Trajectory Prediction via Regime Partitioning and Adaptive Dual-Stream Fusion Network
Trajectory prediction is fundamental to Trajectory-Based Operations (TBO). However, prediction accuracy in Terminal Manoeuvring Area (TMA) is limited by the heterogeneous dynamics of trajectories and the lack of complexity metrics aligned with deep-learning model performance. This paper proposes a complexity-aware trajectory prediction framework comprising two integrated components. At its core is the Adaptive Dual-Stream Fusion Network (ADSF-Net), which combines a Temporal Convolutional Network branch for local kinematic features and a Bidirectional LSTM branch for global temporal intent, integrated through a Dynamic Context Gating mechanism that adaptively arbitrates between the two streams. An altitude-consistent regime-partitioning strategy then decomposes the airspace into homogeneous subspaces, coupled with a Mixture-of-Experts routing strategy that selects an appropriate expert per regime; this design reduces test-set altitude MAE by up to 41.8% relative to a single global model under a flight-level oracle, with a deployment-realistic causal estimator preserving a 28.6% reduction using only past-observed data at each prediction step. A transparent decomposition shows that the dominant share of this gain (39.2 of the 41.8 percentage points) stems from regime partitioning itself. To characterise the resulting regimes and to drive the residual gain from architecture switching, we introduce the Trajectory Statistical-Moment Predictability Index (TSMPI) as a lightweight, model-agnostic diagnostic and routing signal rather than a predictive component; it exhibits a moderate, outlier-robust correlation with per-flight prediction error (Pearson $r \approx 0.30$ , $p \lt 10^{-4}$ ) and provides a principled basis for routing decisions. We additionally derive a moment-based information-theoretic lower bound on the achievable altitude-prediction error that grounds this empirical relationship, identifying dispersion as the dominant complexity driver and clarifying why a purely moment-based signal correlates only moderately with realised error. A sensitivity analysis confirms that TSMPI is robust to weight choice under both local perturbation and uniform random sampling from the Dirichlet simplex.