Novel methods to spatially downscale water demands

To understand water scarcity, it is vital to have reliable water demand data at suitable spatial and temporal resolutions. While more hydrological models are including human influences, and at increasingly fine resolutions, this improvement in process representation is not matched by improved data for driving or validating the models. Water demand data is normally very difficult to access and, when available, is usually at a coarse spatial resolution (often a country level). Downscaling methods for irrigation demand are well developed but domestic and industrial demands are generally naively downscaled using population as a proxy.

This work explores the potential for Machine Learning models in spatial downscaling of industrial demands at a range of resolutions. Various ensemble-tree type models are presented, trained on a recently published high-resolution water abstraction dataset from England, and using easily accessible spatial datasets as explanatory variables. The results are compared to a population-proxy downscaling and demonstrate minor improvements but without achieving the desired level of skill for application in water resource assessments.

Further avenues for exploration are proposed, with the aim of achieving a transferable downscaling method for water demands which can be trained in data-rich regions and applied to data-scarce areas. A successful approach would enhance water resource modelling through improved driving data, and improve our understanding of water scarcity, to support decision making in water allocation under increasingly water-scarce conditions.