Joshua C. Morgan1, Anderson Soares Chinen1, Christine Anderson-Cook2, K. Sham Bhat2, Benjamin P. Omell3, Ryan Hughes1, Goutham Kotamreddy1, Michael Matuszewski3,4, Charles Tong5, David C. Miller3, Justin H. Anthony6, Chiranjib Saha6 and Debangsu Bhattacharyya1, (1)Department of Chemical and Biomedical Engineering, West Virginia University, Morgantown, WV, (2)Los Alamos National Laboratory, Los Alamos, NM, (3)National Energy Technology Laboratory, Pittsburgh, PA, (4)Department of Chemical and Petroleum Engineering, University of Pittsburgh, Pittsburgh, PA, (5)Lawrence Livermore National Laboratory, Livermore, CA, (6)National Carbon Capture Center, Wilsonville, AL
Testing and evaluation of novel CO
2 capture technologies in pilot plants are necessary for reducing the risk and uncertainty of applying these technologies at the commercial scale. However, pilot plant tests require significant time and resources, especially for the comparatively large power plant applications. Therefore, pilot plant test runs should be maximally informative. The typical approach to steady-state design of experiments (DOE) is to use a space-filling approach that considers only the input design space. Even when the output space is considered, a systematic approach is often not followed. Furthermore, typical experiments fail to exploit the knowledge gained as new data are collected. Thus, a sequential DOE with the potential for recourse during the test campaign can be advantageous for accelerating learning from the test runs. To develop a DOE that considers both the output space and enables a sequential DOE requires a mathematical model. As one of the typical test run objectives is to gather knowledge that may be missing or poorly represented in an existing model, it is expected that the existing model is highly uncertain, and new data collected from the test runs are expected to contain measurement uncertainty. Therefore, the sequential DOE should be developed with due regard of the uncertainty in the model as well as the measurements. With this motivation, a sequential Bayesian DOE is developed for steady-state test runs. Since our objective is to minimize the worst prediction variance of the existing model in the design space, the DOE is developed to be G-optimal.
In addition to steady-state DOE, a systematic approach to dynamic DOE for CO2 capture pilot plants is also developed. However, dynamic tests can not only provide significantly more information than the steady-state test runs in a shorter period of time, but they can be instrumental in observing model parameters that cannot be observed through the steady-state test runs alone. Dynamic pilot plant test campaigns, if at all undertaken, typically consist of observations made during variable step changes. However, a dynamic DOE must ensure persistence of excitation so that all frequencies can be observed in the transient data. A D-optimal hybrid dynamic DOE is developed by designing a maximum length pseudo-random binary sequence coupled with a Schroeder-phased input.
The steady-state and dynamic DOEs are implemented in the Pilot Solvent Test Unit (PSTU) at the National Carbon Capture Center (NCCC) in Wilsonville, Alabama. Due to practical limitations of analyzing liquid samples and collecting experimental data, a batch design approach is developed. For tractability of the computational approach, surrogate response surface models are developed for use within the Bayesian framework. This presentation will include a critical analysis of the knowledge gathered during these test runs in comparison to the conventional DOE undertaken by the team in the recent past.
Extended Abstract: File Not Uploaded
