A wind power forecasting method generally provides estimation of future wind power as a single point forecast， while most of the decision-making processes in the electric power systems management require more information than a single value. A new wind power forecasting method is proposed on the basis of discrete time Markov chain models. Aiming at the randomness of power data， a 4-state space is divided on the equal length， and a one-order and two-step hybrid model is built in each state space. The coefficient weights of the hybrid model are obtained by using accelerating genetic algorithm. Since the model analyzes power data directly， it efficiently avoids amplifying errors in converting wind speed forecasts into power forecasts. The hybrid models of four types and the new prediction error formula are presented. Analysis and numerical examples show that the prediction accuracy of hybrid models（ N =102） is better than that of persistence method（PM） model， and the corresponding point prediction and probability distribution estimation are also presented.