Model predictive control of switching systems based on sparse Bayesian identification

Sudden switching of the system poses a challenge for modeling and control. Utilizing an effective method to detect and control the switching system plays an important role in the stability of the system. In this paper, a model predictive control (MPC) framework leveraging sparse Bayesian learning is developed with the idea of multiple online learning on a small amount of data. First, differing from the threshold-based approach in previous literature, fisher information theory is used to detect system switching. The sparse Bayesian learning is then utilized to identify the post-switching system equation, and the resulting model is integrated into MPC for control, thus enabling rapid restoration of the system after mutation. This method is demonstrated on three systems switching with time as well as compared with the linear data-driven model and traditional neural network. The results reveal that the MPC based on sparse Bayesian learning exhibits higher control performance with less training data, shorter training and control program runtime, thereby contributing to the stability of the system.