An Innovation-Based Approach to Adaptive Data-Driven Predictive Control

Recently data-driven predictive control (DDPC) methods have attracted considerable interest, primarily due to the benefit that optimal control inputs can be obtained directly from raw data without identifying an explicit model. However, time-varying characteristics of complex systems can substantially degrade the control performance of existing DDPC methods. In this work, we propose a new adaptive DDPC scheme in which innovations play a critical role. By leveraging the (approximate) white-noise and Gaussian properties of innovations, we develop a data-driven scheme that can effectively detect changes in system dynamics. This naturally motivates a statistically grounded adaptive DDPC framework in which the online adaptation is triggered by the detection of changes in system dynamics. Numerical examples demonstrate the efficacy of the proposed detection scheme and the enhanced control performance of the adaptive DDPC method in contrast to existing heuristic methods.