Extraction of representative attributes as neural-network inputs for power-system transient stability assessment

This paper proposes a rough set (RS) based approach for reduction of the input-dimension of the artificial neural-network (ANN), which is used for power-system transient stability assessment (TSA). In order to apply the symbolic RS methodology, three discretization algorithms, i.e., equal-width, equal-frequency and maximum-entropy, are tested and evaluated to transform the continuous numeric data into discrete representation. The New England 39-bus system is used for TSA test with both the original ANN classifier and the one after reducing its input-dimension. Using the proposed RS reduction techniques, 6 out of the original 11 features are extracted as the most representative attributes. Comparison results show that the 6-input ANN classifier is as effective as the original one with 11-inputs. Meanwhile, the training data set of ANN is compressed by 45.5%.