Uncertainty-aware transformer-based model predictive control for rapid constrained spool-speed tracking

Abstract Rapid spool-speed tracking in turbofan engines is constrained by surge margin, turbine temperature, and rotor-speed limits, while MPC performance depends strongly on prediction-model reliability. This paper proposes an uncertainty-aware Transformer-based MPC framework for time-efficient constrained spool-speed tracking. A conditional encoder-decoder Transformer predicts multi-step engine responses from recent history, future control inputs, and flight conditions, with first-difference features enhancing transient representation. An ensemble of independently trained Transformers estimates model dispersion, which is incorporated into MPC through an uncertainty-dependent constraint margin to moderate aggressive commands when model disagreement increases. The controller is evaluated in closed loop on a high-fidelity component-level engine model against a conventional schedule/PID controller, soft-SAC, and deterministic Transformer-MPC. Across 80 scenarios with 20 repeated runs per scenario, the proposed method achieves a 99.625 % success rate and a mean hitting time of 1.786 s, outperforming the compared controllers while reducing overshoot and constraint-violation metrics relative to deterministic Transformer-MPC. Sensitivity analysis of the tightening factor κ demonstrates the trade-off between response speed and operating-margin preservation, while prediction-interval coverage indicates that ensemble spread provides a useful empirical reliability measure.