Machine Learning for Modeling Water Demand

Abstract

This work shows the application of machine learning (ML) methods to the modeling of water demand for the first time. Classification and regression trees (CART) and random forest (RF), a multivariate, spatially nonstationary and nonlinear ML approach, were used to build a predictive model of water demand in the city of Seville, Spain, at the census tract level. Regression trees (RT) allowed estimation of water demand with an error of 22 L/day/inhabitant and determination of the main driving variables. RF allowed estimation of water demand with error values ranging from 18.89 to 26.91 L/day/inhabitant. The RF method provided better predictions; however, the RT model facilitated better understanding of water demand. This research shows an alternative to the hitherto applied cluster and linear regression approaches for modeling water demand and paves the way for a new set of further scientific investigations based on ML methods.