Hybrid C-LSTM Neural Architecture for Real-Time Adaptive Control of Fixed-Wing AAVs in Variable Turbulence Conditions

Fixed-wing unmanned aerial vehicles (UAVs) performing turn maneuvers in turbulent environments face significant control challenges due to stochastic aerodynamic disturbances. This paper proposes a hybrid turbulence-aware control framework that integrates mathematical modeling, global optimization, and machine learning for adaptive autopilot gain tuning. A semi-invariant-based deterministic reduction of the stochastic lateral-directional model is developed, reducing the dimensionality from 44 to 7 equations while preserving control-relevant statistical characteristics. To determine optimal feedback gains, a Modified Survival of the Fittest Algorithm (MSoFA) is introduced, demonstrating improved global convergence reliability and achieving 5–15% better solution quality compared to conventional evolutionary methods. An offline-generated dataset of optimal gains across varying turbulence intensities and flight altitudes is then used to construct a clustered Long Short-Term Memory (C-LSTM) architecture that combines DBSCAN-based regime identification with recurrent neural approximation. The proposed neural controller achieves 98.3% gain prediction accuracy and reduces computation time by more than three orders of magnitude relative to direct optimization. Extensive numerical simulations confirm stable turn performance over altitudes ranging from 800 to 1500 m and turbulence scales from 150 to 470 m, with terminal yaw-rate deviations below 0.01%. The novelty of the proposed framework lies in the structured integration of (i) semi-invariant-based stochastic model reduction, (ii) globally reliable evolutionary optimization for constrained gain tuning, and (iii) regime-dependent recurrent neural approximation for real-time implementation. To the best of the authors’ knowledge, such an integrated turbulence-aware tuning framework for fixed-wing UAV autopilots has not been previously reported.