This paper proposes an approach for establishing a transient stability clas
sifier and derives a continuous transient stability index, using a three-la
yer feed-forward artificial neural network (ANN), for on-line security asse
ssment in large power systems. With the derived stability index, a never cl
assification scheme creating an 'indeterminate' class, is introduced to min
imize misclassifications and to improve the reliability of the classificati
on results. Several post-fault abstract attributes about the system generat
ors' acceleration rates and kinetic energies provide the basis for the stab
ility classification. In order to derive the transient stability index, a s
emi-supervised backpropagation (BP) learning algorithm, making use of a spe
cially defined error function, is developed. The proposed approach can not
only distinguish whether a power system is stable or unstable, on the basis
of the specific post-fault attributes, but can also provide a relative sta
bility quantifier. Furthermore, as the number of the selected abstract attr
ibutes is independent of the system size, the methodology of the proposed a
pproach can realistically be applied to large power systems. The 10-unit 39
-bus New England power system is employed to demonstrate the proposed appro
ach. The numerical results show that the ANN-based classifier can assess th
e transient stability reasonably well. (C) 1998 Elsevier Science Ltd. All r
ights reserved.