Determination of coagulant dose using deep learning-based forecasting model with 9 years of field operation data

Abstract

Coagulation is one of the most important processes to ensure the proper effluent water quality during the water treatment process. Determination of proper coagulant dose is a time-consuming process involving several water quality factors including pH, turbidity, alkalinity, etc. Traditionally, operators determined coagulant doses heavily relied on their past experiences, and the quality of the settled water could not be reflected immediately. This study provided alternative approaches to determine the appropriate coagulant dose using raw water quality and settled water turbidity data with attention-based deep learning model. Four types of models including multiple linear regression (MLR), random forest (RF), gate recurrent units (GRU), and graph attention multivariate time-series forecasting model (GAMTF) were applied and compared in this work. Long-term multivariate time series characteristics along with graph attention neural network were assembled to determine the coagulant dose and to forecast the settled water turbidity simultaneously. The results showed that GAMTF outperformed the other models with the smallest root-mean-square errors (RMSE), and the highest correlation coefficient (R) and coefficient of determination (R 2 ), and successfully predicted both coagulant dose and settled water turbidity. This study showed the first application of long-term multivariate time series characteristics to train the deep learning model, and the capacity of state-of-the-art graph attention multivariate time-series forecasting model to outperform the other models. This study also proposed an improved decision support system for coagulant dose determination.