Abstract: The accurate and reliable wind power prediction is of great significance to power system scheduling,wind farm efficiency,and the safe and stable operation of the power grid.In order to improve the accuracy of the ultra short term wind power prediction,this paper proposes a combined model based on complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN） and improved dog optimization algorithm（IDOA）to optimize bi-directional long and short term memory（BiLSTM）networks for predicting ultra short term wind power. In this method,the CEEMDAN decomposition is used to reduce the complexity and instability of the original data,and the partial autocorrelation is used to analyze all the decomposed series. The series with greater importance are selected as the input of the IDOA-BiLSTM model. Finally,an ultra short term wind power prediction is conducted through the IDOA-BiLSTM model. Using measured data from a wind farm in Gansu as the dataset, the training model and prediction analysis are conducted. The results show that the proposed ultra-short term wind power prediction model has high prediction accuracy and feasibility for practical application.