关键词: 航班地面保障, 并行级联, 卷积长短时记忆网络, 轻量梯度提升机, 动态预测

Abstract:

To address the issue of insufficient accuracy in multi-node time prediction for flight ground support processes, a hybrid prediction model based on residual cascade and parallel-cascade architecture is proposed. The model integrates a light gradient boosting machine (LightGBM) and a convolutional long short-term memory network (CNN-LSTM), and achieves dynamic prediction of event-chain nodes through dual-start parallel prediction, cascade feature transfer mechanisms, hybrid model parameter configuration and Adam optimization training. Experiments on real flight datasets show that the hybrid model reduces the mean absolute error by 43.5% compared to the pure CNN-LSTM model across seven events, performs comparably to the pure LightGBM model, yet outperforms it on multiple key nodes; the prediction accuracy for cabin door opening reaches an industry-leading level, and the accuracy for cabin door closing attains 97.4%; the prediction accuracy for actual departure time at the end of the event chain reaches 95.3%, representing a 9.9% improvement over the pure CNN-LSTM model, thereby verifying the effective suppression of error accumulation by residual connections.

Key words: flight ground support, parallel-cascade, convolutional long short-term memory network, light gradient boosting machine, dynamic prediction