Parameter selection of multi-class SVM with evolutionary optimization methods for static security evaluation in power systems

摘要

Static Security Evaluation (SSE) is one of the most important real-time studies in power systems. Static security can be assessed by different machine learning methods. In this paper, one of the well-known classifiers, namely the Support Vector Machine (SVM), is used for solving the SSE problem. Proper operation of the SVM heavily depends on the appropriate choice of parameters. The optimization problem aims at determining these parameters. This work presents a study of several heuristic optimization methods for static security. In particular, Modified Particle Swarm Optimization (MPSO), Differential Evolution (DE), Ant Colony Optimization for continuous domain (ACOr), and Harmony Search (HS) are employed for determining the optimal SVM parameters. In addition to studying the performance of various optimization techniques, this work is concerned about viewing the SSE problem as a 2-class, a 3-class, or a 4-class classification problem, where each class is associated with a particular level of security. The performance of each method is presented in terms of classification accuracy and execution speed. It is shown that most optimization methods exhibit a similar performance. However, choosing the best optimization method seems to be dependent on the number of classes, and thus, on the number of security levels required by SSE. The New England 39-bus benchmark system is used for simulation.