Flight Delay Prediction via Learning Long Short-Term Relationship Between Airports

The exponential growth of the global air transport industry has led to severe congestion and flight delays. Accurate predictions are crucial for safe and efficient air traffic management. Deep spatiotemporal methods that account for spatial and temporal dependencies provide a promising solution for flight delay prediction in multiairport scenarios. However, previous approaches relying on traffic flow or geographic distance have struggled to characterize the complex relationships between airports. Recently, causality has emerged as a powerful data-driven tool for this purpose, though existing statistical causal methods focus on long-term delay trends but are insufficient for capturing dynamic short-term relationships. This article proposes a long-short-term spatiotemporal graph neural network LASCGNN (Lag-Attention with Self-corrected Causal Graph Neural Network), which combines instant lag-attention and self-corrected causal inference to accurately predict flight delays. The instant lag-attention module extracts dynamic time lags between time patches at different airports to capture short-term delay relationships. Meanwhile, Granger causality inference, coupled with a self-correction module, constructs multiscale causality graphs among airports and dynamically modifies them to extract long-term delay relationships. Extensive experiments conducted on real data from the top 74 busiest Chinese airports show that LASCGNN outperforms baseline models, with ablation studies demonstrating the effectiveness of its key components. Our results also provide insightful guidance for air traffic management, indicating that airport lag scales are mainly influenced by traffic and geographical factors, with smaller airports more vulnerable to short-term impacts from other airports than long-term ones.