In actual air combat, the target combat intention is realized by a series of tactical actions, and the target state is present to the characteristics of time sequence and dynamic change. The traditional operational intention recognition method only relies on a single moment of reasoning, which is not scientific and effective, and fails to predict the enemy’s intention in advance. Therefore, the bi-directional propagation mechanism and an attention mechanism are introduced on the basis of the gated recur-rent unit (GRU), and a method for predicting the combat intention of aerial targets is proposed. This method is to construct the air combat intention feature set through a layered method, encode to generate numerical time series features, and encapsulate domain expert knowledge and experience into labels. The BiGRU network is used for in-depth learning of air combat features, and the attention mechanism is used to adaptively assign feature weights to improve the accuracy of air target combat intention recognition. In order to realize the advance prediction of the target intention, the air combat feature prediction module is introduced before the intention recognition, and the mapping relationship between the predicted feature and the combat intention type is established. The simulation experiments show that the proposed model can predict the combat intention of the enemy’s air target by one sampling point in advance based on the accuracy of 89.7% intention recognition, and has obviously significance in improving the real-time performance of intention recognition.