NNILM - Neural Non-Intrusive Load Monitoring

This project is a reimplementation of Jack Kelly's rectangles architecture based on Keras and the NILMToolkit.

Network overview

Benchmark comparison

A comparison of the rectangles architecture in this repository with the original implementation by Jack Kelly on the UK-DALE dataset.

Predictions

Predictions for a true positive, a true negative and a false positive example for a dish washer.

Visualization of first convolution layer

Learned filters

The following image shows the 16 learned filters of the first convolution layer.

Activations for a negative sample

The following image shows the activations of the first convolution layer given a negative sample.

Activations for a positive sample

The following image shows the activations of the first convolution layer given a positive sample. The two highlighted filters are thereby interesting. The first filter seems to have learned to detect step changes in the input signal. The second filter seems to have learned to pass the raw input signal to the next layer.

Development

Setup virtualenv

We used Python 2 as programming language, because there are some dependencies that are not compatible with Python 3.

• pip install virtualenv

• python -m virtualenv env

• source env/bin/activate

If you set up virtualenv with Python 3, you have to switch to Python 2: virtualenv --python=/usr/bin/python2.7 env

Install dependencies

Because of a failing dependency hmmlearn, we use the option --no-dependencies. The module hmmlearn is not really needed in this implementation.

• pip install --no-dependencies -r requirements.txt
• pip install -e .

Leave virtualenv

• deactivate

Dataset

It is possible to use any dataset supported by the NILMToolkit. You have to convert the dataset to a .h5 file using the converters provided by the NILMToolkit and then place it into the folder ./data: https://github.com/nilmtk/nilmtk/tree/master/nilmtk/dataset_converters

Run training

Run training for 30 epochs: python nnilm/train.py dish_washer_redd -s0 -e30

Resume training from epoch 30: python nnilm/train.py dish_washer_redd -s30 -e40

Run training on GPU cluster: nohup python nnilm/experiments/train_hpc.py --gpu=2 &

Run prediction

Public dataset: python predict.py dish_washer_eco

CSV prediction: python nnilm/experiments/csv_predictor.py -ddish_washer_redd -idata/aggregated_power.csv -s01-04-2018 -e08-04-2018

Experiments

The folder nnilm/experiments contains files that have been used to run my experiments. It shows how the scripts from this repository can be used in another project.