Prediction of freshwater supplies in the Netherlands

In this project we intend to use hydroinformatics methods – notably machine learning and deep learning – to predict available freshwater supplies, particularly during periods of serious drought. On the basis of historical data, we will study how variations in surface- and groundwater are impacted over time by meteo-hydrological parameters and human activities. A prediction model will be developed to indicate the future availability – for example in 2030 or 2050 – of freshwater supplies and the impact that water management has on them. With these predictions, the water sector will be able to develop strategies for its planning and management activities.

Insight into available freshwater supply

Following three years in a row of extreme drought (2018-2020) in the Netherlands, it is important to investigate how much surface- and groundwater is available for nature, agriculture, industry and the drinking water provision. Especially in the event of water stress, the drinking water utilities, Water Authorities and provinces need to have a clear and comprehensive view of available supplies of freshwater water, so that this water can be optimally distributed. It is vitally important to know how much freshwater can be delivered to different sectors, without there being any damage to their functioning and production as a result of water shortage. Several studies have moreover shown that climate change will aggravate the extreme drought periods, both in terms of their duration and their severity. We therefore not only need to study the availability of freshwater under the current situation, but also produce a reliable estimate for the future, for example, for 2030 and 2050.

Predicting with machine learning

To study the availability of freshwater supplies one can use a combined meteorological and hydrological model that calculates freshwater volumes. However, this model is less suitable for a long-term prediction because the simulation error can accumulate over a long period. This is where data science enters the picture. With data science it is possible to derive a model from the data themselves, so that the error is only dependent on observations and does not accumulate over the period of time. Models built on the basis of data science can clarify how the trend of historical data is determined by different meteo-hydrological factors. The models can then predict future trends on the basis of these anticipated factors. Data science is a typical hydroinformatics method, in which machine learning and deep learning are used to investigate data from multiple perspectives.

Better planning and management decisions

This case will give us an in-depth understanding of the freshwater supplies in the Province of Drenthe. It will allow for answers to questions like: What is the impact of meteorology, hydrology and people on freshwater availability? How does the availability vary under the future scenarios? How frequently will water stress occur in the future? What is the probability that water stress will occur three years in a row? How are the developed models limited to the information? How can it be improved in the future? The answers to these questions will provide useful information to the province, Water Authorities, water utilities and KWR. Confidence with regard to the expected available freshwater supplies will enable better decision-making for the planning and management of the water system.