Machine-Learning Approaches to Power-System Security Assessment

This paper describes ongoing research and development of machine learning and other complementary automatic learning techniques in a framework adapted to the specific needs of power system security assessment. In the proposed approach, random sampling techniques are considered to screen all relevant power system operating situations, while existing numerical simulation tools are exploited to derive detailed security information. The heart of the framework is provided by machine learning methods used to extract and synthesize security knowledge reformulated in a suitable way for decision making. This consists of transforming the data base of case by case numerical simulations into a power system security knowledge base. The main expected fallouts with respect to existing security assessment methods are computational efficiency, better physical insight into non-linear problems, and management of uncertainties. The paper discusses also the complementary roles of various automatic learning methods in this framework, such as decision tree induction, multilayer perceptrons and nearest neighbor classifiers. Illustrations are taken from two different real large scale power system security problems : transient stability assessment of the Hydro-Québec system and voltage security assessment of the system of Electricité de France.