06786495

2846 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 6, NOVEMBER 2014

Preventive Control StabilityVia Neural Network Sensitivity

Mauricio C. Passaro, Member, IEEE, Alexandre P. Alves da Silva, Senior Member, IEEE, andAntonio C. S. Lima, Member, IEEE

Abstract—This paper discusses the power systems stabilitymargin improvement by means of preventive control based ongeneration re-dispatch using a neural sensitivitymodel. Thismodeluses multilayer perceptron networks with memory structure in theinput layer. The training of this model is made with temporal datasamples from time domain simulations, incorporating informationabout the dynamic behavior of the system, unlike the methodsproposed in the literature in which the pre-fault system data areused instead. The sensitivity is used as a guideline in selectingthe most effective set of generators in the reallocation of theamount of active power capable of increasing system security. Theeffectiveness of the proposed methodology has been demonstratedthrough the application to a large system.

Index Terms—Neural network application, power system dy-namic stability, power system security.

I. INTRODUCTION

E LECTRIC power systems are strategic for the economicactivity and its constant evaluation is very important for

ensuring continuity of energy supply to society. In recent years,these systems have gone through several changes. The processof deregulation of the electricity industry in many countrieshas brought new challenges for the interconnected systems op-eration. As they were operating close to their limits [1], thisnew scenario led to the occurrence of several blackouts world-wide. For instance, in 1996 a blackout affected the West CoastSystem (WSCC) [2], and in 1999 [3] and 2009 [4] a systemwideblackout affected most of the Brazilian National Power Grid.The recurrence of these major failures emphasizes the impor-tance of an efficient restoration procedure [5] and the need forthe development of fast, secure and reliable security assessmenttools. Transient stability is without a question a critical problem,although solution via analytical techniques alone does not allowpreventive or corrective actions in a timely manner.

Manuscript received August 22, 2013; revised December 02, 2013 andMarch21, 2014; accepted March 29, 2014. Date of publication April 10, 2014; date ofcurrent version October 16, 2014. This work was supported in part by INERGE,FAPEMIG, CNPq, and CAPES. Paper no. TPWRS-01085-2013.M. C. Passaro is with the Department of Electrical Engineering, Federal

University of Rio de Janeiro, Rio de Janeiro, RJ, Brazil, and also with theBrazilian Independent System Operator (ONS), Rio de Janeiro, RJ, Brazil(e-mail: [email protected]).A. P. A. da Silva is with GE Global Research—Brazilian Technology Center,

Rio de Janeiro, RJ, Brazil (e-mail: [email protected]).A. C. S. Lima is with the Department of Electrical Engineering, Federal Uni-

versity of Rio de Janeiro, Rio de Janeiro, RJ, Brazil (e-mail: [email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TPWRS.2014.2314855

Transient stability is, according to many authors [6], the moststudied problem in power systems dynamic performance. Thereis a vast amount of publications [7] dedicated to improve tran-sient stability and they have in common the use of active powerredispatch to achieve this goal [8]. The main issue is to deter-mine the amount of generation to be relocated and the selec-tion of generators. The integration of computational intelligence(CI) techniques with analytical methods can provide significantimprovement in the security assessment process, given that sta-bility studies databases available in electric utilities planning de-partments contain useful information.Nowadays, the strategic planning of the Brazilian Indepen-

dent System Operator (ONS) has focused on the implementa-tion and operation of a computational tool to analyze static anddynamic security in real time [9]. Based on the results of state es-timation or based on data from short-term planning studies, theprocess is to determine the conditions under which the systemmeets the criteria established by the operating procedures [10].The results obtained by simulation and presented in the form

of security regions or nomograms have been helping planningand operation teams to streamline the analysis for different oper-ating situations and to support decision making from real-timeoperation teams [11]. In this context, it is interesting to inte-grate this analysis tool with CI techniques, specifically neuralnetworks, to provide a set of rules and operational guidelines forpreventive control, in order to improve the dynamic behavior ofthe system after redispatch.This paper describes a methodology based on an applica-

tion of neural networks (NN) for the determination of a sen-sitivity model between stability margins and generation redis-patch. The NN are trained with data from electromechanicaltransient simulations, including information on the dynamic be-havior of the system, unlike almost all the previousmethods pre-sented in the literature [12]. The few exceptions have only dealtwith small-scale problems [13]. Another highlight is the devel-opment of a control strategy based on the proposed sensitivitymodel. Although inaccuracies in models or field measurementsmight affect the results, at this point is not certain how to as-sess these issues and their impact on the system performance.This topic is considered open and left for future research. Nev-ertheless, the models used in the transient stability studies forthe Brazilian system have been able to reproduce large black-outs as well as some minor events [3], [5].The results encourage the use of the proposed model in the

real time operation environment, with the objective of gener-ating preventive control rules based on the dynamic behaviorof the system. These rules may be part of the operation guide-lines of the Brazilian Independent System Operator. In order to

0885-8950 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

PASSARO et al.: PREVENTIVE CONTROL STABILITY VIA NEURAL NETWORK SENSITIVITY 2847

apply the proposed approach to the real-time operation environ-ment, a base load-flow can be generated from forecasted data.The whole analysis can be performed considering the estimatedload and the generation schedule for some minutes ahead [14].

II. PREVENTIVE CONTROL

Traditionally, operational security control is divided into twomain categories: preventive control and emergency control.The purpose of preventive control is to prepare the systemwhen it is still in normal operation, so as to make it capable offacing possible future events satisfactorily. On the other hand,in emergency control, disturbances have already occurred, andthe goal is to control the dynamics of the system to mitigate theconsequences.Regarding the type of control action, the preventive mode is

implemented by generation redispatch, switching network com-ponents such as compensation elements or transmission lines, orby load shedding. In emergency control, the actions are limitedto load or generation shedding, switching shunt capacitors/reac-tors, or network partitioning, coordinated by means of specialprotection schemes.Preventive control applies sensitivity analysis as the basic

mechanism for decision making. Energy margin analytical sen-sitivity [8], [15], [16] involves a high computational cost. Onealternative is based on neural networks [12]. With neural net-works two ideas have been studied: 1) The use of neural net-works in transient stability assessment with the purpose of clas-sification [17] and preventive control via another procedure,such as generation redispatch and/or load shedding based onoptimization or using decision trees [15]; 2) The use a neuralsensitivity model [12], [19]–[21].The sensitivity evaluation by a neural model provides the par-

tial derivative of a mapping with no explicit analytical formula-tion, and it is dependent on a large data set of systemic informa-tion. Note that neural networks present desirable characteristicssuch as fast response, simplicity in its output format (stable/un-stable or stability margin), and flexibility in managing uncer-tainty. The high computational burden during the training phaseis performed off-line in a planning environment.The proposed neural sensitivity model uses the stability

margin calculated by time domain simulations as an index ofsystemic security [11].

A. Stability Margin

The power system transient stability analysis can be per-formed using the concept of energy and security margin [9].Electric power system transient instability is characterized byseparation of the system into two parts, i.e., a group of unitscalled “critical machines”, which distances itself from the restof the generation system (non-critical machines). This featureallows the system stability study using the equal area crite-rion, where the two groups are represented by two equivalentmachines.Therefore, the power system transient stability can be eval-

uated using the concept of stability margin using the fol-lowing criteria: , the system is considered stable from thestandpoint of transient stability; , the system is consideredunstable from the standpoint of transient stability.

Fig. 1. Power angle curve.

If the system is stable, the stability margin is calculatedfrom the following equation:

(1)

If the system is unstable (acceleration area A1 larger than thedeceleration area A2), the negative margin is numerically equalto the kinetic energy at the point , according to Fig. 1. Thevalue of this margin is calculated by [9].Considering a list of pre-defined contingencies, the stability

margin of the system must satisfy the following relationship[20]:

(2)

where is the stability margin and is the minimum stabilitymargin. If the electrical system is in an unsafe operation condi-tion with respect to contingency , control actions shouldmodifythe stability margin, satisfying the following relationship:

(3)

where is the stability margin related to the i-th contingency.The change required on the stability margin is estimated throughthe sensitivity coefficient:

(4)

where is the sensitivity of the stability margin with re-spect to the control variable P (active power), and corre-sponds to the control variable change.

III. NEURAL NETWORK SENSITIVITY MODEL

The proposed sensitivity model uses a conventional multi-layer perceptron network with a memory structure in its inputlayer. The training, validation and test sets consist of temporalsampled data obtained from time domain simulations.

A. Neural Network

Among the applications of neural networks in power sys-tems, the great majority use the multilayer perceptron architec-ture [22]. According to this publication, about 400 technical ar-ticles on the theme were investigated, and 81.19% of the totalhas used multilayer perceptrons.


Fig. 2. Data acquisition intervals.

B. Temporal Data Sample

Predominantly, both transient stability assessment and pre-ventive control based on neural sensitivity model use pre-faultsystem data to generate training sets for the neural network[19]–[21], [23], [24].In [18] and [22], the use of dynamic data was discouraged. In

[24], it is reported that there is insufficient evidence to determinewhether the static variables are more appropriate than dynamicvariables (temporal data samples) as attributes for neural net-works applied to power systems transient stability analysis. Theneural models proposed by [25]–[27] use temporal data samplesfrom time domain simulations as input variables for feeding thenetwork. However in these papers the neural model aims to clas-sify the system for stability/instability only.The proposed neural sensitivity model uses temporal data

samples for build the training, validation and testing sets, asshown in Fig. 2. The model output is the stability margin forthe operation point represented by the selected input variables.According to Fig. 2, the sampling is done considering three

distinct intervals. The first interval (P1) contains samples ob-tained during the application of a 110 ms fault. The interval (P2)contains information between 50 ms after the beginning of thefault and 100 ms its extinction, i.e., with a duration of 150 ms.The interval (P3) contains post-fault values and ranges from 250to 750 ms. The purpose of dividing the data sampling is the ver-ification of the period that results in a better sensitivity neuralmodel, capable of generating active power reallocation rules forimproving the stability margin.

C. Memory Structure

A simple way to insert short-termmemory in the neural struc-ture is the insertion of time delay on the input variables. Fig. 3shows a memory structure representation applied to the neuralnetwork input layer. Fig. 4 shows the importance of the memorystructure representation in a multi-layer perceptron network.The 9 buses test system proposed in [28] is used, in which fourcontingencies were simulated. Each contingency has a stabilitymargin for a given operating point of MW, MW, 14MW, and 19 MW, respectively.The stability margin estimated by the neural model without a

memory structure in its input layer (star markers) shows a largedispersion around the true margin (calculated by using conven-tional time domain simulation). Including short-term memory

Fig. 3. Time delay in the input layer.

Fig. 4. Stability margin estimation.

in the neural model provides an increase in accuracy for esti-mating the stability margin.

D. Input Variable Selection

The selection of the input data of neural networks applied totransient stability assessment problems, traditionally, is basedon the experience of experts. Such entries typically describeadequately the power system state and the stability evaluationfrom these attributes is conservative. The dimensionality ofthe problem should be as small as possible in order to reducethe computational burden and improve the performance of the


neural model, however without deteriorating its generalizationcapability.While linear analysis methods (such as correlation) are useful

in particular cases, it is essential to consider non-linear relation-ship between different variables. The calculation of mutual in-formation allows the assessment of non-linear dependence be-tween them [28].The mutual information for continuous systems is defined:

(5)

where p is the probability for the possible valuations of thevariables x and y. Tests were performed using the New Eng-land 39-bus test systems. These tests indicated that the vari-ables most relevant to the stability margin are electrical power(Pe) and rotor angle . The latter is treated as a non-control-lable variable while the former is considered as a controllableone. Several compositions of input variables (controllable andnon-controllable, e.g., reactive power, accelerating power andkinetic energy) were used in the neural network training. It wasobserved that the use of electric power (Pe) and rotor angleleads to better results. Also indicated that better sampling pe-riod of temporal data is the post-fault (P3). Probabilities densityfunctions in (5) have been transformed to the discrete domainusing histograms.The use of the respective rotor angles has the objective of

evaluating the neural model performance impact in the trainingprocess and sensitivity evaluation phase. In this case the inputsof the neural network for the uncontrolled variables will not bestimulated and control rules will be obtained only through thechanges imposed on the entries for the controllable variables.

E. Sensitivity Analysis

The interdependence between the power system variables canbe quantitatively determined by sensitivity analysis. Sensitivityis defined as the ratio of change on the dependent variable be-cause of an independent variable variation. This analysis is veryimportant in power systems operation planning studies. It helpsin observing the cause-effect relationship, providing the basisfor system control actions evaluation.In [5] and [10], analytical expressions to obtain the energy

margin sensitivity with respect to system parameters were ap-plied. In [11], the sensitivity was calculated using the Taylor se-ries expansion of the stability margin function. In [19], a neuralsensitivity model that employs a back-propagation algorithmwith fuzzy controller and activation function optimization wasproposed. This model uses pre-fault system data to generate thetraining, validation and testing sets.In the present work, model output (stability margin) is eval-

uated after variations of a given input, keeping the remainingones unchanged. The changes consist of positive and negativesteps of 10, 20 and 30% on the current input value.

F. Modules

The neural model construction process is divided into twostages. The first stage involves the entire systemic data prepa-ration, generation of analysis scenarios, variable selection, net-work training and the creation of preventive control rules. This

Fig. 5. Block diagram of sensitivity model—Rules module.

Fig. 6. Conceptual structure of sensitivity model—Operation module.

stage is executed in planning environment. The correspondingflowchart is depicted in Fig. 5.At the end of the process described in Fig. 5, the sensitivity

neural model is ready to be used in a real time environment.The use of the neural model in a real time environment includesthe following steps: 1) systemic data acquisition via state esti-mation; 2) dynamic simulations for a contingency list definedin the planning environment; 3) temporal data sampling for theselected variables; 4) database building; 5) feed operation pointto the neural model; 6) if the margin obtained by the model isnegative, proceed to redispatch in accordance to the rules estab-lished in the planning stage (see Fig. 6).The set of rules for preventive control is obtained based on the

sensitivity analysis results. For each geo-electrical area, a rankof sensitivities is obtained in order to prioritize the generationunits to be re-dispatched. The most effective generation unit forstability margin improvement when the corresponding outputis increased gets first place. Second place goes to the unit thatimproves the stability margin themost while it has its generationdecreased. These pair of units are the ones to be re-dispatched,keeping the generation-load balance. In the real-time operationof the Brazilian Independent System Operator this informationis quickly obtained using a distributed processing hardware.


Fig. 7. Brazilian National Interconnected System—Equivalent.

IV. APPLICATION

In order to demonstrate the applicability of the proposedmethodology, the Brazilian National Interconnected System(SIN), a large scale power system has been used. The systemconsists of 4077 buses, 260 power plants, 3695 transmissionlines, 2146 transformers, 13 SVC, and 3 HVDC systems (Itaipu,Garabi, and Alumar). Fig. 7 presents a simplified diagram of theinterconnections North-East, North-South, and East-Northeast.In all operation points, a single contingency, three-phase

fault, involving one of the tie lines between the North/Northeastis considered. The fault clearing time is equal to six cycles(100 ms). The bus on which the fault is applied is RibeiroGonçalves, causing the opening of the Ribeiro Gonçalves/Col-inas and Ribeiro Gonçalves/São João do Piauı́ transmissionlines. The operating point of the base case corresponds to amedium load condition. The power flow in the tie lines is 3100MW (Northeast area is importing).

A. Training Data Generation

The creation of the database from which the training, valida-tion and testing sets are extracted was performed by calculatingdynamic security regions, a feature of Organon [9]. The use thenomograms derived from these calculations helps in the prepa-ration of the operating points used for generating training, vali-dation and testing data for the model.The major advantage of this procedure is the possibility of vi-

sual interpretation of critical regions, which allows one to focusthe model synthesis on the regions around the stability border[29]. Operating points used in the preparation of data sets fortraining, validation and testing of the neural model can be seenin Fig. 8. Table I shows the stability margins associated to eachoperating point.

Fig. 8. Nomogram—Operation points.

TABLE IOPERATION POINTS AND STABILITY MARGINS

B. Variables Selection

The variables selection process of the Brazilian Grid aims toindicate the list of more relevant generation units (with min-imal redundancy) regarding stability margin. Initially, 76 powerplants are pre-selected, and the corresponding electrical powers(Pe) and rotor angles are analyzed. The initial set of 76power plants is defined according to the criteria below: 1) opera-tive reserve exceed 90MW (because the margin of the operationpoint used in the tests is MW); 2) unit must be intercon-nected to the system through the main grid (230 kV and above).The previous selection of generators with operative reserve

ensures that only units without technical limits can compose thelist of generators for which sensitivity will be assessed by theneural model.The variable selection process consists of two phases. The

first phase verifies which variables have higher relevance withrespect to stability margin. The second phase reduces the redun-dancy between the previously selected variables. Fig. 9 showsthe input selection process flowchart. Table II shows the vari-ables ranking according to the mutual information value re-garding stability margin.It can be observed a variation of 24.21% between values

for buses 4520 (P.Pedra) and 6420 (Tucuruı́ I), which corre-sponds to the greatest change in the sorted MI list. Therefore,the more relevant variables are selected above that a cutoff. Thepower plants that precede Tucuruı́ I (6420) are discarded, withremaining power plants totalizing 29.After that, the redundancy minimization scheme described in

Fig. 9 is applied. The final list of power plants can be seen inTable III. The list of selected power plants saves 11.84% of theoriginal list. In order to allow an understanding of the spatial dis-tribution of the selected units, they are presented in Table III ac-cording to the corresponding geographical regions (north, north-east, and southeast).


Fig. 9. Relevance maximization and redundancy minimization.

TABLE IIMUTUAL INFORMATION VALUES

TABLE IIISELECTED POWER PLANTS

C. Neural Network Training

The Matlab© Neural Network Toolbox has been applied, andthe learning parameters have been adjusted according to the ex-perience. The sampling rate is 5 ms and temporal data samplesare extracted after the fault (P3).

TABLE IVPAIRS OF POWER PLANTS FOR REDISPATCH

The neural network training, validation and testing sets wereformed from operating points show in Fig. 8. These operationpoints led to temporal data samples totalizing 2060 patterns.

V. RESULTS

The model is validated by comparing its performance in as-sessing sensitivity with time domain simulations, in which thepossible alternatives for relocation are verified.

A. Time Domain Simulation

According to the nomogram shown in Fig. 8, the stabilitymargin for the operating point used in the test is MW. It isobserved that power plants of the northeastern region should in-crease their generation at least 90 MW for the system to operatein a secure condition. This power generation increase should becompensated by decreasing generation in the north or southeastpower plants. A total of 100 MW of generation for relocation isestablished. The list of selected power plants (Table III) estab-lishes combinations of power plants that can be used as controls.The pairs of power plants list can be observed by the Table IV.The cells highlighted in red indicate that the preventive con-

trol via redispatch using these power plants is not effective, i.e.,does not lead the system to a secure operating point. These pairsbelong to the same generation group or do not involve powerplants of the northeastern area.The effective redispatch is compared by monitoring the

systemic kinetic energy as shown in Fig. 10. The pairs ofpower plants leading to lower kinetic energy are 5061-20,5061-754, 5061-3964, and 5061-36. The plants Marimbondo(20), Cachoeira Dourada (754), Cana Brava (3964), and Serrada Mesa (36) belong to southeast region. Among the plantsin the Northeast, Xingó represents the best redispatch optionfollowed by Sobradinho, Paulo Afonso I, and Apolonio Sales

. The sorted list accordingto the maximum kinetic energy system obtained by Fig. 10 isshown in Table V.

B. Neural Model

The neural network topology presents 54 inputs (3 9 con-trol variables with corresponding angles), 27 neurons in thehidden layer and one output. This topology has been obtainedempirically through several tests, in which the numbers of neu-rons and hidden layers has been modified.In fact, cross validation hasmade the estimated test error quite

immune to the number of neurons in the hidden layer. Fig. 11shows the neural sensitivity model output.It can be seen that a generation increase in Xingó (bus 5061)

and a generation decrease in Marimbondo (20), CachoeiraDourada (754), Cana Brava (3964), and Serra da Mesa (36)improve the stability margin (tendency to make it positive).


Fig. 10. Ranking according to kinetic energy (Vke).

TABLE VPRIORITY PAIRS FOR REDISPATCH

Fig. 11. Sensitivity assessment—Xingó x SE Power Plants.

The generation increase in the Northeast implies reducing thegeneration of North or Southeast areas. The sensibilities repre-sent the expected behavior and it can be establish an order ofpriority for redispatch: .The model output indicated to use only power plants in the

Northeast, as shown in Fig. 12. The list of priority obtained is:. In Fig. 13 is depicted the

low sensitivity presented by Tucuruı́ (6425). As indicated bythe time-domain simulation, this power plant had little effect onthe stability margin improvement.

VI. CONCLUSION

This paper proposes a methodology for determining a sensi-tivity model between stability margin and generation redispatch.This methodology is based on neural networks trained with tem-poral data samples originated from transient simulations. Theadopted neural network is a multilayer perceptron with memorystructure in the input layer.The use of temporal data samples and the memory structure

in the neural network input layer provides a model with accurate

Fig. 12. Sensitivity assessment—Power Plants from Northeast.

Fig. 13. Sensitivity assessment—Power Plants fromNorth (Tucurı́ and Xingó).

stability margin estimation capability, and capacity for rankingcandidates for redispatch by sensitivity analysis.The proposed input variable selection procedure has shown

ability to tackle the problem of high dimensionality, inherent tolarge power systems. The application of this procedure to theBrazilian National System has proven its effectiveness.The developed preventive control rules are appropriate and

have been validated by time domain simulations, where theredispatch considering pairs of power plants has been eval-uated. Highly probable contingencies are already includedin market-based decisions. For less probable contingencies,preventive control would not align with market decisions. Theproposed approach can provide useful information with respectto the trade-off between cost and system vulnerability.

REFERENCES

[1] A. P. A. da Silva, “Overcoming limitations of NN’s for on-line DSA,”in Proc. IEEE PES General Meeting, San Francisco, CA, USA, Jun.2005.

[2] C. Y. Chung, L. Wang, F. Howell, and P. Kundur, “Generationrescheduling methods to improve power transfer capability con-strained by small-signal stability,” IEEE Trans. Power Syst., vol. 19,no. 1, pp. 524–530, Feb. 2004.

[3] P. Gomes, “New strategies to improve bulk power system security:Lessons learned from large blackouts,” in Proc. 2004 IEEE Power En-gineering Society General Meeting, Denver, CO, USA, Jun. 2004, vol.2, pp. 1703–1708.

[4] ONS—Operation Daily Report, Nov. 10, 2009 [Online]. Available:http://www.ons.org.br


[5] P. Gomes, A. C. S. Lima, and A. Guarini, “Guidelines for power systemrestoration in the Brazilian system,” IEEE Trans. Power Syst., vol. 19,no. 2, pp. 1159–1164, May 2004.

[6] M. R. Aghamohammadi, A. Maghami, and F. Dehghani, “Dynamicsecurity constrained rescheduling using stability sensitivities by neuralnetwork as a preventive tool,” inProc. IEEEPSCE Power SystemConf.Expo., Seattle, WA, USA, Mar. 2009, pp. 1–7.

[7] D. Z. Fang, Y. Xiaodong, and S. Jingqiang, “An optimal generationrescheduling approach for transient stability enhancement,” IEEETrans. Power Syst., vol. 22, no. 1, pp. 386–394, Feb. 2007.

[8] K. N. Shubhanga and A. M. Kulkarni, “Stability-constrained gener-ation rescheduling using energy margin sensitivities,” IEEE Trans.Power Syst., vol. 19, no. 3, pp. 1402–1413, Aug. 2004.

[9] J. L. Jardim, “Online dynamic security assessment: Implementationproblems and potential use of artificial intelligence,” in Proc. IEEEPES Summer Meeting, 2000, vol. 1, pp. 340–345.

[10] Brazilian ISO, Procedures of Operation Manual, In: Grid Procedures,Module 10, Submodule 10.21, Rev 1.1, 2010 [Online]. Available:http://www.ons.org.br

[11] J. L. Jardim, C. A. S. Neto, and M. G. Santos, “Brazilian system op-erator online security assessment,” in Proc. IEEE PES Power SystemsConf. Expo., Atlanta, GA, USA, Oct. 29, 2006, pp. 7–12.

[12] J. N. Fidalgo, J. A. P. Lopes, and V. Miranda, “Neural networks ap-plied to preventive control measures for the dynamic security of iso-lated power systems with renewables,” IEEE Trans. Power Syst., vol.11, no. 4, pp. 1811–1816, Nov. 1996.

[13] M. Djukanovic, D. J. Sobajic, and Y. H. Pao, “Neural-net based cal-culation of voltage dips at maximum angular swing in direct transientstability analysis,” Int. J. Elect. Power Energy Syst., vol. 14, no. 5, pp.341–350, 1992, Elsevier Science.

[14] J. L. Jardim et al., “A unified online security assessment system,” inProc. CIGRE Biennial Session, Paris, France, Aug. 2000.

[15] V. Vittal, E. Z. Zhou, C. Hwang, and A. A. Fouad, “Derivation of sta-bility limits using analytical sensitivity of the transient energy mar-gins,” IEEE Trans. Power Syst., vol. 4, no. 4, pp. 1363–1372, Nov.1989.

[16] C. R. Minussi and W. Freitas, “Sensitivity analysis for transient sta-bility,” IEE Proc. Gen., Transm., Distrib., vol. 145, no. 6, pp. 669–674,1998.

[17] B. D. A. Selvi and N. Kamaraj, “Transient stability assessment usingfuzzy SVM and modified preventive control,” Int. J. Elect., Comput.,Syst. Engineer, vol. 2, no. 1, p. 64, Mar. 2008.

[18] H. Vasconcelos, J. N. Fidalgo, and J. A. P. Lopes, “A general approachfor security monitoring and preventive control of networks with largewind power production,” in Proc. 14th PSCC Power System Computa-tion Conf., Sevilla, Spain, Jun. 24–28, 2002, p. 1, Session 31, paper 2.

[19] V.Miranda, J. N. Fidalgo, J. A. P. Lopes, and L. B. Almeida, “Real timepreventive actions for transient stability enhancement with a hybridneural network—Optimization approach,” IEEE Trans. Power Syst.,vol. 10, no. 2, pp. 1029–1035, May 1995.

[20] M. R. Aghamohammadi and A. Beik-Khormizi, “Small signal stabilityconstrained rescheduling using sensitivities analysis by neural networkas a preventive tool,” inProc. IEEEPES Transmission andDistributionConf. Expo., New Orleans, LA, USA, Apr. 19–22, 2010, pp. 1–5.

[21] A. D. P. Lotufo, M. L. M. Lopes, and C. R. Minussi, “Sensitivity anal-ysis by neural networks applied to power systems transient stability,”Elect. Power Syst. Res., vol. 77, pp. 730–738, Jul. 2006, Elsevier Sci-ence.

[22] IEEE Committee Report, A Tutorial Course on Artificial Neural Net-works With Applications to Power Systems. Piscataway, NJ, USA,IEEE Press, 1996, 96TP112-0.

[23] Y. Mansour, E. Vaahedi, and M. A. El-Sharkawi, “Dynamic securitycontingency screening and ranking using neural networks,” IEEETrans. Neural Netw., vol. 8, no. 4, pp. 942–950, Jul. 1997.

[24] L. S. Moulin, A. P. A. da Silva, M. A. El-Sharkawi, and R. J. Marks, II,“Support vector machines for transient stability analysis of large-scalepower systems,” IEEE Trans. Power Syst., vol. 19, no. 2, pp. 818–825,May 2004.

[25] K. Omata and K. Tanomura, “Transient stability evaluation using an ar-tificial neural network,” in Proc. 2nd Int. Forum Applications of NeuralNetworks to Power Systems, ANNPS’93, 1993, pp. 130–135.

[26] A. W. N. Izzri, A. Mohamed, and A. Hussain, “An improved methodin transient stability assessment of a power system using probabilisticneural network,” J. Appl. Sci. Res., vol. 3, no. 11, pp. 1267–1274, 2007,INSInet Publication.

[27] R. Ebrahimpour and E. K. Abharian, “An improved method in tran-sient stability assessment of a power system using committee neuralnetworks,” IJCSNS Int. J. Comput. Sci. Netw. Security, vol. 9, no. 1,pp. 119–124, Jan. 2009.

[28] R. Battiti, “Using mutual information for selecting features in super-vised neural net learning,” IEEE Trans. Neural Netw., vol. 5, no. 4, pp.537–550, Jul. 1994.

[29] I. N. Kassabalidis, L. S. Moulin, M. A. El-Sharkawil, R. J. Marks, andA. P. A. da Silva, “Dynamic security border identification using en-hanced particle swarm optimization,” IEEE Trans. Power Syst., vol.17, no. 3, pp. 723–729, Aug. 2002.

Mauricio C. Passaro (M’13) was born in São Paulo,Brazil, on April 25, 1968. He received, in Brazil, theB.Sc. and M.Sc. degrees from the Federal Univer-sity of Itajubá (UNIFEI) in 1994 and 2002, respec-tively, and the D.Sc. degree from Federal Universityof Rio de Janeiro (UFRJ) in 2013, all in electricalengineering.He has worked in Alstom and VA Tech. He is cur-

rently with the Brazilian Independent System Oper-ator (ONS) as a Senior Engineer. His area of expertiseis power systems dynamics and computational devel-

opment of advanced power systems security applications.Dr. Passaro is a Registered Professional Engineer in Brazil (CREA).

Alexandre P. Alves da Silva (M’92–SM’00)received the B.Sc. and M.Sc. degrees in electricalengineering (EE) from the Catholic University of Riode Janeiro, PUC-Rio in 1984 and 1987, respectively,and the Ph.D. degree in EE from the University ofWaterloo, Waterloo, ON, Canada, in 1992.He was a Visiting Scholar at the University of

Washington, Seattle, WA, USA, in 1999. He workedat the Federal University of Itajubá from 1993 to2002, and as a Full Professor at the Federal Univer-sity of Rio de Janeiro (COPPE) from 2002 to 2011,

where he was Chairman of the Electrical Engineering Graduate Program. Hehas published over 200 full papers in journals and conference proceedings.He has worked regularly as a consultant and manager of research projects formajor companies in the Brazilian energy sector. Currently, he is the Leader ofthe Center of Excellence in Smart Systems at the General Electric’s GlobalResearch, Brazil Technology Center, in Rio de Janeiro.Dr. Alves da Silva was the first Editor-in-Chief of the Brazilian Computa-

tional Intelligence Society Journal and he is Chairman of the Intelligent SystemsSubcommittee (PSACE) of the IEEE Power and Energy Society.

Antonio C. S. Lima (S’95–M’00) was born in Riode Janeiro, Brazil, in 1971. He received the B.Sc.,M.Sc., and D.Sc. degrees in electrical engineeringfrom the Federal University of Rio de Janeiro(UFRJ) in 1995, 1997, and 1999, respectively.In 1998, he was a Visiting Scholar with the

Department of Electrical and Computer Engineeringat the University of British Columbia, Vancouver,BC, Canada. From 2000 to 2002, he was with theBrazilian Independent System Operator, ONS, Riode Janeiro, Brazil, dealing with electromagnetic

transient studies for the Brazilian National Grid. Currently, he is an AssociateProfessor with the Electrical Engineering Department, UFRJ, where he hasbeen since 2002.

06786495

Documents

power system security

large system

federaluniversity of

brazilian national power

prefault system data

brazil email

transient stability

senior member