The previous two approaches essentially compress detailed information about individual simulation results into general, more or less global security characterizations.
Additional information may however be provided in a case by case fashion, by matching an unseen (e.g. real-time) situation with similar situations found in the data base. This may be achieved by defining generalized distances so as to evaluate similarities among power system situations, together with appropriate fast data base search algorithms.
In the results provided below we compare the ``
nearest neighbors''
(
) method with decision trees and multi-layer perceptrons. It
consists of classifying a state into the majority class among its
nearest neighbors in the learning set. The main characteristics of
this method are simplicity and high sensitivity to the type of
distances used. In particular, to be practical, ad hoc algorithms must
be developed to choose the distances on the basis of the learning
set.
The idea of applying statistical pattern recognition to power system security assessment dates back to the late sixties with the pioneering work of DyLiacco [21]. Among the recent attempts to apply the nearest neighbor method to security assessment we quote [22].