Deploying the First Neural Network in Energy Transmission Service
Editor's Note: This post was contributed by Rachel Berryman, Deputy Head of AI Center of Excellence at Elia Group.
The Elia Group manages the electricity transmission systems in Belgium and eastern Germany with a reliability level of 99.999%. As the responsible transmission system operator (TSO) for these regions, the Elia Group is also responsible for integrating the increasing share of renewable energy into the grid.?
Energy transition to renewables
Transitioning to renewable energy is necessary for mitigating climate change. Our TSOs are at the center of shifting from the old paradigm, where generation follows consumption, to the renewables paradigm, where demand follows generation. We're moving toward much more intermittent energy generation, including people generating energy themselves with solar panels or with flexibility from their devices.
Energy Data
This transition means more data, for example, from smart meters, connected devices, and real-time markets. Given that supply is more intermittent, we need to tap into the data to make better predictions. Predictive modeling is a mandatory tool for the new electricity paradigm.
As a data science team, we use data to reduce complexity and increase scalability. Energy problems tend to be more difficult than other data science problems in that:?
Why is accessing the data so challenging??
Sourcing systems are located in different security zones. Various technologies are required to access the data, and data governance may also diverge. Sometimes the data is readily available, and sometimes we need to extract it. If you can streamline data availability and access, you have more time for things that add value, such as exploration and analysis.
Our Solution
50 Hertz, the east German TSO in the Elia Group, has an energy data cloud that is an easy-to-use, functional data access platform giving all 50 Hertz employees the ability to explore energy-relevant data.??
The energy data cloud standardizes permissions, including APIs for accessing data from various sources. Users need only one permission to access data, and every dataset has a listed owner.?
Predicting grid losses with a neural network
Once our data scientists could easily access the data, what did they do with it??
One example is forecasting grid losses.?
Energy is unlike other products in that supply has to match demand every second since it can't be stored at scale.?
When energy is generated and transported, around two percent is usually lost because transmission is at very high voltages. However, this amount could be as low as one percent if the distance is short or five percent or more if the distance transmitted is very far. To that end, it is usually a long distance from the wind farms and solar parks to the consumer.
The business case is that a TSO needs to cover grid losses, often purchasing electricity at the last minute for a much higher cost. In 2018, 50 Hertz spent €70m on grid loss cover. The energy transition means that grid losses are increasing (e.g., when energy is transported from offshore wind parks, it has to be transmitted further), and the grid loss range is more volatile.
领英推荐
Since a TSO is a government-regulated monopoly, it's in the interest of all consumers to improve forecasts. To that end, the energy economics data science team started on this task as soon as it gained access to the energy data cloud. The team could integrate far more data points than before (e.g., data from 70 weather stations around Germany). Using this increased data, they trained a simple feedforward neural network with just one hidden layer as an initial model, which provided more accurate predictions than the classical time-series model.
Deployment
Next, they deployed rapidly and independently on the Agile Apps Platform, which runs on OpenShift. This open-source container application platform makes deployment easy for anyone in the company.?
The Day-Ahead forecast has been operating since December 2019. The feedforward neural network has over 100 inputs and a hidden layer with 50 nodes. This project was completed in less than six months. Once everything was deployed on the agile platform, updating the model was easy.
The forecast quality was improved by 7%. In the first year of production, the neural network saved one percent of grid loss cover charges equal to €700,000.
Success factors: Give people the tools they need and get out of their way
How does the Elia Group replicate success across future use cases??
First, we make company data available. We enable data scientists throughout the company, whom we call citizen data scientists, to start their own data science use cases.?
Second, we take advantage of the OpenShift agile app platform because it provides ease of deployment and clear ownership and responsibility: the platform is responsible for having enough computing power for all the apps running on it. But as for the individual app, it is the responsibility of the department and the individual deploying it. In other words, no one has to wait on someone else.??
Finally, the Elia Group has founded the AI Center of Excellence (AI CoE), a facility providing best practices, research support, and training for data science. The center helps with three aspects.?
To learn more about the work that we do, contact the AI CoE at [email protected]
AI Guild Announcements?
Not a member of the AI Guild yet?
Apply online at?https://www.theguild.ai/.
Do you have a use case you would like to share??
Special thanks to?Evan Simpson?for acting as editor of?Deploy It Already.??
Full-Stack Data Scientist
2 年To see Rachel Berryman's presentation at #datalift no.2, click here: https://youtu.be/0BgHZ7oKRXI
Spiking Neural Networks, Simulation-Based Inference and Digital Twins
2 年This was a great read. Sounds like a great prediction use case. Deploying neuronal networks closer and closer to the hardware eventually in areas such as energy storage and demand response will be super interesting and I would love to see more energy related content.