Abstract: Hydrogen sulfide concentration of the tail gas from the methyldiethanolamine (MDEA) desulfurization zone for dry gas from fluid catalytic cracking (FCC) unit must be controlled below 10 μg/g. Based on the 705 sets of historical data from the dry gas desulfurization zone in the FCC unit of a petrochemical enterprise,25 modeling variables were screened from 44 variables by the maximum mutual information coefficient and Pearson correlation coefficient methods,transforming hydrogen sulfide concentration in the tail gas into a classification problem as the target variable. An XGBoost and a DNN binary classification models for predicting target variable were established respectively. The evaluation results of the DNN model are better than that of XGBoost,particularly in the accuracy of classifying substandard samples and the stability of handling complex problems. Combiningthe DNN model with the Genetic Algorithm, the operating variables in thesubstandard samples from the test set were optimized to achieve the standard. Therefore, the DNN model can provide guidance tooptimize the operation of MDEA desulfurization zone in FCC unit.

Key words: hydrogen sulfide, MDEA desulfurization, data mining techniques, genetic algorithm