Prediction of flooding in an absorption column using neural networks

Presents the results of a project aimed at predicting the occurrence of flooding in the absorption column of the Rectisol process at Sasol Synthetic Fuels. The flooding predictor was used as a controlled variable in a multivariable controller with an objective function that maximizes the gas throughput. Thus the prediction of flooding was critical to the success of the project. The technical problem was first to identify a measured variable or a set of variables as a good precursor to flooding. Once suitable precursors had been identified, a statistical flooding predictor was developed. Data from the Rectisol process corresponding to several different flooding scenarios was analyzed. The pressure drop across a section of the column was determined to be a good indicator of flooding. In addition, several other variables were identified as possible precursors. Linear and nonlinear models (neural network models) were developed to predict the pressure drop across the section of the column using the variables identified to be good precursors to flooding. These models were implemented online in the distributed control system. Special flooding tests showed that the neural network differential pressure predictor was accurate. The linear differential pressure predictor, on the other hand, performed quite poorly. This indicated that there were sufficient nonlinear dynamical effects in the Rectisol process to warrant the use of a nonlinear model such as a neural network