3028 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 28, NO. 3, AUGUST 2013 Transient Instability Prediction Using Decision Tree Technique Turaj Amraee, Member, IEEE, and Soheil Ranjbar Abstract—This paper presents a decision tree based method for out-of-step prediction of synchronous generators. For distin- guishing between stable and out-of-step conditions, a series of measurements are taken under various fault scenarios including operational and topological disturbances. The data of input fea- tures and output target classes are used as the input-output pairs for decision tree induction and deduction. The merit of decision tree based detection of transient instability lies in robust classification of new unseen samples. The performance of the proposed method is verified on two test cases including a 9-bus dynamic network and the practical 1696-bus Iran national grid. The simulation results are presented for various input features and learning parameters. Index Terms—Decision tree, out-of-step, power swing, predic- tion, transient stability. I. INTRODUCTION N OWADAYS power systems play an important role in human life. Power system performance depends on its ability to maintain a desired level of stability and security under various fault conditions. Transient or large signal rotor angle stability is one of most important types of stability phe- nomena. Transient rotor angle stability is the ability of power system to maintain its synchronism under a severe fault such as three-phase short circuit [1]. Power system stability under different disturbances depends on installed control equipment to damp electromechanical oscillations. To minimize the spread of an undesired disturbance and the damage to generators, it is required to design appropriate protective schemes. Many techniques have been proposed for out-of-step detec- tion. The most commonly used out-of-step detection technique acts based on the concept of blinders in the relay impedance plane. This method requires some information about the fastest power swing, the normal operation region, the possible swing frequencies, and other system specifications [2]. Another tech- nique for out-of-step detection is based on analytical methods like equal area criterion (EAC). This technique is suitable for a single machine infinite bus (SMIB) system. For large scale power systems with many generators and transmission lines, the analytical methods such as equal area criterion fail to give ef- ficient results. Therefore the time simulations are used for sta- Manuscript received June 30, 2012; revised October 09, 2012 and November 27, 2012; accepted December 31, 2012. Date of publication February 01, 2013; date of current version July 18, 2013. Paper no. TPWRS-00747-2012. T. Amraee is with the Department of Electrical Engineering, K.N. Toosi Uni- versity of Technology, Tehran (14317-14191), Iran (e-mail: amraee@kntu.ac. ir). S. Ranjbar is with the Science and Research Branch, Islamic Azad University, Tehran, Iran (e-mail: ranjbar.slr@gmail.com). 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.2013.2238684 bility and security assessment of large scale power systems. In this method, the set of differential algebraic equations (DAE) of power system are solved using a suitable algorithm. In [3], an out-of-step detection technique has been proposed based on artificial neural network. In this method, the authors have used the kinetic energy deviation, mechanical input power and average acceleration during fault as input features for neural network training. In [4] a neural network based hybrid scheme has been proposed for system wide detection of transient instability. In [5], the application of fuzzy logic using an adaptive nero-fuzzy inference system (ANFIS) has been suggested for out-of-step detection. In this method the angular frequency deviation and the impedance angle measured at the machine terminal are used as inputs to the fuzzy logic. The performance of fuzzy logic and neural network methods highly depends on the user-defined training parameters (e.g., membership function, parameters of defuzzification algorithm, neuron function, etc.). The concept of energy function is an- other technique has been proposed for out-of-step detection [6], [7]. During unstable power swings, the entire power system oscillates in two groups. The series elements that connect these groups are called cutsets. By evaluating the potential energy of the cutset, the stable and unstable conditions are predicted. This technique needs wide area information. A technique for transient instability detection has been proposed in [8] based on the classical equal area criterion (EAC) in the power-angle domain. This technique requires the pre- and post-disturbance information of the system as the inputs to the relay. Also a multivariate polynomial approximation has been presented in [9] for transient stability assessment. In [10], the application of decision tree (DT) theory based on R-Rdot strategy has been proposed for loss of synchronism detection. In this strategy the apparent resistance measured by the relay and its rate of change have been used as the input training features. The same technique has been used in [11] for wide-area response-based control using synchronized phasor measurements. In [12] a K-means clustering pattern recognition technique has been proposed for out-of-step detection. Support vector machine is another technique that has been proposed for transient stability assessment [13], [14]. In this paper a decision tree classifier is proposed to develop an out-of-step predictor. The DT-based scheme extracts the important features using the information gain criterion. The required input features are measured at the relay location well before the instant of out-of-step point. The rest of this paper is organized as follows. In Sections II and III, the fundamentals of out-of-step conditions and its cri- teria are described. The details and structure of the proposed de- cision tree based scheme are presented in Section IV. The simu- lation results of applying the proposed method over a 9-bus dy- namic test system and the practical 1696-bus Iran national grid 0885-8950/$31.00 © 2013 IEEE