Water demand forecasting using machine learning approach
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
National University of Lesotho
Abstract
The increasing challenges related to water security, exacerbated by rapid urbanization,
population growth, and climate variability, necessitate accurate and reliable forecasting
methodologies to support sustainable water resources planning. This study explores water
demand forecasting in Ha-Foso, Lesotho, by evaluating three machine learning models:
Multiple Linear Regression (MLR), Support Vector Regression (SVR), and Artificial Neural
Networks (ANN). Utilizing time series data sets covering meteorological inputs (2012-2022),
population, and water consumption records (2017-2024), the study assesses the influence of
climatic and demographic variables, specifically precipitation, maximum and minimum
temperatures, and population, on domestic water consumption.
The research first used MLR to assess the influence of population, maximum temperature,
minimum temperature, precipitation, and other factors on water demand. Subsequently, the
study evaluated the predictive performance of MLR, SVR, and ANN models. Performance
was evaluated using performance metrics, including the coefficient of determination (R2),
Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute
Percentage Error (MAPE).
The regression analysis consistently identified population as the only statistically significant
predictor of water demand (p<0.001), while climatic variables showed no significant
influence during the study period. In the comparative evaluation, the SVR demonstrated thehighest accuracy and generalization capacity, outperforming ANN and MLR, with the least
error metrics in both the training phase and the testing phase. The 2-year forecast highlighted
the distinct behaviours of each model, with the SVR and ANN models providing more
moderate growth projections compared to the steep, linear increase predicted by the MLR
model.
This study presents the potential of machine learning, particularly SVR and ANN, in
addressing the intricate, non-linear relationships inherent in water demand forecasts,
delivering precise and actionable water demand forecasts for peri-urban settings in Lesotho.
The findings suggest the adoption of advanced architectures and the incorporation of socio-
demographic variables to strengthen predictive capacity. The outputs are expected to support
utility companies such as Water and Sewerage Company (WASCO) in strategic planning,
conservation, and infrastructure investment decisions.