Dynamic Neural Network Model and Parameter Estimator for Hydrochloric Acid Recovery Process
Multi-scale and/or multi-disciplinary approach to process-product innovation
CFD & Multiscale Modelling in Chemical Engineering (T3-4P)
Keywords: Neural network modeling, Neural network parameter estimator, Pickling process, Hydrochloric acid recovery process, Ion exchange resin
This paper describes the development of neural network models and a parameter estimator of a hydrochloric acid recovery process. The hydrochloric acid recovery process consisting of double fixed-bed ion exchange columns is used to remove Fe2+ and Fe3+ ions from pickling liquor of a pickling bath. This results in increasing the acid concentration which then can be reused in the pickling process. Due to the complexity and highly nonlinearity of the process, the modeling of the process based on the first principle is rather difficult and involves many unknown parameters. Therefore, an attractive alternative technique, neural network modeling, has been applied to model this system because of its ability to model a complex nonlinear process, even when process understanding is limited. The process data sets are gathered from a real hydrochloric acid recovery pilot plant and first used for developing a neural network parameter estimator used to estimate unknown/uncertain parameters of the process such as Biot and Peclet numbers. Then, the first principle models of the process coupled with the estimates of the Biot and Peclet numbers are employed to carry out simulation study. The data generated from the simulation study are then verified with those of experimental study and are utilized to train and validate neural network models for the process. Lenvenberg-Marquardt techniques are used to train various neural network architectures, and the accuracy of the obtained models has been examined by a test data set. The optimal neural network architectures of this process can be determined by a Mean Square Error (MSE) minimization technique. The simulation results have shown that the developed neural network models with one hidden layers provide sufficiently accurate prediction of the concentration profile of the process.
Presented Tuesday 18, 13:30 to 15:00, in session CFD & Mutliscale Modelling in Chemical Engineering (T3-4P).
