Machine Learning-Based Soft Sensor for Hydrogen Sulfide Monitoring in the Gas Treatment Section of an Industrial-Scale Oil Regeneration Plant

Monitoring chemical composition is key in several industrial-scale chemical processes. However, traditional composition sensors usually convey drawbacks, including high costs, short lifetimes, and frequent calibration requirements. As an alternative, software (soft) sensors have gained attention in recent years due to their accuracy, ease of training, and potential of integrating widely known machine learning techniques. This study presents the methodology followed to train a soft sensor for hydrogen sulfide monitoring in the gas treatment section of an industrial facility in Italy. In particular, this methodology includes a novel approach for steady-state determination from historical plant data in the presence of several steady states and noise. Unfortunately, only four steady states were found in the plant data, which was insufficient for accurate soft sensor training. As an alternative, these steady states were used to develop and validate a rigorous Aspen HYSYS process simulation. The relative error of the simulation compared to the steady states found was around 1%. This simulation was used to extract additional steady state data, which was sufficient for soft sensor training. The soft sensor, trained following a Kriging Gaussian Process regression, showed a relative error with respect to the rigorous simulation of around 0.7%. The high accuracy of the sensor suggests it is a quick, low-cost, and accurate alternative to estimate the hydrogen sulfide composition and identify potential operating conditions that may lead to off-specification conditions. Future work will focus on field implementation and evaluation of the trained sensor.