Improving chlorine residual predictions in water distribution systems using recurrent neural networks

Abstract

The management of water quality in distribution systems is a pervasive challenge. A high degree of uncertainty in water demand, reaction rates, and conditions of the pipe networks results in significant discrepancies between expected and observed water quality. In an effort to enhance the prediction of chlorine residual within water distribution systems (WDS), this study utilized full-scale WDS data to investigate the capabilities of a hydraulic model EPANET-Water Network Tool for Resilience (WNTR) coupled with process-based chlorine residual and data-driven models. Calculation and analysis of observed chlorine decay rates over 19 weeks of recorded data from a full-scale WDS (n = 19 512) demonstrated significant non-linearities and complex relationships with operational parameters and water quality. Linear regression was applied as a baseline method to model the relationship between water quality parameters and chlorine residual, but its limitations in capturing complex, non-linear interactions prompted a transition towards more sophisticated neural network architectures. Furthermore, EPANET-WNTR coupled with a first-order chlorine residual model showed poor performance in predicting chlorine residuals at a downstream node over the full range of flow conditions with high-frequency. Utilizing a windowing technique to account for sequences representing significant travel times in the dataset, the shift to neural networks, including convolutional neural networks (CNN) and long short-term memory (LSTM) networks demonstrated a significantly enhanced ability to incorporate temporal information and predict chlorine residual. The models achieved mean absolute errors as low as 0.022 mg L⁻¹ and R² as high as 0.952 using a 4-layer LSTM. This research illustrates the effectiveness of adopting data-driven approaches that can capture the relationships and dynamics of water quality parameters based on previous data, marking a significant advancement in water quality management within WDS.