An Experiment Framework for Multi-label NILM with Neural Networks

This repository defines an experiment framework for multi-label classification on NILM and is used to create and evaluate a baseline neural network architecture set.

Data

This project supports the following datasets, using the NILMTK toolkit to read and preprocess the power recordings:

• UKDALE, which records the power demand from five houses in the UK.
• REDD, which contains several weeks of power data from 6 houses in the USA.
• iAWE, which contains power data from 1 house in India.

Neural Network Architectures

To be analysed.

• Fully Convolutional Network (FCN), from Deep learning for time series classification: a review, Fawaz et al. (2019).
• Residual Network (ResNet), from Deep learning for time series classification: a review, Fawaz et al. (2019).
• Convolutional GRU (ConvGRU), from A Dual-input Multi-label Classification Approach for NILM via Deep Learning, Çimen et al. (2020).
• Convolutional LSTM (ConvLSTM), from A Dual-input Multi-label Classification Approach for NILM via Deep Learning, Çimen et al. (2020).
• Fully Convolutional Autoencoder (FCN-AE), a simplification, adaptation from Non Intrusive Load Disaggregation by CNN and Multilabel Classification, Massidda et al. (2020).

Experiment Framework

The experiment framework scenarios used in this project are an adaptation of the following publication:

Symeonidis, Nikolaos & Nalmpantis, Christoforos & Vrakas, Dimitris. (2019). A Benchmark Framework to Evaluate Energy Disaggregation Solutions. 10.1007/978-3-030-20257-6_2.

Project Structure

The project structure is defined as follows:

• 📂 data: Includes modules related to data e.g. loading data using NILMTK, detecting appliance activations and aligning meters.
  • environment: defines the environment params for the training, validation and test sets.
  • generator: defines 2 wrapper functions used to load the data for the experiments, Seq2Seq and Seq2Point.
• 📂 models: Contains 5 neural network architectures used in the experiments
• 📂 experiments: Includes modules to configure and run the 4 experiment scenarios
  • experiments: run experiments, save detailed results
  • metrics: defines the multi-label metrics and reports
• 📂 results: Results of the experiments are saved in this directory.
• 📂 utils: Includes the utility functions
  • path_finder: defines a simple path manager

Requirements

The dependecies of this code are listed as follows:

• python>=3.6
• Cython>=0.27.3
• bottleneck>=1.2.1
• numpy>=1.13.3
• numexpr>=2.6.4
• pytables
• pandas>=0.25.3,<1.0
• matplotlib>=3.1.0,<3.2.0
• networkx==2.1
• scipy>=1.0.0
• scikit-learn>=0.21.2
• jupyter
• ipython
• ipykernel
• nose
• coverage
• pip
• psycopg2
• coveralls
• nilm_metadata
• nilmtk
• hmmlearn
• pytorch
• pyarrow
• seaborn