AlphaBuilding-MedOffice

This is the official repository of AlphaBuilding MedOffice: a realistic OpenAI Gym environment that can be used to train, test and benchmark controllers for medium size office (1AHU + 9VAV boxes)

Overview

We developed AlphaBuilding MedOffice, a virtual testbed of medium size office that is wrapped up with OpenAI Gym. The building model is developed from [EnergyPlus model of medium size office of DOE Commercial Reference Building type]. We are controlling the second floor of this building, which has 1 ATU serving 9 VAV boxes (2 for conference rooms, 3 for closed offices, 4 for open offices).

In addition to the virtual environment, we also implement and open source several RL algorithms (Deep Deterministic Policy Gradient (DDPG)), which could be used as benchmarks for algorithm comparison.

Code Usage

Clone repository

git clone https://github.com/WalterZWang/AlphaBuilding-MedOffice.git
cd AlphaBuilding-MedOffice

Test the environment

Pull the docker image: $ docker pull walterzwang/drl_eplus:9.2
$ cd AlphaBuilding/docker/drl, revise parameter VOLUMN_PATH in the makefile, then $ make run
Run scripts:
- test fmu: $ cd gym_AlphaBuilding, runpython fmuModel/test_fmu.py
- test gym: $ cd gym_AlphaBuilding, runpython test_gym.py
Run DRL algorithms:
- All scripts must be run from the root repository

Repository structure

analysis: code to analyze the result

docker: scripts to develop the docker image

docs: Manuscript submitted version, figs in the paper, and documentation of this environment

gym_AlphaBuilding: scripts to develop the virtual simulation environment

lib: DRL models and utility functions

log: log files for TensorBoard visualization

RL: scripts for DRL

Running

Interface

The co-simulation workflow between the environment and the controller is:

Environment initialization

The first step is to initiate the environment. A list of input parameters to set up the environment is shown as follows:

Input Parameter	Description	Data Type	Default Value
building_path	Path to the FMU model	str
sim_days	Number of days for simulation, start from Jan. 01	int
step_size	Size of each time step for control and simulation, unit [s]	int
sim_year	Year for simulation, used to create time index	int	2015
tz_name	Thermal Time Constant, mean and std	tuple	'America/Los_Angeles'
eprice_path	Path to the utility price signal, not used for energy conservation	str	None

States, actions, and info/comments

Then at each time step, the environment consumes acts (determined by the controller), runs a single time step simulation, and returns states and comments. The contents of acts, states, and comments are shown as follows:

HourOfDay: in the range of 0 to 23

DayOfWeek: in the range of 0 to 6, 0 is Monday, 6 is Sunday

Methods

A list of methods provided by the environment is as follows:

Methods	Example	Input	Returns
Initiate the env	`env=AlphaResEnv(*args)`	Env parameters	An environment instance
Reset the env	`obs=env.reset()`	\	Observations (Initial conditions)
One step simulation	`obs,r,d,c=env.step(a_scaled)`	Control actions	New observations, reward, done flag, comments
Scale states to [-1,1]	`s_scaled = env.scale_state(s_raw)`	Raw states	Scaled states
Scale actions to [-1,1]	`a_scaled = env.scale_action(a_raw)`	Raw actions	Scaled actions
Rescale states to raw	`s_raw = env.rescale_state(s_scaled)`	Scaled states	Raw states
Rescale actions to raw	`a_raw = env.rescale_action(a_scaled)`	Scaled actions	Raw actions

The input of the step function needs to be scaled to [-1, 1]. Users can use the env.rescale method to get the actual value

Application

AlphaBuilding MedOffice can be used to train controllers for commercial buildings. As an example, we demonstrate four applications here:

Feedback

Feel free to send any questions/feedback to: Zhe Wang or Tianzhen Hong

Citation

If you use our code, please cite us as follows:

mingzhe37 / alphabuilding-medoffice Goto Github PK