Abstract
With the continuous growth of civil aviation traffic, traditional manual decision-making approaches for runway configuration struggle to effectively respond to sudden weather changes and temporary flight traffic fluctuations, resulting in low utilization of airport runway resources and significantly impaired operational efficiency. To address this problem, this study proposes a multi-source data fusion framework for airport runway configuration prediction, integrating weather, flight operation and other data. The research begins by identifying three primary data sources: temporal, meteorological, and flight operation that exhibit the highest correlation with runway configuration. Subsequently, a novel Long Short-Term Memory-Transformer (LSTM-Transformer) prediction model is developed, incorporating defined encoder and decoder formats. The model utilizes historical data from the past 12 h, using one-hour intervals as time nodes, to predict optimal runway configurations for the next 6 h. Additionally, an innovative vector encoding method for runway configuration is introduced. To validate the model's effectiveness, a comprehensive simulation dataset was constructed using actual operational data from Hartsfield-Jackson Atlanta International Airport. The study systematically compares three prediction models: LSTM, Transformer, and a novel LSTM-Transformer hybrid, incorporating both linear and vector encoding methods with end-to-end and autoregressive inference approaches, resulting in eight experimental configurations. The results demonstrate that the proposed LSTM-Transformer model, combined with vector encoding and end-to-end inference, achieves optimal prediction accuracy, significantly outperforming traditional models. This approach not only accurately predicts runway configurations but also provides advanced decision support for air traffic controllers, demonstrating substantial practical significance and broad applicability for real-world airport operations.