Repo of the Reinforcement Learning for Grid Control (RLGC) Project

We explore to use deep reinforcement learning methods for control and decision-making problems in power systems.

NOTE: RLGC is under active development and may change at any time. Feel free to provide feedback and comments.

To run the training, you need python 3.5 or above and Java 8. Unix-based OS is required. We suggest using Anaconda 5.0 to create virtual environment from the yaml file we provided.

• To clone our project

  git clone https://github.com/RLGC-Project/RLGC.git
  

• To create the virtual environment

  cd RLGC    
  conda env create -f RLGC_environment.yml  
  

  If you get errors about OpenAI gym you probably need to install cmake and zlib1g-dev. For example on Ubuntu machine, do the following command.

  sudo apt-get upgrade
  sudo apt-get install cmake
  sudo apt-get install zlib1g-dev
  

  After creating environment py3ml, you can activate the virtual environment and do development under this environment.

• To activate virtual environment

  source activate py3ml
  

• To deactivate virtual environment

  source deactivate
  

Note that we need to change the source code of OpenAI Gym Baseline models because To modify the source code,

cd your_path_to_anaconda_directory/envs/py3ml/lib/python3.6/site-packages/baselines/deepq

eg, mine will be

cd ~/anaconda/envs/py3ml/lib/python3.6/site-packages/baselines/deepq

Subsequently you need to modify the learn() function in simple.py script. First, add trained_model = none to the funtion argument list. Then, replace lines of 229 to 255 by following codes.

    episode_rewards = [0.0]
    saved_mean_reward = None
    obs, starttime,durationtime = env.reset()

    #i = 0
    #noise = 0.01 * np.random.randn(4,8,301)
    #np.save("./noise", noise)
    with tempfile.TemporaryDirectory() as td:
        model_saved = False
        model_file = os.path.join(td, "model")
        for t in range(max_timesteps):
            if callback is not None:
                if callback(locals(), globals()):
                    break
            # Take action and update exploration to the newest value
            action = act(np.array(obs)[None], update_eps=exploration.value(t))[0]
            new_obs, rew, done, _ = env.step(action)
            # Store transition in the replay buffer.
            replay_buffer.add(obs, action, rew, new_obs, float(done))
            obs = new_obs #* (1 + noise[:,:,i].flatten())
            #i += 1

            episode_rewards[-1] += rew
            if done:
                #i = 0
                obs, starttime, durationtime = env.reset()
                episode_rewards.append(0.0)

Alternatively, you can simply use the simple.py in the src/py/openAI_gym to replace the same file in the directory <our_path_to_anaconda_directory>/envs/py3ml/lib/python3.6/site-packages/baselines/deepq But you will be at risk of potential compatibility issue with newer version of OpenAI Gym, as we may not test it against future versions of OpenAI Gym. If you have confront such an issue, please report to us.

Remember to remove the original complied cache file as follows:

cd your\_path\_to\_simple.py\_file/__pycache__/
rm simple.cpython-36.pyc

First launch java server which is used to simulate the power grid system. Then launch your training in the virtual environment. We provide two power grid systems in the examples. RLGCJavaServer0.72.jar is the latest release and can be used for the IEEE 39-bus system and RLGCJavaServerSimple.jaris only used for the Kundur 2-area system.

• To launch the java server, open a new terminal

cd ~
cd RLGC/lib  
java -jar RLGCJavaServer0.72.jar 25001

The last parameter is the communication port number between grid system and the training agent. You can switch the port number if necessary.

• To launch the training, you need first activate the virtual environment. Then run the following scripts. trainKundur2areaGenBrakingAgent.py is used for training the generator braking agent for the Kundur 2-area system and trainIEEE39LoadSheddingAgent.py is used for training an agent for regional load shedding in IEEE 39-bus system

source activate py3ml
cd RLGC/src/py  
python trainIEEE39LoadSheddingAgent.py 

During the training the screen will dump out the training log. After training, you can deactivate the virtual environment by

source deactivate

Two Jupyter notebooks (with Linux and Windows versions-- directory paths are specified differently) are provided as examples for checking training results and testing trained RL model.

If you use this code please cite it as:

@article{huang2019adaptive,
  title={Adaptive Power System Emergency Control using Deep Reinforcement Learning},
  author={Huang, Qiuhua and Huang, Renke and Hao, Weituo and Tan, Jie and Fan, Rui and Huang, Zhenyu},
  journal={arXiv preprint arXiv:1903.03712},
  year={2019}
}

If you spot a bug or have a problem running the code, please open an issue.

Please direct other correspondence to Qiuhua Huang: qiuhua DOT huang AT pnnl DOT gov