Temporal Residual Learning for Real-Time Air Traffic Complexity Forecasting

As the Air Traffic Management (ATM) landscape evolves with increasing traffic demand and emerging users, accurate short-term forecasting of air traffic complexity becomes more important in the development of next-generation airspace system. This paper proposes a new Temporal Residual Recurrent Neural Network (TR-RNN) architecture designed to capture the non-linear temporal dynamics of sector complexity. Using a dataset from the Madrid Area Control Center (ACC) comprising over 2 million rows of state vectors, we benchmark the TR-RNN against state-of-the-art tree-based ensemble methods (Random Forest, XGBoost, LightGBM). Our results reveal a performance crossover: while tree-based models benefit from static sector constraints at longer horizons (15–60 minutes), the proposed TR-RNN significantly outperforms them in the short term (5–10 minutes) by effectively modeling the momentum of traffic complexity changes with only univariate information. With sub-millisecond inference latency, the TR-RNN demonstrates the potential for integration into real-time digital assistants and automated conflict detection tools.