Real-time estimation and control of large-scale nonlinear DAE systems
Model-based control incorporates fundamental process knowledge to achieve improved monitoring and control performance. However, on-line model-based control is generally limited to linear models or nonlinear models of low-dimension. Rigorous models of dynamic process are often described by differential algebraic equations (DAEs). Many rigorous DAE models require too much computational effort to be implemented in real-time control applications, where control calculations must be performed on-line (i.e. in a few seconds). The principal focus of this dissertation is to reduce the computational requirements for large-scale model-based estimation and control. This objective is accomplished with a variety of strategies that are combined in an effective way to meet real-time constraints with limited computing resources. The principal strategies are adaptive storage and retrieval off-line to enable efficient on-line control, nonlinear DAE model reduction, and development of an explicit solution to moving horizon estimation (MHE). Both MHE and receeding horizon control (RHC) are developed to meet real-time constraints. In situ adaptive tabulation (ISAT) is used to store and retrieve control solutions. In addition to the adaptation for control applications, ISAT is developed as a general nonlinear function approximator and is shown to outperform neural networks in both interpolation and extrapolation. In addition, ISAT is designed to handle nonlinear functions with discontinuities or regions that are not continuously differentiable. With DAE model reduction, storage and retrieval of control solutions with ISAT, and the explicit solution to moving horizon estimation, real-time nonlinear model predictive control (NMPC) is feasible with large-scale DAE models.