1. Introduction Power system state estimation (SSE) has received increasing attention since its inception in the 1970s. With the continuous improvement of dispatch automation levels, various dispatch automation software urgently require accurate real-time databases as the basis for their calculations. SSE has become an important component of EMS or DMS management software. In the actual operation of power systems, due to errors in measurement and measurement channels, as well as potential interference, some measurements may exhibit significant errors. The presence of poor data can cause great difficulties in estimation, even leading to estimation failure. Therefore, reliable detection of poor data is crucial for the practical application of SSE. Poor data detection generally refers to dividing real-time measurement data into two datasets: reliable data and suspicious data. Then, suspicious data is identified to find all poor data, ensuring successful estimation. Existing methods for detecting defective data can be broadly categorized into two types: one is quantity residual detection, which is computationally simple and intuitive, but suffers from problems such as "residual contamination" and "residual overwhelming," making it unreliable for detecting multiple defective related data; the other is quantity abrupt change detection, which can accurately detect multiple defective related data when the previous estimation result is correct, the network structure remains unchanged, and the load change is small. This paper, based on these two detection methods, proposes fundamental criteria for automatically selecting the appropriate detection method for the current network state based on real-time network information. [b]2 Practical methods for detecting defective data[/b] 2.1 Practical detection methods and improvements (1) Standard residual detection method After obtaining the state variables according to the least squares state estimation algorithm, the standardized residuals can be calculated according to the residual equation as the basis for judging defective data. The residual equation can be written as r＝w·ν (1) where r is the residual (m dimension); w＝E－H(HTR－1 H)－1HTR－1 is the residual sensitivity matrix (m×m dimension); E is the identity matrix; H is the Jacobian matrix at the point; R－1 is a diagonal matrix, which is the measurement weight matrix; ν is the measurement error. The covariance matrix of the residual r is ErrT = wR (2). Take D = diag[wR], and its diagonal element dii is the variance of the corresponding residual component. Define the standardized residual as follows: [img=388,83]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dwjs/2001-10/18-1.jpg[/img] According to the hypothesis testing method, the false detection probability is set to 0.05. Measurements with a standard residual greater than 2.81 are considered suspicious measurements. The limitation of this method is that it cannot overcome the phenomena of "residual contamination" and "residual flooding". This paper uses this standard residual detection method for situations where the network structure changes and the criterion for detecting abrupt changes is no longer applicable due to rapid changes in system load. That is, this algorithm is used as the main detection method when the system load changes. (2) Measurement mutation detection This method was originally used for pre-estimation detection. Here, in order to facilitate the uniformity of the applicable conditions of the detection method and to appropriately increase the role of the correlation between redundant measurements in the current sampling in the discrimination of bad data, a post-estimation mutation detection method combining residuals is proposed. [img=381,261]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dwjs/2001-10/18-2.jpg[/img] Since the true value of the state x cannot be obtained, it is assumed that the state estimation result at the previous sampling time is correct. In this way, the one-step prediction value of the state estimation result of tK-1 times can be used as the true value of tK times detection and substituted into equation (5) to calculate the measurement error. It is assumed that the measurement error ν follows a normal distribution of (0, δ2), and δ is the measurement standard deviation. According to the hypothesis testing method, the false detection rate Pe = 0.05, that is, the measurement of |ν| < 2.81δ is considered to be a reliable measurement. Since Equation (5) omits the higher-order terms of Δx, it weakens the role of large mutations in detection and appropriately increases the influence of residuals, which has a certain inhibitory effect on the measurement mutation caused by normal fluctuations in system load. This method is based on the assumption that the network structure remains unchanged and the system operates smoothly. However, actual systems often have switching operations and rapid increases and decreases in load. Therefore, it is necessary to lock out the mutation detection when the assumption is not met. This is the limitation of this method. 2.2 Autonomously Selected Detection Method The two detection methods introduced above can be used in practical applications, but they have their own applicable conditions. It is difficult to detect measurements that exceed their applicable conditions. In view of the limitations of various detection methods, this paper proposes a practicality criterion for the estimator to autonomously select the detection method according to the actual measurement conditions. The principle and reason for autonomous selection are as follows: (1) Principle 1 If the system has switching operations and the network structure changes, sample sequentially to make state estimation. At this time, standard residual detection is used until the changes in the system state quantities estimated in the two adjacent estimations are not significant. Then, in subsequent estimations, the measurement mutation detection is enabled and the standard residual detection is locked out. Although the standard residual detection cannot identify related multiple bad data, since the distribution of error in sampling is random, the situation where the same related multiple bad data appears in two adjacent state estimates after the switching operation is extremely rare. Therefore, it is feasible to use the system state quantity determined by the above criteria as the benchmark value for mutation detection. (2) Principle 2 If the suspicious measurement detected by the mutation detection method does not contain injection type measurement and does not contain related branch measurement without injection measurement node, it is considered a reliable detection. The reason is that at this time the system control variables (power supply, load) have not changed, and the system state variables will not change significantly under the condition that the network structure remains unchanged. Therefore, the applicable conditions for measurement mutation detection are met. When the system control variables (power supply, load) change rapidly, the system state quantity also changes significantly. The mutation detection method is no longer applicable. At this time, there is no switching operation. Therefore, how to correctly distinguish between normal rapid load change and whether injection type measurement has bad data becomes the key of this method. The criteria for judging such cases will be discussed in detail below. **3. Criteria for Determining Rapid Changes in System Load** 3.1 Network and Measurement Information Adjacency List The network structure and measurement information can be stored in an adjacency list, the specific structure of which is shown in Figure 1. [img=342,153]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dwjs/2001-10/18-3.jpg[/img] Head Node: A sequentially numbered node in the network. The structure of the head node should include information such as whether the node is an injection-type node, whether it is equipped with injection-type measurements, and the injected active and reactive power measurement values. Node Node: A node associated with the head node. The structure of the adjacent node should include the node number, the location of the branch measurement devices with the head node number and the Node node number as endpoints, and the active and reactive power measurement values ​​of the branch. 3.2 Principles and Methods of Judgment If the suspicious measurement set detected by the mutation detection method contains injected measurements as well as related branch measurements of injected nodes without injected measurements, the presence of rapid load changes can be determined by whether the power sum of the cut set containing the injected measurements is less than a threshold. If so, the mutation detection method is locked and the residual detection method is enabled. Specific judgment method: Assuming the suspicious measurement set contains the injected quantity of node i, the associated branches of node i and the injected quantity of node i constitute a cut set, that is, all power measurements in the i-th row of the adjacency list satisfy the following formula: where Pli and Qli are the active and reactive power measurement values ​​of the branches associated with node i; Pni and Qni are the active and reactive power measurement values ​​injected into node i. If there is a branch among the branches associated with node i that has no power measurement installed at either end, assume that this branch is branch l, and its two end nodes are numbered i and j. Then, all branches associated with nodes i and j, excluding the branch in question, and the injected power of nodes i and j constitute a cut set. If there are still branches with unknown power in this cut set, the other endpoint of that branch is also added to the cut set, and so on, until a cut set with all branches having measurements is found. If this cut set satisfies [img=267,53]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dwjs/2001-10/19-1.jpg[/img], where the threshold Dp is the number of branches in the cut set × the maximum active power loss of branches in the network; the threshold Dq is the number of branches in the cut set × the maximum reactive power loss of branches in the network. If the cut set contains two or more suspicious measurements, it is considered that the load of node i has changed significantly, thus causing a significant change in the system state variables. Since this is a normal change in load or generator output, the method for detecting sudden changes in measurements is no longer applicable. Therefore, it is locked and standard residual detection is started until the difference between two adjacent estimation results is not significant, and then it is switched back to mutation detection. If equation (6) is not satisfied, it means that there is bad data in this cut set, and the mutation detection result is considered to be correct. For the associated branches in the suspicious measurement set that have not been equipped with injection measurement nodes, a detectable cut set can be found at the end with injection measurement, and the discrimination method is the same as above. [b]4 Algorithm[/b] The algorithm based on the above principles is as follows: (1) When the estimator is initially started, flag = 0; (2) Perform state estimation and determine whether there is a switch operation. If there is, flag = 0; otherwise, determine whether flag is equal to 1. If flag = 1, go to (6); (3) Start standard residual detection and J(x) detection; (4) After identification (successive estimation identification method), store the final estimated result in the database and determine the maximum value Δx of the state difference between two adjacent estimation results. If max｜Δx｜＜d (threshold), flag = 1; (5) Go to (2) and perform the next sampling; (6) Start mutation detection and obtain the suspicious measurement set K; (7) Search the adjacency table one by one for the suspicious measurements in K. If there are no injection type measurements or related branch measurements that are not equipped with injection type measurements, the estimated results are stored in the database after identification, and then go to (2); if there are such measurements, then distinguish whether the load has changed significantly according to principle 3. If it has changed significantly, flag = 0 and go to (3); otherwise, the estimated results are stored in the database after identification, and then go to (2). 5 Example Analysis A schematic diagram of a 13-node subnet of the power grid in Dagang area is shown in Figure 2, and the experimental analysis data is shown in Table 1. [img=305,196]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dwjs/2001-10/19-2.jpg[/img] [img=300,281]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/dwjs/2001-10/20-1.jpg[/img] In the 15th state estimation, the suspicious measurement set contains node-injected measurements P4 and Q4. Because branch L24 in this node's cut set has no measurement installed, we search whether the power sum of the cut sets {L32, L25, L4} containing nodes 2 and 4 is less than the threshold (power values ​​are expressed in per-unit values, where 0.001 and 0.0015 are the maximum active and reactive power losses of the branch when the network is operating normally, respectively). P4 + P23 + P25 = P4 - P32 - P52 = 0.0281 - 0.0733 - (-0.0432) = 0.002 < 3 × 0.001 Q4 + Q23 + Q25 = Q4 - Q32 - Q52 = 0.009 - 0.0447 - (-0.0318) = 0.0039 < 3 × 0.0015 At this point, it is considered that the load at the 4 nodes has changed significantly, and the abrupt change detection method is no longer applicable. Standard residual detection is initiated. [b]6 Conclusion[/b] This method can effectively utilize the advantages of existing adverse data detection methods, maximizing strengths and minimizing weaknesses. It has achieved good results in practical applications. [b]References[/b] ［1］Yu Erkeng, et al. Power System State Estimation［M］. Beijing: Hydropower Press, 1985. ［2］Li Yi. Detection and Identification of Adverse Data in Power System State Estimation［D］. Tianjin University, 1992. [3] Zhao Haitian, et al. Correlation measurement and detection method for multiple defective data [J]. Proceedings of the CSEE, 1990, 10(6). [4] Zhang Xingmin, et al. Detection and identification of multiple defective data using graph theory method [J]. Proceedings of the CSEE, 1997, 17(1). [5] Jia Peizhang, et al. Optimal estimation and its application [M]. Beijing: Science Press, 1984. [6] Yu Erkeng, Liu Guangyi, et al. Energy Management System (EMS) [M]. Beijing: Science Press, 1998.