In this paper we have proposed a novel approach to voltage security assessment. It exploits non-parametric regression techniques to extract simple and at the same time reliable models of the severity of a contingency, in terms of the incumbent reduction in LPM.
The soundness and computational feasibility of the idea have been illustrated on a realistic model of a large voltage stability limited power system, showing very promising results and in particular the ability to furnish interesting security information. It is however too early to draw definite conclusions about the possible uses of the above regression approach in planning and operation. In particular, we quote the following, still open questions.
Which kind of attributes may provide appropriate accuracy / flexibility compromises for regression models ?In particular, is it possible to use attributes which are independent or controllable system parameters ?
How sensitive are the regression models to the way load power margins are defined, e.g. to stability criteria, models and individual load participations ?
How to devise a search strategy exploiting the approximate models to identify dangerous combinations of outages ?
On the other hand, we recall that the flexibility of the computer based learning approach allows us to account for any required modelling sophistication, and to study the influence on security of any kind of system parameters and models. The regression trees provide interpretable information, which may receive various interesting uses. In this respect the proposed method is a natural complement to the decision tree based approach to voltage security assessment previously proposed [15].