1608

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 4, NOVEMBER 2006

Placement of PMUs to Enable Bad Data Detection in State Estimation

Jian Chen, Student Member, IEEE, and Ali Abur, Fellow, IEEE
AbstractAs the phasor measurement units (PMUs) start populating the substations, their benets for different power system application functions are being debated. One of these functions is the bad data processing that is commonly integrated into the state estimation. Bad data detection is closely related to the measurement redundancy in that bad data appearing in critical measurements cannot be detected. Such measurements, however, can be transformed into redundant measurements by adding few PMUs at strategic locations. In this paper, it will be shown that with few extra PMUs, the bad data detection and identication capability of a given system can be drastically improved. Description of the placement algorithm is given, and case studies carried out on different size test systems are presented. Index TermsBad data processing, critical measurements, meter placement, phasor measurement units (PMUs), state estimation.

I. INTRODUCTION NE of the essential benets of using a state estimator is the ability to detect, identify, and correct measurement errors. This procedure is referred to as bad data processing. Depending upon the state estimation method used, bad data processing may be carried out as part of the state estimation or as a post-estimation procedure. Irrespective of the state estimation method employed, detectability of bad data depends upon the measurement conguration and redundancy. In a given power system, a measurement is classied as either critical or redundant. Redundant measurements can be removed from the measurement system without causing the system to become unobservable. When a redundant measurement is erroneous, this can be detected by statistical tests based on measurement residuals. Removal of critical measurements, however, will lead to an unobservable system, and errors in these types of measurements cannot be detected. A well-designed measurement system should not contain any critical measurements so that bad data processing can be accomplished. It is a simple matter to identify all critical measurements in a given system by methods such as those presented in [1] or [2]. If one wishes to avoid critical measurements, then this can be accomplished by adding few measurements at the proper locations.

In principle, adding any type of measurement will improve redundancy; however, adding a phasor measurement unit (PMU) will potentially make the most impact. PMUs are becoming more common in the substations due to their versatile utilization. A PMU placed at a given bus can provide the voltage phasor at the bus and the phasor currents on several or all lines incident to that bus [3], [4]. It is simple to incorporate synchronized phasor measurements into state estimators along with conventional measurements, and this can be shown to improve state estimation performance [5]. Given a power system that is fully observable by existing measurements, few extra PMUs can be used to convert any existing critical measurements into redundant ones and thus making all bad data in the system detectable. In this paper, a strategic PMU placement algorithm developed for this purpose will be presented. The main goal is to improve the bad data processing capability of state estimators by taking advantage of the PMU technology. The developed PMU placement procedure will be outlined and illustrated by examples. Furthermore, simulation results showing the acquired bad data detection capability as a result of introducing new PMUs will be presented. II. MEASUREMENT JACOBIAN WITH PMUS A simplied linear approximation model is helpful to analyze the network observability. For a given network, the linear model for the real power measurements and bus phase angles can be expressed in the following form: (1) where real power measurement vector, which contains real power ow and real power injection measurements; bus phase angle vector; measurement Jacobian for the real power measurements versus all bus voltage angles; real power measurement error vector. Note that the system observability is independent of the branch parameters as well as the operating state of the system. Thus, for simplicity of observability analysis, all branch im, and all bus voltages are pedances are assumed as j1.0 assumed as 1.0 A PMU can measure both the voltage phasor of its own bus and current phasors on incident branches. Based on the above

Manuscript received August 2, 2005; revised March 15, 2006. This work was supported by the National Science Foundation under Grant No. ECS-0400760. Paper no. TPWRS-00471-2005. J. Chen is with the Department of Electrical Engineering, Texas A&M University, College Station, TX 77843 USA (e-mail: jianchen@ee.tamu.edu). A. Abur is with the Department of Electrical and Computer Engineering, Northeastern University, Boston, MA 02115 USA (e-mail: abur@ece.neu.edu). Digital Object Identier 10.1109/TPWRS.2006.881149

0885-8950/$20.00 2006 IEEE

CHEN AND ABUR: PLACEMENT OF PMUS TO ENABLE BAD DATA DETECTION IN STATE ESTIMATION

1609

assumption, the real part of current phasor can be approximately written as

A. Identication of Critical Measurements A critical measurement is a measurement whose elimination from the measurement set will result in an unobservable system. A power system will be observable only if the measurement Jacobian has full rank. Hence, for an observable n-bus system will be with m-measurements, the measurement Jacobian matrix with a column rank of . a Critical measurements can be identied by various ways that are well documented in the literature [1], [2]. A numerical method that allows easy identication of critical measurements is outlined below.: into its lower trapezoidal and upper Step 1) Decompose triangular factors by using the PetersWilkinson method [6]

So, the linear model measurement Jacobian can be written as follows:

. . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . . . . . . . . . . .

(3)

III. FORMULATION OF THE PMU PLACEMENT PROBLEM The proposed procedure for the placement of PMUs in order to convert all critical measurements into redundant ones will be described by way of a small tutorial example. The benets of having this new measurement conguration are that the system will no longer be vulnerable to the loss of any of those previously critical measurements, and now that they are no longer critical, if they carry bad data, they can be detected. The procedure is formulated as a three-step solution, involving the following stages: 1) identication of critical measurements; 2) nding candidate PMUs that can transform each critical measurement into a redundant one; 3) choosing the optimal set of PMUs among the candidates. that Let us start by forming the measurement Jacobian is partitioned into two submatrices

(2)

where is the permuted matrix derived from by suitably exchanging rows, is the perand are lower and upper mutation matrix, , and the rows triangular factors of dimension of , , and correspond to the essential measurements, redundant measurements, and the candidate PMU measurements, respectively. Essential measurements are dened as the minimum set of measurements that will render the entire system observable [7]. The row exchanges during the decomposition is equivalent to reordering the corresponding measurements. It should be noted that only the rows corresponding to existing measurements will be exchanged because the system is observable. The detailed decomposition method is described in the Appendix. and will be nonsingular for an observable Step 2) can be checked system. Hence, the rank of by checking the rank of the transformed factor given by

where (4) rows correspond to the existing measurements; rows correspond to the phasor measurements associated with all candidate PMUs. Note that the system is assumed to be already observable. If this is not true, a proper PMU placement program such as the one described in [4] can be employed to improve the measurement system and make it fully observable. Candidate PMUs will be at all system buses unless any of the system substations are excluded due to constraints on PMU installations there. Also, note that the reference bus phase angle is excluded from the state vector , and the corresponding column . will be missing in

represents the identity matrix of diwhere . mension Note that since has full rank, and its inverse is multiplied from the right in (4), the row identities . will be preserved in the transformed matrix will still correspond to Hence, each row of the respective measurements. If a column of contains no nonzero element, then the measurement corresponding to its row index will be identied as critical.

1610

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 4, NOVEMBER 2006

measurements associated with this PMU, lowing form:

will take the fol-

Fig. 1. Five-bus test system. (Color version available online at http://ieeexplore.ieee.org.)

The procedure can be illustrated using a simple example. Consider the small ve-bus power system and its measurement conguration shown in Fig. 1. The critical measurements in this system can be easily identied by applying the above procedure. Calculating the transformed matrix of (4) for this example critical

PMU at bus

critical.

As shown above, a PMU placed at bus 1 is assumed to provide three phasor measurements, namely, the voltage phase angle and the current phasors and . Thanks measurement to these three phasor measurements, the critical measurements and will no longer be critical measurements with the addition of one PMU at bus 1. C. PMU Placement Problem The nal step involves the optimal selection of the minimum required number of PMUs from the list of candidates identied at the previous step. that relates PMUs to their Consider an incidence matrix associated phasor measurements and dened as follows: if PMU at bus i includes the meas. j otherwise. For the ve-bus system in Fig. 1, assume that there are ve candidate PMUs corresponding to all ve buses in the system. Each candidate PMU has one voltage phasor measurement and matrix several current phasor measurements, yielding the shown in the equation at bottom of the page. Next, let us dene another matrix that is formed using the in (4) that correspond to binary form of the columns of the critical measurements. It is dened as if otherwise. and meas. ``j" is critcal

Note that the columns corresponding to the two critical measurements and are both null except for the nonzero value of 1 in their respective columns 1 and 4.

B. Identifying the Candidate PMUs for Eliminating Critical Measurements Once the critical measurements are identied, a set of candidate PMUs is selected for each critical measurement. These candidates are chosen in such a way that introduction of any one of them will transform the corresponding critical measurement into a redundant one. Such a candidate list can be obtained easily by simply tracing the column for the critical measurement in the transformed maof (4). Those rows containing nonzero values will intrix dicate the measurements associated with the candidate PMUs. Considering the same ve-bus example above, let us assume that there is a candidate PMU at bus 1 only. Including the phasor

CHEN AND ABUR: PLACEMENT OF PMUS TO ENABLE BAD DATA DETECTION IN STATE ESTIMATION

1611

Note that the matrix provides a compact representation of those critical measurements (columns)that will be transformed into redundant ones by a given candidate PMU (row). For the same ve-bus system that has only two critical measurements (corresponding to columns 1 and 4), the matrix will look as follows:

This will also guarantee network observability in case of the loss of any of these previously critical measurements, in other words, will transform all critical measurements into redundant measurements. An optimization problem whose objective is to place a minimum number of PMUs at best locations so that all critical measurements are transformed into redundant ones can now be formulated. Due to the types of variables involved, it will be an integer programming (IP) problem stated as follows: Minimize Subject to where number of buses in the system; cost of installing a PMU at bus . Thus, the solution vector X of the optimization problem given in (7) above will provide the strategic placements for PMUs, which will eliminate all critical measurements. Solution of this problem for the above ve-bus example yields any one of the buses 1, 2, or 5 as candidates for PMU placement. Only one PMU at any of these three buses can transform both critical measurements into redundant ones. (7)

Since each PMU supplies more than one phasor measurements (typically a voltage phasor and several current phasors), the following matrix product will yield the relationship between candidate PMUs and critical measurements:

D. Algorithm A PMU placement algorithm can then be developed based on the above discussion. The following steps provide one such algorithm. Step 1) Build the measurement Jacobian and the incidence matrix B based on the network data, available measurements, and the assumed PMU candidate buses. and use back-substitutions to obtain Step 2) Factorize . the transformed matrix Step 3) Identify the critical measurements as described in Section III-A. Using this information, form the matrix R. Step 4) Solve the IP problem given by (7). The optimal locations for the PMUs will be those corresponding to the bus indexes of the nonzero entries in the integer solution vector X. E. Extension of Algorithm 1) Optimal Placement for Mixed Measurements: The above presented procedure can be easily revised so that not only PMUs but also conventional measurements are considered as candidate measurements. The required revisions to the algorithm are given below. Since the candidate measurement set will now include two sets of measurements, namely, PMUs and conventional meawill also include surements, the corresponding Jacobian two corresponding submatrices as follows:

(5) Calculating for the above example yields

Note that each row of corresponds to a critical measurement, and each column of corresponds to a candidate PMU. Thus, the following constraint will ensure that for every critical measurement, there will be at least one PMU to make it redundant

(6) where vector, whose entries are all equal to 1; X binary (0/1) vector, whose entries is dened as

if a PMU is installed at bus i otherwise.

1612

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 4, NOVEMBER 2006

The measurement Jacobian form:

will then take the following

where total number of candidate PMUs and candidate conventional measurements; installation cost of a candidate PMU or candidate conventional measurement;

where , , and correspond to the existing, candidate PMU, and candidate conventional measurements, respectively. will be decomApplying the PetersWilkinson method, posed into its lower rectangular and upper triangular factors

binary (0/1) vector, whose entry may correspond to a PMU or a conventional measurement. If the corresponding PMU or measurement is placed, then the entry will be 1, otherwise 0.

where, as in (3), and represent the lower and upper trianand correspond to the essengular factors. The rows of tial and redundant measurements, which are already installed in and correspond to the the system. The rows of candidate PMU and candidate conventional measurements, respectively. Applying the same transformation as done in (4), submatrices , , and will be formed

2) Improving Local Redundancy: Placement of PMUs or conventional measurements for bad data detection is desirable but may not be sufcient for making all bad data identiable. Hence, local redundancy around each bus may be required to remain above a desired level in order to accomplish this. Such a solution can be obtained by making minor changes in the above optimization formulation by including a new redundancy vector W, whose entries are equal to the number of nonzero entries in the corresponding column of . The following revised optimization formulation will yield a solution that will help the system to approach the desired level of redundancy : Minimize Subject to where W vector that corresponds to the existing are redundancy in the system, whose entries decided by the number of nonzero elements in ; corresponding column of modied constraint matrix (10)

Note that the critical measurements can still be identied as . done in Section III-A by checking the null columns in The constraint matrix F will be revised as (8) where ; ; and binary form of the columns of and that correspond to the critical measurements; same incidence matrix as dened in Section III-C. The IP problem of (7) can be formulated using the revised constraint matrix as follows: Minimize Subject to (9)

Note that in forming and , not only those columns corresponding to critical measurements but all columns of and will be considered. They will be transformed into binary form as done in Section III-C. Hence, they will be computed as follows:

is the vector representing the desired redundancy level. Its entries are all equal to the scalar , which is the desired level of local redundancy. An value of 1.0 indicates that the system will have no critical measurement and can detect all bad data. It also means that the system observability will be ensured under loss of any single measurement. An value of 2.0 will imply a more reliable measurement design, where no critical measurement exists, and furthermore, the number of critical pairs are signicantly reduced.

CHEN AND ABUR: PLACEMENT OF PMUS TO ENABLE BAD DATA DETECTION IN STATE ESTIMATION

1613

Fig. 2. Network diagram and measurement conguration for the IEEE 57-bus system. (Color version available online at http://ieeexplore.ieee.org.)

IV. SIMULATION RESULTS A program is developed in order to simulate the proposed method and evaluate its performance. IEEE 57-bus and 118-bus test systems are used for the simulations. The results are presented in three parts. Parts A and B show the placement procedure applied to these two systems. Part C illustrates the benets of such placement for the 118-bus system, where a previously undetected piece of bad data becomes detectable as a result of PMU placement. In Part D, simulation results about improving measurement redundancy and utilization of conventional candidate measurements are presented. The IP problem is solved using the TOMLAB Optimization Toolbox [8]. All simulation cases use the following assumptions. 1) There are no bus limitations for the placement of PMUs, that is, every bus is assumed to be a candidate for PMU placement. This assumption can be easily relaxed without modifying the basics of the presented approach. 2) Installation costs of all PMUs are the same. Due to lack of any such data on costs, all PMUs are given the same weight in the optimization. However, any other cost assignment will work by modifying the cost array. 3) Current phasors along all branches incident to a bus will be measured by the PMU at that bus. These assumptions are made for convenience in carrying out the simulations. The proposed method can work well also without these assumptions. A. IEEE 57-Bus System model, only the real power Using the decoupled linear injection and line ow measurements are considered here. They are chosen to make the entire system observable. The simulated system has a total of 33 real power ow measurements and 32 injection measurements, as shown in Fig. 2. Hence, the number , and the number of of state variables is . measurements is

TABLE I CRITICAL MEASUREMENTS FOR THE IEEE 57-BUS SYSTEM

This measurement conguration contains 13 critical measurements, which are listed in Table I. By using the method presented in the previous section, the optimal placement for PMUs is obtained. The optimal solution yields phasor measurement units at buses 34 and 46. As a result of including these two PMUs, no critical measurements will exist in this system. B. IEEE 118-Bus System In this case, a measurement set is chosen to make the system fully observable. A total of 39 injection measurements and 114 power ow measurements are chosen, yielding a total of state variables and measurements. The locations of these measurements cannot be shown graphically due to the large size of the system. This measurement set contains 29 critical measurements, which are listed in Table II.

1614

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 4, NOVEMBER 2006

TABLE II CRITICAL MEASUREMENTS FOR THE IEEE 118-BUS SYSTEM

61, 80, 83, 86, and 94. Repeating the state estimation with the updated measurement set yields an objective function value of 44443, which is clearly above the threshold of 228.58 (based on 95% condence level and 195 degrees of freedom). Hence, the Chi-squares test will ag the presence of bad data. Furthermore, the largest normalized residual test is also performed and the gross error is identied, having the largest normalized residual . of D. Redundancy Improvement by Both PMUs and Conventional Measurements The extension to the algorithm described in Section III-E2 is implemented and tested. The desired redundancy level is set as 2.0, which will yield a smaller number of critical pairs. In addition to the PMUs, some injection measurements are also considered as candidate measurements in this particular simulation. It is assumed that any bus without an injection measurement is a candidate for injection measurement placement. The cost of a PMU installation is assumed to be twice the cost of placing an injection measurement. The results of simulations are shown in Table III. As the required redundancy increases, so does the number of measurements to be placed. This is evident from the results of Table III, where for the case of the 118-bus system, ve PMUs and 18 extra injection measurements are needed when a redundancy of 2.0 is considered. V. CONCLUSIONS

TABLE III WHEN  = 2:0 IS CONSIDERED FOR IEEE 118-BUS SYSTEM

The solution of the optimization problem yields a total of 13 PMUs to be placed at the following buses in order to transform all critical measurements into redundant . ones: C. Bad Data Detection Capability The same 118-bus system is used to illustrate bad data processing benets gained as a result of PMU placement. In order to show these benets, one of the critical measurements in the system is corrupted by gross errors. State estimation is executed, and the results of bad data processing reveal no bad data. This is expected due to the criticality of the measurement. After applying the proposed PMU placement procedure, the gross error in the same measurement can be detected. Next, the measurement set is populated with several conventional measurements. These are two voltage, 39 pairs of injection, and 114 pairs of power ow measurements. All measurements in the system are assumed to have the same standard per unit. Chi-squares test [7] is used deviation of for bad data detection. A single bad measurement is simulated for the real power ow measurement from bus 3 to bus 5. The true real power is replaced by a gross error ow value of . However, the state estimation fails to as detect this bad data, yielding an objective function value of 0.0260. This value is far below the Chi-squares test threshold of 93.94 (corresponding to 95% condence level and 73 degrees of freedom). After introducing new PMUs using the proposed approach, 13 PMUs are installed at buses 7, 10, 12, 28, 32, 35, 52, 59,

The essential objective of this paper is to illustrate the benets of adding a few phasor measurement units for state estimation, even when the system is initially fully observable without these devices. It is argued that PMUs will provide increased bad data detection and identication capability, which may come in handy during contingencies and existence of bad data in low redundancy pockets of the system. The problem of PMU placement is formulated and solved as an IP problem. The solution will provide the minimum number of strategically located PMUs that will eliminate measurement criticality in the entire system. This implies that any bad data appearing on any single measurement will be detectable. It is shown that depending on the measurement conguration and the system topology, this goal can be achieved by using only a few extra PMUs to transform several critical measurements into redundant ones in the system. The placement problem is then extended to also incorporate conventional measurements as candidates for placement. Furthermore, it is also shown that the same formulation can be used to determine optimal locations when a desired level of local redundancy is considered in the system. This allows design of measurement systems with different degrees of vulnerability against loss of measurements and/or bad data. Several simulation results are provided to illustrate the proposed placement procedure and its effectiveness in enhancing bad data processing. APPENDIX Assume that is a rectangular matrix whose rank is . Therefore, it can be decomposed into a lower-trape-

CHEN AND ABUR: PLACEMENT OF PMUS TO ENABLE BAD DATA DETECTION IN STATE ESTIMATION

1615

zoidal matrix and upper-triangular matrix by using Gaussian elimination, though some rows may need to be exchanged during the decomposition

Because the matrix is supposed to be full rank, the decomposition should complete after steps and generate the upper and lower triangular factors

where lower-trapezoidal matrix, and its diagonal elements are all equal to 1; upper-triangular matrix; permutation matrix that nonzero elements are corresponding to the rows permutation; derived from by suitably permuting its rows. rectangular matrix, steps are For the assumed full rank needed to complete the decomposition. Before the th step, we of the form should have a matrix

REFERENCES
[1] A. Simoes-Costa, T. S. Piazza, and A. Mandel, Qualitative methods to solve qualitative problems in power system state estimation, IEEE Trans. Power Syst., vol. 5, no. 3, pp. 941949, Aug. 1990. [2] A. G. Exposito and A. Abur, Generalized observability analysis and measurement classication, IEEE Trans. Power Syst., vol. 13, no. 3, pp. 10901096, Aug. 1998. [3] T. L. Baldwin, L. Mili, M. B. Boisen, and R. Adapa, Power system observability with minimal phasor measurement placement, IEEE Trans. Power Syst., vol. 8, no. 2, pp. 707715, May 1993. [4] B. Xu and A. Abur, Observability analysis and measurement placement for system with PMUs, in Proc. IEEE Power Eng. Soc. Power Systems Conf. Expo., New York, Oct. 1013, 2004. [5] R. Zivanovic and C. Cairns, Implementation of PMU technology in state estimation: an overview, in Proc. IEEE 4th AFRICON, Sep. 1996, vol. 2, pp. 10061011. [6] G. Peters and J. H. Wilkinson, The least-squares problem and pseudoinverses, Comput. J., vol. 13, no. 4, pp. 309316, Aug. 1970. [7] A. Abur and A. Gomez Exposito, Power System State Estimation: Theory and Implementation. New York: Marcel Dekker, 2004. [8] The TOMLAB Optimization Environment. [Online]. Available: http:// tomlab.biz/. Jian Chen (S05) received the B.S. and M.S. degrees from Xian Jiaotong University, Xian, China, in 2001 and 2004, respectively. He is currently pursuing the Ph.D. degree at the Department of Electrical Engineering, Texas A&M University, College Station. Ali Abur (F03) received the B.S. degree from Orta Do u Teknik niversitesi, g Ankara, Turkey in 1979 and the M.S. and Ph.D. degrees from The Ohio State University, Columbus, in 1981 and 1985, respectively. He is currently a Professor and Chair of the Department of Electrical and Computer Engineering, Northeastern University, Boston, MA. His research interests are in computational methods for the solution of power system monitoring, operation, and control problems.

where, is the element of , and is the element of . is changed in every step and should Notice here that each actually carry a superscript , but it is omitted for convenience. The th step of the decomposition is then as follows. is zero, check the with in the th column 1) If . Then exchange and nd the rst nonzero element the th and th rows. with and store it in 2) Compute . position to , compute a new 3) For each value of from given by and overwrite it on the old .