WaveNILM

WaveNILM implementation in Keras, with Tensorflow backend as described in:

A. Harell, S. Makonin and I. V. Bajić, "Wavenilm: A Causal Neural Network for Power Disaggregation from the Complex Power Signal," ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Brighton, United Kingdom, 2019, pp. 8335-8339. doi: 10.1109/ICASSP.2019.8682543

If using this code in any research, please cite the original paper, seen above.

Dependencies:

• Keras 2.2.2
• Tensorflow 1.8.0
• Sacred 0.7.4

Installation instructions:

The code was developed using pipenv*, and the environment file is included in the repository

• Create a folder for the work environment
• Copy pipfile.lock to the folder
• Run

pipenv install --ignore-pipfile

• Start the work environment using

pipenv shell

• Clone repository to the work environement folder
• Follow instruction in data/readme.md for copying or creating dataset files

* To install pipenv itself run:

pip install pipenv

Running instructions:

Turn on environment using pipenv shell. then run code from src folder as follows:

python waveNILM.py with <\optimizer name here> <config change values here>'

Currently configured optimizers include 'adam' and 'SGD', to include other optimizers, or to modify the optimizer parameters, see src/waveNILM.py'. Other controlable parameters include training length in samples, number of epochs to train, a cross validation flag, and more. For a full list of the parameters see 'src/waveNILM.py'

Some example runing parameters: To run the noisy scenario on the full AMPds2 dataset, with active power as both input and output, with cross validation and adam optimizer simply run

python waveNILM.py with adam

To run the same scenario but only on the first year of AMPds2, run:

python waveNILM.py with adam trn_len=525600

To run a denoised scenario on appliances number 8,14,16, with all other parameters the same, run:

python waveNILM.py with adam app_inds=[8,14,16] noise_mode=0