Toward applications oriented optimal input design with focus on model predictive control

Modern control designs are, with few exceptions, in some way model based. In particular, predictive control has rapidly become a popular control strategy, implemented in a large number of industrial plants. Model predictive control (MPC) uses a model to predict the impact of future control inputs on the controlled plant. The quality of the model can have a large impact on the achievable control performance. It is widely reported that modeling is the single most time and cost consuming part of the commissioning of an industrial MPC and therefore an important research issue. This thesis addresses the need for good modeling for MPC by introducing an optimal input design and identification method tailored to the specifics of predictive control. Parametric models are used and the influence of the individual parameters on the control performance is measured through a cost function. This leads to a set of parameters that are deemed acceptable. Optimal input design is used to ensure, with high probability, that the estimated parameters are in the set of acceptable parameters while keeping experimental cost low. It is shown that optimal input design can lead to a significant reduction of the experimental cost while still guaranteeing acceptable control performance. A toolbox for optimal input design in Matlab is also presented. Real world systems tend to be nonlinear and sometimes it is necessary to model them as such. Input design for two types of nonlinear systems with finite memory is considered. Similarities and differences compared to the linear case are pointed out and exploited. Convex formulations of the optimal input design problem are presented. It is shown by example that the resulting optimal design can differ greatly compared to designs for linear models.QC 2011101