Lots of places describe themselves as being in “the heart of Europe.” But when it comes to energy distribution systems, only one location fits the bill: Laufenburg, Switzerland, population 3,207. This picturesque town on the shores of the Rhine River has not only been a crossroads for Roman and Napoleonic armies, but is today the meeting point for a very different kind of power: the electric lines that connect France, Germany, Austria, Italy, and of course Switzerland itself. It is also home to the atomic clock that synchronizes and sets the frequency — some would say heartbeat — of the entire European Union grid, including Turkey.
Formula for Grid Stabilization
A new machine-learning solution from Siemens is helping Switzerland’s national grid company to accurately predict losses as energy is transferred from country to country. The result: significant cost reductions and enhanced grid stability throughout Europe.
Balancing Production and Demand
Laufenburg is also where Swissgrid, Switzerland’s national grid company, is headquartered. Because of its strategic location, the company plays a key role in ensuring that the balance between electricity production and consumption across many national boundaries is maintained at all times.
But the ability to meet this goal, and thus maintain a steady frequency of 50 Hertz, has a lot to do with making accurate predictions. On the production side, this is becoming more difficult as an increasing amount of power is produced by renewable sources. And on the consumption side, predictions are also tricky because anything from a sudden cold snap to a heat wave can significantly change electricity demand.
Added to these variables is the problem of predicting so-called “transfer losses” — the amount of electricity lost as a result of line resistance as it zips from one country to another. Italy, for instance, purchases electricity from northern Europe, which crosses Switzerland. Such losses, which are the result of a number of variables, such as local weather conditions and the amount of electricity being transferred at a given moment, “average about 1.6 percent of the total load of the Swiss energy grid, which adds up to about 100 megawatt-hours per hour (MWh),” says Dr. Jan Mrosik, CEO of the Smart Grid Division of Siemens’ Infrastructure & Cities Segment. “At an average spot price of €55 per MWh, the monetary equivalent of these losses is €5,500 per hour, or €48.18 million per year.”
In order to make up for transfer losses, Swissgrid must purchase additional electricity on the Swiss Spot Market, a process that begins up to 16 hours in advance every day of the year. Obviously, given the huge amounts of energy involved, the highest level of predictive accuracy is needed. Until recently, the company relied on an algorithm that has a forecast error of approximately 11 percent. But now, thanks to a neural-network-based algorithm developed by Siemens Corporate Technology (CT), Swissgrid expects to lower its forecast error to ten percent — a ten percent improvement, according to Mrosik. “We can reduce both the use of control energy and the amount of unused energy resulting from overestimating transfer losses,” he says.
Increasingly Accurate Predictions
Unlike any other algorithm in this area, the system developed by Siemens Corporate Technology “derives the predicted transfer loss from the load (electricity usage) forecast — two functions normally considered to be separate — in a single, integrated step,” explains Dr. Ralph Grothmann, who, along with Dr. Hans-Georg Zimmermann, developed the algorithm. “It is an integrated model, and that is not only unique but significantly more accurate than competing models.” Thanks to years of experience in refining their understanding of neural networks, the researchers’ load forecasts are typically 97 percent accurate. This accuracy also derives from the fact that CT’s algorithm takes historical data into account as well as variables such as current weather conditions and the fill levels of pumped water storage systems. Once fully implemented, the new hybrid algorithm could add increased stability to the EU’s electricity grid. “The system’s ability to learn is especially important in terms of adapting the grid to energy flows induced by renewable sources such as wind, solar, and hydro power,” says Mrosik.
What’s more, the algorithm can also be used to forecast other factors, such as European energy flows and the specific amounts of energy expected to be produced by wind and solar installations. “Pilot projects carried out by Siemens have demonstrated that our neural networks can predict renewable energy supplies with an accuracy of 90 percent up to 72 hours into the future,” says Mrosik. “That knowledge could really help grid operators to balance energy flows.”