Short-Term Forecasting of Household Water Demand in the UK Using an Interpretable Machine Learning Approach

This study utilizes a rich UK data set of smart demand metering data, household characteristics, and weather to develop forecasting methodology that combines the high accuracy machine learning models with interpretability statistical methods. For this reason, random forest model is used predict daily demands 1 day ahead for groups properties (mean 3.8 households/group) homogenous characteristics. A variety interpretable techniques [variable permutation, accumulated local effects (ALE) plots, individual conditional expectation (ICE) curves] are quantify influence these predictors (temporal, weather, characteristics) on water consumption. Results show when past consumption available, they most important explanatory factor. However, not, combination temporal characteristics can be produce credible similar accuracy. Weather input has overall mild no effect model’s output, although become significant under certain conditions.