Aimed at the problems that traditional trajectory prediction methods based on mathematical or statistical models have a certain of inherent limitations and are difficult to meet increasingly the demands of efficiency, accuracy, and real-time trajectory prediction in the modern aviation field, a novel real-time trajectory prediction method is proposed based on a CNN-LSTM model with an attention mechanism. The proposed model is that multidimensional features are extracted from trajectory data by one-dimensional convolution, reducing the number of input features. Taking the resulting multidimensional time-series data as an input of LSTM, the contextual information can be extracted by LSTM. Moreover, an attention mechanism is employed to assign weights to output from different time-series nodes within the LSTM, focusing on key trajectory information. The experimental validation shows that the proposed model in comparison with the LSTM model and the CNN-LSTM model, produces trajectory predictions to be even more close to match real trajectories. Specifically, the model in this paper achieves a 29.7% reduction in average prediction error compared to the LSTM model and a 25.4% reduction compared to the CNN-LSTM model. In summary, the proposed method significantly enhances the accuracy of trajectory prediction.