Economic Model Predictive Control Using Data-Based Empirical Models
- Author(s): ALANQAR, ANAS W. I.
- Advisor(s): Christofides, Panagiotis D
- et al.
The increasingly competitive and continuously changing world economy has made it necessary to exploit the economic potential of chemical processes which has led engineers to economically optimize process operation to provide long-term economic growth. Approaches for increasing the profitability of industrial processes include directly incorporating process economic considerations into the system’s operation and control policy. A fairly recent control strategy, termed economic model predictive control (EMPC), is capable of coordinating dynamic economic plant optimization with a feedback control policy to allow real-time energy management. The key underlying assumption to design and apply an EMPC is that a rocess/system dynamic model is available to predict the future process state evolution. Constructing models of dynamical systems is done either through first-principles and/or from process input/output data. First-principle models attempt to account for the essential mechanisms behind the observed physico-chemical phenomena. However, arriving at a first-principles model may be a challenging task for complex and/or poorly understood processes in which system identification serves as a suitable alternative. Motivated by this, the first part of my doctoral research has focused on introducing novel economic model predictive control
schemes that are designed utilizing models obtained from advanced system identification methods. Various system identification schemes were investigated in the EMPC designs including linear modeling, multiple models, and on-line model identification. On-line model identification is used to obtain more accurate models when the linear empirical models are not capable of capturing the nonlinear dynamics as a result of significant plant disturbances and variations, actuator faults, or when it is desired to change the region of operation. An error-triggered on-line model identification approach is introduced where a moving horizon error detector is used to quantify prediction error and trigger model re-identification when necessary. The proposed EMPC schemes presented great economic benefit, precise predictions, and significant computational time reduction. These benefits indicate the effectiveness of the proposed EMPC schemes in practical industrial applications. The second part of the dissertation focuses on EMPC that utilizes well-conditioned polynomial nonlinear state-space (PNLSS) models for processes with nonlinear dynamics. A nonlinear system identification technique is introduced for a broad class of nonlinear processes which leads to the construction of polynomial nonlinear state-space dynamic models which are well-conditioned with respect to explicit numerical integration methods. This development allows using time steps that are significantly larger than the ones required by nonlinear state-space models identified via existing techniques. Finally, the dissertation concludes by investigating the use of EMPC in tracking a production schedule. Specifically, given that only a small subset of the total process state vector is typically required to track certain production schedules, a novel EMPC is introduced scheme that forces specific process states to meet the production schedule and varies the rest of the process states in a way that optimizes process economic performance