IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO.

2, MAY 2002

297

Online Preventive Dynamic Security of Isolated Power Systems Using Decision Trees
E. S. Karapidakis and N. D. Hatziargyriou, Senior Member, IEEE
AbstractIn this paper, the application of decision trees (DTs) for on-line preventive dynamic security of isolated power systems is presented. DTs provide the necessary computational speed for on-line performance and the flexibility of providing preventive control. Emphasis is placed on the on-line use of the DTs to test the dynamic security of each generation dispatch scenario and thus to provide corrective advice via generation redispatch. Moreover, the algorithm implemented provides the flexibility of displaying the cost of each redispatch. In this way, the method can help in objective decision-making. Results from the application of the system on actual load series from the island of Crete, where the proposed system is in trial operation, are presented. Index TermsArtificial intelligence, decision trees (DTs), dynamic security, economic dispatch, isolated power systems, preventive control, wind power.

I. INTRODUCTION SOLATED power systems, like the ones operating in large islands, face increased problems related to their operation and control. In most of these systems, the cost of electricity production is much higher than in interconnected systems due to the high operating costs of their thermal generating units, mainly diesel and gas turbines, and the import and transportation costs of the fuel used. Security is also a major concern, since mismatches in generation and load and/or unstable system frequency control might lead to system failures, easier than in interconnected systems. The introduction of increased share of wind power, economically attractive where this potential exists, can result in a further decrease in dynamic security. Thus, next to the more common angle and voltage stability concerns, frequency stability [1], [2] must be ensured. This depends on the ability of the system to restore balance between generation and load following a severe system upset with minimum loss of load. Traditionally, dynamic security has been accounted for by assuring sufficient reserve capacity in the system, so that it can withstand various disturbances. The determination of this reserve capacity was based on off-line stability analysis of credible contingencies under a variety of operating conditions. This was implemented as a fixed emergency reserve (security margin) in the economic dispatch algorithms, allocating generation to the on-line units, so that the system load was supplied
Manuscript received May 15, 2000; revised October 26, 2001. This work was supported by European Commission DGXII as the R&D project MORE CARE: More Advanced Control for Secure Operation of Isolated Power Systems with Increased Renewable Energy Penetration and Storage under Contract ERK5-CT-1999-00019. The authors are with the Department of Electrical and Computer Engineering, National Technical University of Athens, Athens 15773, Greece (e-mail: karapid@power.ece.ntua.gr; nh@power.ece.ntua.gr). Publisher Item Identifier S 0885-8950(02)03822-1.

most economically. Carrying unnecessary reserve capacity however might significantly increase the cost of operation. This is inevitable if a fixed security margin is assumed, since the dynamic performance of the system depends critically on its operating conditions and the dynamic characteristics of the generation response, as shown in [3]. The new operating and planning policies imposed by the current demands for increased efficiency and economy, dictate the need to operate the power system in a more stressed way and to consider more effective means of providing a given level of security. On-line dynamic security assessment (DSA) techniques have been developed aiming to select critical contingencies and assess the dynamic performance of the system for each critical contingency on-line [4]. Only indicative references of a very dynamic activity in this area are provided in this paper due to lack of space. These include techniques based on energy functions [5][7] and artificial intelligence [8][10]. Moreover, in order to operate optimally within the new market conditions, the price of providing a given level of security has to be accounted for. This is directly linked to the provision of remedial actions, in case of insecure situations. For dynamic security, unlike steady-state security, remedial actions can only be preventive leading to load shedding or generation rescheduling [11], [12]. For isolated systems, this means that the cost of load shedding has to be balanced with the cost of providing adequate spinning reserves to avoid it. In [12], a successful methodology for the determination of reserves in isolated power systems is presented, which enables assessment of the costs for the loss of any given unit. It is based on the dynamic simulation of the system after each dispatch taking into account the dynamic response of the generators and the demand and calculates minimum frequency so that load shedding is avoided. The long execution times associated with full dynamic simulations, however, make the on-line application of this method unsuitable for larger systems. The application of decision trees (DTs) is therefore proposed in this paper. Online DSA functions based on DTs, as well as artificial neural networks (ANNs), and kernel regression trees (KRTs) [13][16] have been developed and integrated within an advanced control system tailored for the needs of isolated power systems with increased renewable power penetration within the framework of the CARE European R&D project [17][19]. The control system has been installed on Crete, the largest Greek island, and is currently used to assist its operators by proposing optimal operating setpoints for the various power units, both conventional and renewable. The DSA module has the task of evaluating if a given dispatch scenario is dynamically secure for a number of predetermined disturbances. This paper focuses

0885-8950/02$17.00 2002 IEEE

298

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002

on the application of DTs for the determination of optimal reserves and the provision of corrective advice. Security margins are not included as constraints in the Langrangian relaxation based methods for solving economic dispatch. Instead, based on the DT classification, new unit dispatch is calculated on-line, until a dynamically secure operating state is reached. This technique provides the flexibility of displaying the cost of each proposed solution, it can therefore be used as a valuable on-line decision making aid. Results from the application of the method on actual load series from the island of Crete are presented. It is believed that the fast execution times required by the DTs for classification make the method suitable for on-line control of larger systems. II. OVERVIEW OF METHODOLOGY A. Decision Trees The DT methodology is a nonparametric learning technique able to produce classifiers about a given problem in order to deduce information for new unobserved cases. The DT has the hierarchical form of a tree structured upside down. The construction of a DT is based on a knowledge base (KB) consisting of a large number of operating points (OPs) covering all possible states of the power system under study in order to ensure its representativity. A vector of predisturbance steady-state variables, called attributes, characterizes each OP. The KB is divided in a learning set (LS) used for deriving the classifier structures and a test set (TS) used to evaluate the performance of these structures on new, unobserved OPs. The construction of a DT starts at the root node with the whole LS of preclassified OPs. These OPs are analyzed in order to select the dichotomic test that splits them optimally into a number of most purified, mutually exclusive subsets. For a two-class partition (secure or insecure) the test is defined as (1) where is the optimal threshold value of the chosen attribute . The selection of the optimal test is based on maximizing the additional information gained through the test. The information with respect to gain is based on the entropy of each subset the class partition of its elements, defined as (2) and are the relative frequencies of the secure and where insecure OPs, respectively, in the subset. The selected test is applied to the LS of the node splitting it into two exclusive subsets, corresponding to the two successor nodes. Every subset (node) is characterized by its security index (SI), defined as the percentage of secure OPs belonging to this subset. The optimal splitting rule is applied recursively to build the corresponding subtrees. In order to detect if one node is terminal, i.e., sufficiently class pure, the stop splitting rule is used, which checks whether the entropy of the node is lower than a preset minimum value. If it is, the node corresponds to a sufficiently pure subset (states belong to the same class) and is declared a leaf; otherwise, a test is sought to further split the

node. If the node cannot be further split in a statistically significant way, it is termed a deadend, carrying the two class probabilities estimated on the basis of the corresponding OPs subset. A detailed technical description of this method is given in [10]. In our application of the DTs to DSA, the unit productions and spinning reserves are selected as the initial attributes characterizing each OP. In this way, the rules provided by the DT structures are directly linked to reserves optimization. This is different than previous works of the same authors applying DTs to the DSA of isolated systems [14][16] aiming at the on-line determination of the optimal level of wind power penetration, where the initial attributes were selected accordingly. These different applications reveal the versatility of the applied DT technique. B. Secure Economic Dispatch Economic dispatch analysis determines the power setpoints of the online generating units (3), so as to meet the system load and losses at least cost (3) where total conventional generation; generation of the th unit; number of units. Traditional dispatch algorithms tackle this problem as a constrained optimization problem and base its solution on the concept of equal incremental cost, also known as the Lambda Iteration algorithm: The total production cost of a set of generators is minimized, when all the units operate at the same incremental cost. In order to ensure that the operating setpoints proposed by the economic dispatch algorithm will provide a dynamically secure operating state of the system following prespecified disturbances, the rules extracted by the relevant DTs can be used as additional constraints in the above constrained optimization problem. An alternative implementation is shown in Fig. 1. The economic dispatch and dynamic security functions are performed successively as part of an inner loop, also including load and renewable forecasting and unit commitment functions [17][19]. Following extensive tests on Crete, the iteration step of this inner loop has been set to 20 min with an 8-h optimization horizon. In addition, it was proven that the execution of an outer loop every 4 hours with a 48-h optimization horizon improves significantly the performance of the forecasting and unit commitment functions. Economic dispatch is performed without dynamic security constraints. The generator power setpoints are subsequently input in the DT structure, in order to estimate the security of the proposed operating state. Crossing the DT requires a simple checking of few rules and leads to a leaf of secure or insecure OPs. If the leaf is secure, the units set points are displayed to the operator. In case of an insecure leaf or a deadend with a low SI, an alternative dispatch solution in the neighborhood of the previous solution is sought for, as follows: The test of the last splitting node of the DT is recorded and its threshold value

KARAPIDAKIS AND HATZIARGYRIOU: ONLINE PREVENTIVE DYNAMIC SECURITY

299

lated from the cost functions of the generating units, once the unit productions have been determined. In addition, the security cost can be compared to the cost of load shedding. The unsupplied electric energy can be easily calculated from the operating settings of the under-frequency relays and the load forecasted at each bus affected. Alternatively, it can be estimated from the predisturbance load and the forecasted load as a whole, however its cost is more difficult to determine. For the dispatcher the cost of load shedding can be the price the regulator imposes for energy not served [12]. In the traditional monopoly operation, this cost can be the revenue lost due to the unsupplied electric energy, although this by no means reflects the true cost of load shedding. In any case, the total cost can be calculated from (5) where load shed; cost of kilowatthour in Euros (C); time of load disconnection. III. APPLICATION SYSTEM
Fig. 1. Implementation of dynamic security function in the on-line control software.

is used as equality constraint fixing the power produced by the respective generating unit. Equation (3) becomes (4) and the number of the The total generation is now . This procedure is applied iteratively units for dispatch is until crossing the DT leads to a terminal node with a secure index. It should be noted that the above algorithm is heuristic and does not necessarily lead to optimal solutions from an economic point of view. This is due to the fact that the security region defined by the DT is in general nonconvex (rather a union of convex regions), therefore the method does not provide a global optimum in this region. The algorithm however provides secure OPs subject to the error rates of the TS. This is because, although the KB has been obtained by unconstrained economic dispatch, it does provide a representative sample of both secure and insecure OPs. The error rates should apply therefore equally for the redispatched OPs, as well. In case no secure OP is found, a new Unit Commitment has to be applied. This case can be easily tackled in a relatively simple system, in more complex systems however this is not straightforward and is not dealt with in this paper. C. Cost Analysis The approach described in Section II-B provides the flexibility of displaying the cost of security, i.e., the cost associated with each redispatch. This is easily provided as the difference between the operating cost of the original dispatch and the operating cost of the secure redispatch. These costs can be calcu-

The power system of the island of Crete is the largest autonomous power system in Greece with the highest rate of increase in energy and power demand nationwide. The transmission network consists mainly of 150-kV lines and some 66-V lines. The generation system consists of two major power plants one in Linoperamata and one in Chania located near the major load points of the island. There are 18 thermal, oil-fired generating units with a total capacity of 523.2 mW installed, including six steam units of total capacity 112 mW, four diesel units with 49.2 mW, seven gas turbines with 228,6 mW, and one combined cycle plant with 133,4 mW. In 1998, the peak load exceeded 380 mW, while the lowest load was about 100 mW. One characteristic of the load profile is the large variations (low night valleyshigh evening peaks). The base-load is mainly supplied by the combined cycle and the steam units. The gas turbines normally supply the daily peak load or load that cannot be supplied by units in outage. Gas turbines have a high operating cost, which increases significantly the average cost of the electricity being supplied. Eleven wind parks of a total capacity of 80 mW have already been installed and many more are planned to be installed in Crete in the near future, operated mainly by independent power producers. Due to the high cost of the conventional energy generation, the introduction of increased wind power production proves economically attractive; however, it complicates significantly the operating conditions of the system. Regarding security, all of the wind parks, with few exceptions, are installed at the eastern part of the island, which presents the most favorable wind conditions. As a result, in case of faults on some particular lines, the majority of the wind parks are disconnected. Furthermore, frequency oscillations might easily trigger the under-frequency protection relays of the wind parks, thus causing further imbalance in the system generation/load. A maximum wind power penetration of 30% the load has been used by the system

300

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002

TABLE I SYSTEM GENERATING UNIT GROUPS

Fig. 2. Simulation results of Crete power system.

operators as the respective security margin. However, extensive transient analysis studies conducted in order to assess the dynamic behavior of the system under various disturbances and with different combinations of the generating units have shown that such a fixed security margin does not guarantee the system security and it distorts its economical operation. Thus, under the same contingency the system is shown to collapse with lower than 30% of wind power penetration, while it can survive with higher penetrations [3]. As an example, the change of frequency caused by the outage of a gas turbine, providing 23 mW under two different operating conditions is shown in Fig. 2. Case 1 corresponds to a total load of 207.2 mW supplied as follows: 27 mW by combined cycle (18 mW of spinning reserve), 56.8 mW by the new steam turbines (18.2 mW spinning reserve), 21.3 mW by diesel (27.9 mW spinning reserve), 10.1 mW by the remaining gas turbine (6.1 mW spinning reserve of maximum 16.2 mW), while the wind power is 69 mW, corresponding to 33.3% penetration. It can be seen that the frequency undergoes a severe transient reaching a lowest value of 49.1 Hz; however, the system restores its balance in about 50 s. Case 2 corresponds to a lower load of 199 mW supplied by 27.57 mW of combined cycle (17.43 mW spinning reserve), 69.3 mW of new steam turbines (5.7 mW of spinning reserve), 23.4 mW of diesel (25.8 mW of spinning reserve), and 55.73 mW of wind corresponding to 28% penetration. Although the wind power penetration is lower than the security margin adopted, the system does not manage to regain its stability and is led to frequency collapse. The difference is attributed to the fact that in the first case the spinning reserve is higher (70.2 mW) and provided by faster units (gas turbines), while in the second case by slower units (48.93 mW). The need for spinning reserve optimization can be clearly seen.

For each of the load buses, a perturbation of approximately 30 mW is applied around each of the above operating profiles, while wind parks generation varies from 0 mW to 80 mW. The combination of these four operating profiles covers a load demand from 90 mW to 330 mW. A dispatch algorithm approximating actual operating practices followed in the control system of Crete is applied next in order to complete the predisturbance and wind power , the total OPs. For a given load demand conventional generation (6) is dispatched to the units in operation, depending on their type and their nominal power. The various thermal units are grouped according to their type, as shown in Table I. The attributes characterizing each OP comprise the active power and spinning reserve of all conventional power units. Ten variables are selected as initial attributes. Five attributes correspond to the active production of the conventional unit groups and five attributes to the spinning reserves, respectively. These are coded, as shown in Table I. The spinning reserve of each group is denoted by the extension Res (e.g., CombRes means the spinning reserve of the combined cycle plant). The LS comprises 3748 OPs and the TS 1987 OPs. These numbers were chosen using a trial and error procedure, examining the learning and test errors. These errors are calculated by using the derived security structures to classify the OPs of the LS (used for training) and the independent OPs of the TS, respectively. It should be noted that a larger LS decreases the learning error, it produces however DTs with poor generalization capabilities due to overfitting, i.e., new, unobserved OPs are less successfully classified. For each of the OPs produced, two characteristic disturbances have been simulated using EUROSTAG [20] 1) outage of a major gas turbine; 2) three-phases short-circuit at a critical bus near the wind parks. The first of these disturbances happens very frequently, while the second is particularly severe leading to the disconnection of most wind parks. For each OP the maximum frequency deviation and the rate of change of frequency are recorded. Both of these parameters are checked against the values activating the under-frequency relays used for load shedding and the OPs are labeled accordingly. The security criteria were: Hz and then If the system is insecure else is secure. Using the approach described in this section, 5735 acceptable OPs have been obtained, which are divided into two sets.

IV. DERIVATION OF DECISION TREES A. Creation of the Knowledge Base For the creation of the KB a large number of initial OPs are obtained by varying randomly the load at each load busbar and the wind power at each wind park. These variables are assumed to follow normal distributions around four operating profiles: mW; 1) low-load with a total load mW; 2) medium-load with mW; 3) high-load with mW. 4) peak-load with

KARAPIDAKIS AND HATZIARGYRIOU: ONLINE PREVENTIVE DYNAMIC SECURITY

301

TABLE II EVALUATION OF FIRST DECISION TREE PERFORMANCE

Fig. 3.

Decision tree for short circuit.

TABLE III EVALUATION OF SECOND DECISION TREE PERFORMANCE

Fig. 4. Decision tree for machine outage.

B. Decision Tree Development After several studies, based on the previous LS, the DTs shown in Figs. 3 and 4, are developed. In all nonterminal nodes, information related with the number of the OPs belonging to that node, the splitting test and the SI, defined as the per unit number of secure OPs, are included. In the terminal nodes, the number of the OPs, the SI, and the type of the node are given. Based on the DTs, rules regarding the secure operation of the system in the event of the simulated disturbance are derived. Using the rules of the DT of Fig. 4, it can be easily verified that the OP corresponding to Case 1 of Fig. 2 is classified as secure (led to node number 6), while the one corresponding to Case 2 is classified as insecure (led to node number 7). The quality of the classification rules is evaluated by the global error, the false alarm, and the missed alarm rate obtained, when tested with the independent TS. The global error is defined as the ratio of the misclassified OPs to the total number of the OPs. The ratio of the secure OPs that are classified as insecure to the total number of the secure OPs is defined as the false alarm rate, while the ratio of the insecure OPs that are classified as secure to the total number of the insecure OPs is defined as the missed alarm rate, respectively. The classification performance evaluation for the developed DTs and the SI of each set are shown in Tables II and III. V. PREVENTIVE SECURITY From the developed DTs, interpretable security rules of the nonlinear behavior of the power system in case of the selected disturbances can be extracted. For example, from the first DT it

Fig. 5. Manmachine interface of online DSA.

can be easily observed that the system is classified as secure, if the combined cycle generation exceeds 70.4 mW and has over 30.5 mW spinning reserve. Similarly, the system is classified as insecure, if the combined cycle production is less than 54.4 mW and the gas turbines provide less than 43 mW. This is apparently related to a lower spinning reserve provided by on-line gas turbines. In general, it is shown that the combined cycle attributes appear many times in both DTs, indicating the importance of this base-unit in the secure operation of the power system. DT, ANN, and KRT structures have been integrated in the dispatch center of the power system of Crete. In Fig. 5, the security assessment screen displayed on call is presented. At the top of the screen, the load at critical buses, the current production of the wind parks, and the production of the various thermal units in the two thermal stations are displayed in the form of bar charts. The DSA results for 48 h ahead are displayed in the main screen under the forecasted load curve in the form of lines representing the expected frequency in case of the considered disturbances. The expected frequency is provided by the ANNs

302

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002

Fig. 6. Twenty-four hour diagrams illustrating the load, the security classification, and the operating cost.

Fig. 7.

Effect of dispatch on system frequency deviation.

or KRTs structures. It is shown that the expected frequency deviation in the short-circuit event is unacceptable for most of the time. This contingency has a low probability of occurrence, unless the weather conditions are bad, and the operator might select to ignore it. Between 1:00 and 8:00, however, the expected frequency deviation is shown to be below the 49-Hz threshold for both contingencies simulated. It is characteristic that this period corresponds to the low load period, when a significant wind power penetration can lead to poor dynamic security. In this case, the operator can seek alternative dispatch using the approach developed in Section II. For each redispatch, the frequency variation is updated and the relevant cost is displayed. VI. COST OF SECURITY In this section, results from the application of the secure economic dispatch algorithm on actual load series of Crete are presented. These results have been obtained using the software facilities presented in the Appendix. In Fig. 6, the total load, the corresponding security classification (1 for secure and 0 for insecure) for the machine outage contingency and the operating cost in Euros of a characteristic day are plotted. In the upper diagram, it is shown that, approximately between 9:00 and 10:30, the system is insecure, i.e., at least a significant load shedding will take place. In the lower diagram, the effects of the secure economic dispatch algorithm on the security classification and the system operating costs are shown. The increase of costs during the previously insecure period, provided by the increased and probably faster (more expensive) spinning reserve, is notable. The effect of the two dispatch scenarios on the system frequency deviation, in the case of the machine outage, as obtained by EUROSTAG, is shown in Fig. 7. It is clearly shown that the proposed redispatch will not cause load shedding. The probability of the contingency occurrence however is not considered in this paper. In Figs. 8 and 9, the total generation in milliwatts, and the costs in thousand Euros resulting from the secure economic dispatch and the economic dispatch without dynamic security concerns for the actual data of a whole week are presented, respec-

Fig. 8. Total system generation for one week.

Fig. 9. One week operating cost without dynamic security constraints versus operating cost of dynamically secure operation.

tively. It is clear that the system operation cost rises, when the corrective redispatch procedure of the DTs is executed. It is interesting to compare this rise in cost with the cost of load shedding. For each time period, the amount of the disconnected load can be easily obtained considering the activation settings of the under-frequency load shedding relays and their location in the system, as shown in Tables IV and V For example, the first insecure operating state is detected at 9:04 AM, when the total load demand was 121.9 mW. In the case of the gas turbine outage, the frequency change rate would exceed 0.4 Hz, causing load shedding at four system buses with

KARAPIDAKIS AND HATZIARGYRIOU: ONLINE PREVENTIVE DYNAMIC SECURITY

303

TABLE IV LOW FREQUENCY LOAD SHEDDING

TABLE V FREQUENCY CHANGE LOAD SHEDDING

TABLE VI ECONOMIC RESULTS COMPARISON Fig. 10. Insecure DT classification.

a total load of 9.73 mW. Without corrective actions, the system was classified as secure at 10:30 AM, when the total load demand was 195.5 mW, and the total load of the previous buses was 18.3 mW. Based on the analysis of Section II-C, the net cost of the suggested corrective economic dispatch can be obtained. For the cost of load shedding the revenue lost by the utility due to the unsupplied energy is considered. For the whole week tested, seven time periods of insecure system classification were identified. The load shedding or unsupplied energy cost (UEC) and the additional production cost (APC) incurred by the corrective redispatch are shown in Table VI. The total net cost of the security provided is 5939.25 , which amounts to 19.2% of the total extra system operation cost of 30 935.41 . It should be noted that in Cases 4 and 5, the UEC is lower than the extra generation cost. This result would suggest that based on purely economic criteria the dispatcher could prefer the risk of load shedding rather than applying the redispatch suggestions. It should be noted that in this study, the possible disconnection of wind parks, which could improve system stability in case of insufficient spinning reserve was not considered. The development of suitable DTs is of course possible, if the wind power penetration was included in the initial attribute list, as shown in [16], [17]. In this case, it is assumed that the dispatcher would occasionally accept wind power up to the level the system security was ensured. On the other hand, the attributes selected in this paper allow the secure increase of wind power penetration suggesting unit redispatch to ensure security. Alternatively, the

Fig. 11.

Corrective redispatch at higher cost.

selection of interruptible loads as initial attributes and the preparation of the appropriate KB would provide load-shedding advice. These different DT applications reveal the versatility and flexibility of the method to perform control. In addition, the fast execution time required, make the method particularly suitable for on-line preventive dynamic security applications for large systems, as well. VII. CONCLUSIONS Isolated power systems, like the ones operating in large islands, face increased problems related to their operation and control. In this paper, the application of DTs, in order to determine optimal reserves and to provide corrective advice considering dynamic security of these systems, is shown. Based on the DT classification, new unit dispatch is calculated on-line, until a dynamically secure operating state is reached. This technique provides the flexibility of displaying the cost of each proposed solution weighted against the cost of load shedding; it

304

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 17, NO. 2, MAY 2002

forms therefore the basis for valuable decision-making aid. Results from the application of the method on actual load series from the island of Crete show the accuracy and versatility of the method. Moreover, the fast execution times required for on-line classification of the current operating state by the DTs make the method suitable for large systems, as well. APPENDIX Corrective Security Software The algorithm described in Section II has been implemented in a windows-based program, suitable for off-line DSA. Figs. 10 and 11 depict the DT path corresponding to an insecure state and the alternative path provided by the unit redispatch together with the related operating costs, respectively. ACKNOWLEDGMENT The authors wish to thank all partners in the project, especially Prof. J. A. Pecas Lopes from INESC, Portugal, for his valuable contribution. REFERENCES
[1] P. Kundur, Power System Stability and Control. New York: McGrawHill, 1994. [2] P. Kundur and G. K. Morison, A review of definitions and classification of stability problems in todays power systems, in Proc. Panel Sess. Stability Terms Definitions, IEEE Power Eng. Soc. Meet., New York, Feb. 26, 1997. [3] N. Hatziargyriou, E. Karapidakis, and D. Hatzifotis, Frequency stability of power systems in large islands with high wind power penetration, in Proc. Bulk Power Syst. Dyn. Contr. Symp., Part IVRestructuring, Santorini, Greece, Aug. 2428, 1998. [4] CIGRE TF38.02.13, New trends and requirements for dynamic security assessment, CIGRE, 1997. [5] E. Vaahedi, Y. Mansour, and E. Tse, A general purpose method for on-line dynamic security assessment, IEEE Trans. Power Syst., vol. 13, pp. 243249, Feb. 1998. [6] W. Li and A. Bose, A coherency based rescheduling method for dynamic security, IEEE Trans. Power Syst., vol. 13, pp. 810815, Aug. 1998. [7] A. L. Bettiol, L. Wehenkel, and M. Pavella, Transient stability-constrained maximum allowable transfer, IEEE Trans. Power Syst., vol. 14, pp. 654659, May 1999. [8] M. El-Sharkawi and D. Neibur, Artificial neural networks with applications to power systems, IEEE Power Eng. Society Special Pub. TP-112-0, 1996.

[9] L. Wehenkel and M. Pavella, Eds., Special Issue on Automatic Learning Application to Power Systems of the Revue-E. Brussels, Belgium: SRBE, Dec. 1996. [10] L. Wehenkel, Automatic Learning Techniques in Power Systems. Norwell, MA: Kluwer, 1998. [11] M. La Scala, M. Trovato, and C. Antonelli, On-line dynamic preventive control: An algorithm for transient security dispatch, IEEE Trans. Power Syst., vol. 13, pp. 601610, May 1998. [12] J. W. OSullivan and M. J. OMalley, A new methodology for the provision of reserve in an isolated power system, IEEE Trans. Power Syst., vol. 14, pp. 519524, May 1999. [13] N. Hatziargyriou, S. Papathanassiou, and M. Papadopoulos, Decision trees for fast security assessment of autonomous power systems with large penetration from renewables, IEEE Trans. Energy Conv., vol. 10, June 1995. [14] N. Hatziargyriou, On-line dynamic security assessment of isolated systems, in Proc. Panel Sess. Intell. Syst. Appl. Dynamic Security IEEE Winter Power Meet., New York, Jan. 31Feb. 5, 1999. [15] N. Hatziargyriou, J. A. Pecas Lopes, E. Karapidakis, and M. H. Vasconcelos, On-line dynamic security assessment of power systems in large islands with high wind power penetration, in Proc. 13th PSCC, Trondheim, Norway, June 28July 2, 1999. [16] N. Hatziargyriou, J. A. Pecas Lopes, J. Stefanakis, E. Karapidakis, M. H. Vasconcelos, and A. Gigantidou, Artificial intelligence techniques applied to dynamic security assessment of isolated systems with high wind power penetration, in Proc. Sess. CIGRE, Paris, France, Aug. 2000. [17] CARE: Advanced control advice for power systems with large-scale integration of renewable energy sources, CARE, Contract JOR3-CT960119, Final Rep., Aug. 1999. [18] N. Hatziargyriou, Ed., Special Issue of Wind Engineering, 1999, vol. 23. [19] N. Hatziargyriou, G. Contaxis, M. Papadopoulos, B. Papadias, J. A. Peas Lopes, M. Matos, G. Kariniotakis, E. Nogaret, J. Halliday, G. Dutton, P. Dokopoulos, A. Bakirtzis, A. Androutsos, J. Stefanakis, and A. Gigantidou, Operation and control of island systemsThe Crete case, in Proc. IEEE Power Eng Soc. Winter Meet., Singapore, Jan. 2337, 2000. [20] Eurostag 3.2 Package users guide.

E. S. Karapidakis received the degree in electrical and computer engineering in 1997. He is currently pursuing the Ph.D. degree with the Department of Electrical Engineering, National Technical University of Athens, Athens, Greece. His research interest is in dynamic security assessment using artificial intelligence techniques.

N. D. Hatziargyriou (S80M82SM90) is currently a Professor in the Power Division of the Department of Electrical Engineering, National Technical University of Athens, Athens, Greece. His research interests include dynamic security, artificial intelligence techniques, and dispersed generation. Mr. Hatziargyriou is a member of CIGRE SC38 and is currently Chairman of the IEEE Greek Power Chapter.