(394ab) A Hybrid Framework for Early Flooding Detection in Distillation Columns: Integrating Hydraulic Modeling and Data-Driven Techniques

Distillation columns are crucial for achieving high purity in chemical separations. In the crude oil refining industry, where economic margins are tight, fault monitoring and prediction have become increasingly important. Faults, any deviations from normal operating conditions, reduce efficiency and can even cause asset loss. To meet purity standards, operators sometimes use over-reflux, which raises energy consumption and contributes to flooding by causing excessive liquid buildup. Flooding occurs when the liquid-vapor traffic exceeds a column section’s hydraulic limits and can also be triggered by excessive heating, too much stripping steam, a high flow of low-density feedstock, or a switch from high- to low-density feed. This phenomenon reduces separation efficiency and results in off-spec product quality. Therefore, effective control strategies that predict and prevent flooding, supported by robust fault detection and diagnosis (FDD), are essential for ensuring safety, reliability, and optimal column performance [1].

Empirical equations have been developed to detect flooding using parameters such as surface tension and gas/liquid densities; however, since these parameters cannot be measured directly in real time, oversimplification introduces significant uncertainties. Data-driven models, such as PCA, k-nearest neighbors, and autoencoders, are commonly employed for fault detection in chemical engineering, yet their reliance on available measurementswhich is limited by few faulty samples, operational drifts, faulty sensors, unbalanced classes, and complex system dynamics can lead to overfitting and misclassification [2]. Traditional indicators in distillation columns (e.g., tower pressure differences or purity levels) often signal faults too late for effective intervention. To overcome these challenges, hybrid models that combine data-driven techniques with first-principle approaches are gaining attention for their ability to estimate internal process variables that are difficult to measure directly.

Motivated by the challenges discussed earlier, we enhance fault detection by employing a hydraulic model of the distillation column. Hydraulic models provide key indicators, such as the liquid level on each tray and in each downcomer, that help detect early signs of downcomer flooding [3]. We use a simplified model that assumes the column contains only liquid, applying correction factors to account for the absence of vapor. By combining this model with available measurements of relevant variables, we solve the resulting ordinary differential equations (ODEs) to gain deeper insights into the column's internal dynamics. These extracted variables support our data-driven fault detection approach by enabling the prediction of flooding before it occurs. The method is validated using the ethylene splitter (C2 splitter) within ASPEN Dynamics software and compared against purely data-driven techniques.

References

[1] Jalanko, M., Sanchez, Y., Mhaskar, P. and Mahalec, V., 2021. Flooding and offset-free nonlinear model predictive control of a high-purity industrial ethylene splitter using a hybrid model. Computers \& Chemical Engineering, 155, p.107514

[2] Ochoa-Estopier, L.M., Gourvénec, S., Cahors, R., Behara, N. and Scellier, J.B., 2023. Prediction of flooding in distillation columns using machine learning. Digital Chemical Engineering, 7, p.100098.

[3] Du, Y., Luo, Y., Yang, P., Jia, S. and Yuan, X., 2024. Rigorous design and economic optimization of reactive distillation column considering real liquid hold-up and hydraulic conditions of industrial device. Chinese Journal of Chemical Engineering, 76, pp.211-226.