DEVELOPMENT OF A HYBRID SOFT SENSOR FOR KEROSENE FRACTION FLASH POINT IN A MODEL PREDICTIVE CONTROL SYSTEM OF AN ATMOSPHERIC COLUMN

Abstract

This paper presents the development and experimental validation of a hybrid soft sensor for estimating the flash point of the kerosene fraction within a Model Predictive Control (MPC) system of an atmospheric distillation column. In industrial refining processes, one of the key problems of real-time quality control is the discreteness and significant delay of laboratory measurements, which typically range from 30 to 60 minutes. Such delays make laboratory data unsuitable for direct integration into automatic control loops, leading to suboptimal control decisions and reduced operational efficiency.

To address this limitation, a hybrid modeling approach is proposed. In this approach, the structure of the model is derived from the physical principles governing the atmospheric distillation process, ensuring interpretability and consistency with process dynamics, while the numerical values of the parameters are identified from historical industrial data. The resulting model is formulated as a linear regression with five predictors that reflect the main technological factors affecting the flash point of the kerosene fraction. These predictors include kerosene withdrawal temperature, the steam-to-kerosene flow ratio, feedstock density, column top temperature, and an additional operational parameter representing process variability.

Parameter identification was performed using ridge regression to mitigate multicollinearity and improve model stability. The training dataset consisted of 724 industrial observations collected from a real operating distillation unit. In addition, an adaptive correction mechanism based on an exponential filter with a smoothing coefficient α = 0.3 was implemented to compensate for characteristic model drift caused by changes in feedstock composition and operating conditions.

The developed soft sensor demonstrated high predictive accuracy, achieving an RMSE of 2.15°C and a coefficient of determination R² = 0.88, which lies within the typical error range of laboratory flash point measurements. A comparative analysis with alternative data-driven approaches, including Partial Least Squares (PLS) regression and Artificial Neural Networks (ANN), confirmed the superiority of the proposed hybrid model in terms of robustness, interpretability, and stability.

Simulation of the integrated “soft sensor + MPC” control system showed that the use of the developed estimator improves control performance by approximately 50% and ensures reliable compliance with the critical flash point constraint during operation. These results demonstrate the practical applicability of the proposed approach for real-time quality monitoring and advanced process control in industrial distillation systems.