关键词: 航空运输, 级联预测, 证据推理, 里程碑节点, 轻量梯度提升树, 航班地面保障

Abstract: To address the issues of seasonal concept drift in the time distribution of milestone nodes and the error propagation in multi-node cascading prediction during flight ground support processes, this paper proposes a Cascade Evidential Reasoning-Light Gradient Boosting Machine (CER-LGBM) prediction model. The model consists of two sub-modules: evidential reasoning prediction and cascading rolling update. In the evidential reasoning prediction module, concept intervals are partitioned based on backtracking error sequences, and fixed expert models are trained within each interval. During the testing phase, drift detection is dynamically performed based on the recent error levels of experts in each interval; subsequently, evidential credibility is adjusted adaptively, and the outputs of all experts are weighted and fused. The cascading rolling update module constructs process features by using the real values of newly observed nodes and the prediction errors from the previous round. It controls the update time by temporal consistency constraints, and implements the rolling update prediction for incomplete nodes. Experimental results demonstrate that, in scenarios with drifted data distributions, the accuracy of ±5-minute prediction for 66.67% of nodes reaches 70% or above in the initial prediction phase. As the support process progresses, the prediction accuracy of ±5-minute for all nodes reaches 95% or higher.

Key words: air transportation, cascade prediction, evidential reasoning, milestone node, light gradient boosting machine, flight ground support