NILM: classification VS regression

⚠️ Notice: This project is no longer actively maintained. Feel free to use the existing code, but please note that there won't be further updates or bug fixes.

Table of Contents

1. About The Project
2. Getting Started
   • Create the Environment
3. Datasets
   • UK-DALE
   • Pecan Street Dataport
4. Preprocess the Data
5. Train
   • Reproduce the Paper
   • Thresholding Methods
6. Publications
7. Contact
8. Acknowledgements

About the project

Non-Intrusive Load Monitoring (NILM) aims to predict the status or consumption of domestic appliances in a household only by knowing the aggregated power load. NILM can be formulated as regression problem or most often as a classification problem. Most datasets gathered by smart meters allow to define naturally a regression problem, but the corresponding classification problem is a derived one, since it requires a conversion from the power signal to the status of each device by a thresholding method. We treat three different thresholding methods to perform this task, discussing their differences on various devices from the UK-DALE dataset. We analyze the performance of deep learning state-of-the-art architectures on both the regression and classification problems, introducing criteria to select the most convenient thresholding method.

Getting started

Create the Environment

To create the environment using Conda:

1. Install miniconda

   curl -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh | bash
   

   Say yes to everything and accept default locations. Refresh bash shell with bash -l

2. Update conda

   conda update -n base -c defaults conda
   

3. Clone this repository and cd into the folder

4. Create and activate conda environment (removing previously existing env of the same name)

   conda remove --name nilm-thresholding --all
   conda env create -f environment.yml --force
   conda activate nilm-thresholding
   

Datasets

UK-DALE

UK-DALE dataset is hosted on the following link: https://data.ukedc.rl.ac.uk/browse/edc/efficiency/residential /EnergyConsumption/Domestic/UK-DALE-2017/UK-DALE-FULL-disaggregated

The files needed by this module are ukdale.zip and ukdale.h5.zip. Download both and unzip them in a folder named data inside the root. Once you are done, your local directory should look like this:

nilm-thresholding
|_ nilmth
   |_ [python scripts and subfolders]
|_ data
   |_ ukdale
      |_ [house_1 to house_5]
   |_ ukdale.h5

Credit: Jack Kelly

Pecan Street Dataport

We are aiming to include this dataset in a future release. You can check the issue here: #8

Any help and suggestions are welcome!

Credit: Pecan Street

Preprocess the Data

Once downloaded the raw data from any of the sources above, you must preprocess it. This is done by running the following script:

python nilmth/preprocess.py

This will generate a new folder, named 'data-prep', containing all the preprocessed data.

Train

The folder nilmth contains an executable script to train the models. Run the following line on the root folder (make sure to have the enviroment active and the data downloaded):

python nilmth/train.py

This will train the CONV model using the default parameters. The script stores several outputs in the outputs folder, including:

• .txt files with the model scores over the validation data
• .pth files containing the model weights
• .png files with a section of the validation data and the model's prediction

The list with all the available parameters and their default values is stored in the configuration file.

If you want to use your own set of parameters, duplicate the aforementioned configuration file and modify the paremeters you want to change (without deleting any parameter). You can then use that config file with the following command:

python nilmth/train.py  --path_config <path to your config file>

For more information about the script, run:

python nilmth/train.py  --help

Once the models are trained, test them with:

python nilmth/test.py  --path_config <path to your config file>

Reproduce the Paper

To reproduce the results shown in our paper, activate the environment and then run:

nohup sh run/train_sequential.sh > log.out & 

This will first create a folder named configs, where the different configurations of the models are stored. Then, the script train.py will be called, using each configuration each. This will store the model weights, which will be used again during the test phase:

nohup sh run/test_sequential.sh > log.out & 

Thresholding Methods

There are three threshold methods available. Read our paper to understand how each threshold works.

• 'mp', Middle-Point
• 'vs', Variance-Sensitive
• 'at', Activation Time

Publications

• NILM as a regression versus classification problem: the importance of thresholding

Contact

Daniel Precioso - daniprec - daniel.precioso@uca.es

Project link: https://github.com/UCA-Datalab/nilm-thresholding

ResearhGate link: https://www.researchgate.net/project/NILM-classification-VS-regression

Acknowledgements

• UCA DataLab
• David Gómez-Ullate