Data Driven Methods For Improved Decision Making In The Chemical Process Industries PDF Download

Recent decades have prompted chemical manufacturers to consider new operating paradigms. Globalization and other market trends have reduced profit margins and emphasized the need for processes to operate in a more flexible and agile manner, e.g., to rapidly shift productions targets in response to real-time economic data, or to benefit from participation in short-term electricity markets. The twin prongs of (dynamic) process modeling and mathematical optimization have been key in meeting these challenges. One example is the widely adopted model-predictive control, an optimization-based feedback control framework which uses a dynamic model to determine an optimal sequence of control inputs. More broadly, there is a growing trend toward integrating modeling / optimization problems across the decisional hierarchy, ranging from design and long-term planning to the scheduling and real-time operation of process units. In this dissertation, I propose data-driven solutions that address some of these problems. The first part of this dissertation is concerned with the problem of model quality maintenance for model predictive controllers. I propose a statistical method for locating and estimating plant-model mismatch in these systems, formulated as an optimization problem which minimizes the discrepancy between theoretical and empirical statistics associated with the process variables. The method is capable of detecting and estimating the magnitude of plant-model mismatch in industrially relevant controllers, i.e., using state-space dynamic models and including state estimation. Furthermore, the procedure can be applied to data collected from normal process operation, without requiring costly system re-identification tests. Case studies demonstrate very good performance of the proposed method. The second part of this dissertation is focused on the problem of integrating process scheduling with (nonlinear) dynamics of the control system and the process itself. Such efforts are motivated by the increasing overlap in the time scales of the respective layers in the decisional hierarchy: as scheduling decisions are made more frequently, e.g., as modulation of throughput for participation in demand response; and/or as plant-wide dynamics become slower, e.g., with greater energy/material integration. These trends require integrated solution methods in order to obtain optimal operating policies. First, I propose a framework for explicitly representing the behavior of dynamic systems under model-predictive control within scheduling optimization problems. My approach converts this large-scale bi-level problem into a single-level "mathematical problem with complementarity constraints," in which the optimality conditions of the lower-level MPC problem are embedded directly in the upper-level scheduling problem. Reformulations of the resulting nonlinear optimization problem are proposed to improve computational performance. Two case studies demonstrate that the integrated problem achieves better performance relatively to alternative (i.e., open-loop) formulations, while remaining computationally tractable. Then, I turn my focus toward reduced-order modeling approaches that enable particularly fast solution of integrated scheduling/control problems. I leverage the structure of a class of data-driven nonlinear models, and propose parameterizations of those models that reduce problem sizes and solution times by orders of magnitude without losing any dynamic information. The modeling framework is evaluated using two case studies: scheduling of a multi-product polymerization reactor, and participation of a chlor alkali plant in short-term electricity markets. The reduced computational effort associated with the new framework is then leveraged to solve two-stage stochastic programming problems which account for uncertainty in the problem parameters