Exploring potential dual-stage attention based recurrent neural network machine learning application for dosage prediction in intelligent municipal management

Abstract

Chemical demand prediction is important for water management and the environment. This study aimed to select and apply suitable data-driven models based on real-world big data for dosage prediction towards improved automated control of water treatment plant management. Coagulation is a prominent process in normal water treatment plants (WTP). The chemical reactions are complex and the amount of coagulant dosage required was affected by many factors which makes it difficult to determine the optimal dosage effectively. Additionally, the coagulant process is a typical non-linear, multi-variable, large time-delay, non-stationary, strong coupling and time-varying system. Accurately determining the amount of coagulant added has become one of the most significant challenges. Some studies build a prediction model that only uses current water quality parameters, the previous time sequences were ignored and lacked consideration of multivariate time series and multiple water quality parameters simultaneously, resulting in unsatisfactory prediction accuracy. This study not only takes current water quality parameters into account during the modelling but also considers historical time-series water quality features. We found that the attention-based encoder-decoder of the recurrent neural network framework is an effective model in the area of intelligent water management. In this paper, we studied real-world data with 4 different machine learning models. Compared to the other three potential competitive machine learning algorithm models (random forest, multiple linear regression, and long short term memory), the experiment results demonstrated the best performance for predictive analysis with a highest coefficient of determination (R2) of 0.9908 and lowest values of root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) (1.2524%, 1.1263%, and 1.01%, respectively) in the DA_RNN algorithm. Consequently, this study provides a more reliable and accurate approach for forecasting wastewater coagulation dosage, which is pivotal in terms of the socio-economic aspects of wastewater management.