Predictive modeling of an industrial UASB reactor using NARX neural network

Abstract

A NARX(Nonlinear Autoregressive Exogenous Input) neural network model of an industrial UASB reactor was developed in this research work. A total of 111 days' data were used for the modeling process, specifically, 100 for training and 11 for comparison of predictions. Several designs were generated for the neural network to check the behavior of the predictive model during the training phase. The final design was optimized by observing performance characteristics and regression analysis by using a customized MATLAB script. The model was capable of realizing the dynamics of the system. A 5-6-2 architecture was capable of suitably modeling the UASB reactor and predicting values of biogas production rate and outlet COD concentration. Almost all predictions lied within ±10% deviations. Such a model may be utilized to predict the output of UASB reactor satisfactorily for its supervision, monitoring and control.