An improved Latin hypercube sampling method for power system reliability evaluation is proposed by combining Latin hypercube sampling method with the importance sampling method.In this method，the probabilistic distribution of current sample space of the system is changed by the importance sampling method，and then samples follow the new probability distribution are generated by Latin hypercube sampling.Thus the sampling variance is reduced and massive repeated samples in normal system state are avoided.The proposed method is applied to the reliability evaluation of the power generation section in an IEEE-RTS and modified IEEE-RTS test system.Comparison with the conventional non-sequential Monte Carlo simulation algorithm and the Latin hypercube sampling method shows that the sampling variance is reduced and the sampling efficiency improved with the proposed method with definite calculating accuracy guaranteed，which is suitable for the reliability evaluation of high reliability systems.