Data-driven multi-model minimum variance controller design based on support vectors

Abstract Minimum variance (MV) benchmark is a quantification that is widely applied to compare actual performance of a control loop against its optimal performance. For linear systems, MV benchmark has been well studied, and several well-practiced techniques have been proposed to evaluate the performance of the systems. However, these techniques may not provide satisfactory performance in some situations. For instance, the operational data sampled from a multi-model system does not satisfy stationary conditions that are required for conventional methods. As a result, linear MV benchmark techniques cannot be applied. One possible approach to design MV controller for a multi-model system is to segregate, and label the data using a dynamic clustering method. Then, the MV benchmark can be determined by the multi-model MV control. In this paper, a support vector regression based method is proposed in which all three steps of dynamic clustering, model identification and MV controller design are performed simultaneously through a data driven approach. In addition, a new support vector-based clustering method is proposed by which the data are clustered in residual space of the process model. The optimality and convergence of the proposed algorithm are studied. The proposed multi-model MV controller design technique is demonstrated through a simulated continuous stirred-tank reactor (CSTR) example, which is operated in varying conditions.