Identifier

etd-01222014-142700

Degree

Doctor of Philosophy (PhD)

Department

Chemical Engineering

Document Type

Dissertation

Abstract

Engineers predict, optimize, and monitor processes to improve safety and profitability. Models automate these tasks and determine precise solutions. This research studies and applies advanced and novel modeling techniques to automate and aid engineering decision-making. Advancements in computational ability have improved modeling software’s ability to mimic industrial problems. Simulations are increasingly used to explore new operating regimes and design new processes. In this work, we present a methodology for creating structured mathematical models, useful tips to simplify models, and a novel repair method to improve convergence by populating quality initial conditions for the simulation’s solver. A crude oil refinery application is presented including simulation, simplification tips, and the repair strategy implementation. A crude oil scheduling problem is also presented which can be integrated with production unit models. Recently, stochastic global optimization (SGO) has shown to have success of finding global optima to complex nonlinear processes. When performing SGO on simulations, model convergence can become an issue. The computational load can be decreased by 1) simplifying the model and 2) finding a synergy between the model solver repair strategy and optimization routine by using the initial conditions formulated as points to perturb the neighborhood being searched. Here, a simplifying technique to merging the crude oil scheduling problem and the vertically integrated online refinery production optimization is demonstrated. To optimize the refinery production a stochastic global optimization technique is employed. Process monitoring has been vastly enhanced through a data-driven modeling technique Principle Component Analysis. As opposed to first-principle models, which make assumptions about the structure of the model describing the process, data-driven techniques make no assumptions about the underlying relationships. Data-driven techniques search for a projection that displays data into a space easier to analyze. Feature extraction techniques, commonly dimensionality reduction techniques, have been explored fervidly to better capture nonlinear relationships. These techniques can extend data-driven modeling’s process-monitoring use to nonlinear processes. Here, we employ a novel nonlinear process-monitoring scheme, which utilizes Self-Organizing Maps. The novel techniques and implementation methodology are applied and implemented to a publically studied Tennessee Eastman Process and an industrial polymerization unit.

Date

2014

Document Availability at the Time of Submission

Release the entire work immediately for access worldwide.

Committee Chair

Romagnoli, Jose

Share

COinS