Data-driven stochastic model predictive control for spacecraft attitude tracking

This paper proposes a data-driven attitude control method for spacecraft subject to angular velocity constraints, where only a limited set of input-output measurements from the unknown dynamics is available. Traditional model-based approaches often struggle with unmodeled dynamics and state constraints. To address this issue, we develop a stochastic model predictive control (SMPC) framework based on data-driven system identification. First, under the Koopman operator framework, a linear lifted-state model with uncertainty is constructed from data using the extended dynamic mode decomposition (EDMD) method. To capture the residual error of this approximation, Gaussian process regression (GPR) is employed to probabilistically characterize the model mismatch, delivering state- and control-dependent estimation of the mean and covariance over the prediction horizon. These estimations are incorporated into an SMPC optimization that enforces chance constraints on angular velocity and control torques, maintaining the probability of constraint violation below a specified threshold. The numerical simulations validate the utility of the proposed data-driven SMPC algorithm, demonstrating reliable and accurate attitude tracking while handling system uncertainties and constraints.