Grid2op/grid2op

What it solves

Grid2Op is a platform designed to model and perform powergrid operations. It provides a testbed for sequential decision-making in power systems, allowing researchers and operators to simulate how to maintain a stable grid while handling security issues, maintenance, and changing demands.

How it works

The framework emulates the behavior of powergrids of various sizes and formats by integrating with powerflow solvers (such as PandaPower). It uses a modular architecture that allows users to swap backends, solvers, and observation/action sets. It is specifically designed for Reinforcement Learning (RL), featuring an interface compatible with Gymnasium to enable the training of agents that can modify generator setpoints, perform load shedding, or change the grid's topology.

Who it’s for

It is primarily intended for the Reinforcement Learning community applying AI to power systems, as well as researchers and developers participating in the "Learning To Run Power Network" (L2RPN) competitions.

Highlights

• RL-Ready Interface: Fully compatible with Gymnasium, making it easy to integrate with RL frameworks like Stable Baselines 3.
• Flexible Grid Manipulation: Supports modifying generator voltages, active production, load values, and grid topology (connecting/disconnecting powerlines).
• Modular Backend: Compatible with multiple powerflow solvers including PandaPower, lightsim2grid, and powsybl.
• Coping with Hazards: Includes capabilities to model maintenance, hazards, and attacks that can disconnect powerlines.
• Multi-Environment Support: Natively supports a single agent interacting with multiple environments simultaneously for asynchronous training.