Multi-Channel Structural–Semantic Fusion for Forecasting Air Traffic Control Incidents: Implications for Proactive Air Traffic Safety Management

Effective safety management in air traffic is essential for operational reliability and risk reduction. We propose a multi-channel fusion framework to predict intervals between consecutive air traffic incidents by combining structural, semantic, and temporal information. Inter-incident time series are transformed into complex networks via visibility graphs to learn node embeddings capturing structural recurrence. Semantic features are derived through latent Dirichlet allocation (LDA) and bidirectional encoder representations from Transformers (BERT) embeddings to reveal latent risk-related topics, and an adaptive spectral filter enhances temporal features. These are processed through three modules: a gravity-inspired visibility graph model (GVG), a semantic-aware LSTM (Sem-LSTM), and a spectral-enhanced temporal convolutional network (Spec-TCN). An attention mechanism fuses all modules to predict incident intervals. Using 1298 real-world incidents from China’s Central and Southern Region for validation, the model achieves a mean absolute error of 1.42 h and sMAPE of 17.5%. SHAP analysis indicates that structural similarity and incident topics jointly drive prediction. By integrating interval predictions with topic cues, we construct a safety management framework enabling proactive decision-making. This framework delivers a practical bridge from interval predictions to proactive air traffic control (ATC) decisions.