Imputation of flight ground handling node data based on graph attention networks

Efficient flight ground handling is crucial for airline operations, yet sensor failures, human errors, and data interruptions often lead to missing timestamps. Traditional imputation methods and basic machine learning models fail to capture nonlinear dependencies and heterogeneous interactions. This paper proposes a graph attention network-based flight timestamp imputation model (GCFTI-GAT). The model represents each ground handling node as a fully connected graph, integrating flight context attributes with learnable node embeddings and employing multi-head attention for dynamic weight adjustment. A masked loss function is introduced to optimize only missing entries. Experiments on a dataset of 30,000+ flights with 50% missing data show that GCFTI-GAT achieves a 6.37-minute average absolute error, outperforming the best baseline by over 50%. This approach has significant implications for ground handling and transportation scheduling.