AI in water management

AI in water management means training a model on measurements that water authorities already have: water levels, flows, rainfall, lab results and treatment plant process data. You can use it to predict a water level or flow, identify an anomalous measurement or suggest a treatment plant setting. These are among the AI applications we build, often for the public sector.

Water also features on our regional pages: bridges and waterways in Rotterdam, national water authorities in Utrecht, and networks and water in Groningen.

With a model that combines measured flow, rainfall, forecast rainfall and evaporation. Waterschap Limburg has published such a model for the flow of the River Roer in the Algorithm Register: a linear regression that provides hourly forecasts to support decisions. A specialist interprets the result and makes the decision.

Rijkswaterstaat uses Delft-FEWS, a system developed by Deltares, to forecast conditions along the Rhine and Meuse. A custom model can supplement that system where it falls short locally. The challenge is that high water levels are rare, so there are few examples of the situation that matters most. Read more about forecasting.

Usually more than they use: telemetry from pumping stations and weirs, monitoring wells, water quality lab results and treatment plant process data. Through Waterinfo, Rijkswaterstaat publishes current, forecast and historical measurements for national waters, including water levels, wave heights, flow data and water quality.

For data exchange, there is the Aquo standard, maintained by Informatiehuis Water, a partnership between provinces, Rijkswaterstaat and water authorities. Using its definitions makes it easier to combine data series from different databases. To learn where these sources come together, see our data and AI platform page.

A treatment plant collects measurements continuously, making it well suited to a model. For Waternet, we developed a machine learning model to minimise nitrous oxide emissions at a wastewater treatment plant. We are now developing a model to reduce the use of iron chloride in the pretreatment of river water for drinking water.

Water authorities are paying considerable attention to nitrous oxide. STOWA has published a handbook on monitoring nitrous oxide at municipal treatment plants, and the Lachgas Versnellingsprogramma is bringing continuous measurement to more facilities. Those measurements provide the historical data a model can learn from.

By monitoring it. A replacement sensor, a pumping station operating differently or a wetter winter changes the input, making forecasts less accurate. This is called drift. For Rijkswaterstaat, we built a generic package that monitors all its models in production for data drift, prediction drift and outliers.

Decide in advance who receives an alert and who approves a new version. You can read how this works technically on our MLOps page. Management can sit with your team or with us through management and continuous improvement.

The Water Framework Directive is the European directive on clean and healthy surface water and groundwater. Member states agreed to achieve good status by 2027 at the latest, with progress tracked through water quality measurements.

A model does not make water cleaner, but it can help identify an anomalous measurement or show which measure appears to have an effect. Lab samples are only taken occasionally, so the historical record at each monitoring point is often limited. Start by asking which decision about monitoring or measures needs to improve.

Annex III of the AI Act lists AI used as a safety component in the supply of water, gas, heating or electricity as high-risk. According to the European Commission, those obligations will apply from 2 December 2027. Your legal or compliance team must assess whether a model for a pumping station or treatment plant falls within that category; we can help you take stock of your models.

Public authorities publish their algorithms in the Algorithm Register, including the model for the flow of the River Roer. Read about how to do this responsibly on our Responsible AI and AI Act pages.

When a fixed rule in the pumping station controls already works well, when the decision rarely arises or when insights from the measurements are enough. In those cases, a Power BI dashboard using the telemetry is more useful than a model.

If the data series have gaps or no one knows when a sensor was replaced, start with the data foundation. If a simpler solution is a better fit, we will say so.