This study presents a new intelligent controlled islanding scheme based on wide area measurement systems data to avoid the wide area blackout. Three offline, online and real-time parts are applied to solve three problems including where and when to implement islanding and what to do after separation. New security-based criteria are used to determine the initial stable coherent groups. The boundaries of islands are obtained adaptively considering different operating points by using the weighted time varying graph structure of the network. To reach more stable islands, reactive power is considered by using a self-tuned online fuzzy factor in graph weights. The number of necessary islands with their locations is determined in online part by monitoring the dominant inter-area oscillations between the initial groups (IGs). Then, the network is split into islands with the objective of minimum power flow disruption. To detect the unavoidable islanding cases correctly, a new parallel adaptive neuro-fuzzy inference system (ANFIS) structure is designed. In a parallel structure, for each of two adjacent IGs a distinct ANFIS is also applied to consider variable stability margins between groups. Simulation results confirm that the blackout can be avoided in a large power grid by using the proposed method.