Monte Carlo simulation (MCS) is an efficient tool in the reliability analysis of large composite generation and transmission systems. In practice, however, planners take greater interest in evaluating only their operating regions, which generally encompasses a reduced set of consumer buses. In this case, supply failure (i.e., load shedding) is often a rare event, which substantially increases the computational cost of the MCS. To deal with this situation, the use of variance reduction techniques was proposed in numerous works, which present good results, but still find some difficulty in modeling the physical restrictions of the transmission network. Computational intelligence techniques have also been employed in order to increase the efficiency of composite reliability algorithms. Machine learning approach encompasses most of these techniques. Generally, its task is to directly reduce the execution of state adequacy analysis through classifiers. Such classifiers need to be trained from samples with good representation of the failure, which restricts its application, given that the failure of interest of system agents is, in general, a rare event. In this work, a new application of the importance sampling technique, via crossentropy method, that adequately models transmission network restrictions, is proposed. Additionally, seeking to further improve the tool’s performance, it is proposed to couple a state classifier in order to differentiate states between failure or strained success (i.e., success states in which system/area/bus is close to failure). The simulation technique employed is the quasi-sequential MCS with the purpose of properly modeling renewable sources. The method is evaluated by the means of a modified version of the IEEE Reliability Test System - 1996, which presents composite failure characteristics as those of real systems.