Exploring nontraditional LSTM architectures for modeling demethanizer column operations

Marta Mandis

Primo

;Roberto Baratti;Jorge Chebeir;Stefania Tronci

;Josè Romagnoli

2024-01-01

---

Abstract

Digital twins have recently attracted attention as a new technology that can facilitate the digital transformation of process industries. It may provide live, or near real-time, information and insights into the process and may be used for monitoring, control and optimization purposes. In this study, a digital twin has been developed for modelling the demethanizer column of a NGL separation plant. Based on a non-conventional Long Short-Term Memory (LSTM) neural network arrangement, the surrogate model has been trained and validated using data obtained by the process simulator Aspen HYSYS. Model prediction can be obtained using only readily available variables as input data, ensuring easy and cost-effective implementation. Measurement noises have been considered in order to mimic real-world measurements in a real plant. In both steady-state and transient conditions, the developed demethanizer digital twin accurately reconstructs the separation operation, including compositions, temperatures, and pressures in the reboiler and all column stages.