Deep learning-based coagulant dosage prediction for extreme events leveraging large-scale data

Abstract

The escalating frequency and severity of extreme weather events, attributed to climate change, present significant challenges for water treatment plants (WTPs). Addressing these challenges requires transitioning to automated processes for real-time responses. This study uses a deep learning model to predict coagulant dosage and settled water turbidity, particularly under abnormal conditions such as extreme weather conditions and operational changes. Real-time monitoring data from a WTP in South Korea included input parameters such as raw water quality indicators and operational settings, with output parameters being coagulant dosage and settled water turbidity. The data were preprocessed and used to train the deep learning model, which incorporated a Convolutional Neural Network for feature extraction and a Gated Recurrent Unit for time series analysis. The results showed robust predictive capabilities for coagulant dosage under both typical and extreme weather conditions (R2 = 0.87 and 0.86, respectively) and reasonably accurate predictions for settled water turbidity (R2 = 0.73 and 0.56, respectively). These findings highlight the model's potential for automation in WTPs, even under extreme weather conditions. However, the model's performance was compromised in the case of operational changes involving chemical transitions, as these were influenced by subjective decisions, thereby impacting data distributions. Compared to existing methods, our approach offers strong predictive capability for coagulant dosage and settled water turbidity even during extreme events, enhancing real-time operational efficiency. This study underscores the importance of utilizing large-scale data in water treatment process modeling to improve deep learning model's responsiveness to unforeseen events across various conditions. © 2024 Elsevier Ltd