The wind speed is 50 km/h. While conditions are ideal at EnBW’s onshore wind farms in Langenburg and Winterbach in the state of Baden-Württemberg, a heavy data storm is raging on the screens of Meik Schlechtingen and Josef Feigl in Hamburg, around 600 and 700 kilometers away respectively. This is where the information from the wind turbines converges, one stream of data after another. Each turbine contains up to 1,500 sensors – and they report hundreds of thousands of values per minute. The pair are working with artificial intelligence (AI) to subdue the data storm and thus enable them to detect faults and anomalies in the turbines at an early stage. By doing so, they are already saving EnBW around ten million euros per year.
The needle in the haystack
Josef Feigl sits back in his office chair. The doctor of business mathematics is fascinated by the technology. Each of the rotor blades is around 60 meters long and weighs around 25 metric tons. All the same, even a relatively weak wind speed of 12 km/h could drive the turbine at a height of 150 meters. To his left sits his colleague Meik Schlechtingen. He recalls: “Until around 2010, the maintenance of wind turbines still worked largely on a reactive basis. Back then, data analysis was like looking for a needle in a haystack. Analysts used to search manually for abnormalities. If there was any indication, an engineer drove to the wind turbine and looked into it on-site. It was like looking for a needle in a haystack.”
Machine learning – training for the AI
The two men do things differently nowadays. They develop models and train the AI to detect and report any deviations from the norm – this is known as machine learning. And it always works particularly well when it involves very large and complex amounts of data. “Where humans would no longer see the wood for the trees, AI welcomes the amount and complexity of the data,” smiles Feigl. That’s because the more precise the underlying information, the more reliably the AI can tell whether there is a fault.
How does AI work and how does it help people?
“Listening in” on wind turbines using artificial intelligence
Plenty of data – the two men can work with this. Every ten minutes, each wind turbine sends its operating data. This includes data on the temperature of bearings and gear oil or the speed of the generator. Plus there is information sent from other sensors. “We use it all to listen in on the system – almost in the same way as a doctor listens to the chest with a stethoscope. It enables us to detect hairline cracks in gears at an early stage, for instance,” explains Feigl. The real challenge, however, lies not in eavesdropping, but in understanding the data, locating the fault and estimating how serious the problem is. “Only humans can do that – which is why we look at any unusual data together with engineers and technical experts. Then we send out a service team, if necessary,” adds Schlechtingen. The experts have to use different procedures, depending on the type and amount of data. With the tool used to measure the wind speed, known as the anemometer, the matter is relatively straightforward, says Feigl. The information from two or three sensors would be sufficient to analyze the fault situation here. “Anemometers are important. There are always two per wind turbine so that the turbine can safely switch on and off in cases where, for example, there is too much wind. If the two anemometers record different values, we look at that.” For other error messages, the data situation is much more complex.
Live first aid
Yet the two men and their team intervene if necessary, even before the engineers are on-site. If, for example, they detect that a serious fault is imminent and there is risk of a complete system failure, they restrict the wind turbine’s output. This protects the damaged component and gives the service team the necessary time to replace it, enabling them to cut costs and improve the availability of renewable energy.
At the heart of the data flow
Yet how do the two men read that from the thousands of pieces of information? Schlechtingen points to the computer and explains the data flow that streams onto the screen: “The sensors send us information. We correlate this with other parameters that we simultaneously record: outside temperature, wind speed, rotational speed, output and so on. We feed the model with these values – and it then learns the dependencies between the individual measured parameters. This is how the model simulates a certain normal behavior.” If the data deviates from the norm, the system notifies the specialists and they carefully examine the values.
450 wind turbines are monitored with the aid of AI
Schlechtingen, Feigl and their colleagues now work with over 22,000 models that can be trained in seconds. The EnBW team currently monitors 450 wind turbines with the aid of AI. Why would they need so many models for considerably fewer wind turbines? Feigl nods. He is used to the question: “We operate many different wind turbines, made by different manufacturers and fitted with different components. Furthermore, they are all located in different places and are thus exposed to different environmental conditions. Every wind turbine is ultimately unique.” They need several models to cover all of them.
Thousands of computers at the touch of a button
In addition, the wind and weather are generally capricious. New wind turbines must therefore firstly go through a complete annual cycle and various operating states before the models are able to provide precise data. Only then can a model filter out any seasonal fluctuations, for example. Technical progress and increasing computing power help the two of them above all else: “When I first started, we had a powerful computer in my office that processed data day and night,” recalls Schlechtingen. “These days, at the touch of a button, we can buy the computing power of hundreds or thousands of machines that calculate things for us.”
The limits of AI
Ultimately, however, the amount and quality of data is always decisive. “It works like a mathematical equation,” says Feigl. “If you miscalculate at the start, the overall result will not be right.” This is what the experts call garbage in, garbage out. In other words, if the AI is fed with incorrect information, it will also produce incorrect results. Another problem with AI is that even though it may be able to help with large and complex amounts of data, it is unable to abstract information. “This means that our tens of thousands of models are specialists. Each one is familiar with a small part of the wind turbine. For example, one model will detect damage to the measurement chain, while another will detect problems with the anemometers – but no model can do both. One of our aims for the future is therefore to get the many different models to work better together,” says Feigl. And Schlechtingen adds: “It would help if the AI could precisely pinpoint the abnormalities. We are working on that.”
Both agree: Development in the field of AI research will make quantum leaps in the near future.