Machine learning and ant-based strategies

Improved digital twins are enabling precise performance forecasts in transformer design.

Thanks to simulations using especially realistic digital twins, transformers will be able to meet customer requirements even faster and with greater precision in the future. Two researchers from Siemens Corporate Technology have developed new approaches in this area. Machine learning and strategies borrowed from ants are taking digital twins closer and closer to the real thing.


By Frank Krull 

Laws of nature, design data, material values, functional workflows: These are the ingredients needed to create digital twins of real technical systems. However, they are “not enough to produce a truly accurate representation,” say Siemens Corporate Technology researchers Denis Smirnov and Yayun Zhou. “This isn’t enough for many simulation tasks.” That’s why they are currently working on ways to make digital twins – with which the properties of transformers are simulated during the design phase – even more realistic using principles of machine learning and special optimization methods. 

Faster and with greater precision

Kerwin Stretch, who heads the Global Technology Center for Distribution Transformers within Smart Infrastructure from Siemens would like to use this methodology to meet customer requirements even faster and with greater precision. “The performance parameters that a transformer must meet vary tremendously depending on the use case and place of deployment,” explains Stretch. “We’re talking about values such as the maximum temperature that cannot be exceeded without damaging the transformer, or the maximum power losses that can occur when idling, that is, when no power consumer is connected. Until now, it has required a great deal of time and expertise to come up with the best design solution, since all parameters are interrelated. A rapid simulation with realistic digital twins would significantly speed up the process.” 

Learning from comparative values

Smirnov, an expert in machine learning, found that distribution transformers, in particular, provided a very favorable starting point for the use of his tools. “For years, we have been recording the discrepancies between the simulated and the real performance values for each transformer here,” Smirnov explains. “The result is a huge database containing tens of thousands of comparative values that we can include in the make-up of the twin using machine learning methods.” What happens in this process differs little in essence from the way people would build on their own experiences based on comparing values to refine their concept of the twin. But a computer performs the process so much faster, which means that highly complex connections can be discovered which a human would never recognize. 

Much better hit rate

Smirnov and Stretch have already thoroughly tested the twin, making improvements using machine learning, to calculate simulations of no-load losses. “We found that the calculations using the new twin deliver much better results than those with the twin we used previously,” Smirnov recounts. “Not only were the new simulation results closer to the real values in 88 percent of cases for the 7,000 transformers for which we performed the calculations, but the discrepancies between the new simulated values and the real values were reduced by half compared to the earlier simulations.” 

No need for hardware upgrades

The new simulation has now been incorporated into the simulation software currently being used to design distribution transformers. Both calculations for no-load losses are now displayed for the designers. “With no need for hardware upgrades,” Smirnov emphasizes. “The existing computers acquire the machine learning function through a training process. Although that costs you time, the subsequent calculations will take just a few milliseconds.” “If we consider how much faster and more accurately customer requests can be fulfilled in this way, the cost is actually extraordinarily low,” observes a satisfied Stretch. He thus very much hopes that he can use simulation with machine learning to calculate further parameters in the near future. 

Twins beyond compare

Even so, databases containing tens of thousands of comparative values for simulated and real performance parameters are available for just a small number of transformers. Industrial transformers aren’t included, for instance, since only a few of these are manufactured each year. But Zhou, an expert in mathematical optimization at Corporate Technology, has nevertheless managed to find a way to make the twins more realistic in the case of these giants, which can often be more than ten meters high and weigh more than 500 metric tons. She started with twins with model descriptions that included variable parameters. For example, the core dimensions, the type and number of windings, and other parameters can be freely chosen within certain limits. 

Optimization, ant-style

For these twins, Zhou and her co-workers Harald Held and Meinhard Paffrath developed a method that determines the best combination of parameters for the required performance values using a mixture of simulation and optimization. Because it would take too long to simulate all conceivable combinations and compare them against each other, the optimization procedure used is based on the way colonies of ants find the shortest route between their nest and a sizeable food source. Zhou explains the principle: “The ants use a scent to mark their path. Because the ants using the shortest path cross it more often and therefore mark it more regularly than those on the longer paths, over time the shortest route will become more strongly marked than all others.”

Sharing the benefits

Smirnov has already expressed interest in extending the use of Zhou’s process to include his distribution transformers. “It’s a pity that I can’t offer my process for industrial transformers in return,” he says regretfully. “But there’s nothing we can do about it as long as the amount of values available for comparison is insufficient.” Zhou is already looking forward to working with Smirnov to test whether his twins can be refined a little bit further using her procedure. She is very confident that they can.

5. Aug. 2019

Frank Krull is a physicist and journalist and works in the communications department of Siemens.

Subscribe to our Newsletter

Stay up to date at all times: everything you need to know about electrification, automation, and digitalization.