In order to determine it a ratio is low or high, the F-Snedeccor 2 /s 2 M error is computed and distribution is used (see Fig. 6). The ratio s Method located in the X-axis. If it is lower than Fcrit , then the performances of the methods are similar (the hypothesis H0 is accepted). Method , then 2 /s 2 M error On the contrary, if the ratio s is higher than Fcrit there is statistical evidence that the methods perform different. Note that the value Fcrit depends on the degrees of freedom used and on the condence level employed. These values are provided in the last column of the ANOVA table, for a condence level of 2 /s 2 M error 95%. Note also that Fig. 6 provides an example of value s . In this case (Fig. 6), the hypothesis H0 is accepted, concluding that the methods exhibit the same performance. References
[1] A. Abur, A. Gmez-Expsito, Power System State Estimation. Theory and Implementations, Marcel Dekker, New York, 2004. [2] F.C. Schweppe, J. Wildes, Power system static state estimation. Part I. Exact model, IEEE Transactions on Power Apparatus and Systems 89 (1) (1970) 120125. [3] H.M. Merrill, F.C. Schweppe, Bad data suppression in power system static state estimation, IEEE Transactions on Power Apparatus and Systems 90 (1970) 27182725. [4] D.M. Hawkins, Identication of Outliers, Chapman and Hall, London, 1980. [5] D.A. Belsley, E. Kuh, R.E. Welsch, Regression Diagnostics: Identifying Inuential Data and Sources of Multicollinearity, John Wiley & Sons, New York, 1980. [6] R.D. Cook, S. Weisberg, Residuals and Inuence in Regression, Chapman and Hall, New York, 1982. [7] A.C. Atkinson, Plots, Transformations, and Regression: An Introduction to Graphical Methods of Diagnostic Regression Analysis, Oxford: Clarendon Press, New York, 1985. [8] S. Chatterjee, A.S. Hadi, Sensitivity Analysis in Linear Regression, John Wiley & Sons, New York, 1988.

[9] V. Barnett, T. Lewis, Outliers in Statistical Data, 3rd edition, John Wiley & Sons, New York, 1994. [10] Conrmation of multiple outliers in generalized linear and nonlinear regressions, Computational Statistics & Data Analysis 25 (1) (1997) 5565. [11] J.W. Wisnowski, D.C. Montgomery, J.R. Simpson, in: A comparative analysis of multiple outlier detection procedures in the linear regression model, Computational Statistics & Data Analysis 36 (3) (2001) 351382. in: Estimation of missing values in possibly partially nonstationary [12] A. Luceno, vector time series, Biometrika 84 (1997) 495499. in: Detecting possibly non-consecutive outliers in industrial time [13] A. Luceno, series, Journal of the Royal Statistical Society, Series B: Statistical Methodology 60 (1998) 295310. [14] A. Luce no, in: Multiple outliers detection through reweighted least deviances, Computational Statistics & Data Analysis 26 (1998) 313326. [15] N. Billor, S. Chatterjee, A.S. Hadi, in: Iteratively re-weighted least squares method for outlier detection in linear regression, Bulletin of the International Statistical Institute 1 (2001) 470472. [16] E. Castillo, A.S. Hadi, A.J. Conejo, A. Fernndez-Canteli, in: A general method for local sensitivity analysis with application to regression models and other optimization problems, Technometrics 46 (4) (2004) 430445. [17] T.-C. Cheng, in: Robust diagnostics for the heteroscedastic regression model, Computational Statistics & Data Analysis 55 (2011) 18451866. [18] P.J. Rousseeuw, A.M. Leroy, Robust Regression and Outlier Detection, John Wiley & Sons, New York, 1987. [19] P.J. Rousseeuw, K. Van Driessen, in: A fast algorithm for the minimum covariance determinant estimator, Technometrics 41 (1999) 212223. [20] E. Caro, A.J. Conejo, in: State estimation via mathematical programming: a comparison of different estimation algorithms, IEE Proceedings-Generation Transmission and Distribution 6 (6) (2012) 545553. [21] R. Mnguez, B.G. Reguero, A. Luce no, F.J. Mndez, in: Regression models for outlier identication (hurricanes and typhoons) in wave hindcast databases, Journal of Atmospheric and Oceanic Technology 29 (2012) 267285. [22] L. Guangyi, Y. Erkeng, Y. Song, in: Novel algorithms to estimate and adaptively update measurement error variance using power system state estimation results, Electric Power Systems Research 47 (1) (1998) 5764. [23] S. Zhong, A. Abur, in: Auto tuning of measurement weights in WLS state estimation, IEEE Transactions on Power Systems 19 (4) (2004) 20062013. [24] R. Pires, A. Simoes Costa, L. Mili, in: Iteratively reweighted least-squares state estimation through givens rotations, IEEE Transactions on Power Systems 14 (4) (1999) 14991507. [25] R. Jabr, B. Pal, in: Iteratively reweighted least-squares implementation of the WLAV state-estimation method, IEE Proceedings-Generation Transmission and Distribution 151 (1) (2004) 103108. [26] D. Wiens, in: Designs for weighted least squares regression, with estimated weights, Statistics and Computing (2012) 111. [27] E. Caro, A.J. Conejo, R. Mnguez, in: Power system state estimation considering measurement dependencies, IEEE Transactions on Power Systems 24 (4) (2009) 18751885. [28] E. Caro, A.J. Conejo, J. Morales, R. Mnguez, in: Calculation of measurement correlations using point estimate, IEEE Transactions on Power Delivery 25 (4) (2010) 20952103. [29] E. Caro, A.J. Conejo, R. M nguez, M. Zima, G. Andersson, in: Multiple bad data identication considering measurement dependencies, IEEE Transactions on Power Systems 26 (4) (2011) 19531961. [30] P.W. Holland, R.E. Welsch, in: Robust regression using iteratively reweighted least-squares, Communications in Statistics Theory and Methods A6 (1977) 813827. [31] W.J.J. Rey, Introduction to Robust and Quasi-Robust Statistical Methods, Springer, Berlin, Heidelberg, 1983. [32] H. Xuming, S. Portnoy, in: Reweighted LS estimators converge at the same rate as the initial estimator, The Annals of Statistics 20 (4) (1992) 21612167. [34] H. Scheff, The Analysis of Variance, Wiley, New York, 1959. [35] S. Chakrabarti, E. Kyriakides, G. Ledwich, A. Ghosh, in: On the inclusion of PMU current phasor measurements in a power system state estimator, IET Generation, Transmission & Distribution 4 (10) (2010) 11041115. [36] E. Caro, R. Singh, B.C. Pal, A.J. Conejo, R.A. Jabr, in: Participation factor approach for phasor measurement unit placement in power system state estimation, IEE Proceedings-Generation Transmission and Distribution 6 (9) (2012) 922929.

3 The number of degrees of freedom is the number of independent values in the nal calculation of a statistic that are free to vary.