AI-Driven Wastewater Reuse System: Optimizing Urban Water Management

Abstract

Accurate prediction of coagulant dosage is a fundamental aspect in wastewater treatment plants, traditionally managed using methods such as the jar test. These manual approaches have limitations, they are usually time-consuming and lack real-time responsiveness to dynamic changes in raw water quality. This research suggests an AI-driven system strategy for automating and optimizing predictive modeling of coagulant dosage for turbidity removal. Two datasets with different coagulants, PAC (Polyaluminum Chloride) and Alum (Aluminum Sulfate), were utilized, consisting of pH, turbidity, color, and respective dosage values. Various learning approaches were investigated, including Multi-Layer Perceptrons, ensemble learning approaches (stacking, bagging, boosting), and few-shot learning with Large Language Models such as GPT-4o, Gemini 2 Flash, and LLaMA3. Performance was assessed based on the mean squared error (MSE) and the correlation coefficient (r). The findings show that AI models enhance efficiency and accuracy in controlling coagulant dosage. Specifically, for PAC coagulant prediction, the ensemble learning using boosting offers the best performing model with a low MSE (0.01) and high correlation coefficient (0.5431), balancing between predictive performance and small model size, making it optimal for embedded system deployment. On the other hand, for Alum coagulant, ensemble learning using stacking achieved the highest prediction accuracy with low MSE (0.0056) and high correlation coefficient (0.839), but considering model size constraints, boosting remains the best model due to its lightweight size and competitive accuracy.