With the rapid growth in global air traffic volume, accurate airspace traffic prediction has become critical for enhancing aviation safety and promoting sustainable airspace management. However, most existing approaches have demonstrated success primarily in short-term forecasting, whereas long-term traffic prediction remains challenging. This difficulty arises because, as the temporal granularity grows, the predictive model’s ability to learn traffic dynamics declines, and positional information is more easily lost over long sequences, resulting in diminished forecasting accuracy. To address these challenges, this paper proposes a novel intrinsic dynamics capture architecture, termed IDCformer, which capitalizes on the intrinsic characteristics of airspace flow to achieve long-term sequence prediction. IDCformer comprises three core modules: a Trend and Seasonal Extraction module (TSE), enhanced feature representation and position-aware Patch Time Series Transformer (PatchTST), and a Local Self-Attention module (LAT). Specifically, the TSE module preprocesses the input data to stabilize the data and extract long-term dynamics; second, the position-aware PatchTST alleviates the issue of temporal order loss in long sequences by integrating convolutional positional signals; finally, the LAT provides hierarchical refined processing to capture local fluctuations, thereby improving the accuracy of long-term forecasting. Experimental results based on real-world air traffic data indicate that our method surpasses other state-of-the-art models in predictive performance. Furthermore, this paper investigates the capacity of IDCformer to incorporate external information; the findings demonstrate that when external data are introduced as additional input features, IDCformer’s long-term prediction performance is further enhanced, illustrating its potential for effectively leveraging multisource information.
Liu et al. (Mon,) studied this question.