Synergizing Phenomenological and AI-Based Models with Industrial Data to Develop Soft Sensors for a Sour Water Treatment Unit

Abstract

Sour waters are one of the main aqueous byproducts generated during petroleum refining and require processing in sour water treatment units (SWTUs) to remove contaminants such as H2S and NH3 in compliance with environmental legislations. Therefore, monitoring the composition of SWTU effluxents, including acid gas, ammoniacal gas, and treated water, is essential. This study aims to present an AI (artificial intelligence) hybrid-based methodology to develop soft sensors capable of real-time prediction of H2S and NH3 mass fractions in the effluents of SWTUs and validate them using real data from industrial units. Initially, a new database based on the dynamic simulation of a two-stripping-column SWTU phenomenological model, developed in Aspen Plus Dynamics® V10, was generated, aiming at non-faulty runs, unlike our previous work. Ensemble methods (decision trees), such as gradient boosting and random forest, and support vector machines were compared for soft sensor creation using these simulated data. The best outcome was the development of six soft sensors based on random forest with R2 greater than 0.87, MAE less than 0.12, MSE less than 0.17, and RMSE less than 0.41. Variable importance analysis revealed that the temperature of the second stage of Column 1 significantly influences the thermodynamic equilibrium of H2S and NH3 separation from sour waters, being critical for five of the six soft sensors. After this initial stage using data from the phenomenological model, data from an industrial-scale SWTU were used to develop real soft sensors. The results proved the effectiveness of the conjugated use of a physical model and industrial data approach in the development of soft sensors for two-column SWTUs.