Multi-resolution spatio-temporal modeling for accurate terminal-area flight trajectory prediction

Accurate terminal-area flight trajectory prediction is essential for modern air traffic management. However, it remains challenging because trajectory data exhibit rapid local fluctuations interwoven with longer-term trends across multiple dimensions. Existing sequence-modeling methods struggle to represent complex multiscale spatio-temporal patterns, which limits predictive accuracy. To address these limitations, we propose WISAM, a novel forecasting framework that more effectively captures spatio-temporal dependencies in trajectory data. Specifically, inspired by BiLSTM architectures, we propose the wave-like information extraction module (WIM) that reorganizes one-dimensional sequences into two-dimensional representations, and we introduce the cyclic acceleration network (CAN), which simultaneously models fast local fluctuations and long-term trends. We also present the attention-enhanced spatio-temporal pooling module (AE-STPool) that uses grid-based partitioning together with multiple attention mechanisms to aggregate spatial features from trajectories. Furthermore, we propose the novel AdGELUMish activation function to strengthen the nonlinear modeling capacity of the prediction head. Experiments on a public terminal-area trajectory dataset show that WISAM consistently outperforms six state-of-the-art baselines; compared with the strongest competing method (WITRAN), WISAM reduces MAE by approximately 12% and RMSE by over 26%. Moreover, zero-shot evaluations on trajectories from previously unseen airports demonstrate cross-airport generalization without retraining. These results suggest that WISAM can provide reliable multi-step trajectory predictions that may support terminal-area analytics and safety-related decision-support applications.