A large-scale evaluation of machine learning algorithms in mid-term water demand forecasting

Abstract

ABSTRACT Water utilities currently face challenges in accurately measuring user-end consumption as they have to visit every conventional meter that is installed in their service area. This study explores the use of accessible machine learning models to forecast water demand at two crucial temporal scales, addressing practical needs such as customer billing. To draw meaningful conclusions, these models were tested and trained from actual user-end water consumption data from over 2.1 million customers over a 10-year period. These data were provided by the Water and Sewage Company of Greece (EYDAP). The large-scale experiment included the use of statistical models (ARIMA and SARIMA), deep learning [long short-term memory (LSTM)], and clustering techniques (k-NN algorithm). The model with the best performance was ARIMA, outperforming all the other models including the more complex ones. The LSTM model did not perform as expected in predicting water consumption, it excelled only in predicting the total amount of water consumed. These forecasts can be valuable tools for water utility companies, aiding in tasks such as customer billing and water balance calculations. This paper covers all the necessary steps that must be taken to achieve meaningful user-end water consumption forecasts, from raw data preprocessing to hyperparameter tuning for machine learning algorithms.