Abstract: Wind energy is an unstable energy with random fluctuation. Large-scale integration of wind power into the power grid will have a great impact on the stability of the power grid. Effective prediction of wind power range will greatly improve the economy and stability of the power grid. Aiming at the nonlinear and non-stationary characteristics of wind power data, a multi-step interval prediction method of ultra-short-term wind power based on CNN-BILSTM was proposed. Firstly, the wind power data fluctuates slightly up and down to form the initial upper and lower limits of CNN-BILSTM model. Secondly, variational modal decomposition（VMD） was used to decompose the data of upper and lower limit into several sub-components to reduce the non-stationary characteristics of wind power time series. Then, the sub-components were input into CNN-BILSTM model to obtain the prediction interval of wind power. Finally, the improved coverage width criterion was taken as the objective function optimization interval to obtain the wind power prediction interval at the given confidence level.The actual operation data of a wind farm was compared with CNN-GRU, CNN-LSTM, KELM and SVR models. The verification results show that the multi-step interval prediction method of CNN-BILSTM ultra-short-term wind power based on VMD can effectively improve the accuracy of ultra-short-term interval prediction of wind power.