Enhancing Coagulation Prediction in Water Treatment Using a Similarity Score–Based Piecewise Machine Learning Model

Coagulation is vital in water treatment, but predicting coagulant doses using mechanistic models is challenging due to complex processes. Traditionally, jar testing is used, but it struggles with varying water quality. This study explores machine learning (ML) techniques to predict effluent quality and coagulant requirements. It found that random data splits yield higher similarity scores and better performance than time-based splits. Random forest regressor (RFR) and linear support vector regressor (Linear SVR) performed well in high and low similarity scenarios, respectively, while multilayer perceptron (MLP) showed consistent performance across all conditions. A hybrid ML model combining RFR, Linear SVR, and MLP outperformed individual models, reducing errors by 1%–45%. Converted to ONNX format, the model was integrated into the digital twin of a water purification plant in Texas, improving real-time predictions. This study demonstrates the potential of ML in optimizing water treatment and reducing the need for traditional testing methods.