ComfortLearn

ComfortLearn is an OpenAI Gym-like environment that leverages historical Building Management System (BMS) data from real buildings and existing longitudinal thermal comfort datasets for Occupant-Centric Control strategies and benhmarking.

This platform is developed by Matias Quintana.

Contributing

We welcome contributions in different forms, from bug reports to pull requests. We encourage the community to use, improve, and extend ComfortLearn by adding new zones and energy functions models and sharing the implementation of control algorithms, as well as the inclusion of more longitudinal thermal comfort datasets. For major changes, please open an issue to discuss how you would like to change the platform.

Requirements

Refer to requirements.txt for a list of Python library dependencies. You may install the required libraries by executing the following command:

pip install -r requirements.txt

Documentation

Files and Folders

• main.py: Main file to execute the environment and the simulation.
• comfortlearn.py: Contains the ComfortLearn environment and the class Occupant and functions zone_loader() and occupant_loader().
• common/: Contains files for preprocessings and utility functions.
• img/: Contains all generated images and figures.

Modifyable by the user

• zones_state_action_space.json: JSON file containing the possible states and actions for every zones building.
• zone_attributes.json: JSON file containing the attributes of the zones. These zones can be dependant on the empiricald thermal comfort datasets available and are located within the respective data/ folder.
• energy_models.py: Contains the class Zone which is called by the ComfortLearn class.
• agent.py: File that contains the agent class that will learn to control the Zone(s) temperature set-point.
• reward_function.py: Contains the function reward_function_sa() to help the controller find an optimal control policy.
• occupant_data_preprocessing_<DATASET>.ipynb: Notebook with preprocessing and cleaning of a respective <DATASET>.
• zone_data_preprocessing.ipynb: Notebook with preprocessing and cleaning of Zone(s) data and weather data.
• results_loop.ipynb: Notebook that analyses the simulation results.
• train_pcm.py: Defines and trains Personal Comfort Models (PCM) for occupants in a dataset in data/.
• data/: Contains the different thermal preference datasets.
• configs/: Contain configuration files with the following:

Variable	Values	Description
config_name	String	Experiment name
seed	Int	Seed value
data_path	String	Empirical datasets folder path
num_new_occupants	Int	Number of occupants as agents
zone_selection	random, forced	Occupant movement between zones
occupant_timing	fixed, stochastic	Occupant's enter and leave time
occupant_tolerance	Float	Number within $[0, 1]$ for filtering occupants
occupant_tol_file	String	CSV with occupants and their tolerance path
occupant_preference	String	CSV with occupants thermal preference path
occupant_background	String	CSV with occupants metadata path
occupant_pcm	Pickle	Path to Pickle file with occupants PCM
zone_attributes	String	Zone(s) attributes JSON path
weather_file	String	Weather CSV path
zones_states_actions	String	Zone(s) state action space JSON path
max_steps	Int	Number of steps for the simulation
cost_function	List of Strings	Name of the evaluation metrics to use
central_agent	Boolean	True: Central agent, False: Multi-agent
save_memory	Boolean	For RL's agent's buffer
verbose	Boolean	True: Debugging mode
agent_type	String	Name of the class of agent to use

PCM

Each occupant in the occupant dataset had a Random Forest model trained on its data using a 5-fold cross-validation with F1-micro as evaluation metric.

Details about the modeling are stored within the config/folder. Example:

python train_pcm.py configs/train_pcm_<DATASET>.yaml

This procedure is taken from Quintana, M., Schiavon, S., Tartarini, F., Kim, J., & Miller, C. (2022). Cohort comfort models--Using occupants' similarity to predict personal thermal preference with less data. arXiv preprint arXiv:2208.03078.

Simulation

To execute one full run of the simulation, simply run:

python main.py <your_config_file>.yaml

You can also manually set the seed directly as an argument:

python main.py <your_config_file>.yaml 10

In this case the seed used throughout the simulation is 10

Bash scripts are also included to run simulations with different zones and/or different occupant profiles.

If you want to run a specific experiment, make sure to create a yaml file with the different paramters for the envionrment and agent. Examples of this can be found in the configs folder. The script ./run_all_zones_loop.sh 010 runs simulations on both temperature zones with occupants with a tolerance of 10%

Case study: Occupant profiles in different temperature zones

We present a real-world case study of ComfortLearn where we use historical data from a real building and data from existing longitudinal thermal comfort datasets. The building and thermal comfort datasets were obtained in Singapore (tropical climate), classified by ASHRAE as climate zone 1A. In total, two rooms with different ventilation mechanism from an educational building were chosen. These rooms are university classrooms that host students throughout the academic year and environmental data is continuously being logged by the building's BMS system. The period of analysis is one full work week, 2021-10-18 to 2021-10-22, middle of the fall semester, and time steps of 15 minutes are chosen. The dataset can be download here.

More simulation parameters and their values chosen for this case study are:

Parameter	Value
Occupant arrival time	Sampled from $\mathcal{N}(\mu=8, \sigma^{2}=2)$
Occupant departure time	Sampled from $\mathcal{N}(\mu=17, \sigma^{2}=2)$
Num. of occupants	10
Tolerance parameter	[0.10, 0.30, 0.60, 1.00]
Num. Temperature zones	2
Time duration	One full week
Time steps	672 (15 minutes intervals)
Iterations	30

Results

Daily average UNC values for each occupant averaged over 30 simulations for occupants with tolerance of $0.6$. The lower the better.

Time-series of UNC values throughout a week for a single occupant of tolerance $1.0$ (top) and $0.6$ (bottom). Darker line is the mean over time with shaded region represented by 2-standard deviation and a dashed line at a UNC value of $0.5$. The lower the better.

Learning

As of the release of ComfortLearn, building control strategies have not yet been benchmarked nor implemented. The case study presented so far focuses on evaluation buildings historical data.

Nevertheless, the files that would require modification for such implementation are the following: Main editable files for own submissions:

• agent.py
• reward_function.py

Cite ComfortLearn

Quintana, M., Nagy, Z., Tartarini, F., Schiavon, S., & Miller, C. (2022). ComfortLearn: Enabling agent-based occupant-centric building controls. Third ACM SIGEnergy Workshop on Reinforcement Learning for Energy Management in Buildings & Cities (RLEM) (RLEM ’22), 4. https://doi.org/10.1145/3563357.3566167

License

MIT License

Copyright (c) 2022 Building and Urban Data Science (BUDS) Group and National University of Singapore (NUS)

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.