AI consultant in Groningen

Not on a case study we can share. Our water and energy case studies come from elsewhere, including the Datalab team at Waternet, which supplies drinking water in and around Amsterdam. What we learn there about models in production is not tied to one region.

Our only office is on Westerdoksdijk in Amsterdam. To meet, we can come to you, you can visit us, or we can speak online.

Making better use of available capacity is one of the approaches in the province of Groningen’s grid congestion programme, alongside measures such as energy hubs and smart charging. A forecast of consumption and generation for each grid connection shows when you will stay within your contracted grid capacity and which demand can be shifted.

To do that, you need sufficiently detailed measurement data for each connection. If you share capacity with other businesses in an energy hub, you also need to address shared data and who manages it: data governance.

Possibly. The AI Act classifies AI systems as high-risk if they are intended as safety components in the supply of water, gas, heating or electricity (Annex III, point 2). The question is whether a failure could make that supply unsafe: a model that controls a valve is more likely to raise this question than one that proposes a schedule.

Following the digital omnibus, the European Commission says the Annex III rules will apply from 2 December 2027. Your own legal counsel or compliance department must assess how the law applies to your use case. We can help prepare an inventory and risk classification, and set up AI governance. You can find an outline of the legislation under AI Act.

Usually three sources: sensor time series from SCADA, the system used to monitor and control equipment remotely; fault and maintenance history from the maintenance management system; and reference data for each asset, such as its type, year of construction and location. Linking them is often the hardest part: a replacement pump does not always get a new identifier in the measurement data.

Having many comparable assets under one operator helps, because a model can then compare patterns across assets. From Groningen, the Noorderzijlvest water authority manages locks, weirs, pumping stations, flood defence closures and treatment plants, for example. We describe the approach on our predictive maintenance page.

Since 1 January 2026, non-residential buildings with a heating or air-conditioning system rated above 290 kW have been required to have a building automation and control system (GACS). According to RVO, the threshold will fall to 70 kW from 2030. Such a system must continuously track and analyse energy consumption and detect losses in efficiency.

That measurement data is also useful for energy management: identifying anomalies in each building and deciding which building to make more sustainable first. Where the grid is congested, reducing peak demand also frees up capacity. We discussed how to move from compliance to energy management in the webinar Meer dan GACS.

European, national and regional funding, including from Nij Begun, will support a Dutch AI factory in Groningen: an expertise centre and a supercomputer that, according to the Dutch government, will be fully available in early 2028. The government lists energy networks as one of its applications. Its waste heat could eventually be used by WarmteStad to heat homes in north-west Groningen.

For most applications in grid management, water and buildings, computing power is not the constraint. The challenges are reliable measurement data, an owner for each model and management after go-live. You can read about our work in this sector on our data and AI for the energy sector page.

With MLOps: a way of working that lets you version, test, deploy and monitor models. A model registry and CI/CD templates give every subsequent model the same route into production. This is how the Datalab team at Waternet has set things up.

In the energy sector, data often changes because the physical world changes: a sensor is replaced, a new pump is installed or a winter is particularly cold. Drift monitoring shows when this causes a model’s performance to change and retraining is needed. We describe this approach under MLOps.

When a fixed control setting can solve the problem, such as a switching time, a heating curve or an alarm threshold. Or when the measurement data is not yet coming in reliably: connecting it to a dashboard is then the first step, and that alone often shows you where things are going wrong.

If you already have an application in mind and only need someone to make sensor data from SCADA and building management systems accessible, consider a data engineer on secondment through our secondment service.