Abstract: For the prediction of offshore wind power, the traditional wind power prediction models seldom take into account the difference in output power caused by the changes in wind directions and atmospheric conditions. In order to improve the prediction accuracy, this paper constructs a power-direction model based on wind directions and power losses while considering the atmospheric stability, and proposes an offshore wind power prediction method based on the encoder-decoder framework. This method can update the wake effect losses according to the Pd model, effectively suppress the predicted power fluctuation, and distinguish the wake effects under different atmospheric stratification stabilities. First, the prediction models like the long-short term memory (LSTM) neural network are used to verify the atmospheric stability and the effectiveness of the Pd model. Then, the encoder-decoder is used to predict the wind power of the actual offshore wind farm. The experimental results show that the encoder-decoder method, which considers the atmospheric stability and uses the Pd model, has a 2.39% lower root mean square error than that of a single encoder-decoder prediction method.