Prediction of multi-sectoral water withdrawals through time using machine learning methods
Abstract
Around the globe, growing population and demand for water create a serious risk to long-term sustainability. Effective planning and management of water supplies requires accurate predictions and forecasts of water use under different conditions. However, the complex array of factors that influence water use (including socio-economic, technical, regulatory, and climatic conditions) make the development of predictive models highly challenging. Machine-learning models have the ability to capture these complex, nonlinear relationships, but their application has typically been limited to municipal water demand forecasting, excluding other uses such as industrial and agricultural withdrawals. This work leverages a dataset of over 1.35 million monthly observations of water withdrawal for over 4,000 self-supplied facilities in Virginia dating back to 1982. This data is used to develop and test multiple machine-learning models for water withdrawals in the agricultural, aquaculture, commercial, energy, industrial and municipal sectors based on climatic and socio-economic predictor variables. In addition to standard machine-learning methods, it also demonstrates a novel approach to machine-learning in panel datasets (multiple users over time) based on a generalized version of hierarchical regression. These methods are shown to provide more accurate predictions of water withdrawals then classic statistical models as well as methodologies commonly applied in practice. Additionally, variable importance diagnostics can be used to characterize nonlinear, interdependent relationships between water withdrawals and predictor variables.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H23F..04S