Predicting Effluent Quality in Full-Scale Wastewater Treatment Plants Using Shallow and Deep Artificial Neural Networks

This research focuses on applying artificial neural networks with nonlinear transformation (ANNs) models to predict the performance of wastewater treatment plant (WWTP) processes. The paper presents a novel machine learning (ML)-based approach for predicting effluent quality in WWTPs through explaining the relationships between the multiple influent and effluent pollution variables of an existing WWTP. We developed AI models such as feed-forward neural network (FFNN) and random forest (RF) as well as deep learning methods such as convolutional neural network (CNN), recurrent neural network (RNN), and pre-train stacked auto-encoder (SAE) in order to avoid various shortcomings of conventional mechanistic models. The developed models focus on providing an adaptive, functional, and alternative methodology for modeling the performance of the WWTP. They are based on pollution data collected over three years. It includes chemical oxygen demand (COD), biochemical oxygen demand (BOD 5 ), phosphates ( P O ₄ − 3 ), and nitrates ( N O ₃ − ), as well as auxiliary indicators including the temperature (T), degree of acidity or alkalinity (pH), electric conductivity (EC), and the total dissolved solids (TDS). The paper presents the results of using SNN- and DNN-based models to predict the effluent concentrations. Our results show that SNN can predict plant performance with a correlation coefficient (R) up to 88%, 90%, 93%, and 96% for the single models COD, BOD5, N O ₃ − , and P O ₄ − 3 , respectively, and up to 88%, 96%, and 93% for the ensemble models (BOD 5 and COD), ( P O ₄ − 3 and N O ₃ − ), and (COD, BOD5, N O ₃ − , P O ₄ − 3 ), respectively. The results also show that the two-hidden-layers model outperforms the one-hidden-layer model (SNN). Moreover, increasing the input parameters improves the performance of models with one and two hidden layers. We applied DNN (CNN, RNN, SAE) with three, four, and five hidden layers for WWTP modeling, but due to the small datasets, it gave a low performance and accuracy. In sum, this paper shows that SNN (one and two hidden layers) and the random forest (RF) machine learning technique provide effective modeling of the WWTP process and could be used in the WWTP management.