نمذجة وتوقع الكلور المتبقي لمياه الشرب باستخدام الذكاء الاصطناعي في مدينة كربلاء

Modelling and predicting of Residual Chlorine for drinking water using AI in karbala city

Abstract
Ensuring the safety of drinking water is vital for public health, with effective disinfection playing a central role. Chlorination is widely applied, as it provides a residual protective effect within the distribution network. However, maintaining an optimal residual level is critical: low values reduce disinfection efficiency, while excessive levels accelerate pipe corrosion and pose health risks. This study investigated water quality at the central water plant in Karbala (Al-Hussein neighborhood), where chlorine is used for disinfection. To address this, thirty AI models were trained using five algorithms, and Compliance with the World Health Organization (2017) guidelines and the Iraqi drinking water standard No. 417 (2009) was assessed. Physical and chemical analyses were conducted at the plant and along two main pipelines. Results showed that most parameters were within acceptable limits, except turbidity and residual chlorine, related to insufficient turbidity treatment at the plant. Such increases might negatively affect the condition of the pipes within the network and inhibit the efficacy of the disinfectant. As for the residual chlorine percentage, it has consistently exceeded the permissible limits at almost all points. This offers a cumulative harm to customers while causing damage to the pipes through its interaction with the pipe walls. And these results validate the necessity for the plant to set up a monitoring system that can control and make proactive decisions about chlorine levels at specified monitoring locations. In terms of modeling, the algorithms that were used are : Random Forest (RF), Extreme Gradient Boosting (XGBoost), General Regression Neural Network (GRNN), Multi Layaer Perceptron (MLP), and Transformer. By training each algorithm on three different scenarios by using Python based on the inputs.
Because the R² values were almost similar and high in all scenarios, the Mean Squared Error (MSE) was employed as the primary metric for the final evaluation and comparison of the models. The study results demonstrated that the Transformer achieved MSE values ranging from 0.053 to 0.101 across both directions and scenarios, while the GRNN model exhibited a range of 0.009 to 0.219. Both models demonstrated excellent performance and closely competed for the first rank. The second rank was attained by the XGBoost model with an MSE range of 0.102 to 0.335, followed by the RF with values between 0.197 and 0.595. The MLP ranked last, recording the highest error values (0.117–0.625), reflecting the weakest and least stable performance across the different scenarios. This study concluded that the process of water purification quality in the plant needs to improve, and smart AI models can be applied as a supportive tool in the proactive prediction of residual chlorine to avoid any human errors during the application of chlorine dosage.