• Introduction

  Given that it is difficult to reproduce some NILM algorithms, we create this code repositories implementing some state of the art (or classical) energy disaggregation algorithms by the means of prevailing deep learning toolkit Pytorch [1] and noted NILM toolkit nilmtk [2].

• Set up

  • Create your own virtual environment with Python > 3.6

  • Configure deep learning environment with pytorch (GPU edition) ≥ 1.3.0 + cuDNN

  • Install other necessary dependencies, such as Matplotlib, Scikit-learn etc.

  • Clone this repository (Please notice that code in the folder \nilmtk\. is slightly different from the code in [2], so please clone this folder too)

  The environments we used are listed in the file environment.yml. If you use conda, you can use conda env create -f environment.yml to set up the environment.

• File Directory

  Since the original nilmtk toolkit may seem to be redundant for testing Neural NILM algorithms (Deep Learning method), we sorely use it for the generation of power data within a specific period of time. Thus you can only focus on these files or folders: \nilmtk\api.py, \nilmtk\loss.py, \nilmtk\disaggregate, \tutorial\experiment_example.ipynb and \tutorial\code_example.ipynb

  The whole file directory is as follow (We omit some unimportant details)：

  ├── README.md							                   
  ├── environment.yml						//Environment dependencies
  ├── nilm_metadata          				
  │   └── *								//Some details are omitted
  ├── tutorial                     
  │   ├── experiment_example.ipynb     	//How to carry out your own NILM experiment
  │   ├── code_example.ipynb			    //How to write your own algorithms or metrics
  ├── nilmtk                     
  │   ├── api.py							//The core api to carry out NILM experiment
  │   ├── loss.py							//The evaluation Metrics
  │   ├── diaggregate
  |   |	├── __init__.py
  |   |	├── attention_pytorch.py		   //Seq2Seq with Attention
  |   |	├── bilstm_pytorch.py			   //BiLSTM
  |   |	├── dae_pytorch.py				   //Denoising AutoEncoder
  |   |	├── disaggregator.py			   //Base Class
  |   |	├── energan_pytorch.py			   //EnerGAN
  |   |	├── seq2point_pytorch.py		   //Seq2Point
  |   |   ├── attention_cnn_pytorch.py       //CNN_Attention
  |   |   ├── seq2seqcnn_pytorch.py          //CNN_Seq2Seq
  |   |   ├── bilstm_pytorch_multidim.py     //Multiple input features BiLSTM
  |   |   ├── dae_pytorch_multidim.py        //Multiple input features DAE
  |   |   ├── seq2point_pytorch_multidim.py  //Multiple input features Seq2Point
  |   |	└── sgn_pytorch.py				   //SGN
  │   └── *								   //Some details are omitted
  

• Algorithm Details

  In the folder \nilmtk\disaggregate, you may find several NILM algorithms, they are listed as follow:

  • Denoising AutoEncoder [3]
  • BiLSTM [3]
  • Seq2Point [4]
  • Seq2Seq [4]
  • Seq2Seq with Attention [5]
  • SGN [6]
  • EnerGAN [7]
  • CNN_Attention[8]

And several NILM algorithms with '_multidim' suffix, such as bilstm_pytorch_multidim, ae_pytorch_multidim, seq2point_pytorch_multidim. They are original algorithms with multiple input features(P or P + Q or P + S O or P + Q + S), which are not included in nilmtk[2]

Notice that our implementations of BiLSTM[3] and EnerGAN[7] are slightly different from original papers, and the experiment results have shown that our method will result in improved accuracy. To avoid confusing users, we will list our implementation details as follow:

• For BiLSTM, the input of the first fully connected layer is the concat of all the hidden states instead of the last hidden state which was the way Kelly used.
• For EnerGAN, we incorporate reconstruction loss(L1 loss) while training Generator, which is proved to be valid in most pix2pix tasks[9].

• Tutorial

  • Refer to \tutorial\experiment_example.ipynb to know how to carry out your own NILM experiment.

  • Refer to \tutorial\code_example.ipynb to know how to write your own energy disaggregation algorithms or evaluation metrics.

• References

  [1] https://pytorch.org/

  [2] https://github.com/nilmtk/nilmtk

  [3] KELLY J, KNOTTENBELT W. Neural NILM: Deep neural networks applied to energy disaggregation[C]//The 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments. November 4-5, 2015, Seoul, South Korea: 55-64.

  [4] ZHANG C, ZHONG M, WANG Z, et al. Sequence-to-point learning with neural networks for non-intrusive load monitoring[C]//The 32nd AAAI Conference on Artificial Intelligence. February 2–7, 2018, Louisiana, USA: 2604–2611.

  [5] WANG Ke, ZHONG Haiwang, YU Nanpeng, et al. Nonintrusive Load Monitoring based on Sequence-to-sequence Model With Attention Mechanism[J]. Proceedings of the CSEE,2019,39(1):75-83.

  [6] SHIN C, JOO S, YIM J. Subtask Gated Networks for Non-Intrusive Load Monitoring[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2019, 33(1): 1150–1157.

  [7] KASELIMI M, VOULODIMOS A, PROTOPAPADAKIS E, et al. EnerGAN: A GENERATIVE ADVERSARIAL NETWORK FOR ENERGY DISAGGREGATION[C/OL]//ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Barcelona, Spain: IEEE, 2020: 1578–1582.

  [8] XU Xiaohui, ZHAO Shutao, CUI Kebin. Non-intrusive Load Disaggregate Algorithm Based on Convolutional Block Attention Module[J/0L]. Power System Technology, 2021: 1-8.

  [9] ISOLA P, ZHU JY, ZHOU T, et al. Image-to-Image Translation with Conditional Adversarial Networks[C/OL]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 5967–5976.

  ---

• Future Work To Do

  Override some of the functions which are repeated in the current framework with sealed APIs.