Artificial neural network based models for predicting the effluent quality of a combined upflow anaerobic sludge blanket and facultative pond: Performance evaluation and comparison of different algorithms

The main objective of this work was to test different artificial neural network (ANN) based models, i.e. the ANN feed forward back propagation (ANN-FFBP), deep feed forward backpropagation (DFFBP), and deep cascade forward back propagation (DCFBP) models, for predicting the effluent quality of an upflow anaerobic sludge blanket-facultative pond (UASB-FP) system. The overall removal efficiency in the UASB-FP was >84% at organic loading rates of similar to 26 kg d(-1). The chemical oxygen demand (COD), ammonical nitrogen (AN), total suspended solids (TSS), biochemical oxygen demand (BOD), total Kjeldahl nitrogen (TKN), and total phosphorus (TP) were inputs to each model, while the water quality characteristics of the UASB-FP effluent was used as the output. The dataset of 180 samples, collected over a one-year period, was utilized to train, test, and validate the developed models. Compared to ANN-FFBP and DFFBP, the DCFBP network demonstrated the strongest capacity for prediction. The correlation coefficient R-Train and the root-mean-squared error (RMSE) for the selected DCFBP model (3 hidden layers and 11 neurons/layer) in the training data set were 0.997 and 6.018, respectively. The sensitivity analysis of the DCFBP model shows that the model's performance is very sensitive to BOD followed by AN, COD, TP, TSS and TKN, respectively. The results of this study will be helpful to wastewater treatment (WWTP) plant managers in their pursuit of data-driven UASB-FP based WWTP management.