Optimization of the local model network structure for predictive control

Abstract

In this work the structure of the network consisting of local models is optimized via iterative algorithm. The initial position of the local models in the operating space is obtained using the clustering algorithm. Clustering of dynamic data is used to unable multiple local data regions to be identified as a function of similarity between the dynamic data within the local data regions. The number of clusters is further reduced by merging clusters together based the modelling performance and similarity criterion. The reduced model is then used in the GPC framework as a linear approximation of the process in the current operating point. Simple structure of the local model network enables its linearization along the future trajectory. The approach is illustrated by a simulation study of a continuous stirred tank reactor.