Comparison of machine learning algorithms to predict daily water consumptions

Abstract

This paper focuses on a comparison of machine learning algorithms for predicting the cumulative daily water consumption. The data are collected from an internet-based platform that provides usable data. A pre-processing has been designed for checking the integrity of data, i.e., detecting missing data and abnormal consumptions. In order to optimize the water uses in distribution networks, monitoring and forecasting consumption are good solutions. Five models, namely the Polynomial Regression (PR), Nonlinear AutoRegressive (NAR), Support Vector Regression (SVR), MultiLayer Perceptron (MLP) and Long Short-Term Memory (LSTM) are designed and compared to find the most accurate solution to forecast daily water consumption. The performance of these models is based on the Root Mean Square Error (RMSE) calculated from desired values. The water consumption for the next five days is predicted with no prior information but only centralized past measurements. Results show a predicting precision with NAR of about 5 and 23 l/day in respectively domestic and industrial installations where up to 1500 and 2700 l/day can be used.