Overcoming limitations of NNs for on-line DSA

One of the most challenging problems in on-line operation of power systems is dynamic security assessment. Several methodologies have been proposed to solve this problem. However, most of them require a high computational burden. Analytical techniques for stability analysis do not allow the operators to take preventive or corrective measures in due time. One possible solution to overcome this drawback is the application of pattern recognition techniques. Artificial neural networks have shown outstanding precision for classification and regression tasks. The major shortcomings of the pattern recognition approach, via neural networks, are the inference opacity and the curse of dimensionality. Black-box results are not acceptable when neural networks are to be used in safety critical applications. Besides, the pattern recognition approach has to deal with thousands of variables in large-scale power systems. This paper tackles both, issues. An algorithm for qualitatively justifying a neural network inference, through the extraction of production rules, is adapted to the voltage stability problem. The curse of dimensionality in transient stability analysis via pattern recognition is overcome by using support vector machines.