Neuro-Fuzzy Decision Trees for Dynamic Security Control of Power Systems A.K. Bikas, E. M. Voumvoulakis and N. D. Hatziargyriou Department of Electrical and Computer Engineering NTUA Athens, Greece bikas.gr@gmail.com , emvoumv@power.ece.ntua.gr , nh@power.ece.ntua.gr Abstract—This paper addresses the problem of dynamic security classification as well as security control of power systems., using class pattern recognition. More specifically, Neuro-Fuzzy Decision Trees (N-FDTs) are proposed i.e. fuzzy decision tree structure with neural like parameter adaptation strategy, in order to classify the security status of a power system. The method is applied on a realistic model of the Hellenic Power System, investigating two cases. The first case focuses on stressed operation of the power system and proposes corrective load shedding to avoid voltage instability. The second state investigates the scenario of large scale wind power integration to the system, and proposes wind power shedding as a preventive means to avoid Keywords-Dynamic Security Assessment, Corrective Control, Preventive Control, Load Shedding, Neuro-Fuzzy Decision Tree I. INTRODUCTION Difficulties in expanding the generation and transmission system force modern Power Systems to operate often close to their stability limits, in order to meet the continuously growing demand. Security is defined as the capability of maintaining the continuous operation of a power system under normal operation and following significant perturbations [1]. Dynamic Security Assessment are methodologies for evaluating the stability and quality of the transient processes between the pre- contingency and post-contingency states. In this case, DSA aims at ensuring that the system will be stable after the contingency occurrence and that the transients caused by such a contingency will be well damped, of small amplitude and with little impact on the quality of service. In the field of Dynamic Security Assessment (DSA) [2], [3] much attention has been paid to preventive, as well as corrective control. Preventive control refers to a set of actions that are applied when a potentially dangerous violation is detected through DSA. Corrective actions are applied to offset a security violation after the occurrence of a threatening contingency.[4], [5], [6], [7]. Analytical methods assess the security of a power system, by performing numerical simulation of a contingency in a given state. However, the nonlinear nature of the models expressing physical phenomena and the growing complexity of modern power systems make on-line security assessment a very challenging task. Thus, for large power systems in which many contingencies must be assessed, even with multiple- CPU computing, full simulation methods require times often unsatisfactory for online analysis. In this aspect, Intelligent Systems (IS) are seen to have features that can bring benefits in comparison to analytical methods [8]. Once developed, IS technologies as neural networks (NN) and decision trees (DT) can provide solutions very fast. Furthermore, IS are learning systems, i.e. they have the ability to recognize, if a system condition has previously occurred and predict its security accordingly. Similarly, if properly designed, an IS can adapt to new conditions by learning from situations previously seen. Finally IS can provide a high degree of discovery, i.e they have the ability to uncover salient, but previously unknown, characteristics of, or relationships in, a system. In this paper, an automatic learning framework for DSA and corrective control determination has been developed. A dataset is obtained from a large set of dynamic simulations of the defined disturbances, comprising Operating Points (OPs) that cover the behavior of the power system in the examined operating region. Each OP is described by a vector of pre- disturbance steady-state variables, called features, that can be either directly measured or indirectly calculated quantities, and is characterized as secure or insecure on the basis of the post- disturbance values and the security criterion. For each insecure OP a number of corrective control actions are simulated and their effect on the security status of the system is recorded. The derived dataset of OPs is split into a Learning Set (LS), used to derive security evaluation structures, and a Test Set (TS) used for testing the developed structures. A N-FDT is trained offline to assess the security status of the system. Inverse reading of the paths along the nodes of the constructed N-FDT provides the necessary security control. The method is applied to the Hellenic Power System investigating two study cases. The first case focuses on peak loading of the system with large power transfer from the generating areas in the North and West of Greece to the main load center in the Athens metropolitan area. The model determines the amount of load that must be shed in order to avoid voltage instability under the event of a unit loss. The second case studies the operation of the system projected to 2012, assuming large scale wind power integration [15]. The IS