Development of an integrated adaptive control system for technological units of an oil refinery based on multimodel predictive control and self-learning digital twins

Abstract

The article is devoted to the development and pilot-industrial validation of an integrated adaptive control system for oil refining process units, aimed at overcoming the limitations of classical DCS/PID loops and model degradation in traditional linear APC solutions under nonstationarity, nonlinearity, and equipment parameter drift. An architecture is considered that combines multimodel nonlinear predictive control with a receding horizon and an ensemble of self-learning digital twins functioning as dynamic predictors and virtual quality analyzers under conditions of significant laboratory analysis delays. Primary oil processing (EDU-AVT-6) and the catalytic cracking section are described as key technological examples; processing of large volumes of high-frequency archival telemetry and sparse quality data required data cleaning, synchronization, and computational support procedures compatible with industrial real-time requirements. The digital twins are implemented in a hybrid manner: recurrent neural network models (LSTM/GRU) are supplemented with simplified physicochemical relations of mass and heat balances to enhance robustness when operating conditions move beyond the training scenarios. A continuous adaptation mechanism is introduced that initiates retraining when forecast error increases, as well as a setpoint validation module that constrains control actions by technological regulations. It is shown that for a distillation column under feed disturbances, adaptive NMPC provides a significant reduction in settling time, suppression of overshoot, and a decrease in integral error criteria compared to cascade PID and standard MPC, while simultaneously reducing specific steam consumption. For virtual quality analyzers, high explanatory power of the models (R² > 0.92) and errors comparable to laboratory reproducibility for a number of indicators are demonstrated, making closed-loop quality control possible without waiting for analyses. At the catalytic cracking level, a redistribution of the material balance toward light fractions is recorded with a reduction in gas yield, coke formation, and catalyst losses, and correlation analysis links regime stabilization with economic effects. The described results and computational latency indicators confirm the practical feasibility of the approach and its relevance for improving energy efficiency, target product yields, and control reliability at modern oil refineries.