This study presents an effective simulation framework with importance splitting (ISp) method for estimating small failure probability of dynamic structure with multi-correlated stochastic excitations, and addresses its application to large wind turbine extreme response predictions. The ISp method, also referred to as subset simulation with splitting, splits important sample paths into multiple branches at various stages in simulation. It permits the estimation of a rare event through estimations of intermediate subset events. The framework presented combines the ISp method with multivariate autoregressive (MAR) modeling of stochastic excitations, which is very efficient in generating conditional samples and facilities the application of ISp method to different kinds of excitations. This framework is applied to evaluate extreme responses of a 5MW onshore wind turbine with mean recurrence interval of 50-year. The translation process approach is also addressed for estimating large extremes with a newly introduced translation function instead of using Hermite polynomial model.