Robust monitoring of solvent based carbon capture process using deep learning network based moving horizon estimation

Solvent based CO₂ capture is a promising technology for the decarbonisation of energy and industrial sectors. On-line monitoring of the key parameters within the CO₂ capture process can provide insight into the operating status of the system and lay the foundation for the controller design. However, limited by current measurement technologies, some important parameters such as CO₂ concentration, solvent loading etc. are still difficult to be measured accurately in real time, especially under complex operating conditions. Therefore, this paper presents an intelligent model-based robust soft sensor to monitor the key operating parameters within the CO₂ capture process. The robust monitoring approach is consisted by two parts, one is a Long Short-Term Memory (LSTM) Network based surrogate model identified using the dynamic operating data generated from a high-precision simulator, which can represent the transient behavior of the carbon capture process over a wide range; while the other part is a moving horizon estimator which fully uses the combination of available process measurements and the developed LSTM network prediction to enhance the reliability of the estimation in the cases of unknown interferences, noises and measurement faults. Simulation results on a large-scale monoethanolamine-based carbon capture process model show that the proposed monitoring method can accurately estimate the clean gas CO₂ concentration and lean/ rich solvent loading over a wide operating range, even in the presences of noises and measurement faults.