Static Security Assessment of Electrical Power Systems Using Neural Classification Techniques
by
Samuele Grillo
University of Genova - Department of Electrical Engineering
Coauthors: Angelo Alessandri (University of Genova) Stefano Massucco (University of Genova) Federico Silvestro (University of Genova)

The deregulation of the electric energy market has recently brought to a number of issues regarding the security of large electrical systems. The occurrence of contingencies may cause dramatic interruptions of the power supply and so considerable economic damages. Such difficulties motivate the research efforts that aim to identify whether a power system is insecure and to promptly intervene. In this paper, we shall focus on neural methods for the purpose of a reliable and rapid contingency evaluation.

The standard approaches to the security assessment of electrical power systems are usually classified as either static or dynamic. More specifically, the static security analysis (SSA) is the post-contingent steady state evaluation of the power system by neglecting the transient behavior and any other variations that may depend on the load-generation conditions. On the contrary, if one accounts for the transition from the pre-contingent state to the post-contingent one, in the literature it is usually referred to as dynamic security analysis (DSA).

An electrical power system connects energy sources to consumers. It is a large and complex system composed of generators (with related controllers), loads, network devices, protection schemes, and so on, all dynamically coupled according to the Kirchoff laws that govern voltage and power flows. The power balance can be computed by solving the steady-state algebraic equation via numerical solvers. Basing on the steady-state equilibrium point for all the possible contingency, one at a time, the SSA provides an evaluation of occurrence and severity of a fault on the operation of the power system according to a given security ranking. The SSA is preliminary to the DSA, which results from a local stability analysis of the equilibrium point. The main difficulty in performing SSA is the high computational requirements, which reduces the possibility of effectively using of SSA tools on line in Power System Control Centers and motivated the search for suitable severity and performance indexes.

In order to deal with such numerical burdens pattern recognition techniques, as well as, in this context, neural networks were considered. Neural networks are well-suited to deal with pattern recognition in an efficient way, as they can be trained off line and used on line to classify outages thanks due their generalization capabilities. To this end, Kohonen and multilayer neural networks were proposed for the purpose of SSA and DSA.

In order to perform SSA, a new approach is proposed in this paper based on the idea of constructing a contingency classifier resulting from the optimization of a given performance index proposed by authors. Such index depends on both false alarms and miss-detections in general since we want to ensure the possibility of tuning directly the classifier by weighting differently the number of false alarms and miss-detections. This is a novelty with respect to the methodologies available in SSA application. Depending on the context, one can give more importance to false alarms or vice-versa to miss-detections. Unfortunately, the design of such classifier is difficult, as the cost function is nondifferentiable. To this end, we need to rely on the use of neural networks that are trained to minimize the cost function with descent techniques that do not use derivatives. More specifically, we selected the weights using a Simultaneous Perturbation Stochastic Approximation (SPSA) technique. Comparisons with standard learning applied to pattern recognition will be presented using as benchmark a realistic power system proposed by IEEE.

Date received: March 13, 2008