for Non-Intrusive Load Monitoring

Summary: This is a code repository for the paper Concurrent Loads Disaggregator for Non-Intrusive Load Monitoring (arXiv:2106.02352). We developed the Synthesizer of Normalized Signatures (SNS) algorithm to simulate the aggregated consumption of a various number of appliances. This algorithm takes signatures of individual appliances (i.e. current and voltage) and converts them into European system (311 V, 50 Hz etc.). Then, the schedule (i.e. which appliances will work together) is randomly sampled from uniform distribution. Finally, corresponding signatures are being aligned by the reference voltage and summed up. This repo also includes implementation of the neural network architecture proposed. We wrap the training & testing pipeline into the Pytorch Lightning framework.

Keywords: signal decomposition, energy disaggregation, nilm, simultaneous appliances, concurrent loads, synthetic data, neural network

Clone this repository:

git clone https://github.com/arx7ti/cold-nilm.git

The required packages can be installed via standard Python package manager:

pip install --upgrade pip
pip install -r requirements.txt

Tree of the project

notebooks
├── Results.ipynb
└── Statistics.ipynb
configs
├── sns.ini
├── tuner.ini
└── model.ini
source
├── __init__.py
├── wrappers.py
├── synthesizer.py
├── data.py
├── model.py
└── utils.py
main.py

This repository contains following configuration files

1. configs/sns.ini defines parameters of the SNS algorithm to fully reproduce the dataset simulated in the paper;
2. configs/tuner.ini defines hyper-parameters space for the model selection procedure as well as meta parameters of the model (w_max, augmentation etc.);
3. configs/model.ini defines optimal hyper-parameters obtained after model selection procedure. This file will be rewritten automatically once the tuner is done.

NOTE: For logging purposes during the training you can add neptune_api_key and neptune_project keys to the [extra] section of the configs/tuner.ini or configs/model.ini and pass the argument -tl neptune.

The main.py has 3 sections:

1. run-sns to generate the synthetic datasets for train, validation and test:

python main.py run-sns -j -1

2. run-tuner [optional] to find the best model hyper-parameters. It took us about 1 day to iterate over the 100 trials on 2 RTX 2080Ti GPUs and Ryzen 3900X CPU.

python main.py run-tuner -j 5 -g 0.5 -t 100 -e 100 -es 10 -gp 5 -dl y

3. run-model can be used to train the model from scratch with hyper-parameters set in the configs/model.ini:

python main.py run-model -j -1 -g 2 -e 100 -es 10 -dl y --train y

To resume the training after an interruption pass the following flag: --resume y.

To test the model with weights given in the configuration file:

python main.py run-model -j -1 -g 1 

Download weights from Google Drive [64 MB]

You are also allowed to test the model under different values of w_max:

python main.py run-model -j -1 -g 1 --override-w-max 3

Once the model is tested the plots will be available in the model's folder.

Several Jupyter notebooks are available in the notebooks directory:

1. notebooks/Results.ipynb comprises values obtained in the Results section of the paper;
2. notebooks/Statistics.ipynb shows more detailed statistics on the UK-DALE and REDD datasets.

Fig. 2 Example of the synthesized measurement signal and its spectrogram

Training from scratch takes 1.5-2.5 hours with use of 2 RTX 2080Ti. Example of the output after the model's test:

test_weighted_mean_f1 = 82.86%
test_mean_f1_w=1 = 94.63%
test_mean_f1_w=2 = 92.40%
test_mean_f1_w=3 = 89.09%
test_mean_f1_w=4 = 86.24%
test_mean_f1_w=5 = 83.15%
test_mean_f1_w=6 = 80.84%
test_mean_f1_w=7 = 78.57%
test_mean_f1_w=8 = 76.31%
test_mean_f1_w=9 = 74.58%
test_mean_f1_w=10 = 72.79%
...
Under the optimal threshold = 0.6552

Comparison table

Fig. 3 Distribution of mean F1-score across different number of concurrent appliances

Fig. 4 Distribution of F1-score across different appliances

• configs/model.ini: the model hyper-parameters define a hash value to prevent the models for being overlapped during multiple training/testing attempts. The threshold option doesn't contribute to the hash value, this parameter is being computed during the validation and established for the test procedure. It can be replaced by your own value, that is the test results will be given under your threshold. Option weights is used only with argument --train n and will override the optimal weights if any (best checkpoint still preserved, but not used during that test).

• run-model: --override-w-max will override w_max of the ModelCOLD class only for the current test (even the test which goes right after the training).

Paper is under the review, the preprint is available at:

@misc{kamyshev2021cold,
      title={COLD: Concurrent Loads Disaggregator for Non-Intrusive Load Monitoring},
      author={Ilia Kamyshev and Dmitrii Kriukov and Elena Gryazina},
      year={2021},
      eprint={2106.02352},
      archivePrefix={arXiv},
      primaryClass={eess.SP}
}

[1] Faustine, Anthony, and Lucas Pereira. "Multi-Label Learning for Appliance Recognition in NILM Using Fryze-Current Decomposition and Convolutional Neural Network." Energies 13.16 (2020): 4154.