Optimization of coagulant dosage in WTP using deep learning models with big data

Abstract

The water supply sector has been experiencing a decline in workforce, and 314 water treatment plants (WTPs) of the 483 WTPs in South Korea, which were mostly built during the period of economic development, are now over 20 years old and deteriorating. Furthermore, the decrease in population and the rise in labor costs have led to a continuous reduction in water treatment plant personnel, resulting in a deterioration of operational expertise. Additionally, the deteriorating condition of facilities and the challenges in securing reinvestment for facility and operational improvements have further exacerbated the situation. The increase in WTP accidents can be attributed to the declining expertise of water supply personnel. To effectively respond to global climate change, innovative transformations are required in the water supply sector, including improvements in operational processes.

Following the COVID-19 pandemic, the most critical aspect in the water supply sector is the automation of operations and the enhancement of real-time accident detection capabilities, aiming to replace manual workforce with automation. Future water treatment plants need to implement automated operations using big data and AI technologies.

This study focuses on the automation and optimization of coagulant dosing, which has been a challenging aspect of WTP operation. Coagulant dosing accounts for more than 60% of the chemical treatment costs, and the production costs have been continuously increasing. In the past and present still, coagulant dosing has been carried out by overdosing for stable operation (called sweep floc coagulation method), but complete automation control based on feedback using physicochemical domain knowledge has shown limitations and issues.

Physicochemical models in water treatment, although capable of simulation based on existing domain knowledge, face challenges in real-time automation and optimization. In contrast, deep learning models can achieve complete automation and optimization with using big data. In this study, the acquisition of real-time big data, Prophet outlier removal for time series data, data normalization, and Optuna hyperparameter tuning for deep learning model optimization were researched.

To predict coagulant dosing and sedimentation basin turbidity in water treatment, one-dimensional convolutional neural network (Conv1D) models were established to enhance the correlation analysis between time series data, while gated recurrent unit (GRU) models were developed for time series interpretation. The study presented coagulant dosing predictions (training: 2016-2019, prediction: 2020) using five years of real-time big data and compared and analyzed the accuracy and characteristics of the Conv1D and GRU models against physicochemical models through verification of sedimentation basin turbidity prediction.

Lastly, to optimize coagulant dosing, the predicted dosage of coagulant was determined while maintaining sedimentation basin turbidity below 1.0 NTU. For the low turbidity season, the predicted dosage of coagulant obtained through deep learning was gradually reduced by 5% to 20% increments, and the sedimentation basin turbidity was predicted. By considering the results for both the high turbidity and low turbidity seasons, the final optimization of coagulant dosage was achieved. Furthermore, the reduction in coagulant dosage was used to calculate the cost reduction in chemical production of WTP, resulting in an estimated 12% reduction in production chemical unit cost.