Abstract: In order to accurately control the yaw angle of the wind turbine and fully utilize wind energy to increase the power generation of the unit, a deep learning based wind direction ultra-short-term prediction method based on historical data is proposed. Firstly, variational modal decomposition（VMD）is used to decompose wind direction data into multiple subsequences. Considering that the residual component after decomposition still retains a large number of signal features, adaptive noise complete set empirical mode decomposition（CEEMDAN） is further used to perform secondary decomposition for the residual component. On this basis, combined with features such as wind speed and environmental temperature, a bidirectional long short-term memory network with attention mechanism（Bi-LSTM-Attention） is used to predict wind direction in ultra-short-term. Using real SCADA data from a wind farm in Hebei, 5-minute wind direction prediction experiments are conducted, and compared with other methods, the results show that better prediction performance can be got with the proposed method.