The very fast growth of computing power makes it feasible to generate very large statistical data bases for security analysis of large scale power system problems, in particular by exploiting trivial parallelism to carry out the simulations.
On the other hand, much progress has recently been achieved in the field of computer based learning methods, be it in the context of non-parametric statistical techniques, neural networks or machine learning. It becomes thus feasible to apply these techniques to real large scale power system problems. In this attempt a tool box approach, combining various complementary learning approaches should be developed.
In this paper, we have considered the problem of voltage security, both emergency state detection and preventive security assessment, and we have provided a comparative study of a set of ``orthogonal'' supervised learning algorithms.
Further, to appraise the validity of our conclusions, we have included the results obtained with one of our data bases by an independent research team. Our own results appear to be consistent with theirs. Note also that similar investigations have been carried out in the context of other security problems, in particular transient stability [24][23]. In these applications the main qualitative conclusions concerning the complementary features of the different methods remain valid.
We may thus conclude that decision trees provide a mature, flexible and effective tool for the analysis of large data bases of security information. The regression methods, such as multi-layer perceptrons or projection pursuit, show high promise to further improve accuracy, in particular by exploiting security margins or indices. Finally, the nearest neighbor method, while full of potential, needs further development to allow its systematic use. Of course, other methods out of the scope of the present investigation, in particular non-supervised clustering techniques, may provide interesting complementary tools, as well.