mk-home-automation is an additional service for the MK Home ecosystem. It allows to receives the protocoled Tradfri light information from mk-home-server and trains as model to predict weather a specific light bulb is turned on or off for a specific weekday and daytime.

It uses the tracked data of the TradfriDataCollection module of mk-home-server to do so. And trains a gradient boost model based on your own behavior of switching your connected lights in your home.

While this service alone only supports the training and prediction of light states for specific weekday/daytime combinations, the Automator module of mk-home-server enables to automatically control lights based on the prediction of this project in different modes.

To learn more about how mk-home-automation works look at section X of this README.

To learn more about how to use the automation functions in combination with mk-home-server look at the specific section it the README of that project.

• Python 3.8
• Rquired pip packages installed on your system (see Pipfile)

Optional:

• pipenv
• pm2
• Docker
• Jupyter

There are different options to run this project. While it is created to work natively with a python environment on raspberry pi, it is also possible to use it with pm2, Docker or directly on your system in combination with pipenv.

To discover or enhance the base machine learning model and data aggregation you also can use the attached jupyter notebook in this repository. The important data modelation, and feature optimization used in the application code is originated there.

Because the training and prediction functionality of this project is encapsulated in a REST API (using Flask), it is recommended to use gunicorn to run it reliable and safe in production. This purpose is already included into the pm2 and Docker setup shown below.

To run mk-home-automation with pm2 you need to install pipenv and pm2. After doing so, jump into the root folder of this project and install the required python version and pip packages which are needed with pipenv:

pipenv install

After having setup all the requirements you can just start the application and the build-in REST server with:

pm2 start ecosystem.config.yml

To run this service permanently use:

pm2 save

To build the project via Docker run from the root folder:

docker build -t mkha .

To start it then use:

docker run -it -p 9090:9090 mkha:latest

If you want to configure the outgoing port or other things, feel free to change it inside the Dockerfile. Other project related configuration you can find and change in the .env.docker file.

If you want to run the project directly via pipenv use:

pipenv install

And to start the project:

pipenv run gunicorn --bind 0.0.0.0:9090 --pythonpath 'app' server:app"

When using mk-home-automation for development purposes, you can use pipenv and start the application directly with:

pipenv run python app/server.py

If not all dependencies are installed correctly for the virtual environment remember to use pipenv install first.

When having DEBBUGING set to True in the .envfile, flask also is doing hot reloading on code changes.

To explore or modify the steps of creating and optimizing the model used in this project, you can directly use the attached jupyter notebook in this repository.

To run it just go in the related directory with:

cd jupyter

and start the notebook with:

jupyter notebook

For now all configuration for the project is maintained in .env files. For development and production use the .env file, for the docker setup use the .env.docker file.

The functionality of training the tracked light interactions from mk-home-server or predicting the on/off state of a specific light at a concrete weekday/daytime combination is exposed through an REST API by the application.

To use it there are two routes integrated in the service which are accessible by everyone who can connect to this service.

To train the model use the /train route of the API with a POST request and an attached csv file with all the light interaction data comes from the mk-home-server application.

After training a new model based on the input data, the created model get saved in the file system to reuse it later on.

To predict a light state of specific light at a provided weekday/daytime combination use the /predict route with a POST request.

You also need to provide the needed information through a JSON body attached to the request. That information must have the following form:

{
    "instanceid": 65537,
    "type": 2,
    "weekday": "Thursday",
    "time": "16:23"
}