In the study of atmospheric dynamics, atmospheric wind field data is an important parameter for analyzing the dynamic characteristics, dynamic processes and atmospheric interactions of the atmosphere. However, affected by various factors, when MST (Mesosphere-Stratosphere-Troposphere Radar) detects the upper middle layer area, it often fails to obtain complete wind field data of the middle layer. To solve this problem, this study introduces the Deep Long Short-Term Memory (LSTM) time series prediction model to complete the missing wind field data. Based on the traditional LSTM time series prediction model, by increasing the number of network layers to capture the deeper relationships between data, a prediction model based on the deep LSTM recurrent neural network is constructed, and the parameters of the model are optimized. The horizontal wind field data of the atmospheric middle layer of the Wuhan MST radar (Chongyang Station) is selected as the input for model training and testing. The experimental results show that the deep LSTM recurrent neural network exhibits excellent prediction performance in wind speed prediction, and the predicted wind speed is highly consistent with the measured data. Furthermore, the deep LSTM model has strong adaptability in data fitting, can effectively supplement the missing data, and shows good predictive ability.