Sliding window neural network based sensing of bacteria in wastewater treatment plants

Ensuring the performance of wastewater treatment processes is important to guarantee that the final treated wastewater quality is safe for reuse. However, bacterial concentration present along the different stages of treatment process is not easily measured routinely for the plant operators. In this paper, a moving horizon sensing approach based on neural networks is proposed to estimate the bacterial concentration in wastewater sampled along different stages of the plant. Due to the difficulties to measure the bacteria and the lack of a sufficiently large dataset, a Wasserstein generative adversarial network (WGAN) is designed to generate synthetic data. The Wasserstein critic loss is computed on a held-out validation set to evaluate the WGAN. Then, the generated data is used to train a long short term memory (LSTM) neural network that is developed to predict the biomass concentration and update the LSTM weights by a sliding window learning approach. Two datasets for WWTP are used to test the proposed method: first, effluent concentrations simulated using a benchmark simulation model no.1 (BSM) based on membrane bioreactor (MBR), where three different weather profiles of influent data were considered then, sampled data from MBR plant at King Abdullah University of Science and Technology (KAUST). Finally, the prediction results indicate that WGAN successfully generates realistic samples that are used to train the LSTM neural network. In addition, estimation performance of the proposed method is compared with a multilayer perceptron neural network (MLP-NN). Results showed that the proposed method improves the bacteria estimation performance compared to MLP-NN. (C) 2021 The Authors. Published by Elsevier Ltd.

Automated summary

This study proposes a neural network-based approach to estimate bacterial concentration in different stages of wastewater treatment processes, using WGAN to generate synthetic data. The method was tested with two datasets, showing that WGAN successfully generates realistic samples for training LSTM neural network, which outperforms traditional methods (MLP-NN) in bacterial estimation performance.