(622j) Constrained Nonlinear State Estimation of An Acid Gas Removal Process As Part of An Integrated Gasification Combined Cycle (IGCC) Power Plant with CO2 Capture

In future IGCC plants with CO₂ capture, the acid gas removal unit (AGR) needs to satisfy strict emission standards for CO₂ and H₂S. With the current target of capture of CO₂ and the sulfur species in the AGR, CO₂ concentration in the clean syngas is expected to be 2-5% (vol) while the H₂S concentration (in ppmv) is expected to be in the single digits. Since in a full-scale IGCC plant, the amount of syngas is large, small deviations from the target values can result in significant amounts being vented to the atmosphere.  The reliability and precision of the typical sensors measuring gas compositions are usually low. In addition, the typical sampling periods for these sensors are in the orders of several minutes. A large noise-to-signal ratio further complicates the situation. In addition, a number of states in this system is very important for effective control of the plant. To address this problem, a constrained nonlinear state estimator is developed in this work.

The AGR plant, modeled in this work, is a two-stage unit for selective removal of H₂S and CO₂ using a physical solvent SELEXOL^®. In the first stage, H₂S is captured in an absorber and then recovered in a stripper before sending it to the Claus unit for sulfur recovery. An intermediate stripper column, known as the H₂S concentrator, ensures that the concentration of H₂S in the Claus feed satisfies certain limits. In the second stage, CO₂ is captured in an absorber and then recovered by multi-stage flash blocks before being sent for compression. An intermediate flash vessel ensures that the H₂ slip in the stripped-off CO₂ does not exceed some stipulated guide-lines. In the chilled solvent process, the solvent is cooled to about 4^oC by using a NH₃ refrigeration process.  For calculating accurately the thermodynamic properties of the system, the simulation is implemented using the perturbed chain statistical associating fluid theory (PC-SAFT) EOS where the model parameters are regressed from experimental data.  

For linear systems with linear constraints, various algorithms can be considered that can lead to the same optimal constrained linear state estimates. However, constrained state estimation in a nonlinear system is usually not optimal and largely depends upon the algorithm being considered for a particular system. The estimation problem becomes complicated due to the presence of a multimodal probability density function. For the AGR process under consideration, state constraints naturally arise due to non-negative concentrations of the chemical species. In addition, this system has more than one thousand states. So an efficient algorithm is also needed for solving the problem with reasonable computational effort. In this work, the nonlinear constrained state estimation problem is solved by using a variant of the unscented Kalman filter (UKF) that uses a Jacobian free algorithm. A nonlinear programming problem (NLP) is considered in the update step considering constraints on individual updated sigma points and updated states. With certain simplifications, the NLP problem is reduced to a quadratic programming (QP) problem.  Computational performance, estimation errors, and constraint errors of this constrained UKF are compared with the results from the unconstrained Kalman filter and unconstrained UKF.