With the growing stress on today's power system, it is operated much closer to its stability limit. Under such circumstances it is highly desirable that one must be able to assess the security and stability of the electric power system when exposed to disturbances/faults. In the post-fault transient analysis of interconnected systems, the transient energy margin which is a complex function of prefault system conditions, structure of fault (type and location) and network topology at the specified fault clearing time gives a quantitative idea about the stability of the system. High adaptation capabilities of artificial neural networks make them capable of synthesizing the complex mapping that transform the input features in to a single-valued space of energy margin. Appropriate input feature selection has a direct bearing on the consistency and accuracy of mapping. This issue has been addressed in the present paper by comparing the prediction results based on approaches (Sobajic and Pao, 1989), (Sobajic and Pao 1992), (Jeyasurya, 1993) in the time domain, energy domain and its corresponding time domain calibration. Subsequent to the above comparison, the much haunting question of whether to train the network in the energy or time domain has been answered satisfactorily. It has been observed that fault clearing time is a key parameter that anticipates the success of possible calibration of energy margin results into the time domain. Test cases for prediction have been collected from many different operating conditions in power systems. Multilayer perceptron model with ADAPTIVE LEARNING ALGORITHM is used to carry out the present studies. © 1997 Taylor & Francis.