Abstract: The continuous increase of wind power penetration rate has brought great challenges to the calculation of power balance,grid operation scheduling and system frequency stability. In order to quantify the uncertainty of regional wind power generation, a direct multi-step probability prediction model for wind power clusters is proposed. First, in order to effectively exploit the spatialtemporal correlation between the data of every wind farm and the total power output of clusters, the maximum mutual information coefficient（MIC） method is used to select the reference station and key input characteristic variables. Then, based on the sequencesequence network architecture, combined with the probability prediction characteristics of quantile regression（QR） method and the advantages of WaveNet which can capture a wide range of time-dependent characteristics, a multi-horizon quantile（MQ）-WaveNet model is designed for probability prediction of wind power clusters, which can realize the MQ probability prediction of wind power in multiple steps for wind power clusters. Finally, the operation data of 12 adjacent large-scale wind farms in the southeast wind region of Hami, Xinjiang in China are selected for case study. The results show that the proposed model can effectively predict the power output of wind power clusters only by using the key feature variables of reference stations. The model has low complexity and is easy to be applied in engineering.