283454 Surrogate-Based Optimization of CO2 Capture Process with Aqueous Amines
An increase in the level of greenhouse gases, mainly carbon dioxide, methane and nitrous oxide, in the atmosphere has become more severe in recent decades. The environmental and global warming concerns accelerated the research and development of carbon dioxide capture technologies to control and reduce the emissions of this primary greenhouse gas. As a result, there are three main technologies to capture CO2 from the combustion of fossil fuels: pre-combustion, post-combustion, and oxy-combustion. In the post-combustion capture, CO2 is recovered from the flue gas, whereas in pre-combustion capture, carbon is removed from the fuel prior to combustion. In oxy-combustion capture, the fuel is burned using an almost pure oxygen stream resulting in an almost pure CO2 stream. We focus on the post-combustion capture technologies because they can be readily used to retrofit existing plants and have the highest potential to reduce the CO2emissions in the short to medium term.
We developed surrogate-based optimization approaches to study the process synthesis and optimization of post-combustion carbon capture technologies, i.e., a conventional amine-based absorption. Surrogate- based optimization approaches can be categorized into two groups based on their implementations. In the first class, the optimization module uses local search space to estimate an appropriate direction to reduce the objective function (assuming minimization), e.g., response surface methodologies. In the second category, the simulation is used to build surrogate models, e.g., the artificial neural network (ANN), of the objective function and/or the constraint over the whole decision space, and the optimization is performed using these surrogate models.
In this work, the surrogate-based optimization frameworks using both categories were developed: response surface methodology and artificial neural network as the objective function. 1365 kg mole/h of gas turbine flue gases at a pressure of 1.1 bar and a temperature of 40 oC was processed in a conventional amine-based absorption plant for CO2 recovery. The simulation was developed using Aspen HYSYS Version 7.1. The amines property package was used as the thermodynamic model. Li-Mather electrolyte model was selected as the equilibrium model.
In the response surface methodology, the process simulation was run that corresponds to the operating conditions and design variables generated by an experimental design. The combination of the decision variables as the inputs and the CO2 capture cost as the outputs of the simulation was used to obtain the first order regression model for the objective function. We used the R2 with the cutting value of 0.5 and adjusted-R2 statistics in order to assess the fitness of the first and second order response models. If it was concluded that the first-order model represents the data set, the steepest descent was performed to move the system in the direction of maximum decrease in the objective function. On the other hand, if it was found that the data set cannot be represented accurately with the first order model, the second-order model was utilized to find the local minimum CO2 capture cost and the corresponding design and operating conditions. The impact of different amine absorbents and their concentrations on the optimum CO2 cost was analyzed. The results suggested that CO2 recovery with 48 wt% DGA required the lowest CO2 removal cost. Sensitivity analysis for the optimum amine-based CO2 capture plant was used to study the impact of flue gas feed composition and utility cost fluctuations on the minimum CO2 capture cost. The CO2 capture cost showed an inversely proportional relationship with the amount of CO2composition in flue gas while the utility cost did not show significant impact.
In the artificial neural network approach, the ANN was utilized to mimic the objective function behavior, cost of CO2 removal, in terms of the decision variables. Feed forward neural network was used because of its mathematical simplicity. A back propagation (BP) algorithm was employed to train the network. We developed an algorithm to determine the appropriate sample size for constructing accurate artificial neural networks as surrogate models in optimization problems. In the algorithm, two model evaluation methods—cross-validation and/or bootstrapping—were used to estimate the performance of various networks constructed with different sample sizes. The CO2 capture process with 48 wt% DGA solution resulted in the lowest CO2 removal cost. The minimum CO2 capture costs obtained from both approaches are similar for all solvents, except in the case of TEA, where the ANN approach is able to locate a lower CO2 capture cost compared to the RSM algorithm results. This is because the RSM algorithm searches for the minimum CO2 capture cost in the local region of large number of absorber stages, while the ANN approach was able to identify the region of lower CO2 capture cost over the entire domain of the number of absorber stages for a better prediction of the minimum CO2 capture cost.