Big Data-driven Predictive Control Using Multi-view Clustering

This work presents a big data-driven predictive control (BDPC) approach using multi-view clustering to approximate nonlinear system behaviors (represented by a set of input-output variable trajectories) with local linear sub-behaviors (represented by Hankel matrices). The nonlinear behavior space is partitioned based on two views: Euclidean distance of trajectories, and the angle of linear subspaces that trajectories belong to. Subsequently, a BDPC controller is designed to locate the online trajectory into the most relevant linear sub-behavior and determine control actions subject to optimization in every receding horizon. Finally, the BDPC approach is illustrated using an example of controlling the Hall-Héroult process.