Robust Predictive Control of Networked Systems under Bounded Communication Delays

Abstract

Networked control systems (NCSs) are increasingly used in motion-control and cyber-physical applications in which sensing, computation, and actuation are connected through digital communication channels. Although networking provides flexibility and enables remote computation, communication delays create a mismatch between the plant's physical state and the state available to the controller. When combined with bounded disturbances, this mismatch can cause a conventional model predictive controller (MPC) to optimize a trajectory that is feasible for its delayed model but infeasible for the actual plant. This paper develops a low-complexity robust MPC formulation for linear networked systems subject to bounded communication delays and additive disturbances. The proposed approach does not augment the plant with a sequence of delay-dependent states. Instead, the delayed measurement is propagated through the original plant model to reconstruct the current state, while the resulting reconstruction error is analytically bounded using the maximum communication delay and disturbance magnitude. The uncertainty is subsequently propagated across the prediction horizon and introduced into the MPC problem through stage-dependent constraint tightening. Consequently, the online controller remains a standard quadratic program with the original plant dimension, while communication uncertainty is incorporated through an offline-computed uncertainty sequence. A constrained position-velocity model is used to evaluate the method under a maximum communication delay of five sampling periods and bounded disturbances. The robust controller reduces position RMSE from 0.9723 to 0.3174 and accumulated absolute error from 13.2166 to 3.6708 while reducing the accumulated position constraint violation from 8.3093 to zero. The results demonstrate that explicitly separating nominal prediction from delay-induced uncertainty can provide a practical compromise between robustness, constraint protection and computational simplicity.