This is the official implementation of the paper Stabilizing Voltage in Power Distribution Networks via Multi-Agent Reinforcement Learning with Transformer (KDD2022 Research Track).

The Transformer-based Multi-Agent Actor-Critic Framework (T-MAAC) is based on MAPDN. Please refer to that repo for more documentation.

We suggest you install dependencies with Dockerfile and run the code with Docker.

docker build . -t tmaac

We use load profiles and PV profiles provided by MAPDN.

1. Download the data from the link.
2. Unzip the zip file and there are 3 folders as following:
   • case33_3min_final
   • case141_3min_final
   • case322_3min_final
3. Go to the directory [Your own parent path]/T-MAAC/environments/var_voltage_control/ and create a folder called data.
4. Move the 3 folders in step 2 to the directory [Your own parent path]/T-MAAC/environments/var_voltage_control/data.

You can train the model using the following command.

source activate mapdn

## running in case141
python train.py --alg icstransmaddpg --alias example_0 --mode distributed --scenario case141_3min_final --qweight 0.01 --voltage-barrier-type l2 --save-path trial

## running in case322
python train.py --alg icstransmatd3 --alias example_0 --mode distributed --scenario case141_3min_final --qweight 0.1 --voltage-barrier-type l2 --save-path trial

The meanings of the arguments:

• --alg is the MARL algorithm, e.g. maddpg, matd3, icstransmaddpg, icstransmatd3.
• --alias is the alias to distinguish different experiments.
• --mode is the mode of environment, e.g. distributed.
• --scenario is the power system on which you like to train, e.g. case141_3min_final, case322_3min_final.
• --qweight is the q_weight used in training. We recommend 0.01 for case141 and 0.1 for case322.
• --voltage-barrier-type is the voltage barrier function in training, e.g. l1, l2, bowl.
• --save-path is the path to save the model and configures.

It is worth noting that the difficulty of voltage control problem varies during different months of a year. For example, during the midday summer, excessive active power from intense sunlight is injected into the grid, creating a more significant challenge for the voltage control task than in winter. Thus, a series of fixed scenarios must be chosen to evaluate algorithms fairly.

We randomly select 10 episodes per month, a total of 120 episodes, which constitute the test dataset test_data.csv. Each episode lasts for 480 time steps (i.e., a day). And fixed 10 episodes of 120 episodes are selected to evaluate performance in training phase. (see def evaluation() in models/model.py)

python test.py --save-path trial/model_save --alg icstransmaddpg --alias example_0 --scenario case141_3min_final --qweight 0.01 --voltage-barrier-type l2 --test-mode test_data

The meanings of the arguments:

• --alg is the MARL algorithm, e.g. maddpg, matd3, icstransmaddpg, icstransmatd3.
• --alias is the alias to distinguish different experiments.
• --scenario is the power system on which you like to train, e.g. case141_3min_final, case322_3min_final.
• --qweight is the q_weight used in training. We recommend 0.01 for case141 and 0.1 for case322.
• --voltage-barrier-type is the voltage barrier function in training, e.g. l1, l2, bowl.
• --save-path is the path to save the model and configures.
• --test-mode is the test mode, e.g. single, test_data. test_data means use test_data.csv to evaluate algorithms.
• --test-day is the day that you would like to do the test. Note that it is only activated if the --test-mode is single.
• --render indicates activating the rendering of the environment.