Aiming at the problems of large errors and insufficient adaptation to meteorological changes in existing ballistic drop prediction methods, a ballistic dataset containing meteorological conditions is established and a CNN-BiLSTM-BiGRU ballistic drop prediction method based on a self-attention layer is proposed in this paper. The self-attention layer and residual connection are introduced into the constructed combined model to strengthen the model's ability to dynamically focus on the information at different moments when processing the input sequences, and to alleviate problems such as gradient explosion in the network. It also uses the input representation of multi-dimensional time series data to reduce the ballistic drop prediction error by combining multiple information such as historical ballistic trajectory data and target characteristics. The simulation results show that the prediction effect of the CNN-BiLSTM-BiGRU network model based on the self-attention layer is better than the other models, and the maximum error of the range prediction accounts for 0.156% of the true value, and the maximum error of the lateral deviation prediction accounts for 5.904% as of the true value. The method provides an important reference for the field of ballistic drop prediction.