1574 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 4, NOVEMBER 2006 Method Combining ANNs and Monte Carlo Simulation for the Selection of the Load Shedding Protection Strategies in Autonomous Power Systems Emmanuel J. Thalassinakis, Member, IEEE, Evangelos N. Dialynas, Senior Member, IEEE, and Demosthenes Agoris, Member, IEEE Abstract—This paper describes an efficient computational methodology that can be used for calculating the appropriate strategy for load shedding protection in autonomous power sys- tems. It extends an existing method that is based on the sequential Monte Carlo simulation approach for comparing alternative strategies by taking into account the amount of load to be shed and the respective risk for the system stability. The extended methodology uses artificial neural networks (ANNs) for deter- mining directly the parameters of the most appropriate load shedding protection strategy. For this purpose, the system inputs are the desirable probabilistic criteria concerning the system security or the amount of customer load interruptions. Using this methodology, the utility engineers can adopt a specific strategy that meets the respective utility criteria. The methodology was tested on a practical power system using a computer simulation for its operation, and the obtained results demonstrate its accuracy and the improved system performance. Index Terms—Frequency response, load shedding, Monte Carlo methods, neural nets, power system protection, simulation. I. INTRODUCTION U NDERFREQUENCY (u.f.) load shedding protection con- stitutes the last line of defense for a power system when the frequency declines and its stability is in danger. The system can return to a safe state by shedding an appropriate amount of load, which is determined by the settings of the u.f. load shedding relays [1]–[6]. This protection is very important for autonomous power systems that have no interconnections with neighboring systems assisting them, and they often experience severe frequency drops after the loss of one or more generating units [7], [8]. However, there is no widely used method for de- termining the settings of this protection, which are the number of u.f. levels, the respective time delays, and the amount of the load to be shed. The electric power utilities adopt different ap- proaches that are mainly based on operating experience, and they use the real system to test the settings. A computational method has been published [9] that is based on the Monte Carlo simulation approach [10], [11] and can Manuscript received July 12, 2005; revised February 16, 2006. Paper no. TPWRS-00412-2005. E. J. Thalassinakis is with the Islands Region Department, Public Power Cor- poration, Iraklion, Greece (e-mail: mthalassinakis@mekriro.gr). E. N. Dialynas is with the School of Electrical and Computer Engi- neering, National Technical University of Athens, Athens, Greece (e-mail: dialynas@power.ece.ntua.gr). D. Agoris is with the Department of Electrical and Computer Engineering, University of Patras, Patras, Greece (e-mail: dagoris@cres.gr). Digital Object Identifier 10.1109/TPWRS.2006.879293 be used for comparing and selecting the most appropriate load shedding strategies. Three main groups of probabilistic indexes are calculated (A, B, C). Group A consists of the reliability in- dexes of generating units. The system frequency response in- dexes (group B) depict the risk that the system enters after the loss of the balance between generation and load demand. The load shedding indexes (group C) represent the price that is paid in customer load interruptions to keep the desirable level of security. However, it is not only necessary to compare alternative strategies and select the most appropriate one but also to de- termine directly the strategy that satisfies predefined desirable indexes concerning the allowable risk of the system or the permitted amount of customer load interruptions. The ability of artificial neural networks (ANNs) to learn by training any complex, nonlinear input/output mapping makes them very attractive to handle problems that are difficult to solve analyti- cally [12], [13]. This paper describes a methodology and creates a model that are based on the combination of ANNs and the Monte Carlo simulation for setting the u.f. load shedding relays. The neces- sary patterns to train the ANNs were obtained from the applica- tion of the existing method [9]. This paper is organized as fol- lows: Section II formulates the model and describes two inverse problems. Section III analyzes the building blocks of the model and describes the procedure for the model solution. Section IV presents the results from the application of the methodology on a power system based on the autonomous power system of the Greek island of Crete. II. FORMULATION OF THE MODEL A. Existing Method The existing computational method [9] deals with the problem of comparing alternative load shedding strategies by analyzing the consequences of each strategy (frequency re- sponse, system risk, power interruptions), and its block diagram is shown in Fig. 1. This method applies the sequential Monte Carlo simulation approach for simulating the system operation and response at the following three levels that correspond to different time domains. • Level 1—Operation of the Dispatching Center: It concerns the available generating unit’s commitment and their eco- nomic dispatch for each consecutive hour of the year in order to satisfy the hourly load demand and the spinning reserve requirements. 0885-8950/$20.00 © 2006 IEEE