Robust Nonlinear Soft Sensor for Online Estimation of Product Compositions in Heat-Integrated Distillation Column

This paper proposes the development of a robust nonlinear soft sensor for online estimation of product compositions in a Heat-Integrated Distillation Column (HIDiC). Traditional composition analyzers, such as gas chromatographs, are costly and suffer from long measurement delays, making them inefficient for real-time monitoring and control. To address this, data-driven soft sensors are developed using tray temperature data obtained from a high-fidelity dynamic HIDiC simulation. The study investigates both linear and nonlinear modeling strategies for composition estimation, including principal component regression (PCR), artificial neural networks (ANNs), and, for the first time in HIDiC modeling, a Bidirectional Long Short-Term Memory (BiLSTM) network. The objective is to evaluate the capability of each method for accurate estimation of product compositions in a HIDiC. The results demonstrate that the BiLSTM-based soft sensor significantly outperforms conventional methods and offers strong potential for enhancing real-time composition estimation and control in HIDiC systems.