A forum is available for comments, suggestions and help on discord: https://discord.gg/cYsYrPT.
The Grid2op documentation can be found here: https://grid2op.readthedocs.io/en/latest/
If you would like to get familiar first with the problem and make your hands on quickly, a sandbox competition on a smaller case remains open as well: https://competitions.codalab.org/competitions/24493
Power grids transport electricity across states, countries and even continents. They are the backbone of power distribution, playing a central economical and societal role by supplying reliable power to industry, services, and consumers. Their importance appears even more critical today as we transition towards a more sustainable world within a carbon-free economy, and concentrate energy distribution in the form of electricity. Problems that arise within the power grid range from transient brownouts to complete electrical blackouts which can create significant economic and social perturbations, i.e.de facto freezing society. Grid operators are still responsible for ensuring that a reliable supply of electricity is provided everywhere, at all times. With the advent of renewable energy, electric mobility, and limitations placed on engaging in new grid infrastructure projects, the task of controlling existing grids is becoming increasingly difficult, forcing grid operators to do “more with less”. This challenge aims at testing the potential of AI to address this important real-world problem for our future.
Visualization of a scenario on the wcci grid and environment. Your submission results will come up with such a visualization for your agents.
Try your own submission to get your own visualization!
You are ready to get started and we are looking forward for your first submission in the Participate section to become the control room operator of the future !
In this section, we give you the instructions to help you:
The challenge is based on an environment (and not only a dataset) in which an agent can learn from interactions. It runs under the Grid2op platform.
The Grid2op platform can be installed as any python package with:
pip install grid2op
We also strongly recommend to use the "baseline" python package that will be updated during the competition with:
pip install l2rpn-baselines
A starting kit is available for you to download in the Participate section on codalab, along with the proper game environment for the competition. Several notebooks should help you understand how to properly run the Grid2op platform using chronics to train and test your agents.
The starting kit also gives details about how to check that your submission is valid and ready to run on the competition servers.
Once grid2op is installed, you can get the competition data (approximately 4~4.5Go) directly from the internet. This download will happen automatically the first time you will create the environment of the competition from within a python script or shell:
env = grid2op.make("l2rpn_wcci_2020")
You can visit the "Good to Know" section for more information and parametrization, for example to allow faster learning. The general help of the platform is available at https://grid2op.readthedocs.io/en/latest/ . As probably most of you are not familiar with power systems in general, we have made some introductory notebooks of the problem we are tackling and the grid2op platform. These notebook are available without any installations thanks to "mybinder" at the following link https://mybinder.org/v2/gh/rte-france/Grid2Op/master.
Essentially, a submission should be a ZIP file containing at least these two elements:
In the starting kit, and script is here to help you create and check your submission is valid:
python3 check_your_submission.py --help
/!\ This is a code submission challenge, meaning that the participant has to submit his code (and not his results).
Upon reception of the challenger's submission, will be read by Codalab and the code will be run on the competition servers. The detailed structure of the submission directory can be found in the starting kit.
Then, to upload your submission on Codalab:
Codalab will take some time to process the submission and will display the scores on the same page once the submissions have been processed. You may need to refresh the page. As explained in the rules, if your submission takes more than 20 minutes to run, a timeout error will be raised and your submission will be ignored.
In the "Submit / View Results" sub-section in the Participate section, you can see the status of your submission. Once it is processed, you can review your submission, see the score it obtained and the time it took to run. When clicking on the blue cross next to your submission, different logs are available for additional information. You can also download your submission again if you want to. More importantly, you can get logs of your agent's behavior over different scenarios in the folder "output from scoring step". Several indicators over all the scenarios that the agent was run on can be visualized in the html file.
To compare your score to the ones of the other participants, please go on the Results page. The Leaderboard is displayed there. Be aware that only your last submission's score is considered there.
Throughout the competition, the environment will be the same. From the start an until the end, the version of grid2op 0.9.1.post1 is used. The python packages you can use on codalab are listed as the "challenge" dependency of grid2op. To replicate the environment of the competition you can install it this way:
pip install grid2op[challenge]==0.9.1.post1
If you are familiar with docker, all the code that is used by codalab is used through docker. You can get the exact version used by the competition with:
docker pull bdonnot/grid2op:0.9.1.post1
The following prizes are sponsored by Geirina and State Grid of China.
The top two teams will be awarded $3,000 each in the form of travel expenses to either:
To get the prizes, the winners will be required to share their model open-source following the L2RPN baseline template.
This challenge is governed by the general ChaLearn contest rules.
This challenge starts on May 20th 2020 and ends on July 1st 2020. Prizes for winners are listed in the Prizes section.
This challenge would not have been possible without the help of many people.
Protocol and task design:
Data format, software interfaces, and metrics:
Environment preparation and formatting:
Baseline methods and beta-testing:
Other contributors to the organization, starting kit, and datasets, include:
Our special thanks go to:
We put at your disposal a starting kit that you can download in the Participate Section. It gives you an easy start for the competition, in the form of several notebooks:
If you need any help, do not hesitate to contact the competition organizers on the dedicated discord forum server that we opened for the competition: https://discord.gg/cYsYrPT
The default backend is pandapower, a well-known open-soure library in the power system community. However, it can be a bit too slow when it comes to running thousands of simulations. For that aim, the lightSim2Grid simulator (https://github.com/BDonnot/lightsim2grid) was developped in C++, imitating pandapower behavior and reproducing its results for our current power grid modelization. A speedup factor of 30 can be achieved, which should be of great use when training an agent.
NB: at the moment this simulator is not available on Microsoft Windows based machine (unless you manage to compile the SparSuite framework on Windows). For the other platform, installation instructions are provided in the abvo mentionned github.
Once installed you can use it this way:
from lightsim2grid.LightSimBackend import LightSimBackend
backend = LightSimBackend()
env = grid2op.make("l2rpn_wcci_2020", backend=backend)
# and now you can do the code you want to do
The performance increase can be rather large. On a desktop (Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz), just using regular agent (without any call to external librariries), for the "l2rpn_wcci_2020" is it possible to perform 1000 steps in 72.8s (13.7 it/s) whereas the same agent, with LightSimBackend only takes 2.7s (370.4 it/s) to perform the same number of steps (speed ups around 27:1 in favor of LightSimBackend)
/!\ ATTENTION /!\ We want to emphasize that the default grid2op backend is PandaPowerBackend. This default backend will be used to score your agent. There might exist some slight differences (we never noticed difference higher than 1e-7 - 1e-6) between PandaPowerBackend and LightSimBackend.
As operators do in real life, you can simulate the effect of diffent actions before taking a decision (a step in Grid2op framework). This comes at a cost in terms of computation time, but allows to validate the relevance of your action on the short term.
To understand all the features of Grid2op framework and use it to its full potential, you will find most of the answers on how to use it through its documentation: https://grid2op.readthedocs.io/en/latest/.
You can specify your own reward, a function that can be different from the score of the competition. We believe that reward design is an important aspect of the competition, and a participant should think about which reward is best to let its agent learns and explore.
To do so you simply need to change the "reward"
from grid2op.Reward import L2RPNReward
env = grid2op.make("l2rpn_wcci_2020", reward_class=L2RPNReward)
As always more information on this feature can be found at https://grid2op.readthedocs.io/en/latest/reward.html
Different parameters to configure an environment allow to modulate the difficulty for an agent to deals with that environment.
You can report to the full description of the parameters used in each level at the end of this section for more information.
For instance, it is possible to inhibit line disconnection when overloaded, hence avoiding any blackout and allowing an agent to operate and learn until the end of scenario. This easy mode could be a prefered mode when your start training your agent. By modifying the environment parameter you can hence design a learning curriculum for your agent, making the environment more and more difficult to eventually operate in the full environment setting.
For this competition, 4 difficulty levels are available. For example, you have easier environment with
env = grid2op.make("l2rpn_wcci_2020", difficulty="0")
# in this case the environment does not simulate the powerline disconnection when there are overflows for example.
Increasing order of difficulty are (see the addendum for a detail on every level):
To inspect and study some particular scenarios and compare the behavior of different agents, the Grid2Viz interface is a great tool to try and use (https://github.com/mjothy/grid2viz)
To generate all the chronics of the environment for the competition, we used the chronix2grid package. If you want to generate additional chronics, you can use it yourself https://github.com/mjothy/ChroniX2Grid/tree/master/chronix2grid
Once your submission has run on the platform, you can visualize how your agent behave. Two main plots are available, here is an example:
To get them, you can go to your submission, clik on the "+" sign and then "Download output from prediction step"
You can visualize how many time steps and the score per scenarios, the cost of maintaining your powergrid at each time step and a gif image that sums up the result of your agent on one given scenario. For example the score per scenarios:
In this section we will list the different values taken for the Parameters that are used to make default difficulty levelts. For more information about the real definition of these attributes, you can visit: https://grid2op.readthedocs.io/en/latest/parameters.html
Yes in case that all these settings are not enough you can definitne your own set of parameters at the creation of the environment. For this you can do:
from grid2op.Parameters import Parameters
param = Parameters()
param.NB_TIMESTEP_OVERFLOW_ALLOWED = ...
param.NB_TIMESTEP_COOLDOWN_LINE = ...
# change any other attribute of the parameter class
env = grid2op.make("l2rpn_wcci_2020", param=param)
# and now the created environment is configured with you parameters
The objective of the competition is to design an agent that can sucessfully manage to operate a powergrid. Operate a powergrid here means: find ways to modify how the objects are interconnected together (aka "changing the topology) or modify the productions to make sure it stays safe (see "Conditions of Game Over").
More information are given in the 1_Power_Grid_101_notebook prodived in the starting kit.
If you have any question, we are here to answer you on the official discord: https://discord.gg/cYsYrPT
As any system, a power grid can fail to operate properly, as illustrated on the challenge website. This can occur under conditions such as:
These conditions can appear when power lines in the grid get disconnected after being overloaded. When a line get disconnected, it loads gets distributed over other power lines, which in turn might get overloaded and thus disconnected as well, leading to a cascading failure (blackout).
When the power in a line increases above its thermal limit, the line becomes overloaded. It can stay overloaded for few timesteps before it gets disconnected, if no proper agent action is taken to relieve this overload (2 timesteps are allowed in this challenge, see the Parameters class in grid2op), this is what we call a "soft overload". If the overload is too high, the line gets disconnected immediately (above 200% of the thermal limit in this challenge). This is a 'hard' overload. At some point this can lead to a very rapid cascading failure in a single timestep, if some lines already got disconnected and other lines get quite loaded.
Actions can consist on:
Those parameters are accessible through the "Parameters" class of grid2op. During training, you can modify some of these parameters to relax some constraints and initialize your training better.
Observations about the state of the grid can be retrieved from the environment to be used for your agent. Please read the table in the grid2op documentation. You can recover information over current productions, loads, and more importantly about the flows over the lines and the topology of the grid. You are free to use whatever observation available, make the best of it!
Some parameters of the environment can easily be modified before running an agent on it. By doing so you can actually modulate the difficulty of a given problem, and define some learning strategy as explained in section "Good To Know". For the competition, you will be tested with the followind default parameter values:
Your agent is evaluated on 10 scenarios of different lengths starting at different times.
You can have a look at the 2_Develop_And_RunLocally_An_agent notebook provided on the starting kit.
Definition of a cost function
The cost function that an agent will be evaluated on represents the cost of operations of a power grid, as well as the cost of any blackout that could occur. Let explains the details of that in the following.
1) cost of energy losses
To begin with, we will recall that transporting electricity always generates some energy losses Eloss(t) due to the Joule effect in resistive power lines at any time t:
At any time t, the operator of the grid is responsible for compensating those energy losses by purchasing on the energy market the corresponding amount of production at the marginal price p(t). We can therefore define the following energy loss cost closses(t):
2) cost of redispacthing productions after actions on generators
Then we should consider that operator decisions when taking an action can induce costs, especially when requiring market actors to perform specific actions, as they should be paid in return. Topological actions are mostly free, as the grid belongs to the power grid operator, and no energy cost is involved. However, redispatching actions involve producers which should get paid. As the grid operators ask to redispatch energy Eredispatch(t), some power plants will increase their production by Eredispatch(t) while others will compensate by decreasing their production by the same amount to keep the power grid balanced. Hence, the grid operator will pay both producers for this redispatched energy at a cost credispatching(t) higher than the marginal price p(t) (possibly by some factor):
3) total cost of operations
If no flexibility is identified or integrated on the grid, operational costs related to redispatching can dramatically increase due to renewable energy sources as was the case recently in Germany with **an avoidable 1 billion €/year increase**.
We can hence define our overall operational cost coperations(t):
Formally, we can define an "episode" e successfully managed by an agent up until time tend (over a scenario of maximum length Te) by:
where ot represents the observation at time t and at the actions the agent took at time t. In particular, o1 is the first observation and otend is the last one: either there is a game over at time tend or the agent reached the end of the scenario such that tend = Te.
An agent can either manage to operate the grid for the entire scenario or fail after some time tend because of a blackout. In case of a blackout, the cost cblackout(t) at a given time t would be proportional to the amount of consumption not supplied Load(t), at a price higher than the marginal price p(t) by some factor beta:
Notice that Load(t) >> Eredispatch(t) , Eloss(t)
which means that the cost of a blackout is a lot higher than the cost of operating the grid as expected. It is even higher if we further consider the secondary effects on the economy (More information can be found on this blackout cost simulator: https://www.blackout-simulator.com). Furthermore, a blackout does not last forever and power grids restart at some point. But for the sake of simplicity while preserving most of the realism, all these additional complexities are not considered here.
Now we can define our overall cost c for an episode e:
We still encourage the participants to operate the grid as long as possible, but penalize them for the remaining time after the game is over, as this is a critical system and safety is paramount.
Finally, participants will be tested on N hidden scenarios of different lengths, varying from one day to one week, and on various difficult situations according to our baselines. This will test agent behavior in various representative conditions. Under those episodes, our final score to minimize will be:
For a naive agent (a "do-nothing" agent that does not actually take any action) the cost function can get really high (in the order of billions of $) in our scenarios since a blackout most likely occur in a scenario.
Comparing two agents that scores on a billion scale is not easy (eg. it is not clear that 33025056 is worst than 33025053). So, we decided to apply linear transformations to improve the readability and better represent the ability of an agent to be robust and performant:
- -100.0 No step played, maximum blackout penalty for all steps of all scenarios.
- 0.0 for the "do nothing" baselines.
- 80.0 for playing all the scenarios completely while keeping losses equal to the difference of productions and consomations of the scenario.
- 100.0 for the best possible agent: an agent that handles all scenarios with topology, that is without additional redispatching cost, and which reduces the losses to 70% of the initial electricity losses of the do nothing agent (NB this is probably un achievable for most scenarios).
This means that:
- the score should be maximized rather than minimized
- having a score of 100 is probably out of reach
- having a positive score is already pretty good!
On the first figure the see the operationnal costs (highligthed in red) of a few "interesting" controlers (note that these controler are "theoretical" controlers). Some might not be feasible in practice. We compare the score with the operational cost of four of them:
- "Com. Game Over" is the worst possible controler. He does a complete game over for all the scenario
- "Do nothing": is the agent that does nothing. It serves as baseline.
- "No Game Over" is a controler (theoretical) that would not game over, but would not take any action.
- "No Game Over + loss optim." is a controler that does better than the previous one in the sense that it will also takes care (and succeed) in managing the losses. To make sure we have an upper bound on it, we supposes that such a controler is 30% more efficient than the "No Game Over" controler in reducing the operational cost. [NB for most scenarios, this is probably out of reach]
figure: representation of the operationnal costs of the few (theoretical) controlers of interest.
In reality though, we want to emphasize the fact that keeping the grid is a problem, but reducing the operationnal cost is also really interesting. To this end, we decided to to assign different scores as decribed in the image bellow:
- "Com. Game Over" has the worst possible score of -100.00. If this score is displayed, it means your submission is probably not valid.
- "Do nothing": is the "reference" agent. It has the score of 0. (NB. In all cases do nothing agent has a score of 0 regardless of its capacity to succeed to manage completely a scenario. This means that you can have a slightly negative score in such cases if your agent did worst than the do nothing at managing the scenarios (in terms of operation costs) but it manage to get to the end of it.)
- "No Game Over" is assigned a score of +80.00. (NB in case the do nothing successfully manage all the scenario, this part is "skipped" see the note bellow)
- "No Game Over + loss optim." is assigned a score of +100.00.
In case of a scenario that a "do-nothing" agent can handle until the end, you score will be 0 if you finish the scenario and don't do better at managing the losses than a do nothing. So in addition of being robust, managing efficiently the electricity losses will be especially rewarded for some scenarios.
figure: representation of the operationnal costs of the few (theoretical) controlers of interest as well as their asscoaited scores.
For this competition, there exists 10 hidden scenarios of 3-day long, distributed over the months in a year and over the days in a week.
Scenarios have been cherry picked to offer different levels of difficulty, can start at arbitrary time steps ( but chronics starts always at middnight!). Time interval between two consecutive time step is fixed and will always be 5 mins.
You can use any rewards you want in grid2op, different from our cost function for competition evaluation, both at training time (when you train your agent on your computer) or at test time.
To change the reward signal you are using, you can, at training time, specify it at the creation of the environment:
from grid2op.Reward import GameplayReward
env = grid2op.make("l2rpn_wcci_2020", reward_class=GameplayReward)
We invite you to get have a look at the official grid2op documentation about rewards at https://grid2op.readthedocs.io/en/latest/reward.html
|Starting Kit||2.391||#1 Development phase|
Start: May 20, 2020, midnight
Description: Development phase: you can try your models in this phase
Start: June 30, 2020, midnight
Description: Test Phase your model will be tested only once on this dataset
Oct. 1, 2020, midnight
You must be logged in to participate in competitions.Sign In