Automatic induction of fuzzy decision trees and its application to power system security assessment

Abstract

In the context of power system security assessment, automatic learning methods have been proposed along many different paradigms (statistics, neural nets, symbolic artificial intelligence, fuzzy systems). In order to solve various learning problems, different automatic learning methods are complementary and the best way to exploit large real data bases is to use a data mining tool-box approach. In this paper we argue that fuzzy decision tree induction promises to be a particularly attractive tool in this framework. It is indeed able to handle large scale problems, while combining interpretability with smooth interpolation capabilities. We describe a new algorithm able to infer fuzzy decision trees in domains where most of the input variables are numerical and output information is best characterized as a fuzzy set. It comprises three complementary steps: growing (selecting relevant attributes and fuzzy thresholds); pruning (determining the appropriate tree complexity); refitting (tuning the tree parameters in a global fashion). The basic features of the method are discussed from theoretical and practical viewpoints. Simulation results obtained in the context of transient stability assessment of a real system are provided in order to support our claims.