Welcome to Grid2Op’s technical documentation!¶
Grid2Op is a pythonic, easy to use framework, to be able to develop, train or evaluate performances of “agent” or “controller” that acts on a powergrid in different ways.
It is modular and can be use to train reinforcement learning agent or to assess the performance of optimal control algorithm.
It is flexible and allows the power flow to be computed by the algorithm of your choice. It abstracts the modification of a powergrid and use this abstraction to compute the cascading failures resulting from powerlines disconnection for example.
abstract the computation of the “cascading failures”
ability to have the same code running with multiple powerflows
parallel execution of one agent / controller on multiple independent scenarios (multiprocessing)
fully customisable: this software has been built to be fully customizable to serve different purposes and not only reinforcement learning, or the L2RPN competition.
Grid2Op is a python module that aims to make easier the research on sequential decision making applied to power systems.
This package adopt the “reinforcement learning” point of view and is compatible with the openAI gym programming interface (see section Compatibility with gymnasium / gym for more information).
Applied to power system, the “reinforcement learning” framework ask:
a “controller” (named Agent) to take an “action” on the powergrid (for example for L2RPN competitions in 2019 and 2020 these actions consist in modifying the connectivity of the powergrid).
the “environment” (a.k.a the “real world”) applies that action on the powergrid, applies some other modifications and return the next state.
The goal of grid2op is to model “sequential decision making” that could be made by human operators, for example changing the configuration of some “substations” as demonstrate in the figure below:
Any kind of “controller” can be implemented using this framework even though it has been inspired by the “reinforcement learning” community. You can implement some heuristic “controllers” (some examples are available in the Agent module description), “controllers” that comes from the Optimization community (for example “Optimal Power Flow”) or “Model Predictive Control”. One of the goal of Grid2Op is to allow everyone to contribute to closing the gap between all these research communities.
Main module content¶
- Available environments
- Make: Using pre defined Environments
- Content of an environment
- Possible workflow to create an environment from existing time series
- Input data of an environment
- Optimize the data pipeline
- Using “optimization” technique
- Compatibility with gymnasium / gym
- Model Free Reinforcement Learning
- Model Based / Planning methods
- Elements modeled in this environment and their main properties
- A grid, a graph: grid2op representation of the powergrid
- Creating a new backend
- Opponent Modeling
- Rules of the Game
- Time Series Handlers
- Utility classes
- Voltage Controler
If you still can’t find what you’re looking for, try in one of the following pages:
Still trouble finding the information ? Do not hesitate to send a github issue about the documentation at this link: Documentation issue template