Modeling of expanded granular sludge bed reactor using artificial neural network

Abstract

In present study, a three-layer backpropagation neural network (BPNN) model was developed to predict the performance of an expanded granular sludge bed (EGSB) reactor. Six related variables such as influent chemical oxygen demand (COD) concentration, hydraulic retention time (HRT), alkalinity (ALK) concentration, pH, volatile fatty acid (VFA) concentration and oxidation reduction potential (ORP), were selected as inputs of the model. All input values were converted to the range (−1, 1) before passing them into the network. Activation function of hidden layer and output layer were “tansig” and “purelin” individually. Several comparisons were conducted to obtain an optimal network structure. Dividerand function was chosen to divide the operating data into training group, testing group and validation group. The Levenberg Marquardt algorithm (trainlm) was found as the best of the ten training algorithms. Other model parameters such as number of neurons in the hidden layer (X1), initial adaptive value (X2) and initial value of weights and biases (X3) were optimized using response surface methodology (RSM). The optimum conditions for minimum mean squared error (MSE) were as follows: X1 (12), X2 (6.0) and X3 (1.0). The precision of optimum ANN model was assessed by means of various statistics such as MSE, determination coefficient (R2), coefficient of variation (CV) and MSE. The result indicated that the proposed ANN model exhibited superior predictive accuracy for the forecast of COD removal performance by EGSB system. Finally, the results of connection weights method demonstrated that VFA concentration (50.37%) had a remarkable impact on reactor performance.