Prediction and multi-objective optimization of pilot-scale carbon capture system based on multi-source monitoring information and novel data-driven model

With the increasing global demand for carbon emission reduction, predictive research on the performance of carbon capture systems is of great significance. This paper constructs performance prediction models for a carbon capture experimental system embedded with physical constraints, utilizing three machine learning algorithms: Random Forest (RF), Back Propagation Neural Network (BPNN) and Convolutional Neural Network (CNN). Twelve critical operating parameters were selected as input parameters based on the Pearson correlation coefficient, with CO₂ output volumetric flow rate, capture efficiency, regeneration energy consumption, and thermal efficiency as output parameters. The results indicate that the RF algorithm is the most suitable among the three machine learning algorithms, with all output variables achieving R² values exceeding 0.98 and average relative prediction errors below 0.6 %. Furthermore, a “prediction-optimization-decision” collaborative optimization framework was established to simultaneously maximize CO₂ output volumetric flow rate, capture efficiency, and thermal efficiency, while minimizing regeneration energy consumption. The optimal system operating points were determined as 1.33 m³/h, 90.54 %, 91.71 %, and 4.54 GJ/t, respectively. Comparison with experimental optimization results shows that the errors in all key parameters are below 5 %, establishing a benchmark for the design of efficient, low-cost, and sustainable carbon capture experimental systems.