ANN sensitivity analysis for identification of relevant features in security assessment

This paper deals with a problem of identification of the best subset of variables that should be used for dynamic security assessment of a power system, when this task is provided by artificial neural networks (ANN). The approach described here exploits ANN output sensitivities relatively to the inputs and correlation degrees, to identify the most relevant system variables to be used for an effective security assessment task. The ANNs are initially trained with all low-correlated candidate features, which enables the sensitivity analyses for the initial set of system features. Derivatives of the ANN output relative to each input are obtained by exploiting the chain rule, similar to the one used for weights adaptation on backpropagation algorithm. A description of the application of this approach in a real system is present in the paper. Results obtained in the dynamic security assessment problem of the network of the Island of Crete were quite successful.