Development of an Intelligent Soft Sensor Using Time-Series Neural Networks for Real-Time Composition Prediction in CDUs

Abstract

This paper investigates the creation and application of a real-time soft sensor aimed at improving process stability and control by predicting product composition in Crude Distillation Units (CDUs). Creating a machine-learning model capable of continuously analysing and predicting product composition during crude oil distillation is utilized for better operation and control. The research is founded on experimental data obtained from validated dynamic simulations of the CDU process using Aspen-HYSYS software. Time Series Linear Regression (TSLR), Time Series Partial Least Squares (TSPLS), and Time Series Neural Networks (TSNN) are methodologies employed in various soft sensor models developed. Performance metrics, such as root mean square (RMS) and the coefficient of determination (R-squared), facilitate the assessment of these models. The Time-series Neural Network (TSNN) distinguishes itself as the optimal model for estimating distillation endpoints in CDUs, achieving the lowest Root Mean Square (RMS) error of 0.8006 and the highest coefficient of determination (R-squared) due to its superior accuracy and predictive capabilities. The TSNN soft sensor, integrated with the Aspen HYSYS modelling plant, provided real-time estimations of diesel molar flow and composition. The model, integrated with the simulated plant, was trained on real-time data from an Aspen HYSYS simulation of a crude oil distillation unit, enabling continuous live estimations.