Foaming prediction in the post-combustion CO2 capture plants (amine-based) by utilizing machine learning techniques

Abstract

Human activities have increased the emitted greenhouse gases into the atmosphere. Among greenhouse gases, the excess of CO2 in the atmosphere has caused severe environmental issues such as global warming and ozone depletion. Several international agreements, such as the Paris Agreement, have been signed to reduce CO2 emissions and their impacts on the environment. These agreements aim to reduce the CO2 footprint regarding all human activities (move toward CO2 neutralization). For this reason, there have been substantial efforts to develop new methods and technologies applied to the exhaust gas from industrial activities to reduce the concentration of greenhouse gases (in a way that is not harmful for human and environment) prior to release them into atmosphere. Post-combustion flue gases contain significant amount of CO2. One common method for capturing CO2 released from post-combustion flue gases is to use an amine-based (i.e., solvent) CO2 capture plant. Although these plants have demonstrated good performances, the problem of foaming within their columns (absorber and desorber) reduces the plant's efficiency. Due to the complexity of the process, there is no physical model that can simulate foaming within a post-combustion CO2 capture (amine-based) plant. Therefore, the main goal of this report is to develop, for the first time, a data-driven model that can simulate and predict the mentioned undesirable characteristic. The data used in this report was provided by technology center Mongstad (TCM) that is the external partner of this project. The data includes the time series (10968 hourly time-steps) of 35 features (i.e., physical properties of the process) regarding the CO2 capture process of TCM post-combustion CO2 capture plant (amine-based). It is worthwhile to mention that the solvent used in the TCM plant was CESAR1. Comprehensive data preprocessing was done in order to tag foaming/non-foaming time-steps as well as to ensure the high data quality before feeding the data to models. Furthermore, a correlation-based method was used for feature selection in order to avoid feeding statistically similar features to the models (retained features are 14). An artificial neural network (ANN) was employed to build a predictive foaming model. The developed model showed a promising performance in predicting the foaming in the plant where the model also was validated by comparing the results with a decision tree model results. Overall, the findings of this report provide practical insights that may allow for the prediction of foaming occurrence within post-combustion CO2 capture (amine-based) plants. The latter can lead to the implementation of preventive measures that can increase CO2 capturing efficiency, and thus favoring the environmental sustainability.