Abstract: ‍ ‍The trajectory prediction plays a crucial role in ensuring the safety and efficient operation of air traffic. The predicted trajectory information serves as input for decision-making tools such as trajectory optimization and conflict alerts， and the accuracy of prediction depends on the model’s ability to extract features from the trajectory sequence. The trajectory sequence data is a multidimensional time series with rich spatio-temporal characteristics， where each variable exhibits long-term and short-term temporal patterns， and there are also spatial dependencies among these variables. To fully capture these spatio-temporal features， this paper proposes an Encoder-Decoder model based on the fusion of spatio-temporal features， referred to as the Spatio-Temporal Encoder-Decoder （STED） model. In the Encoder， a dual-channel network consisting of gated recurrent unit （GRU）， convolutional neural network （CNN）， and attention mechanisms （AT） is employed to separately extract the spatio-temporal features of the trajectory. The Decoder concatenates and fuses the spatio-temporal features， and utilizes GRU to learn and recursively output the fused features， thereby achieving multi-step prediction of future trajectory information. The performance of the algorithm is validated using real trajectory data. Experimental results demonstrate that the proposed STED network model achieves high accuracy in short-term trajectory prediction within a prediction range of 10 minutes. It outperforms data-driven trajectory prediction models such as LSTM， CNN-LSTM， and AT-LSTM in terms of accuracy. Furthermore， the STED network model predicts a trajectory point with an average time consumption of 0.002 seconds， demonstrating excellent real-time performance.