A new algorithm for Kohonen layer learning with application to power system stability analysis

In certain classification problems, input patterns are not distributed in a clustering manner but distributed uniformly in an input space and there exist certain critical hyperplanes called decision boundaries. Since learning vector quantization (LVQ) classifies an input vector based on the nearest neighbor, the codebook vectors away from the decision boundaries are redundant. This paper presents an alternative algorithm called boundary search algorithm (BSA) for the purpose of solving this redundancy problem. The BSA finds a fixed number of codebook vectors near decision boundaries by selecting appropriate training vectors. It is found to be more efficient compared with LVQ and its validity is demonstrated with satisfaction in the transient stability analysis of a power system.