Predictive Dynamic Model of a Carbon Capture System: Pilot Scale Validation at National Carbon Capture Center

Predictive dynamic models are essential for effectively scaling up CO₂ capture systems, predicting their transient response, and developing efficient control strategies for optimal performance in the face of real world disturbances such as changes in flue gas flowrate, composition, and temperature. Because typical amine-based systems are highly non-ideal and can exhibit large nonlinearities, validation of dynamic models using experimental data is crucial to ensure their predictivity. The U.S. Department of Energy’s Carbon Capture Simulation Initiative (CCSI) process modeling team has been working on developing such a predictive dynamic model. The team recently collaborated with the National Carbon Capture Center (NCCC) in Wilsonville, AL to obtain dynamic data from an MEA test campaign for validating its dynamic model. The resulting model demonstrates how a reliable predictive capability can be developed to help reduce risk during subsequent scale up.

The dynamic model is based on a highly predictive steady state model developed by the CCSI process modeling team and has been implemented in both Aspen Plus^® and gPROMS^®. The physical property, mass transfer, and hydraulic models have been validated using comprehensive lab-scale, bench-scale, pilot-scale, and wetted wall column data. A number of user models have been developed to complement the inaccuracies in the library models available in these commercial software. While most of the test runs reported in the literature are steady-state and focus on a narrow operating range, the operating conditions in the test runs conducted at NCCC were varied widely to enable the full range of model performance to be validated. The dynamic runs were conducted by introducing carefully-designed step changes in the solvent, flue gas, and reboiler steam flowrates and recording the transients of all key variables. In addition, the dynamic runs maintained persistence of excitation of the process in order to provide information across the entire spectrum of data including both high and low frequency information. The measured data include transient response of all the sensors in the pilot plant including the gas composition sensors. Due to lags in the liquid solvent measurement apparatus, lean and rich liquid solvent samples were manually collected every 5 minutes and analyzed in the NCCC laboratory for composition and loading. Since a number of sensors provided noisy and erroneous data and a few key measurements were not available, a dynamic data reconciliation approach was developed for gross error detection and identification concurrent with dynamic model validation. Because the process model was implemented in commercial software where the equality and inequality constraints stemming from the process models are not analytically available, a novel dynamic optimization technique was developed using a sequential optimization approach. Due to the large size of the observed data space and because the high-dimensional model includes the entire CO₂ capture system and balance of plant, a number of approaches were developed to keep the computational expense tractable. The results show interesting transient response that would be crucial to consider for developing efficient control strategies.