In the computer based learning framework, security is assessed by first generating large data bases of extensively simulated power system situations via available numerical security analysis tools. These data are then exploited by learning algorithms to derive the underlying relationships among input parameters (or attributes) describing power system states and security information in the form of either discrete classes or continuous indices. The resulting compressed information may provide a better understanding of the considered security problems, as would be required in planning or operational planning, and/or effective decision making tools as would be required in operation. In short, security assessment within this framework needs a data base generation software and appropriate learning algorithms.
The data base generation is not our concern in this paper. We merely mention that in practice data bases are generated in a pragmatic trial and error fashion, exploiting information provided by experts in charge of security studies to screen sufficiently well the operating conditions. The approach may be largely independent of the learning algorithms applied, but is strongly problem and system dependent. It is also worth noting that the response time of data base generation may be reduced significantly by exploiting available computing power in parallel for the repetitive simulations required. We refer the interested reader to refs. [2][1] describing data base generation approaches for real systems.
As concerns the computer based learning approaches, we believe that different methods are complementary rather than competitive. Indeed, the learning algorithms proposed in the literature have different capabilities and more or less restricted application domains. On the other hand, security assessment has several practical needs difficult to conciliate such as efficiency, interpretability and reliability. An important aspect of research thus concerns the exploration of different methods on realistic data bases.
In this paper, we consider the problem of voltage security and compare features of a sample of as ``orthogonal'' as possible learning algorithms taken from the three main classes of methods : top down induction of decision trees from machine learning, multi-layer perceptrons from artificial neural networks, and nearest neighbor rules from statistical pattern recognition.
The paper is organized as follows.
Section 2 sketches the overall computer based learning framework for security assessment and discusses the three classes of learning algorithms.
In section 3 we introduce the voltage security problem and the three test systems used (an academic 7-bus system; a 320-bus EHV model of the EDF Brittany subsystem; and a 1250-bus EHV+HV model of the latter), and summarize the results obtained from extensive simulations. In particular, we outline the results obtained with 22 different algorithms applied to the data base corresponding to the academic system within the independent ESPRIT research project StatLog [3].
Finally, section 4 discusses important practical requirements for computer based learning methods and how the various presented algorithms may contribute to fulfill them.