americast / drl_hvac

Optimising electricity expenditure in an HVAC system under dynamic electricity pricing scheme and weather conditions using a DDPG model.

DRL-based control for HVAC systems

Optimising electricity expenditure in an HVAC system under dynamic electricity pricing scheme and weather conditions.

Requirements

Python 3.6 or 3.7 required. Matlab R2014b+ required.

Python package requirements:

pudb
gym
torch

Matlab addon requirements:

optimtool
Control System Toolbox

Install the python packages using

$ pip3 install pudb
$ pip3 install gym
$ pip3 install torch==1.3.1+cpu torchvision==0.4.2+cpu -f https://download.pytorch.org/whl/torch_stable.html # CPU-only variant

Also, install the Matlab Engine API for Python. Instruction can be found here.

Dataset

Data is generated via the Matlab files. They are called directly from Python.

Installation

Go to the DRL/custom_gym directory and issue $ pip3 install -e .

Training

All the models shall be saved in the DRL/models/ directory.

DRL-based controller

Go to the DRL directory and issue $ python3 driver.py -p <port no> to train. Seven RL models run in parallel at ports provided at command-line flag, as well as the next six ports. A small help section may be accessed using python3 driver.py -h.
The code will run for 100 episodes, and keep plotting the reward as DRL/models/figs/updates_<zone>.png at every step.
The best model for each zone shall be saved in the DRL/models/ directory.

DRL optimiser for MPC-based controller

Go to the DRL directory and issue $ python3 driver.py -c to train.
The code will run for 100 episodes, and keep plotting the reward as DRL/models/figs/updates_combo.png at every step.
The best model for each zone shall be saved in the DRL/models/ directory.

Inference

All the results shall be stored in DRL/results/ directory.

DRL-based model

Add a flag -i to the command to perform inference i.e., $ python3 driver.py -p <port no> -i.

DRL optimiser for MPC-based controller

Again, add a flag -i to the command to perform inference i.e., $ python3 driver.py -c -i. The results shall be stored in DRL/results/ directory.

Results

Only MPC-based controller.

Add a flag -m to obtain results from the MPC-based controller i.e., $ python3 driver.py -m from the DRL/ directory. The results shall be stored in DRL/results/ directory.

Obtain combined plots

In order to obtain plots (and make the results easily interpretable), issue the command python3 plot_gen.py -c -m. Again, all the commands are to be issued from the DRL/ directory. Here flag -c ensures the results from the DRL+MPC-based are considered while generating the plots. -m adds results from the only-MPC based controller.

Trained models and results

Trained model files and interpretable results (plots only) have been provided in the DRL/models/ and DRL/results/ directory respectively. In order to obtain the full results, run the commands as instructed above.

Description of the files

DRL/driver.py: Calls the required modules: (i) DRL-based controller: the seven RL models one by one in parallel, generates weather data for each step and triggers the matlab code to start, or (ii) MPC-based: Calls Matlab directly and runs them, (iii) DRL+MPC: Calls the combined python module which calls Matlab as and when needed, or (iv) Testing: Performs inference for either of them.
DRL/caller.py: The matlab code sends the state data to this file, which in turn sends them to the RL models which wait for the data over TCP sockets.
DRL/main.py: Calls all the other modules and runs them by and by.
DRL/model.py: Contains DL models for actor and critic.
DRL/DDPG.py: Calls and updates the actor and critic modules.
DRL/utils.py: Addon implementations (noise, memory and normalisation).
DRL/plot_gen.py: Generates plots for easy interpretation of the results.
DRL/custom_env.py: Environment (the step function contains action implementation and reward) for running the DRL-based controller.
DRL/custom_env_combo.py: Environment for running the DRL+MPC model. DRL/combo.py: Runs the DRL+MPC model

drl_hvac's Issues

article

Hello, may I ask if the code is reproduced? If it is reproduced, is there a corresponding article?