Wednesday, November 7, 2007 - 1:13 PM
398c

Advanced Process Engineering Co-Simulation Using Cfd-Based Reduced Order Models

Yi-dong Lang1, Lorenz T. Biegler1, Sorin Munteanu2, Jens Madsen2, and Stephen E. Zitney3. (1) Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, (2) Ansys-Fluent Inc., Lebanon, NH 03766, (3) Collaboratory for Process and Dynamic Systems Research, National Energy Technology Laboratory, P.O. Box 880, Morgantown, WV 26507-0880

The process and energy industries face the challenge of designing the next generation of plants to operate with unprecedented efficiency and near-zero emissions, while performing profitably amid fluctuations in costs for raw materials, finished products, and energy. To achieve these targets, the designers of future plants are increasingly relying upon modeling and simulation to create virtual plants that allow them to evaluate design concepts without the expense of pilot-scale and demonstration facilities. Two of the more commonly used simulation tools include process simulators for describing the entire plant as a network of simplified equipment models and computational fluid dynamic (CFD) packages for modeling an isolated equipment item in great detail by accounting for complex thermal and fluid flow phenomena. The Advanced Process Engineering Co-Simulator (APECS) sponsored by the U.S. Department of Energy's (DOE) National Energy Technology Laboratory (NETL) has been developed to combine process simulation software with CFD-based equipment simulation software so that design engineers can analyze and optimize the coupled fluid flow, heat and mass transfer, and chemical reactions that drive overall plant performance. The process/CFD software integration was accomplished using the process-industry standard CAPE-OPEN interfaces.

A potential barrier to the widespread use of process/CFD co-simulation is that the integration of high-fidelity equipment models may lead to unacceptable co-simulation turnaround times, especially for cases in which one or more CFD models are embedded in the iterative flowsheet solution process. One promising solution is the use of reduced-order models (ROMs) that approximate the CFD-based equipment simulations, while keeping the computational cost manageable. ROMs are equipment models that contain only relevant states or a reduced number of irrelevant states. The advantage is dramatically increased simulation speeds and lower memory requirements.

To obtain ROMs for equipment items with arbitrary geometries, we propose a novel state-space ROM strategy. In contrast to the popular order reduction method of proper orthogonal decomposition (POD), this work focuses on a principal component analysis (PCA) approach. First, the statistical Latin hypercube sampling (LHS) technique is used for experimental design to determine a distribution of plausible sets of operation and boundary conditions for the CFD model. Second, using off-line CFD simulations with the designed cases, we obtain a database consisting of snapshots of the flow-field variables (state variables). Third, selecting the snapshots of one state variable of interest, we formulate a snapshot matrix and implement PCA on it to derive the ranked principal components (PC) and the coordinates of the data elements in the transformed vector space (scores). Fourth, using a neural network formulation, we train a mapping from the inputs of CFD model into the scores. Finally, we assemble the ranked PC and mapping score to develop the ROM.

Two case studies are considered, both based on a 2-D FLUENT CFD model of a gas turbine combustor. One case has a single input, the other one has ten inputs. The PCA-based ROM can be executed within seconds and leads to a CPU savings of two to three orders of magnitude. Replacing the CFD model in the process flowsheet with the ROM through the standard CAPE-OPEN unit operation interface therefore allows APECS process/CFD co-simulations to be solved more efficiently and effectively.