Neuro-fuzzy Soft Sensor Estimator for Benzene Toluene Distillation Column

The distillation is widely used separation technique in oil and gas refineries. Accurate measurement of the composition of separated constituents is necessary to estimate the purity of the products. Composition measurement using online analysers causes process delay and requires large initial investment. As a solution to this problem, soft sensor estimators can be used to determine the composition of separated product. In this work soft sensor estimators are used for predicting top and bottom compositions in benzene toluene distillation column. More sensitive tray temperatures, re-boiler duty and reflux rate (measured variables) of distillation column were used to predict top and bottom composition (unmeasured). Data used for soft sensor based estimation are generated using process simulation software HYSYS. NARX based ANFIS algorithm was proposed for soft sensor modelling. In this method, most influential inputs for soft sensor modelling were selected using exhaustive search. Neural network model and ANFIS model are also compared using statistical criteria like root mean square error and correlation coefficient (R2) values. It has been shown by the results that ANFIS performs better while comparing neural network method and ANFIS with the same number of iteration.