(377e) A Hybrid CNN-Based Framework for Optimizing Amine Gas Treating Processes

Amine gas treating is an important step in natural gas processing, where accurate modeling and optimization are essential to achieve efficient operation. This work presents a novel hybrid approach that integrates first-principles modeling with machine learning techniques to develop a real-time optimization framework for this process. A key innovation is the use of a 1-D Convolutional Neural Network (CNN) as a surrogate to model the complex spatial relationships in the absorber trays' temperature profile. This allows formulating key constraints in the optimization problem relating to the estimated temperature at various trays. The 1D-CNN is particularly well-suited for this task due to its ability to capture local patterns, which allows effectively modelling the complex dynamics of the absorber trays and improve the accuracy of our predictions. In conjunction with the CNN model, several other surrogates are developed using Lasso polynomial regression. A parameter estimation framework is then developed and solved using differential evolution. This framework estimates unmeasured process variables, such as the composition of H2S and CO2 in the sour feed gas, and the concentrations of amine components, by minimizing the error between predicted and measured process variables. The estimated parameters are then used to inform the process optimization framework, which is formulated as a two-stage problem. The first stage minimizes amine circulation rate while satisfying temperature and loading constraints, while the second stage maximizes sour gas feed rate while adjusting the circulation rate as needed. The application of this methodology to a real-world problem yields promising results, with an average increase in plant throughput of 3.7% and a reduction in amine circulation rate of 2.8%, resulting in energy savings of 2.2%. This work demonstrates the potential of hybrid modeling approaches to improve the efficiency and productivity of complex industrial processes, and highlights the value of CNNs in modeling spatially correlated process variables in process engineering problems.