A deep learning approach to forecast the influent flow in wastewater treatment plants

Abstract

For the management and operation of a Wastewater Treatment Plant (WWTP), the influent flow is one of the most important variables. Hence, this paper presents an evaluation of multiple Deep Learning models to forecast the influent flow in WWTPs for the next three days, taking into account previous influent observations as well as historical climatological data. Long Short-Term Memory networks (LSTMs) and one-dimensional Convolutional Neural Networks (CNNs), following a channels last approach, were conceived to tackle this time series problem. The best candidate LSTM model was able to forecast the influent flow with an approximate overall error of 200 m3 for the three forecast days. On the other hand, the best candidate CNN model presented a slightly higher error, being outperformed by LSTM-based models. Nonetheless, CNNs, which are typically applied in the computer vision domain, also showed interesting performance for time series forecasting.