Data-Driven LightGBM Attitude Controller for Low Earth Orbit Satellites With Explicit Gravity Gradient, Geomagnetic, and Aerodynamic Disturbance Modeling

Abstract

This paper presents a data-driven spacecraft attitude controller based on LightGBM gradient-boosted trees, specialized for Low Earth Orbit (LEO) satellites under realistic environmental disturbances. We augment the sixth-order Cayley–Rodrigues attitude model with an analytical gravity gradient torque, prove that exact feedback linearizability is preserved (Theorem 1), and replace the white-noise disturbance assumption of a prior formulation with a tilted-dipole geomagnetic model and a Jacchia–Roberts-style atmospheric density model. The regressor is trained on 300,000 state–torque pairs sampled i.i.d. over the operating envelope and labelled by the feedback-linearization expert; a comparison across six function approximators shows the framework to be approximator-agnostic, with LightGBM retained for its structural leaf-sum stability certificate (Lemma 2) and balanced accuracy–latency profile. A total stability theorem (Theorem 3) with an input-to-state corollary extends the closed-loop guarantee to the bounded perturbation structure of LEO. Closed-loop simulations at 500 km and 800 km give 5.04% overshoot and 31.4 s settling time, within 0.6 percentage points of the exact feedback linearization baseline, and an analytic near-origin handover reduces the steady-state pointing error below 0.001°, independent of the training seed. A robustness suite — actuator saturation, sensor noise, friction, impulsive disturbance, elliptical and inclined orbits, and initial attitudes up to 170° — shows graceful degradation throughout, and a sensitivity sweep confirms that the transients are invariant to space-weather parameters. A momentum-accumulation study shows where the structured disturbance model matters, and a timing benchmark grounds the deployment case in statically boundable worst-case execution time rather than raw speed.