Machine Learning for Physical Simulation Challenge

Machine Learning for Physical Simulation Challenge

# PowerGrid Dataset

# Context

To tackle climate change, renewables energies such as solar and wind need to be massively integrated in the years to come into the power grid to achieve decarbonation. This brings complex new operational challenges.

# Industrial Challenge

Due to their intermittency, many more scenarios need to be studied to mitigate risks near real-time to keep the grid safe without cascading failures. In particular, the operators need to anticipate what could happen in terms of congestions on the grid in case some unexpected outage occurs on some power lines under different wind and solar conditions. Hence the need for powerflow simulation is exploding at least 1000-fold in control rooms.

# Objective

This competition aims at unlocking new methods for speeding-up these simulations by several order of magnitude, while ensuring some realistic criteria, in order to run this more complex risk assessment tomorrow. You will apply this on a grid of similar size of one controlled by a human operator with an energy mix that include 30% wind and solar as expected on the French grid in the near future.

# Competition Dataset

A demo version of the usecase datasets is provided in input_data_local for 3 different scales of Power Grid environments. Two environments l2rpn_case14_sandbox and l2rpn_neurips_2020_track1_small includes 14 and 38 nodes respectively and provided for participants to test (optional) their solution. Environment l2rpn_idf_2023 is the challenge environment. All the final solutions (which are submitted on codabench) should be trained and evaluated on this environment. See the description of provided jupyter notebooks below to see how to use and import these datasets.

A configs folder is provided which includes the configurations (parameters) related to benchmarks and augmented simulators (aka models). The users could change the simulators configurations to change the hyperparameter of existing models. More details on how to use these configs is provided in the notebooks.

Finally, the trained_models folder contains a set of trained baseline models, which are used to show the evaluation procedure and scoring.