Estimating the great error in wind power prediction contributes to optimizing scheduling of power system which contains wind power and improving the ability of the power grid to accommodate large-scale wind power plant. According to the analysis of the error distribution of historical wind power prediction, an approach to estimating the great error based on clustering wind power fluctuation process is proposed. Firstly, wind power prediction data is divided into diverse fluctuation processes by swinging door algorithm, and on this basis, cluster prediction errors of the same distribution by analyzing the correlation between the fluctuation and the amplitude of wind power and the distribution of prediction errors. Then this paper fits probability density distribution of the prediction errors and estimates the great error adopting slide bandwidth kernel density estimation method. Finally, the wind power data of BPA in the United States is taken as example, the effectiveness of this method is validated by comprehensively analyzing different methods. This work is supported by National Natural Science Foundation of China (No. 51207115).