The invention provides a short-term wind power prediction method based on double decomposition, includes the following steps: S1 importing the original wind power time series, S2 decomposing the original wind power into several sub-modes by using the variable mode decomposition technique, S3 using the sample entropy method to calculate the sample entropy value of the sub-mode obtained by S2, the sub-modes with larger sample entropy and the sub-modes with smaller sample entropy being obtained, S4 using wavelet packet decomposition technology to decompose the sub-modes with larger sample entropyvalue in S3, S5 using on-line robust limit learning machine to set up prediction models for the sub-modes with small sample entropy obtained by S3 and the new sub-modes obtained by S4 quadratic decomposition respectively, and the final wind power prediction results being obtained by superposition. The invention has the advantages of fast calculation speed, better generalization ability, higher prediction accuracy and better adaptability to the time-varying system in practical application.