Feature selection for static security assessment using neural networks

Addresses the issue of the curse of dimensionality with respect to building neural network classifiers for large-scale power systems. Rather than using all the available measurement variables as classifier inputs, the authors use statistical techniques to extract features with maximum first- and second-order discriminatory information. The selected features are then used as inputs for training and testing the layered perceptron classifier. Classification in the resulting lower-dimensional space leads to reduced complexity and enhanced accuracy. The resulting compact classifier is easier to build in terms of both hardware and software. The concepts were proved through simulations on the extended IEEE-8 bus and 30-bus systems