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

This paper proposes advanced soft sensor models based on machine learning and deep learning for real-time estimation of top and bottom product compositions in a Heat-Integrated Distillation Column (HIDiC). Conventional composition analyzers, such as gas chromatographs, are expensive and suffer from significant measurement delays, making them less efficient for real-time measurement and control. As a cost-effective alternative, soft sensors can be developed using process data from a high-fidelity dynamic HIDiC simulation, with tray temperatures as the model inputs. This research develops and evaluates both linear and nonlinear modeling strategies for composition estimation in a HIDiC, including Principal Component Regression (PCR), Artificial Neural Network (ANN), and marking the first application of its kind in HIDiC modeling a Bidirectional Long Short-Term Memory (BiLSTM) network. While PCR and ANN achieved reasonable accuracy, their performance was limited by an inability to fully capture the temporal dependencies and complex nonlinearities inherent in the distillation process. In contrast, the BiLSTM model, leveraging its deep learning architecture and temporal memory capabilities, successfully learned long-range dependencies and intricate dynamic patterns in the process data. Comprehensive performance evaluation based on Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²) demonstrated that the BiLSTM model outperformed the traditional models significantly. The results confirm that the BiLSTM-based soft sensor not only enhances prediction accuracy but also represents a novel and effective approach for real-time composition estimation in HIDiC systems, offering great potential for advanced monitoring and control applications.