active_env is a Python toolkit that applies
reinforcement learning algorithms in an electric power system. The project was
originally a master thesis written in partnership with Statnett,
the Norwegian Transmission system operator.
The agent is allowed to modify the power consumption at nodes in a
distribution system with high peak demand and high production from solar power.
The goal of the agent is to reduce the number of current and
voltage violations in the grid by increasing/ decreasing the consumption
at appropriate times. The increase/decrease in power consumption is intended
to be a simplified program for demand response that exploits the residential
flexibility in the grid.
pandapower is used
for power flow calculations and
stable-baselines
for reinforcement learning.
active_env inherits the prerequisites from stable-baselines and pandapower.
You need python >=3.5. You can find installation guides here:
ActiveEnv is custom gym environment that can interact with reinforcement algorithms in stable-baselines. A PPO agent is trained in this example:
from stable_baselines.common.vec_env.dummy_vec_env import DummyVecEnv
from stable_baselines import PPO2
from stable_baselines.common.policies import MlpPolicy
from active_env.envs.active_network_env import ActiveEnv
env = ActiveEnv()
env = DummyVecEnv([lambda: env])
model = PPO2(MlpPolicy, env, verbose=1)
model.learn(total_timesteps=100000)active_env has a custom gym environment called ActiveEnv that can interact
with reinforcement algorithms in stable-baselines. Check out the
notebook tutorial
for an introduction to the environment, and its functionality.
The master thesis can be found here in the file solberg2019.pdf