Abstract:
Water is one of the fundamental necessities of human life. Ensuring clean and safe water for
communities is essential. The main steps in the water treatment process include coagulation
and flocculation, sedimentation, filtration, and disinfection. This study focuses on introducing a
machine learning-based approach to optimise coagulant dosage in the water treatment process
at the Demodara Water Treatment Plant. Currently, the traditional jar test is used to determine
coagulant dosage; however, it is time-consuming and labour-intensive. In this study, historical
jar test data from 2019 to 2025 were collected and preprocessed to develop predictive models.
Water quality parameters such as pH, turbidity, and colour were applied to six machine learning
models, such as Random Forest, XGBoost, LightGBM, CatBoost, Artificial Neural Networks
(ANN) and K-Nearest Neighbours (KNN) to predict the optimal dosage of Poly Aluminium
Chloride (PAC), the coagulant used at the Demodara Water Treatment Plant. All models were
trained using dedicated training and testing datasets. Hyperparameter tuning was conducted
using grid search with cross-validation, and model performance was assessed using mean absolute
error (MAE), mean squared error (MSE), and the coefficient of determination (R2). Among
these models, the highest predictive accuracy was achieved in Random Forest. The R2 values
resulting from the Random Forest were 0.93, 0.89, and 0.93 for pH, turbidity, and colour, respectively.
Corresponding MAEs were 0.05 for pH, 2.93 for turbidity, and 11.04 for colour,
while MSEs were 0.01, 25.94, and 416.29, respectively. The best-performing model was integrated
into a graphical user interface to support real-time plant operations. This machine
learning-based approach enhances consistency in treatment quality, reduces chemical waste,
and minimises human intervention. It will provide operators with an interpretable, convenient,
and scalable solution to improve the efficiency and reliability of the water treatment process.
Future work will focus on incorporating real-time sensor data to enable dynamic dosage adjustments,
further increasing the practicality of ML integration in real-world settings.
