Disturbance Estimator-Based Deep Distributed MPC for Satellite CEACS

Abstract

The rising interest in small satellites for space missions highlights their vital role in Earth observation, communication, and scientific research. The Combined Energy and Attitude Control System (CEACS) improves these satellites by reducing size and mass while increasing payload capacity using flywheels for simultaneous energy storage and attitude control. This research advances the deep-learning-based fuzzy model predictive control (D-FMPC) for the satellite attitude regulation, especially for CEACS. We propose an adaptive controller that integrates a neural network-based disturbance estimator with the D-FMPC, resulting in an adaptive deep-learning fuzzy-MPC (AD-FMPC). To design the disturbance estimator, both feedforward neural network and nonlinear autoregressive network with exogenous inputs (NARX) configurations have been investigated, with training performed using Levenberg–Marquardt and Bayesian regularization algorithms. The Bayesian regularization algorithm, with its ability to incorporate probabilistic priors and improve robustness, enables superior performance in the disturbance estimation. Comparative analysis shows that the NARX network trained with Bayesian regularization delivers the most reliable and precise results, making it the optimal choice for integration with the D-FMPC. Further numerical evaluations show that the AD-FMPC significantly improves CEACS attitude pointing accuracies compared with the D-FMPC in the presence of unknown disturbance torques, parametric uncertainties, and attitude actuator constraints. In fact, this proposed attitude control solution can still be considered as a standard satellite attitude option that is also applicable for those satellites with the conventional attitude control actuators.