�����˿�����

If a pump and a valve that are located close to each other in a power station but have no direct functional connection happen to fail at the same time, that might not be pure coincidence. Something serious could have happened that the local monitoring systems didn’t detect. Water damage caused by an inconspicuous leak from a pipe, perhaps, or a fire not yet picked up by the smoke detectors on-site. An AI system would think so, and it would alert the standby team accordingly – as long as it had been previously trained using a knowledge graph of the power station. A knowledge graph is a special database capable of being evaluated by a machine. It contains not just the properties of the power station components but also all the connections between them, including those that are non-functional. For the pump in the example, that includes knowledge of all the other components close-by.

 views knowledge graphs as a key prerequisite turning the vision of an autopilot for complex large-scale power stations into reality. The controller needed for the purpose requires artificial intelligence. Unlike with humans, however, it’s difficult to train an AI system on an entire power station all at once. “It’s easier to split the power station up into sections using knowledge graphs and make each portion machine-usable,” says Weustink, who develops strategies and technologies for digitalizing energy systems at �����˿�����. “These sections can then be controlled via artificial intelligence systems that are networked together.”

Knowledge graph technology hasn’t yet reached the point where it can enable autopilots to be created for entire power stations. Generating the graphs is still a very labor-intensive process, and some essential tools for querying the knowledge from the graphs still need to be developed. There are some gaps in the development stages that Weustink can’t fix by himself. Thankfully that’s not necessary, because knowledge graphs in power stations aren’t just relevant to autopilots; there are other potential applications like optimizing the engineering of new power stations and servicing. So Weustink has no shortage of allies at �����˿����� who share his eagerness to make knowledge graphs usable for power station applications.

Data streams between functional components and areas in a power plant represented using knowledge graphs. The level of complexity equals that of the neural network in the human brain. The segment of a combined cycle power plant shown here comprises approximately 10,000 components and more than 50,000 connections. Representing entire power stations quickly increases this to several million components and an even larger number of connections.

 is one of these enthusiastic colleagues. An engineer specializing in energy and environmental engineering, she shares Weustink’s preference for complex digitalization challenges. In recent years, she and her team have developed a data integration system that builds on knowledge graphs. It brings together data from a number of sources, enriches it with meaning, and links it with data from other sources that shares the same meaning. It also harmonizes and simplifies access to and views of this data. The data is therefore accessible to humans via free-text searches like those used by Google and also for software programs to perform automated analyses, for example.

�����˿����� now has a database for in-house use that brings together the data from 50 power station projects based on 12 data sources using the data integration system. “The knowledge graph used here currently contains about half a billion data points,” Soller explains. The database now has more than 1,000 users who are happy to report that, thanks to the integrated view of the data, they get answers to complex questions much more quickly than from previous databases. Weustink is an especially big fan: “The database is really fantastic!” He also wants to access this database with his autopilot and incorporate the data generated at his end.

The data integration system has also recently started being used in a customer project that involves networked construction planning for a new power station construction project in Hong Kong. The data integration system lays the groundwork for utilizing software-based Building Information Modeling (BIM), not only to create the building envelope but also for laying out, installing, and commissioning the entire interior – in other words, the actual power station system. Customers can then monitor the progress of their new building project with a level of accuracy that was previously impossible with facilities of this size,” says Soller. “This makes it much easier for them to ensure their projects remain on schedule and within budget.”

Weustink still needs a little patience before he can turn his vision of the autopilot into a reality using the knowledge graphs. It’s already possible to build an AI system that’s capable of assigning the simultaneous failure of a pump and a valve to a local fault using the data integration system. However, Weustink notes: “That still isn’t enough for the autopilot. We need to add more information to the knowledge graphs: for example, from the control system, the electrical and pipeline installation, and maintenance records.” 

The expert teams at �����˿����� are putting all their efforts into overcoming this hurdle, in addition to minimizing the work needed to generate the graphs and develop the query tools. “We can’t overlook the fact that stages like the data integration system are already generating many benefits, both for constructing and for operating power stations. In some cases, the benefits can be even greater than those offered by the autopilot,” emphasizes Weustink.