Surrogate-Assisted Hierarchical Intelligent Control of an MEA-Based CO₂ Capture Process under Flue-Gas Disturbances

The operation of MEA-based carbon capture systems is complicated by nonlinear interactions among flue-gas conditions, solvent circulation, solvent regeneration, and energy consumption. This study proposes a surrogate-assisted hierarchical intelligent control framework for maintaining CO₂ capture performance while minimizing regeneration energy and solvent use. An Aspen Plus model of the absorber–regenerator system was employed as a virtual process for generating data over a broad operating region. Nine input variables, including flue-gas flow rate, CO₂ concentration, gas temperature and pressure, solvent flow rate, MEA concentration, lean-solvent temperature and loading, and reboiler duty, were varied using Latin hypercube sampling. Of 5,000 simulated operating conditions, 4,436 physically feasible and converged cases were retained. The dataset was divided into training, validation, and testing subsets in proportions of 70%, 15%, and 15%, respectively. A multi-output artificial neural network with three hidden layers containing 64, 64, and 32 neurons was developed to predict six indicators. On the test dataset, the model achieved determination coefficients ranging from 0.987 to 0.996. For CO₂ capture efficiency, the model obtained an RMSE of 0.61 percentage points and an R2R^2R2 of 0.996. The average prediction time was 1.7 ms, compared with 8–15 s required for one converged simulation. The surrogate model was integrated into a hierarchical architecture in which a high-level intelligent supervisor optimizes lean-solvent flow rate, reboiler duty, and solvent temperature, while local PID controllers provide rapid setpoint tracking. Closed-loop tests included changes in flue-gas flow rate, inlet CO₂ concentration, gas temperature, combined disturbances, model uncertainty, and measurement noise. Compared with conventional PID control, the proposed system reduced the average settling time from 18.4 to 7.8 min and decreased the environmental-limit violation duration from 6.2 to 0.8 min. The specific regeneration energy decreased from 3.90 to 3.43 GJ t⁻¹ CO₂, representing a 12.1% reduction, while the average capture efficiency increased to 90.3% and product purity reached 96.4%. The average optimization time of 46 ms confirms the feasibility of real-time supervisory implementation.