Context

This project allows to forecast drinking water demand for one week at an hourly scale for the Bruson-Le Sappey test area

Results

The best forecasts were obtained with LightGBM, which proved to be fast and easy to use. After training on data from January 2022 to October 2024, the predictions were tested at the end of 2024, yielding a mean relative error of 18.3% for November and 21.9% for December, compared to an average hourly consumption close to 0.015 m³.

It is interesting to note that demand is influenced by several factors, including the hour, day of the week, and month of the year, while temperature has little importance. Moreover, the predictions remain robust even at year-end, despite the fact that this holiday period differs from the rest of the year.

The method can be quickly tested on other sectors: in particular, it would be interesting to transfer it to an area with more water meters and a greater tourist influence.