TABLE OF CONTENTS In the previous section we have presented an overview of various ANN techniques for static security assessment. In the area of security analysis, supervised and unsupervised ANNs have different objectives. Unsupervised approaches divide the operating space into classes of operating points thus pre-processing the data set by reducing it to a limited number of typical cases. These cases can then be evaluated either with standard methods or with supervised learning. Supervised approaches approximate the security boundaries of the operating space thus memorising data points of a high-dimensional function and interpolating between them.
We further have presented two applications of ANN to static security assessment. A supervised MLP was trained as a security predictor. The probabilistic performance measure for this ANN looks at the likelihood of misclassifying a potential contingency. An unsupervised Kohonen network was trained for the clustering of operating states. Here the weight vectors present the probability distribution of the input vectors and the likelihood of misclassification is strongly related to the size of the clusters and thus to the number of neurons of the feature map. In the last section we have outlined how these concepts can be integrated in a utility EMS.
These neural networks are capable of dealing with stochastic variations of the scheduled operating point and with inaccurate data, whereas classical methods need 'accurate' data for example obeying Kirchhoff's laws. Especially for clustering approaches like the Kohonen map state estimation might become a redundant task. Since clustering is quite robust with respect to...
