Machine learning for selecting time series models for drinking water pollutants in Meoqui, Chihuahua, Mexico

Abstract

The objectives were to ascertain potential variations in the levels of As, Pb and F in drinking water in Chihuahua, Mexico, across years, to evaluate time series models to forecast patterns, and to determine whether these levels comply with the Mexican Official Standard (NOM). The models implemented were: ARIMA, Facebook Prophet, GLMNET, and a hybrid model combining Facebook Prophet and an Extreme Gradient Boosting (XGBoost) model. In the context of the ARIMA and Facebook Prophet models, the 'hold-out (data)' method was employed. Conversely, for the GLMNET and hybrid models, a cross-validation procedure with five folds was utilised. Data (from the Junta Municipal de Aguas y Saneamiento – JMAS-) from Meoqui, Chihuahua, Mexico were collated from January 2010 to December 2023. The maximum levels of As stipulated by NOM have been adhered to since 2021. Since 2019, the maximum permissible levels of fluoride (F) have consistently been adhered to and the maximum levels of lead (Pb) have not been exceeded during the period under analysis. Levels of As and F have decreased in recent years, potentially due to corrective measures implemented by the relevant authorities. Projections suggest that the levels of As, Pb and F will remain below the maximum thresholds established by NOM. Facebook Prophet provides the most accurate predictions of the concentrations of As, Pb and F. These findings will provide the competent authorities with the necessary evidence to implement sustainable water resource management practices and guarantee the quality of water consumed by the population.