The invention discloses an importance sampling Monte Carlo power system reliability evaluation method based on a geometric optimization-minimum variance method. First, a geometric optimization method in the field of planning and a variance minimization model in importance sampling are combined together, and a variance minimization model in reliability evaluation is solved through geometrical optimization, and then, reliability evaluation is conducted using obtained importance sampling parameters. Compared with the traditional reliability evaluation method based on importance sampling, the reliability evaluation method of the invention has the characteristics of smaller convergence variance, faster convergence and higher evaluation precision. By improving the evaluation speed and evaluation precision of system reliability evaluation and especially by capturing small-probability/highly-influential rare events, the reliability nuance between different power grid planning schemes can be evaluated quickly, and a quantitative and accurate auxiliary reference basis can be provided for the selection of power grid planning schemes.