A Demethanizer column Digital twin with non-conventional LSTM neural networks arrangement

This work aims to develop a digital twin for a demethanizer column and provide a useful tool for monitoring and quality control of the NGL recovery process. For this purpose, a digital data-driven model was proposed to mimic real dynamics of a cold residue reflux (CRR) unit through the incorporation of physical knowledge. A non-conventional LSTM network arrangement was developed considering training test and validation data sets generated by the process simulator Aspen HYSYS®. This simulation model was built by considering realistic measurement noises to mimic the actual measures in a real plant. The obtained surrogate model was evaluated considering its ability to recreate the operation of the actual distillation column, estimating the temperature and composition transient profiles of the bottom column product and of every stage of the column. Overall, the model developed with the proposed LSTM network arrangement proves capable of successfully reconstructing the actual profiles of all the considered variables.