The probability prediction of wind power is an effective method to analyze the uncertainty of future wind power. To improve the accuracy of wind power probability prediction, a ultra-short-term wind power probability prediction method based on variational mode decomposition (VMD) and improved quantile regression gated recurrent unit (QRGRU) is proposed. Firstly, VMD is used to decompose the original wind power sequence into mode functions with different characteristics. Then, a probability prediction model based on QRGRU is established for each mode function. The network structure constraint is used as the penalty term of the objective function to improve the stability of the QRGRU weights in the iterative correction process. Finally, the predictive value of each mode function is superimposed under different quantile conditions, and the probability density function of future wind power is obtained by using a non-parametric kernel density estimation method. Based on the actual measurement data of a wind farm, a specific calculation example is analyzed. The results show that the proposed method can take the coverage of the interval into account, reduce the width of the interval and perform better predicting results in different prediction steps.