This repository contains a processed sample of the NILM-UY dataset and an implementation of the algorithms proposed in NILM: Multivariate DNN performance analysis with high frequency features. Instructions on how to get access to the full raw data is also included.
- Pre-processing algorithms for the UK-Dale dataset.
These algorithms are based on Neural NILM (Jack Kelly). These functions are useful for reading and pre-processing the
h5file from UK-Dale. - Training and evaluation scripts for the models proposed in the paper.
The pre-processing notebook serves as an example on how to process the UK-Dale dataset for training the dissagregation models.
The training and evaluation procedure is divided into three notebooks.
-
Training notebook. This notebook contains the code for training a single model for one appliance. We also included an script for training all the architectures for all the considered appliances.
-
Metrics notebook. This notebook allows you to load a previously trained model and calculates the metrics reported in the paper. AUC, Recall, Precision, Accuracy, False Positive Rate, F1-Score, Reite, MAE.
-
Rolling windows evaluation. This notebook can be used for evaluating the previously trained models with the rolling windows approach. Contains the code for loading the whole power time series and making predictions in a rolling window fashion.
- Pre-processed UK-Dale dataset ready for being used with the code base provided in this project. The dataset is already temporarily splitted: data_ini.pickle and data_fin.pickle
- Pre-processed NILM-UY dataset. The uruguayan dataset already pre-processed with a sampling period of 6 seconds. This data is also ready for being used with the code provided in this project. datos_uruguay.pickle
- Trained weights of the models pesos.zip
Alternative Google Drive link:
Data
The raw NILM-UY dataset collected in Uruguay contains aggregated and disaggregated data:
- High sampling frequency aggregated data from two homes. (140 gb)
- Individual power measurements per appliance. With a sampling period of 1 minute. (37 mb)
If you are interested in accessing this data for research purposes send us an email at cmarino@fing.edu.uy or emasquil@fing.edu.uy and we can provide you with download links.