DEVELOPMENT OF DATA-DRIVEN MODELS FOR INFLUENT PREDICTION AT WASTEWATER TREATMENT PLANTS

Abstract

Influent flow rate is essential to the operation and management of wastewater treatment plants (WWTPs). To support safe operation and effective management of WWTPs, a number of process-driven models were previously built for predicting the influent flow rate. However, in order to capture the complex nonlinear relationships in wastewater systems, these process-driven models require large-scale monitoring and complicated parameter tuning. In this research, to address those drawbacks, data-driven models are investigated for influent flow rate prediction. Three data-driven models, including multilayer perceptron (MLP), long short-term memory (LSTM) network, and random forest (RF), are introduced and developed. The developed models are applied to three WWTPs in Canada for influent flow rate prediction to demonstrate their applicability. Influent flow rate prediction with two temporal resolutions (i.e., daily and hourly) are provided. The results show that the proposed models have an overall good performance, especially the RF model. For both temporal resolutions, the performance of RF models is stable and satisfactory. In addition, an uncertainty analysis approach for the RF model is developed to provide more robust predictions. To the author’s knowledge, this is the first Canadian study of wastewater influent flow rate prediction based on advanced data-driven techniques. The high temporal resolution prediction and the probabilistic prediction approach proposed in this research represent a unique contribution to methodologies related to wastewater modeling. This research can provide valuable support for WWTPs to improve operational efficiency and management effectiveness.