Accurately predicting air traffic flow in terminal areas is critical for balancing demand and capacity, particularly under challenging weather conditions. However, the complex interactions between weather patterns and air traffic make reliable predictions difficult. To address this issue, we propose a novel approach called the Multimodal Spatial-Temporal network for Weather-Aware prediction (MST-WA), designed to enhance air traffic flow prediction (ATFP) in terminal areas. Our method begins by constructing a spatial–temporal graph that captures the topology of the terminal area, including airports, routes, and fixes as nodes. A weather-aware module is then introduced, leveraging a Residual Network (ResNet) and attention mechanism to model the deep spatial–temporal correlations in the Weather Avoidance Field (WAF). The proposed model architecture integrates five key branches: arrival flow, departure flow, graph network topology, weather conditions, and flow constraint control, with predictions generated via an attention-based Long Short-Term Memory (LSTM) network. Experimental results using real-world data from Guangzhou Baiyun Airport, China, show that MST-WA outperforms baseline models in ATFP. Furthermore, a case study in convective weather scenarios demonstrates the model’s adaptability and effectiveness. We believe that the proposed model can serve as a valuable tool for air traffic controllers, enhancing decision-making and improving overall air traffic management.
Read full abstract