A new method for power system reliability evaluation called self-adapting stratified and importance sampling（SASIS） is presented. With the SASIS， the system state space is partitioned into one contingency-free state subspace and various contingency order state subspaces. As contingency-free state subspace sampling is completely avoided， the SASIS converges fast in the system with high reliability. The number of sampling is optimally allocated among the contingency order state subspaces and the probability density function is steadily rectified. This method will markedly increase the calculating efficiency while eradicating the problem of low efficiency with the Monte Carlo method in high efficiency systems as reported in the past. Compared with other Monte Carlo methods， the results of the IEEE-RTS test system show that the method proposed is rational and highly effective and free from degradation.