The adaptive under frequency load shedding (UFLS) can be used to stabilize the system frequency by shedding an appropriate amount of load depending on the disturbance occurred. However, not only the total load to be shed is not minimized in this scheme, but also it may be inaccurate calculation of power deficit if the system is subjected to cascading failures during the shedding process. In this paper, a novel UFLS scheme is presented to minimize the amount of load shedding and keep the minimum frequency within the acceptable limit during severe disturbances and cascading failures. The shed load is minimized using genetic algorithm (GA) according to an objective function based on the system frequency. Multi scenarios of contingences are performed on the proposed system to collect the training patterns for an artificial neural network (ANN). The ANN can online determine the minimum amount of load shedding and keep the frequency within a safe margin. Although using the proposed scheme may cause the minimum frequency of the system to be below the minimum frequency of the adaptive scheme but the proposed method can perform minimal load shedding in various disturbance scenarios. Results are provided in the form of time domain simulations via MATLAB/SIMULINK.