Multi-resolution temporal encoding enhanced air traffic flow prediction with frequency-aware decomposition

Air traffic flow prediction (ATFP) is a cornerstone of modern air traffic management (ATM) systems, supporting critical operations such as controller decision-making and flight schedule optimization. Air traffic flow exhibits notable periodic variations across different time scales due to flight planning and scheduling regularity, alongside unforeseen fluctuations due to operational complexity and weather variability. However, most existing ATFP methods fail to consider the multi-scale periodicity and the intricate short-term fluctuations, thereby limiting their prediction accuracy and reliability in real-world scenarios. Toward this gap, in this work, a novel multi-resolution temporal encoding enhanced ATFP framework is proposed to improve the performance of the ATFP task. A multi-resolution temporal encoding is proposed to capture latent temporal correlations across multiple periodic scales by explicitly incorporating the prior periodic knowledge into the ATFP modeling process. A frequency-aware decomposition mechanism is designed to disentangle the flow series into time-variant series and time-invariant series to better capture short-term fluctuations and long-term progression. Furthermore, an attention-based Kolmogorov-Arnold Network is designed as the predictor to fully exploit high-level contextual information and generate high-precision predictions. Experimental results on a real-world air traffic flow dataset demonstrate the superiority of the proposed framework, achieving over 20% relative improvements across all prediction horizons and evaluation metrics compared to the best baseline models. In addition, the effectiveness of all technical modules is confirmed through extensive ablation and generalization studies. Most importantly, the proposed framework shows considerable interpretability that is validated by in-depth visualization analyses, providing a reliable tool for modern ATM systems.