(339n) Machine Learning Models for Water Use Patterns Analysis in Small Rural Agricultural Communities for Informed Design and Deployment of Membrane-Based Water System

The supply of safe and clean drinking water has been of global concern from a human health perspective for which surface water such as rivers and reservoirs, brackish groundwater are being increasingly utilized. In particular, remote disadvantaged communities in rural agricultural areas are being increasingly confronted with a loss of valuable local potable groundwater resources due to contamination and excessive salinity. In this regard, membrane-based water treatment has been proposed to upgrade these water resources and protect against different contaminants. The design and operation of such systems, however, requires detailed understanding of water use patterns for these communities in order to properly both size the required treatment capacity and its detailed temporal variability. Accordingly, the current study presents a robust machine-learning approach to describing and forecasting water use data in small communities relying on an extensive database of multi-year hourly water use acquired from multiple communities, located in Salinas Valley, California, via installed smart water meters. Autoregressive moving average (ARMA) models were developed from historical water consumption data for the periods of October 2015 – October 2019 to forecast water demand and a detailed assessment of water usage trends at various time scales. The models included input regarding population density, categorical information, and climate metrics (e.g., temperature and rainfall). Confirmatory analysis via self-organizing maps (SOM) clustering approach was then conducted to identify similarities among the remote communities in terms of their water usage at seasonal, monthly, weekly and daily levels. ARMA models, with good predictive performance (i.e., R2 > 0.85), provided accurate description and ability to forecast variations in water consumption. Data analysis revealed similarities among the three studied communities with respect to their water usage at specific times of the day, week, monthly and seasonally. The ability to describe water use patterns via the developed machine learning models provided guidance for the design, deployment, and operational scheduling of the needed water treatment systems in small communities.