Learning to Run a Power Network
Following the success of the Horizon prize to predict flows in the French long distance high voltage electricity transmission grid, managed by the company “Réseau de Transport d’Électricité” (RTE), we are organizing a new challenge. The goal of this challenge is to test the potential of Reinforcement Learning (RL) to control electricity transportation in power grids, while keeping people and equipment safe. Hence, it is the "gamification" of a serious problem: operating the grid is becoming increasingly complex because of the advent of less predictable renewable energies, the globalization of energy markets, growth in consumption and concurrent limitations on new line construction. To make smart grid happen, it is becoming urgent to optimize more tightly the grid operation, considering a broader range of topological changes used more frequently, without compromising security.
Will you help rejuvenate our aging Electricity Fairy ?
The Electricity Fairy Paintaing in Paris
The Learning To Run Story
Previously on Learning to Run ...
A first Reinforcement Learning challenge organized at NIPS 2017 by Stanford caught our attention and inspired us: The Learning To Run Challenge - or how to learn a controller to make an agent walk and run without falling! This was a tough coordination and planning problem.
See the video for some fun and some impressive results obtained with RL! For more details, visit the github page.
... Coming soon on Learning to Run !
A power grid can be seen as well as a system that needs to be controlled with coordination to ensure proper stability while operating under a challenging hazardous environment. Otherwise the system will fall into a BLACKOUT !
See a didactic video explaining a near blackout event in Europe in 2006.
A power system in Action
Ensuring Production and Consumption Balance at all time
Consumption varies during the day depending on human habits. Production needs to match consumption at all time to ensure a proper balance between offer and demand. How and where the energy is produced can vary as well along the day, adding some complexity to power grid operations.
For more information and interactive navigation through historical days of consumption, you can visit RTE ECO2MIX website.
French Electrical Consumption over a day
French Electrical Production and Energy mix
Operating the grid to safely transport electricity
A power grid has to transport electricity safely from productions to consumptions over the course of the day. Hour after hour, operators tries to minimize energy losses through the grid while making sure no power line is overloaded. Otherwise, lines will get disconnected for safety reasons. This can trigger a possible snow ball effect which can effectively lead to a Blackout!
Here is an example of a power grid in action over time with no human intervention. After some iterations, some lines begin to be overloaded. Few time steps, later they get disconnected. After few additional iterations, more lines are disconnected, eventually leading to a blackout where productions cannot supply consumptions anymore.
Renderer of a running power grid (IEEE14) under the pypownet environment
Operating your first power grid interactively
Previsouly you saw what the role of power grid is and what can badly happen to a power grid when operated. Let's now see what you can do to operate the grid and avoid or solve those dangerous situations.
The University of Illinois has created a simple educational interactive power grid applet to understand how a power grid works and how you can operate it. You can complete the 5 short challenges that are given to deepen your understanding on power grid operations.
Note that the less costly operation to redispatch the power flows in real life is to change the topology of the grid by opening or closing switches, whereas changing productions are costly. Subsations can be seen as the articulation joints of the power grid body. To best operate the grid and let the electricity flows smoothly, you need to coordinate those articulations, similar to learning to run challenge. The goal of the challenge is to find the best topologies configuration over time to manage the power flows while insuring the system stability.
101 power grid interactive applet
Timeline for the Competition (Finished!)
May 15th, 2019: Beginning of the competition with the release of public RL environment. Participants can start submitting agent models on Codalab platform and obtaining immediate feedback in the leaderboard on validation scenarios.
May 27th, 2019: Potential release of a new baseline to foster competition if several participants are already doing better than this baseline.
June 15th, 2019: Start of the testing days on unseen test scenarios.
June 19th, 2019: End of the competition, beginning of the post-competition process
Jul 1rst, 2019: Announcement of the L2RPN Winners.
Jul 14th, 2019: Beginning of IJCNN 2019.
To be continued at NeurIPS 2020
The competition was tight for the top 5 teams! Geirina Team eventually ranked 1rst and Learning_RL team ranked 2nd, making them the winners of this first episode of L2RPN Challenge! Congratulations to them, they did remarkable team work to reach this performance!
They both explain their respective approach in the following videos:
PAPERS TO GO FURTHER - Join us!
Below is the list of accepted and submitted papers from our Apogee project Team since 2017.
It should help you have a deeper understand of the challenges we are tackling and yet the approaches we have been exploring. We hope this could give you draw some inspirations to join us make further advances in the field of AI for Smart Grids!
Most of those papers can be found on arxiv and HAL libraries.
- IERP 2017: Introducing Machine Learning for power system operation support
- ESANN 2018: Fast Power System Security Analysis with Guided Dropout
- IJCNN 2018: Anticipating contingengies in power grids using fast neural net screening
- ISGT Europe 2018: Optimization of computational budget for power system risk assessment
- ISGT Europe 2018 & NIPS 2019 Workshop: Guided machine learning for power grid segmentation
- MedPower 2018: Expert System for topological remedial action discovery in smart grids
- ESANN 2019: Leap Net for power grid perturbations
- IJCNN 2019: Graph Neural Solvers for Power Systems
- EGC 2019: Semi-supervised labelling, Towards an Extended Expert Approch
- ECML 2019: Interpreting atypical conditions in systems with conditional autoencoder
In review for 2019:
- Neurocomputing Jounal: LEAP Nets for System Identification and Application to Power Systems
- PSCC 2020: Unsupervised Graph Neural Solver for Power Flow Computation
- PSCC 2020: Learning to run a power network by training topology controllers
Réseau de Transport d'Électricité (Electricity Transmission Network), usually known as RTE, is the electricity transmission system operator of France. It is responsible for the operation, maintenance and development of the French high-voltage transmission system, which at approximately 100,000 kilometres (62,000 mi), is Europe's largest. RTE R&D is one of the strongest in the world in the field of power grids and has many research collaboration around the world, especially in Europe and in the USA. RTE is now a member of the Linux Foundation Energy Initiative in which it open-sourced many simulators and applications.
ChaLearn is a non-profit organization with vast experience in the organization of academic challenges. ChaLearn is interested in all aspects of challenge organization, including data gathering procedures, evaluation protocols, novel challenge scenarios (e.g., competitions), training for challenge organizers, challenge analytics, result dissemination and, ultimately, advancing the state-of-the-art through challenges.