Since wind power point forecast errors are unavoidable, probabilistic forecasts can fully describe the uncertainty of wind power, and then provide further guidance for the dispatching department’s decision-making. The current wind power probabilistic prediction methods are still incomplete in mining its physical change process. Therefore, this paper constructs a new short-term wind power probabilistic prediction framework that considers the spatiotemporal dependence of errors by mining the spatiotemporal characteristics of historical wind power data and numerical weather prediction (NWP）. Firstly, the point prediction results are obtained through the gated recurrent unit (GRU）; then, a multi-position NWP is introduced, and a multi-level error scene division method considering the characteristics of space-time dependence is proposed; finally, the Bootstrap sampling method is used to reconstruct the error to form a new adaptive modeling a sample set of short-term wind power probabilistic prediction at different confidence levels is carried out. The experimental results show that the effect of the overall framework has been verified in a wind farm in Northeast China under the probabilistic prediction considering the spatial and temporal dependencies. Compared with the same confidence level, the prediction accuracy is effectively and significantly improved, and the evaluation index PICP is improved by 0.53% and 0.44%, 0.32%, PINAW shrinks by 2.40%, 2.14%, 0.06%, which proves the feasibility and effectiveness of the proposed method.