AI enhanced model predictive control for optimizing LPG recovery through integrated computational modeling design of experiments and multivariate regression

Liquefied Petroleum Gas (LPG) recovery in debutanizer columns presents challenges in balancing operational efficiency and process stability under varying conditions. Conventional control strategies often fail to sustain optimal recovery. This study integrates process modeling and control, using Aspen HYSYS for steady-state simulation and dynamic implementation of model predictive control (MPC). Response surface methodology (RSM) was applied to steady-state simulation results to analyze key process variables. Feed molar flow rate was the most influential factor, while pressure-related variables showed minor but statistically significant effects. The quadratic model and 3D response surfaces confirmed key interactions. A regression decision tree model was developed in MATLAB to support deployment of artificial intelligence-enhanced MPC (AI-enhanced MPC). MPC improved LPG recovery from 99.73 to 99.85%, reduced reboiler duty from 1,557,000 to 1,550,000 kcal/h, and reflux flow from 281.2 to 271 kgmole/h. AI-enhanced MPC further increased recovery to 99.9%, reduced reboiler duty to 1,501,956 kcal/h, condenser duty to 2,415,726 kcal/h, and reflux flow to 262.6 kgmole/h, indicating superior energy efficiency and control precision. Although feed molar flow remained dominant, both control systems regulated its impact via pressure, temperature, and reflux. Product temperature dropped from 49.88 °C to 49.24 °C, and pressure from 12.39 to 11.95 bar, indicating enhanced thermal stability. The novelty of this study lies in integrating RSM with both conventional and AI-enhanced model predictive control, forming a hybrid framework enabling steady-state optimization and dynamic control for improved LPG recovery. The proposed framework supports industrial LPG recovery by improving energy efficiency, product quality, and dynamic stability.