Deep Neural Networks Applied to Energy Disaggregation
Please note that this code was never finished! It is incomplete! Please do not use unless you really, really know what you are doing! Please see my Neural NILM prototype code to see exactly what I implemented in my Neural NILM paper! Your best bet if you want to re-implement Neural NILM would probably be start from scratch using a modern framework like TensorFlow, and perhaps use nilmtk to help load the datasets.
This is a complete re-write of the Neural NILM Prototype.
Here is a short iPython notebook to show how to load activations from UKDALE.
Please also note that my neuralnilm code is only provided as a proof of concept and as a reference. It isn't yet in a state where it's easy for other people to run the code. Unfortunately I won't have time to work on the Neural NILM code or to provide support anytime in the forseable future. So the code is provided 'as is' without any support at the moment. Maybe one day I'll have time to support the code or to update the code.
That said, if you're really determined to read and understand this code then here's a quite description of the basic architecture:
Each 'experiment' is defined in a Python script in
experiment_definitions/
. Each of these files must have a run
function which, well, runs the experiment!
experiment_definitions/
also should include a text file called
job_list.txt
which is just a list of experiment names like e575
.
The script scripts/run_experiments.py
looks at job_list.txt
to
find the next job to run. After each job completes,
run_experiments.py
removes that job from the job_list
. In this
way, it is possible to add, remove or re-order jobs from the
job_list
while an experiment is running.
The re-usable code which makes up the bulk of the NeuralNILM package
lives in neuralnilm/neuralnilm
. net.py
includes a Net
class
which handles the construction of each neural network. trainer.py
includes a Trainer
class which trains each net and sends metrics to
the Mongo database.
neuralnilm/neuralnilm/data
includes the mechanisms for loading and
transforming data. This is designed to be very modular so you can
easily mix and match data sources, pre-processing steps and networks.
For a quick overview of how this all works, take a look at some of the
python files in the experiment_definitions
directory.
neuralnilm/neuralnilm/monitor
contains the mechanisms for loading
metrics and metadata from the Mongo database and plotting it.
See requirements.txt
for the required Python packages. Also requires Mongo DB.
You need to create a ~/.neuralnilm
text file to specify some
parameters. Please see example_config_file
.
See http://www.mkyong.com/mongodb/mongodb-allow-remote-access/