1236

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 3, AUGUST 2012

Preventive Dynamic Security Control of Power Systems Based on Pattern Discovery Technique
Yan Xu, Student Member, IEEE, Zhao Yang Dong, Senior Member, IEEE, Lin Guan, Rui Zhang, Kit Po Wong, Fellow, IEEE, and Fengji Luo, Student Member, IEEE

AbstractThis paper presents a statistical learning-based method for preventive dynamic security control of power systems. Critical operating variables regarding system dynamic security are rst selected via a distance-based feature estimation process. An unsupervised learning procedure called pattern discovery (PD) is then performed in the space of the critical variables to extract the subtle structure knowledge called patterns. The patterns are geometrically non-overlapped hyper-rectangles, representing the system dynamic secure/insecure regions and can be explicitly presented to provide decision support for real-time security monitoring and situational awareness. By formulating the secure patterns into a standard optimal power ow (OPF) model, the preventive control against dynamic insecurities can be efciently and transparently attained. The proposed method is validated on the New England 39-bus system considering both single- and multi-contingency conditions. Index TermsDynamic security, feature estimation, knowledge discovery, pattern discovery, preventive control.

I. INTRODUCTION

ODERN power systems are being pushed to operate ever closer to their security limits due to the rapid growth of electricity demand and the unmatched infrastructure constructions. When exposed to a severe contingency, the system is more prone to losing its dynamic security, which can result in catastrophic consequences such as cascading outages and/or widespread blackouts [1]. To protect the power system against dynamic insecurities, preventive actions based on dynamic security assessment and control (DSA&C) are carried out to 1) evaluate the systems dynamic security conditions under a set of credible contingencies and 2) modify the systems

Manuscript received February 24, 2011; revised June 24, 2011, October 02, 2011, and November 30, 2011; accepted December 20, 2011. Date of publication February 27, 2012; date of current version July 18, 2012. This work was supported in part by a Hong Kong RGC GRF Grant # 515110 and the National Natural Science Foundation of China (NSFC) under the project # 50407014 and an open grant from Intelligent Electric Power Grid Key Laboratory of Sichuan Province of China. Paper no. TPWRS-00168-2011. Y. Xu, Z. Y. Dong, R. Zhang, and F. Luo are with the Centre for Intelligent Electricity Networks (CIEN), The University of Newcastle, Callaghan NSW 2308, Australia (e-mail: daniel.xu@uon.edu.au; zydong@ieee.org; r.zhang@uon.edu.au; tracyluofengji@gmail.com). L. Guan is with the Electric Power College, South China University of Technology, Guangzhou, China (e-mail: lguan@scut.edu.cn). K. P. Wong is with the School of Electrical Electronic and Computer Engineering, The University of Western Australia, Perth, Australia, and also with The University of Newcastle, Callaghan NSW 2308, Australia (e-mail: kitpo@ieee.org). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TPWRS.2012.2183898

operating state to withstand the harmful contingencies should they really occur. Conventional approaches for DSA&C are analytical, which model and solve the problems based on a large set of nonlinear differential-algebraic equations [2][5]. In recent years, the intelligent system (IS) strategy has been identied as a promising alternative to overcome some inadequacies of the conventional approaches and provide additional value [6], [7]. In the area of DSA, neural network (NN) [8], support vector machine (SVM) [9], decision tree (DT) [10], and extreme learning machine (ELM) [11] techniques have been applied with satisfactory performance. By learning from a dynamic security database, the nonlinear relationship between the power system operating parameters (input) and the corresponding security index (output) can be extracted and reformulated in an IS. In the online application phase, the systems security can be determined in real-time as soon as the input is fed. The IS-based DSA not only provides much faster online assessment speed, but also outperforms conventional approaches in terms of data requirement, generalization capacity, and extensibility [7]. Along with the well-established DSA applications, the IS technique has also shown encouraging potential in designing dynamic security controls (DSCs) [14][18]. In the literature, it is noted that most of the reported methods rely on DT technology. By exploiting the classication rules of the DT, control schemes against dynamic insecurities can be derived. In [14], a DT-based online preventive control of isolated power systems is presented. In [15], DT is utilized for predicting one-shot stabilizing controls. In [16], load shedding strategies are inferred from inverse reading of the DT. In [17], a hybrid method using NN and DT is designed for preventive generation rescheduling. In [18], the dynamic secure boundaries are approximated by a linear combination of DT rules, which are then applied for generation rescheduling and load shedding. Compared with conventional methods, the DT-based DSC strategies are faster, more exible, and interpretable. However, it is also known that the DT induction is a supervised learning process, i.e., it fully relies on the prior knowledge of the data to grow the tree. Under some particular conditions such as class imbalance, the tree tends to be sensitive to little changes of the data. Besides, when a tree is found complex/deep in structure, some subjective judgments are needed to prune the tree such that explicit rules can be acquired for use. Under a similar knowledge extraction and utilization framework, this paper applies an alternative, yet more robust, statistical learning technique called pattern discovery (PD) [19] in developing a preventive control method for transient instability prevention. PD belongs to unsupervised learning;

0885-8950/$31.00 2012 IEEE

XU et al.: PREVENTIVE DYNAMIC SECURITY CONTROL OF POWER SYSTEMS BASED ON PATTERN DISCOVERY TECHNIQUE

1237

in an unbiased and comprehensive manner, it statistically discovers the hidden structure in a database and provides objective, transparent, and interpretable knowledge called patterns for specic use [19][21]. The patterns are geometrically non-overlapped hyper-rectangles in a Euclidean space, easy to present and interpret; when discovered in a power system critical feature space, they can represent the dynamic secure/insecure regions, providing decision support for real-time security monitoring and situational awareness. By explicitly formulating the patterns into a standard optimal power ow (OPF) model, the preventive control can be achieved transparently and efciently. In the rest of this paper, PD is introduced in Section II; the proposed methodology is presented in Section III; as a case study, Section IV examines the method on the New England 39-bus system; and conclusions are drawn in Section V. II. PATTERN DISCOVERY PD is proposed by Wong et al. [19], [20]; as an efcient knowledge extraction technique, it is unique for being able to discover nonlinear and multimodal patterns of high order very fast, and rank them according to their statistical signicance for interpretation, comparison, and assessment so that greater understanding of the data can be achieved and thereby better decisions can be made. The PD has been successfully applied to numerous statistical learning problems from both academic and engineering areas, mitigating data rich information poor (DRIP) embarrassment effectively [19][21]. In previous work, PD has been employed for rule-based DSA [12], [13]. A. Overview Based on the argument that the knowledge for a certain class (or group) is the signicant event associations inherent in the data of that class or group, PD aims to search the statistically signicant subset in a Euclidean space in an unbiased and exhaustive manner [19]. Generally, the discovery process consists of residual analysis and optimization, where the former identies signicant organized information by statistically scaling the degree of the difference between actual occurrence and expected occurrence of instances, and the latter is responsible for searching out all the subtle information in the space. The fundamentals and computation procedure of PD are introduced as follows, and more mathematical details and proofs could be found in [19] and [20]. B. Key Denitions Consider a continuous data set in the -dimensional Euclidean space , let represent its feature set, and each feature , takes on values from its domain . The following denitions are made for PD [19], [20]. Event: an event, , is a Borel subset [28] of , while a Borel subset geometrically forms an -dimensional hyper-rectangle in , dened by (1)

where is a one-dimensional semi-closed interval along the th feature, . Volume: the volume of an event, , is the hyper-volume occupied by the Borel subset. Let represent the length of the th interval of event , , the volume of is calculated by (2) Observed frequency: the observed frequency of an event , , is the actual number of instances that fall inside the volume occupied by . Pattern: a pattern is a statistically signicant event. Let be a test statistic corresponding to a specied discovery criterion and be the critical value of the statistical test at a signicant level of . An event is considered to be signicant, i.e., a pattern, if it satises the condition (3) to test the signicance of the Residual: as the statistic pattern candidates [19], [20], the residual of an event is the difference between its actual occurrence, i.e., observed frequency, and its expected occurrence: (4) where is the expected occurrence, or expected frequency, under the pre-assumed model estimated by the given data set. C. PD by Residual Analysis and Recursive Partitioning Typically, the expected frequency is estimated under uniform random distribution [20] (there are also other criteria for this [19]). This provides implications that if instances are randomly distributed in the space, there is no signicant structure information in the space. Equivalently, the larger the residual, i.e., the difference between observed frequency and uniform random distribution, is of an event, the more structure information exists in it. Based on above denitions, PD is a process of searching all the signicant events in the instance space and can be generally viewed as an optimization problem [19]. For this paper, a residual analysis combined with recursive partitioning procedure is adopted to discover patterns [20]; the procedure consists of recursively partitioning the instance space with residual evaluation of each hyper-rectangle, until all the signicant events (patterns) are identied. Its main computation steps are as follows: 1) Divide the instance space into events, where is the number of feature dimension and is the number of partitions for each feature. 2) Rene the boundaries of each event, by adjusting the event boundaries to coincide with the minimum and maximum coordinates of the contained events. This step makes the events non-overlapped on boundaries except that they have overlapped instances. 3) Calculate the residual value of each event using the following equation [20]: (5)

1238

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 3, AUGUST 2012

provide decision support for real-time dynamic security monitoring and situational awareness. Once an insecure operating point (OP) is detected through DSA, the preventive control can be activated by driving the insecure OP into the secure region. To be systematic and economical, the computation process consists of formulating a pattern as explicit generator output constraints into a standard OPF model and solving the OPF. In doing so, the preventive control is transparent and efcient. To be specic, this paper focuses on transient stability, but it should be noted that the methodology could also be extended for other dynamic security problems, e.g., voltage and frequency stability, and/or other control measures, e.g., discrete controls, since PD is generic and is not limited to a specic physical phenomenon. B. Database Preparation The efciency of the proposed method relies strongly on the dynamic security database where the patterns are discovered. Typically, such a database consists of a large number of instances, each associating a pre-contingency OP and the corresponding dynamic secure index [6]. The OP is characterized by features, such as steady-state operating parameters; the dynamic security index can be a discrete class label, e.g., secure or insecure, or a continuous value, e.g., stability margin, with respect to a contingency. To be effective, the database should reect the possible and representative operating region that the system shall go through. In practice, the database can be acquired from historical DSA archives and/or be generated via exhaustive ofine simulations. To generate a database for the upcoming period of interest, given the forecasted load levels, a range of OPs can be produced based on the knowledge of generation scheduling information [11]. The number of OPs required for satisfactory performance on a specic power system could be experimentally determined. Under the considered contingencies, the dynamic security indices of the produced OPs are computed through time-domain simulations. Since PD is unsupervised learning, the class imbalance and possible over-tting problems are not the concern during the database generation stage. C. Feature Space Selection In the context of transient stability control, we select generator active power outputs as the features to characterize an OP. This is rational and effective because of the strong inherent coupling relationship between the generator rotor angles and the active power outputs. In the mean time, it is also necessary to select the critical generators to constitute the feature space for knowledge discovery because 1) it is usually the case that only a subset of generators is responsible for the loss of synchronism [4], [5], and 2) the irrelevant generators contain very little knowledge and may be noise that hinders the knowledge extraction. In the literature, the trajectory sensitivity technique [4] and the single machine equivalent (SIME) approach [5] have been used for critical generator identication, and a sensitivity analysis is performed in [18] to determine effective generation shifting pairs. These approaches only address single OP/contingency scenarios. In this paper, a distance-based feature estimation algorithm called RELIEF [22][24] is adopted to statistically identify the critical generator features.

Fig. 1. Example of PD by residual analysis and recursive partitioning.

where is the residual of the event , is the count of the observed contained samples of event , is the expected frequency of event , and is the estimated asymptotic variance of the numerator, which can be calculated by [20] (6) where is the total count of the instances in the whole space. 4) Evaluate the signicance of each event according to the following criterion: at the percent signicant level, an event is signicant if ; is negative signicant if ; is insignicant if , where is the value of the standard normal deviation, , such that . 5) For each signicant event, repeat steps 1) to 4) to acquire more subtle description of the pattern, until termination conditions [20] are met. Fig. 1 illustrates an example of PD in a 2-D sine distribution instance space. The upper window shows the rst partition and the lower window shows the last partition. It can be seen that the structure knowledge in the instance space has been well identied as the rectangles. III. PROPOSED METHODOLOGY A. General Description Based on PD technique, this paper develops a preventive control method for dynamic insecurity prevention. In the ofine stage, from a dynamic security database, a critical generator feature space is rst selected through a distance-based feature estimation procedure. PD is then performed in the feature space to extract patterns in an unbiased and comprehensive manner. The dynamic secure/insecure regions of the power system are represented by the unions of corresponding patterns. During the online implementation stage, the secure/insecure regions can

XU et al.: PREVENTIVE DYNAMIC SECURITY CONTROL OF POWER SYSTEMS BASED ON PATTERN DISCOVERY TECHNIQUE

1239

The idea behind RELIEF is to evaluate the quality of features according to how well their values distinguish among instances near each other. It not only considers the difference in features values and classes, but also the distance between the instances. Under such a criterion, the good features can cluster similar instances and separate dissimilar ones in the distance space. As benets, the selected feature space has potential structure information suitable for pattern discovery. The original RELIEF algorithm [22] consists of iteratively updating the weight for each feature by the following equation:

(12) (13) Then in (8) can be reobtained using Bayes rule:

(7) is the instance sampled in the th where denotes a feature, iteration, is the nearest instance from the same class as (called nearest hit) while is the nearest instance from the different class with (called nearest miss), and is the number of sampled instances guaranteeing the weights between . Function calculates the difference between the values of feature for two instances and : (8) From the statistics point of view, the weight of feature an approximation of the difference of probabilities is

(14) Depending on the form of the dynamic security index being continuous or discrete, RELIEF algorithms can be employed to estimate the generator features in the power system dynamic security database. Typically, the features having largest weights should be selected as critical features. In this paper, a continuous transient stability index (TSI) is adopted to represent the dynamic security degree of an OP under a contingency [18]: (15) where is the maximum angle separation of any two generators during the transient period. To translate the TSI into a binary security class, a positive TSI corresponds to secure, while a negative TSI corresponds to insecure. D. Pattern-Based Dynamic Secure/Insecure Regions

(9) To deal with noises, incomplete data, and multi-class problems, RELIEF was extended with the following weight updating equation [23]:

PD is performed in the critical generator feature space to extract the patterns. For each pattern, a security class label is assigned according to the possibility of different instances occurred in the pattern. For pattern , its class label can be determined by the following rule: (16) and , respectively, denote the number of secure where and insecure instances in , and is the threshold to measure the occurrence possibility of a class in a pattern (it is set as 50% in this paper). The patterns, after assigned security labels, are the structure knowledge of the power system transient stability characteristics in the critical generator feature space. The secure and insecure regions under the contingency, and , can be represented by the unions of corresponding patterns:

(10) is the prior probability of class where is a class label, , and is an user-dened parameter. Instead of nding the nearest one hit and miss, (10) nds nearest hits and misses to average their contribution in updating the weight, thereby reducing the risk of miss estimation; the introduction of aims to estimate the ability of separating each pair of classes. For regression problems, i.e., the class of the instance is continuous, rather than judging whether two instances belong to the same class or not (absolute difference), the probability of difference is used [24]. This probability is modeled with the relative distance between the predicted class values of the two instances. Function is reformulated using the probability that predicts values of two instances: (11)

(17) where and denote the th secure and th insecure patterns, respectively. Due to the explicit mathematical and geometrical forms of the patterns, the secure/insecure regions can provide rule-based decision support for dynamic security monitoring and situational awareness. Typically, an OP located in the insecure region would imply an insecure status. When the dimension of the feature space is low, say two or three, the monitoring can be visualized.

1240

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 3, AUGUST 2012

E. Computation Procedure To preventively control an insecure OP, the action involves driving it into the pattern-based secure region via generation rescheduling. Since the patterns can be mathematically formulated as a set of inequality constraints on the features, they can be seamlessly incorporated into a standard OPF model, by which the generation rescheduling can be systematically and economically determined. An OPF model minimizing the generation costs is employed in this paper, which takes the following form: (18) subject to

(19) (20) (21) (22) (23) where , , and are generation cost coefcients of the th generator, is the active output of the th generator in the dispatchable generator set , is the reactive output of the th reactive source in the reactive source set , is the voltage magnitude of the th bus in the bus set , is the load of the th bus, and are conductance and susceptance between the th and the th bus, respectively, and is the apparent power across the th branch in the branch set . Specically, the computation of the preventive control consists of formulating the generation output limits dened by a secure pattern into (20) and solving (18)(23). Note that the inclusion of a pattern into the OPF adds no complexity to the OPF model, so it can be solved by any classical programming algorithm. As the secure region may consist of many separate patterns, one needs to determine which pattern to apply in the OPF. To avoid over-control, the nearest secure pattern in the distance space could be rst selected for use, but it should be noted that, due to the nonlinear nature of the problem, the term nearest cannot guarantee the optimality in view of OPF computations. The resulting OP may not be completely secured because of insufcient security margin provided by the nearest secure pattern. In this regard, a subsequent DSA must be executed to verify the resulting OP. If the new OP is still insecure, another secure pattern should be applied. The whole computation procedure is described in Fig. 2. Once an insecure OP is detected, its rst-nearest secure pattern is formulated into the OPF and a new OP is obtained through the OPF solution. If the new OP is secure, the computation stops; otherwise, the next-nearest secure pattern is applied, and the process repeats until a secure OP is obtained. The term th nearest is determined by ranking the Euclidean distance between the OP and the centroids of the patterns. The Euclidean distance between two points and in an -dimensional space is calculated by (24)
Fig. 2. Computation owchart of the proposed preventive control.

where and are the values of the th feature of instance and , respectively. The centrioid of a pattern is calculated by (25) is where is the value of the th feature of the centrioid, value of the th feature of the th instance in the pattern. Note that the features in distance calculation need to be normalized into a same range, e.g., [1]. IV. SIMULATION RESULTS The proposed method is tested on the New England 39-bus system whose one-line diagram is shown in Fig. 3. The system data are obtained from [25] and the generation cost coefcients are taken from [4]. In the simulations, the power ow and OPF are solved using MATPOWER [26], DSA is performed through time-domain simulation [27] (simulation time is 4 s), and the PD is realized in JAVA programming platform. A. Database Generation For the test system, its total active power load at basic operating state is 6097.1 MW. An initial OP is obtained by an OPF solution, and its total generation cost is 60 920.39$/h. A three-phase fault at bus 22 cleared by tripping line 2221 after 0.14 s is considered rst (named Fault 1). DSA shows that the initial OP is insecure under this fault: the generators lose synchronism (see Fig. 4) and its TSI is -90.7. A database is articially generated for use. The P and Q demand at each load bus are varied randomly within range of the initial OP; under each load distribution, a set of OPs with

XU et al.: PREVENTIVE DYNAMIC SECURITY CONTROL OF POWER SYSTEMS BASED ON PATTERN DISCOVERY TECHNIQUE

1241

Fig. 6. RELIEF feature estimation result (Fault 1).

Fig. 3. One-line diagram of the New England 39-bus system.

Fig. 7. Non-critical feature space by G32 and G33.

Fig. 4. Rotor angle swing curves of the initial OP under Fault 1.

Fig. 8. Different classes of instances and discovered patterns.

Fig. 5. Histogram of TSI distribution of the generated database (Fault 1).

B. Critical Generators Regressive RELIEF is applied to evaluate the 10 generator features. The weight of each feature is shown in Fig. 6, where it can be seen that generators G36 and G35 are much larger than others in weight, indicating they are most signicant to the transient stability under Fault 1. Note that some generators have negative weights, which means that they have negative distinguishing capabilities, i.e., they are mixing rather than separating the instances in the space. G36 and G35 are selected to form the critical feature space for PD.

various generation scenarios are produced in two ways: 1) randomly vary the generation outputs of each generator and solve the power ow model, and 2) randomly alter the generation cost coefcients of each generator and solve the OPF model. Note that only converged cases are saved for use. In doing so, 1000 OPs are produced and their TSIs under Fault 1 are then obtained through DSA. Fig. 5 shows the TSI distribution of the OPs where 113 are secure and 887 are insecure.

1242

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 3, AUGUST 2012

Fig. 10. Rotor angle swing curves of the 1st new OP under Fault 1.

Fig. 9. Secure/insecure regions and corresponding centriods.

To better show the effectiveness of the RELIEF algorithm, the critical feature space and a non-critical feature space constructed by G32 and G33 are visually compared. The non-critical feature space is portrayed in Fig. 7 and the critical feature space can be found in Fig. 8 (note that the patterns are also shown as the black rectangles). It can be clearly observed that in Fig. 7, the OPs of different classes are heavily overlapped, indicating no (or very little) structure information existing in the space. By marked contrast, the OPs in Fig. 8 are appreciably separated, implying evident structure knowledge about the security characteristics. C. Pattern Discovery Results Totally 28 events are discovered in the critical feature space where 27 of them are evaluated as signicant, i.e., patterns. These 27 patterns encompass 996 instances in total and occupy 75% volume of the whole instance space. As shown in Fig. 8, the patterns form non-overlapped rectangles in the critical generator space. It is worth mentioning that the PD is quite computationally efcient as it only costs several CPU seconds on the generated database. The patterns are assigned security labels and their centroids are calculated and displayed in Fig. 9 where and o denote the cerntroids of insecure and secure patterns, respectively. Four patterns are secure and 23 are insecure. D. Preventive Control Results The secure/insecure regions under Fault 1 are represented by the corresponding patterns. Since the dimension of the feature space is two, it can provide an illustration of visualized dynamic security monitoring. As shown in Fig. 9, the initial OP (shown as the red triangle) is found located in the insecure region, which is consistent with its actual security status. To eliminate the risk of insecurity, preventive control should be armed to drive it to the secure region. Following the computation process described in Fig. 2, the nearest pattern of the initial OP is applied rst, whose region is in the space of G35 and G36 (normalized value). By formulating them into the OPF model and solving the OPF, the 1st new OP is obtained. DSA shows that the 1st new OP remains insecure (see Fig. 10); however, it is important to note that the TSI of the 1st new OP

Fig. 11. Rotor angle swing curves of the 2nd new OP under Fault 1.

becomes , which has been largely improved over the initial OP. The next nearest secure pattern is then applied, yielding the 2nd new OP, and DSA shows that this OP becomes secure, its TSI is 40.5, and its rotor angles are shown in Fig. 11. This new secure OPs generation cost is 61 003.14$/h, which is 82.75$/h higher than the initial OP due to the control cost. To further test the proposed method, all the 887 insecure Ops generated in the database are controlled using the 4 secure patterns. It is found that all of them can be secured after the control, further conrming the effectiveness of the secure patterns. Investigation shows that 14 Ops are secured by their 1st secure pattern, 737 Ops by their 2nd nearest secure pattern, 108 Ops by their 3rd nearest secure pattern, and 28 Ops by their 4th nearest secure pattern. E. Multi-Contingency Control In the literature, most of the reported methods consider only one single contingency [16][18]. But in practice, depending on the DSA results, it is usually needed to include more than one contingency in the preventive control. Generally, the most likely and harmful contingencies that lead to insecurities should be controlled. In this section, multi-contingency is considered in the proposed method. The idea is to identify a general critical feature set for all the faults, and the patterns extracted in the general critical feature space can be the collective secure/insecure regions under the faults. Then the multiple faults can be simultaneously controlled in one computation run. Another harmful disturbance is considered: it is a three-phase fault at bus 28 cleared by tripping line 2829 after 0.10 s (named Fault 2). Under Fault 2, the TSI of the initial OP is and for the OPs in the database, 236 are secure and 764 are insecure. A general TSI is taken as the average value of the TSIs of

XU et al.: PREVENTIVE DYNAMIC SECURITY CONTROL OF POWER SYSTEMS BASED ON PATTERN DISCOVERY TECHNIQUE

1243

Fig. 12. RELIEF feature estimation result (Fault 1 and Fault 2).

Fig. 13. RELIEF feature estimation result (Fault 2).

Fault 1 and Fault 2. It is considered that this general TSI can account for the composite security measure under the two faults; thus, the two faults can be simultaneously involved in estimating features. Regressive RELIEF is applied to evaluate the generator features under the general TSI. According to the result in Fig. 12, G38, G36, and G35 should be selected as critical generators for the two faults. While G36 and G35 are critical features of Fault 1, it implies that G38 is critical to Fault 2. Interestingly, this is consistent with the feature estimation result considering only Fault 2 (see Fig. 13). In the space of G38, G36, and G35, 22 patterns are discovered, where 3 are secure and 19 are insecure. The union of the secure patterns represents the collective secure region under the two faults. The initial OP is then controlled, and it is found that after the 1st nearest secure pattern applied, a new secure OP has been obtained, whose TSIs under Fault 1 and Fault 2 are 45.6 and 48.2, respectively, and generation cost is 61 184.2$/h. The rotor angle swing curves are shown in Fig. 14. F. Discussions With the discovered patterns, the initial OP is stabilized under both single- and multi-contingency conditions, providing the advantage of interpretable and transparent control mechanism and high computation efciency. In the meantime, it is also clear that the success of the control relies strongly on the acquisition of secure patterns. In the case studies, due to the high severity of the two considered contingencies, most of the generated OPs are insecure, leading to only a small number of secure patterns discovered: 4 for Fault 1 and 3 for Fault 2. In practice, efforts can be made to enrich the secure patterns, e.g., by simulating more secure OPs and/or making use of historical DSA archives. This can effectively enhance the quality of the method. As the PD process is quite efcient, the database enrichment/updating can be online performed using up-to-date operating information.

Fig. 14. Rotor angle swing curves of the new OP after multi-contingency control (upper window-Fault 1, lower window-Fault 2).

V. CONCLUSIONS PD is an effective unsupervised learning technique to discover interpretable and unbiased knowledge from a large-scale database and has been applied for solving many real-world problems [19][21]. Based on PD, this paper develops a method for preventive dynamic security control of power systems. The idea is to extract patterns from a feature space characterized by critical generators. The patterns are a set of non-overlapped hyper-rectangles representing the dynamic secure/insecure regions of the power system. For online use, they can provide decision support for security monitoring and situational awareness. The preventive control involves driving the insecure OP into the secure region, which is transparent and efcient. In practice, this can be achieved systematically and economically by formulating the secure patterns as generation output constraints into an OPF model. Case studies are conducted on the New England 39-bus system considering both single- and multi-contingency conditions, and the simulation results have veried the proposed method. Discussions on improving the quality of the method are also given. As a generic knowledge discovery approach, PD can also be exploited for tackling other problems in power engineering, where DRIP is encountered. ACKNOWLEDGMENT The authors would like to thank the anonymous reviewers for reviewing this paper and providing numerous valuable and inspiring comments and suggestions for improving this paper. REFERENCES
[1] Y. V. Makarov, V. I. Reshetov, A. Stroev, and I. Voropai, Blackout prevention in the United States, Europe, and Russia, Proc. IEEE, vol. 93, no. 11, pp. 19421955, Nov. 2005.

1244

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 3, AUGUST 2012

[2] P. Kundur, Power System Stability and Control. New York: McGraw-Hill, 1994. [3] D. Gan, R. Thomas, and R. Zimmerman, Stability-constrained optimal power ow, IEEE Trans. Power Syst., vol. 15, no. 2, pp. 535540, May 2000. [4] T. B. Nguyen and M. A. Pai, Dynamic security-constrained rescheduling of power systems using trajectory sensitivities, IEEE Trans. Power Syst., vol. 18, no. 2, pp. 848854, May 2003. [5] D. Ruiz-Vega and M. Pavella, A comprehensive approach to transient stability control: Part INear optimal preventive control, IEEE Trans. Power Syst., vol. 18, no. 4, pp. 14461453, Nov. 2003. [6] K. Morison, On-line dynamic security assessment using intelligent systems, in Proc. 2006 IEEE PES General Meeting, Jun. 2006. [7] Z. Y. Dong, Y. Xu, P. Zhang, and K. P. Wong, Real-time stability assessment of electric power systems with intelligent systems, IEEE Intell. Syst. Mag., to be published. [8] Y. Mansour, E. Vaahedi, and El-Sharkawi, Dynamic security contingency screening and ranking using neural networks, IEEE Trans. Neural Netw., vol. 8, no. 4, pp. 942950, Jul. 1997. [9] L. S. Moulin, A. P. Silva, M. A. EI-Sharkawi, and R. J. Marks, Support vector machines for transient stability analysis of large-scale power systems, IEEE Trans. Power Syst., vol. 19, no. 2, pp. 818825, May 2004. [10] K. Sun, S. Likhate, V. Vittal, V. Kolluri, and S. Mandal, An online dynamic security assessment scheme using phasor measurements and decision trees, IEEE Trans. Power Syst., vol. 22, no. 4, pp. 19351943, Nov. 2007. [11] Y. Xu, Z. Y. Dong, K. Meng, R. Zhang, and K. P. Wong, Real-time transient stability assessment model using extreme learning machine, IET Gen., Transm., Distrib., vol. 5, pp. 314322, Mar. 2011. [12] T. Wang, L. Guan, and Y. Zhang, A modied pattern recognition algorithm and its application in power system transient stability assessment, in Proc. 2008 IEEE PES General Meeting, 2008. [13] Y. Xu, L. Guan, Z. Y. Dong, and K. P. Wong, Transient stability assessment on China Southern Power Grid System with an improved pattern discovery-based method, in Proc. 2010 Int. Conf. Power Syst. Tech. (POWERCON2010), Hangzhou, China, Oct. 2010. [14] E. S. Karapidakis and N. D. Hatziargyriou, On-line preventive dynamic security of isolated power systems using decision trees, IEEE Trans. Power Syst., vol. 17, no. 2, pp. 297304, May 2002. [15] K. Mei and S. M. Rovnyak, Response-based decision trees to trigger one-shot stabilizing control, IEEE Trans. Power Syst., vol. 19, no. 1, pp. 531537, Feb. 2004. [16] E. M. Voumvoulakis and N. D. Hatziargyriou, Decision trees-aided self-organized maps for corrective dynamic security, IEEE Trans. Power Syst., vol. 23, no. 2, pp. 622630, May 2008. [17] K. R. Niazi, C. M. Arora, and S. L. Surana, A hybrid approach for security evaluation and preventive control of power systems, in Proc. 2003 IEEE Power Engineering Conf., 2003, pp. 193199. [18] I. Genc, R. Diao, V. Vittal, S. Kolluri, and S. Mandal, Decision treesbased preventive and corrective control applications for dynamic security enhancement in power systems, IEEE Trans. Power Syst., vol. 25, no. 3, pp. 16111619, Aug. 2010. [19] A. K. C. Wong and Y. Wang, Pattern discovery: A data driven approach to decision support, IEEE Trans. Syst., Man, Cybern. C, Appl. Rev., vol. 33, no. 1, pp. 114124, Feb. 2003. [20] T. Chau and A. K. C. Wong, Pattern discovery by residual analysis and recursive partitioning, IEEE Trans. Knowl. Data Eng., vol. 11, no. 6, pp. 833852, Nov./Dec. 1999. [21] Discover*e, Pattern Discovery Technologies Inc. [Online]. Available: http://www.patterndiscovery.com/discovere.html, [22] K. Kira and L. Rendell, A practical approach to feature selection, in Proc. 9th Conf. Machine Learning, 1992, pp. 249256. [23] I. Kononenko, Estimating attributes: Analysis and extension of relief, in Proc. European Conf. Machine Learning, 1994, pp. 171182. [24] M. Sikonja and I. Kononenko, Theoretical and empirical analysis of ReliefF and RRelief, Mach. Learn., vol. 53, pp. 2369, 2003. [25] M. A. Pai, Energy Function Analysis for Power System Stability. Norwell, MA: Kluwer, 1989. [26] R. Zimmerman, C. Murillo-Snchez, and R. Thomas, MATPOWER steady-state operations, planning and analysis tools for power systems research and education, IEEE Trans. Power Syst., vol. 26, no. 1, pp. 1219, Feb. 2011. [27] K. Meng, Z. Y. Dong, K. P. Wong, Y. Xu, and F. Luo, Speed-up the computing efciency of power system simulator for engineeringbased power system transient stability simulations, IET Gen., Transm. Distrib., vol. 4, pp. 652661, May 2010. [28] S. C. Port, Theoretical Probability for Applications. New York: Wiley, 1994.

Yan Xu (S10) received the B.E. and M.E. degrees from South China University of Technology, Guangzhou, China. He is now pursuing the Ph.D. degree at the University of Newcastle, Callaghan, Australia. He was previously with the Hong Kong Polytechnic University, Hong Kong. His research interests include power system stability and control, power system planning, smart grid technologies, and intelligent system applications to power engineering.

Zhao Yang Dong (M99SM06) received the Ph.D. degree from the University of Sydney, Sydney, Australia, in 1999. He is now Ausgrid Chair Professor and Director of the Centre for Intelligent Electricity Networks, the University of Newcastle, Callaghan, Australia. He previously worked with The Hong Kong Polytechnic University, the University of Queensland, Australia and Transend Networks, Australia. His research interests include smart grid, power system planning, power system security, load modeling, renewable energy systems, electricity market, and computational intelligence and its application in power engineering. Dr. Dong is an editor of the IEEE TRANSACTIONS ON SMART GRID.

Lin Guan received the B.S. and Ph.D. degrees from Huazhong University of Science and Technology, Wuhan, China, in 1990 and 1995, respectively. She is now a Professor of Electric Power College, South China University of Technology, Guangzhou, China. Her research interests include power system stability and control, power system planning and reliability, and application of articial intelligence in power systems.

Rui Zhang received the Diploma from Ngee Ann Polytechnic, Singapore, in 2005 and the B.E. degree from the University of Queensland, Brisbane, Australia, in 2009. She is now pursuing the M.Phil. degree at the University of Newcastle, Callaghan, Australia. Her research interests include power system operation and control, data-mining techniques, and their applications to power engineering.

Kit Po Wong (M87SM90F02) received the M.Sc., Ph.D., and D.Eng. degrees from the University of Manchester, Institute of Science and Technology, Manchester, U.K., in 1972, 1974, and 2001, respectively. Since 1974, he has been with the School of Electrical, Electronic and Computer Engineering, The University of Western Australia, Perth, Australia, except from 2002 to 2011, when he was with Department of Electrical Engineering, Hong Kong Polytechnic University. He is a con-joint professor of the University of Newcastle, Callaghan, Australia. His current research interests include computation intelligence applications, and power system analysis, planning and operations and smart grid. Prof. Wong received three Sir John Madsen Medals (1981, 1982, and 1988) from the Institution of Engineers Australia, the 1999 Outstanding Engineer Award from IEEE Power Chapter Western Australia, and the 2000 IEEE Third Millennium Award. He was General Chairman of IEEE/CSEE PowerCon2000 and IEE APSCOM in 2003 and 2009. He was an Editor-in-Chief of IEE Proceedings in Generation, Transmission and Distribution and Editor (Electrical) of the Transactions of Hong Kong Institution of Engineers. He currently serves as Editor-in-Chief for IEEE PES Transactions letters. He is a Fellow of IET, HKIE, and IEAust.

Fengji Luo (S11) received the B.E. and M.E. degrees from Chongqing University, Chongqing, China. He is now pursuing the Ph.D. degree at the University of Newcastle, Callaghan, Australia. His research interests are power system computations and software developments.