MSCNN-BiLSTM-AM: an integrated deep framework for trajectory forecasting

This paper proposes a novel trajectory prediction method, MSCNN-BiLSTM-AM, designed to address the key challenges of traditional methods in accurately capturing trajectory dynamic features and resisting data interference in complex flight environments. The method incorporates an Improved Squeeze-and-Excitation Network (ISENet), combining Local Maximum Pooling (LMP) and Local Average Pooling (LAP) in a dual-channel design, enabling precise capture of key trajectory features. The innovative multi-scale spatiotemporal coupling architecture deeply integrates CNN’s spatial feature extraction capabilities with BiLSTM’s temporal modeling advantages, establishing an explicit connection to the aircraft’s dynamic equations. Experiments on the Automatic Dependent Surveillance-Broadcast (ADS-B) real data and Northwestern Polytechnical University (NPU) simulation datasets show that the model improves various metrics by 28.6%-41.2% (Height RMSE = 49.239 m, Speed RMSE = 5.752 m s⁻¹) compared to the optimal benchmark method in conventional scenarios. Under extreme conditions with 7% outliers and 8% data missing, the model still demonstrates excellent performance with MAE = 0.1557 and MSE = 0.0065, achieving a robustness improvement of 59.6%. The results indicate that the proposed method outperforms the benchmark methods and provides a reliable technical solution for intelligent trajectory prediction of hypersonic aircraft.