Data assimilation and control system for adaptive model predictive control

Model-based control of complex systems is a challenging task, particularly when the system model involves many uncertain elements. To achieve model predictive control of complex systems, we require a method that sequentially reduces uncertainties in the system model using observations and estimates control inputs under the model uncertainties. In this work, we propose an extended data assimilation framework, named data assimilation and control system (DACS), to integrate data assimilation and optimal control-input estimation. The DACS framework comprises a prediction step and three filtering steps and provides adaptive model predictive control algorithms. Since the DACS framework does not require additional prediction steps, the framework can even be applied to a large system in which iterative model prediction is prohibitive due to computational burden. Through numerical experiments in controlling virtual (numerically created) fusion plasma, we demonstrate the effectiveness of DACS and reveal the characteristics of the control performance related to the choice of hyper parameters and the discrepancies between the system model and the real system.