Advanced-step Stochastic Model Predictive Control using Random Forests

Stochastic Model Predictive Control (SMPC) can improve the system performance under probabilistic uncertainties by shaping the predicted probability distribution functions of system states. The scenario approach incorporates a large number of scenarios into an online optimization problem. It may cause substantial feedback delays, which in turns leads to system performance degradation. In this paper, an advanced-step SMPC (asSMPC) is proposed to address this issue. During the background stage (the period between control implementation and the next sampling event) of each time step, it generates many predictions of the next-step state and solves parallel SMPC problems to form a dataset. The machine learning algorithm - Random Forests (RF) is adopted to construct an approximate control law for the next time step. Simulation studies have shown that the proposed asSMPC approach can achieve control performance similar to that of scenario-based SMPC without computational delays.