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

Assessing the impact of climate change on water demand is a challenging task. This paper proposes a novel methodology that quantifies this impact by establishing a link between water demand and weather based on climate change scenarios, via Coupled General Circulation Models. These models simulate the response of the global climate system to increasing greenhouse gas concentrations by reproducing atmospheric and ocean processes. In order to establish the link between water demand and weather, Random Forest models based on weather variables were used. This methodology was applied to a district metered area in Naples (Italy). Results demonstrate that the total district water demand may increase by 9–10% during the weeks with the highest temperatures. Furthermore, results show that the increase in water demand changes depending on the social characteristics of the users. The water demand of employed users with high education may increase by 13–15% when the highest temperatures occur. These increases can seriously affect the capacity and operation of existing water systems.

Understanding, comparing, and accurately predicting water demand at different spatial scales is an important goal that will allow effective targeting of the appropriate operational and conservation efforts under an uncertain future. This study uses data relating to water consumption available at the household level, as well as postcode locations, household characteristics, and weather data in order to identify the relationships between spatial scale, influencing factors, and forecasting accuracy. For this purpose, a Gradient Boosting Machine (GBM) is used to predict water demand 1–7 days into the future. Results show an exponential decay in prediction accuracy from a Mean Absolute Percentage Error (MAPE) of 3.2% to 17%, for a reduction in group size from 600 to 5 households. Adding explanatory variables to the forecasting model reduces the MAPE up to 20% for the peak days and smaller household groups (20–56 households), whereas for larger aggregations of properties (100–804 households), the range of improvement is much smaller (up to 1.2%). Results also show that certain types of input variables (past consumption and household characteristics) become more important for smaller aggregations of properties, whereas others (weather data) become less important.

Water demand forecasting is an essential task for water utilities, with increasing importance due to future societal and environmental changes. This paper suggests a new methodology for water demand forecasting, based on model stacking and bias correction that predicts daily demands for groups of ~120 properties. This methodology is compared to a number of models (Artificial Neural Networks–ANNs, Generalised Linear Models–GLMs, Random Forests–RFs, Gradient Boosting Machines–GBMs, Extreme Gradient Boosting–XGBoost, and Deep Neural Networks–DNNs), using real consumption data from the UK, collected at 15–30 minute intervals from 1,793 properties. Results show that the newly proposed stacked model that comprises of RFs, GBMs, DNNs, and GLMs consistently outperformed other water demand forecasting techniques (peak R² = 74.1%). The stacked model’s accuracy on peak consumption days further improved by applying a bias correction method on the model’s output.