In the age of climate change, energy must be used more efficiently, and utilities have to invest in advance innovations. In this case E.ON developed a sustainable district heating and cooling service called E.ON ectogrid™. E.ON ectogridTM maximizes the reuse of energy within the system and drastically reduces the energy supplied by up to 75 percent. The dynamic controls of E.ON ectocloud™, the digital platform that handles and controls the energy system in buildings and facilities of different sizes, are based on machine learning (ML) and Internet of Things (IoT). E.ON relies on Azure Machine Learning, Azure Data Factory, and Azure DevOps for the orchestration to ensure that it can prepare the ML models for international rollout as quickly as possible and develop and refine them at any time as needed.
The challenge: Scale smart grid controls internationally
In today’s world, waste heat should no longer go unused. That’s why energy supply is moving toward heat pumps and chillers, which can generate heating or cooling for buildings from low temperature excess energy. With a model project in Lund, Sweden, E.ON is taking this even further. In its low-temperature E.ON ectogrid™, a network of heat pumps and chillers produce all heating and cooling for the buildings. Each building sends its excess heat or cold to the grid and the balancing unit so that other buildings can use this excess when needed. Through this shared energy use and storage with dynamic thermal balancing, E.ON ectogrid™ is aided by the intelligent controls in E.ON ectocloud™, optimizing the delivery of energy, and minimizes operating costs, as well as environmental impact.This concept places high demands on IT: around 10,000 sensors monitor the buildings, generating some 2.5 million readings per day. The system also integrates weather forecast and energy market data. Continuously adapting heat use requires sophisticated detection, optimization, and forecasting algorithms: “We collect data in near real time, and we have to assign the buildings a score every 30 minutes,” says Mayur Sand, Digital Product & Transformation Manager and Team Lead ML at E.ON. “It’s very important for us to be able to process all the complex thermodynamic calculations in parallel, as there are forecasts both at the building level and at the overall level, and the weather changes all the time. This means we have to respond as quickly as possible to give the systems the correct parameters.” To do this, E.ON uses E.ON ectocloudTM, a cloud- and ML-based control infrastructure developed specifically for this system.
In 2020, E.ON began marketing the now proven E.ON ectogrid™ internationally: “In the next five years, we’ll be seeing E.ON ectogrid™ in a number of different regions: in Milan, in Amsterdam, and in Poland, France, and Germany,” Sand says. One challenge the company will face, according to Henrik Karström, Product Owner E.ON ectocloudTM, will be adapting the ML models to new local conditions: “In Milan, we can use waste heat from the sewer system,” he says. “In Poland, there’s a factory that operates several hours a day. We have waste heat during those hours, but none at night.” This is where the processes for developing the custom ML models required reach their limits. Initially, E.ON’s data scientists trained the ML models on local devices. “We saw a huge gap between the results our data scientists were getting and the implementation of those results in production,” says Sam Julian, Head of Data Engineering & AI Solutions at E.ON. They came to a realization: “We can’t offer a scalable solution in the cloud using local methods and techniques,” Julian says. “To scale the solution while also adapting it to fit differing needs, we had to completely change our thinking.”
The solution: End-to-end DevOps processes with cloud-native Azure services
So the project team decided to switch the ML DevOps processes to a uniform cloud-native basis. Their data science colleagues initially weren’t very enthusiastic about this, preferring to continue to their work in the familiar way. That’s when Sand turned to Microsoft: “We conducted a proof of concept with Microsoft and presented it to the data scientists and engineers. After all, we needed everyone’s support. It was a very intense debate.” Microsoft assembled an interdisciplinary team, including experts from Seattle, and, as part of its Accelerator program, explained to the E.ON team how they could make the most effective use of Microsoft services—Azure Machine Learning, Azure Data Factory, and numerous others. The project team used Microsoft’s reference architecture, together with a solution accelerator that was developed specifically for the “many models” scenario, and customized it for E.ON’s specifications. In this way, they succeeded in convincing their colleagues: “We had lengthy discussions, but now they are major proponents of this approach,” Sand says.
After all, the cloud-native environment offers significant advantages: “The nice thing about the Azure Machine Learning stack is that we have a coherent workspace for data scientists or data engineers as well as for the DevOps pipelines,” Julian says. “For the ML training, we use the parallel setup and get the result immediately, with no system administration whatsoever. With Azure ML, the data science team can focus entirely on the business benefits.”
The cloud infrastructure also proves its worth in the daily operation of traditional district heating that has been around for a long time: “With district heating, the environmental impact and the cost of operation are greatest whenever the reserves have to be activated,” Karström explains. “Our algorithms let us adjust the building temperature in one-degree increments. The IoT solution combined with the ML algorithms enable considerable savings here.” And what is particularly important, as Karström points out: “We now have much greater stability. It would be impossible to optimize the operation of this kind of grid without a stable ML infrastructure.”
The business benefits: Significantly faster processes
The E.ON team’s core business is analyzing figures, so it’s no surprise that Sand can put precise figures on the savings: “Azure Data Factory and Azure ML enabled us to speed up the data feed for the more than 400 buildings by about 25 percent; now that our calculations can be performed in parallel, we can train our models even 50 percent faster. We’re also 25 percent faster when it comes to assigning building scores—and we’re just starting out on this journey. I think we can save much more if we introduce more complex algorithms and datasets.” The project team is currently migrating additional algorithms to the Azure infrastructure and training the data scientists. “They’re thrilled to have a tool that makes their lives easier,” Sand says. This also makes it easier to recruit staff: “Since everything is in the cloud now, it’s easier to bring new data science professionals on board using standardized tools,” Karström says—an important point, as data scientists are few and far between.
According to Julian, thanks to the multi-client-capable cloud environment, there’s no longer anything standing in the way of a quick international rollout: “If we want to launch the solution in Italy or Poland, we can replicate it with minor changes and roll it out in no time at all. It takes just a couple of hours to be operational in a new location.” Going forward, the E.ON team has concrete plans for some technical aspects, too: “The frequency of data collection and scoring will move toward real time,” Sand says. E.ON will integrate new forecasting models, for instance for “peak shaving” (absorbing load peaks): “In the ideal case, buildings can be preheated before a cold front moves in, so they don’t all need heat at once.”
The cloud-based ML infrastructure enables E.ON to use smaller and smaller differences in temperature at shorter and shorter intervals to more and more effectively optimize heat supply. Thus, the way to greater sustainability is through a smart grid of heat pumps—and the Azure cloud, as Azure ML permits a cool view of data even in traditional district heating grid.“Azure Data Factory and Azure ML enabled us to speed up the data feed for the more than 400 buildings by about 25 percent; now that our calculations can be performed in parallel, we can train our models even 50 percent faster.”
Mayur Sand, Digital Product & Transformation Manager and Team Lead ML, E.ON
Follow Microsoft