Parallel Genetic Algorithms Applied to Damping Controllers Tuning on a Linux Cluster of PCs CARMEN L. T. BORGES ENRIQUE C. VIVEROS GLAUCO N. TARANTO Department of Electrical Engineering Federal University of Rio de Janeiro PO Box 68516, 21945 – 970 BRAZIL carmen@dee.ufrj.br Abstract: - The coordinated tuning of power system stabilizers (PSS) consists of an optimization problem where the objective function is to maximize system damping. Since conventional optimization methods tend to obtain a local optimum instead of a global one, the application of Genetic Algorithms (GA) to this problem has been considered. However, the computational effort required by the GA approach is very high and may become prohibitive for large-scale systems. This paper presents the implementation of two categories of Parallel Genetic Algorithms (PGA) applied to the PSS tuning problem, namely: master-slave and multi-population. In the master- slave PGA, the evaluation of the many chromosomes within a sole population is performed in parallel, whereas in the multi-population PGA, many populations are evaluated concurrently on different processors. Different communication topologies, migration rates and substitution strategies have been exploited in the multi- population PGA in order to evaluate the parallel speedup. The parallel platform utilized is a Linux cluster of PCs composed of 24 microcomputers interconnected by a switched Fast-Ethernet network. The results obtained show high speedup and good parallel efficiency on actual power system models. Key-Words: - Power System Stabilizers (PSS), Parallel Genetic Algorithms (PGA), Cluster Computing, Master- Slave PGA, Multi-Population PGA. 1 Introduction An Electrical Power System (EPS) is frequently submitted to non-predicted disturbances that modify its operation condition. Since the EPS components do not have linear behavior, its final state is highly dependent on the initial conditions prior to the disturbance and the characteristics of the disturbance itself. In the case of small disturbances, like variation in load demand, linear models may approximately predict the EPS behavior. Power System Stabilizers (PSS) are controllers designed to enhance the so-called small-signal stability of EPS, by damping electromechanical oscillations. There is a myriad of methods for PSS tuning, especially for the case when one PSS at a time is tuned for one operating condition. However, one hardly finds procedures for coordinated tuning of PSS when various operating conditions are to be taken into consideration during the tuning process. The problem of coordinated tuning of PSS can be setup in the form of an optimization problem, whose objective is to maximize the minimum system damping for the various operating conditions taken into consideration. The design of multiple power system controllers requires that several specifications be accomplished in order to ensure the system operation with adequate margins for a certain number of operating conditions. The large number of controlled devices in modern power systems, associated with the increased utilization of existing equipment, calls for a more rigorous and systematic decentralized control design procedure. The efficiency of Genetic Algorithms (GA) when applied to the coordinated PSS tuning problem is demonstrated and detailed in [1]. However, GA may require high computation effort when searching for optimal solution for large-scale EPS. One way to speedup the computation requirements is to use parallel processing. This paper presents several parallel GA implementations on a cluster of PCs. The implementations exploit the impact of communication topologies and migration strategies on the performance of the parallel GA. 2 Problem Formulation Power system damping controllers are usually designed to operate on a decentralized way. Input signals from remote sites are considered not reliable enough and avoided. Considering the performance of the control system for several different operating conditions ensures robustness of the controllers. Tuning of power system damping controllers