Torch-NILM is the first NILM-specific deep learning toolkit build on Pytorch and Python. The purpose of the toolkit is to help researchers design and execute huge amount of experiments and comparisons with ease. It contains 3 APIs that cover various needs that are encountered when creating a new architecture such as benchmarking, hyperparameter tuning and cross validation. Furthermore, Torch-NILM comes with a handy reporting module that exports a results report in xlsx format alongside with comparative base plots to visualize the results.
The build-in benchmark methodology contains a series of scenarios with escalating difficulty to stress test the generalization capabilities of the tested models. In addition, Torch-NILM provides a set of powerful baseline models to conduct comparisons.
The toolkit is compatible with NILMTK.
The complementary paper "Torch-NILM: An Effective Deep Learning Toolkit for Non-Intrusive Load Monitoring in Pytorch" can be found here.
In case you use Torch-NILM to conduct research, please consider to cite our paper:
@Article{en15072647,
AUTHOR = {Virtsionis Gkalinikis, Nikolaos and Nalmpantis, Christoforos and Vrakas, Dimitris},
TITLE = {Torch-NILM: An Effective Deep Learning Toolkit for Non-Intrusive Load Monitoring in Pytorch},
JOURNAL = {Energies},
VOLUME = {15},
YEAR = {2022},
NUMBER = {7},
ARTICLE-NUMBER = {2647},
URL = {https://www.mdpi.com/1996-1073/15/7/2647},
ISSN = {1996-1073},
ABSTRACT = {Non-intrusive load monitoring is a blind source separation task that has been attracting significant interest from researchers working in the field of energy informatics. However, despite the considerable progress, there are a very limited number of tools and libraries dedicated to the problem of energy disaggregation. Herein, we report the development of a novel open-source framework named Torch-NILM in order to help researchers and engineers take advantage of the benefits of Pytorch. The aim of this research is to tackle the comparability and reproducibility issues often reported in NILM research by standardising the experimental setup, while providing solid baseline models by writing only a few lines of code. Torch-NILM offers a suite of tools particularly useful for training deep neural networks in the task of energy disaggregation. The basic features include: (i) easy-to-use APIs for running new experiments, (ii) a benchmark framework for evaluation, (iii) the implementation of popular architectures, (iv) custom data loaders for efficient training and (v) automated generation of reports.},
DOI = {10.3390/en15072647}
}
For any enquiries, please contact the main authors.
For the installation users are advised to use Anaconda. Torch-NILM requires Python 3.7+. The provided torch-nilm.yml file contains the necessary dependencies. With the following command the appropriate environment is created (for more details check here).
conda env create -f torch-nilm.yml
Activate the environment:
conda activate torch-nilm
Defining an experiment requires only a few lines of code. A template of setting an experiment is provided in set_experiment.py.
In order to set up the experiment the following configurations should be provided:
a) The experiment_parameters contain the general experiment configurations that are essential for all the supported experiments. A small description is provided bellow. For more information go to the corresponding doc string in torch_nilm/lab/nilm_experiments.py or consult the paper.
- EPOCHS: 'The number of training epochs of a model'
- ITERATIONS: 'The number of iterations each experiment should run. It is helpful for calculating
statistics'
- INFERENCE_CPU: 'If _true_ the inference will executed on the CPU'
- SAMPLE_PERIOD: 'The sampling period in seconds'
- BATCH_SIZE: 'The batch size needed for training and inference of the neural networks'
- ITERABLE_DATASET: 'If _True_ the data will be provided to the network in an efficient way.
More in https://pytorch.org/docs/stable/data.html'
- PREPROCESSING_METHOD: 'The preprocessing method. Four methods are supported: sequence-to-sequence
learning, sliding-window approach, midpoint-window method and sequence-to-subsequence approach.'
- FILLNA_METHOD: 'The method to fill missing values. Zero filling and linear interpolation are supported'
- FIXED_WINDOW: 'The length of the input sequence'
- SUBSEQ_WINDOW: 'The length of the output sequence when sequence-to-subsequence preprocessing method is chosen.'
- TRAIN_TEST_SPLIT: 'The ratio of data to used for training/inference'
- CV_FOLDS: 'The number of folds when cross validation experiment is chosen'
- NOISE_FACTOR: 'The percentage of the added noise can controlled with a noise factor, a factor to multiply a
gaussian noise signal, which will be added to the normalized mains timeseries.'
After the declaration of the experiment_parameters list the user should save the list as an ExperimentParameters object:
experiment_parameters = ExperimentParameters(**experiment_parameters)
b) The devices list contains the desired electrical devices/appliances to run the experiments for. Currently, five appliances are supported:
ElectricalAppliances.KETTLE
ElectricalAppliances.MICROWAVE
ElectricalAppliances.FRIDGE
ElectricalAppliances.WASHING_MACHINE
ElectricalAppliances.DISH_WASHER
c) The experiment_categories list contains the desired benchmark categories to be executed. The categories are based on the benchmark method proposed in [1]. Two categories are supported:
SupportedExperimentCategories.SINGLE_CATEGORY
SupportedExperimentCategories.MULTI_CATEGORY
d) The model_hparams list contains the hyperparameters for the desired models to train. The user can add only the desired models.
model_hparams = [
{
'model_name': 'VAE',
'hparams': {'window_size': None, 'cnn_dim': 256, 'kernel_size': 3, 'latent_dim': 16},
},
{
'model_name': 'NFED',
'hparams': {'depth': 1, 'kernel_size': 5, 'cnn_dim': 128,
'input_dim': None, 'hidden_dim': 256, 'dropout': 0.0},
},
{
'model_name': 'SimpleGru',
'hparams': {},
},
{
'model_name': 'SAED',
'hparams': {'window_size': None},
},
{
'model_name': 'WGRU',
'hparams': {'dropout': 0},
},
]
After the declaration of the model_hparams list the user should save the list as an ModelHyperModelParameters object:
model_hparams = ModelHyperModelParameters(model_hparams)
e) In order to execute hyperparameter tuning with cross validation, the user should define the hparam_tuning list. That list contains the versions of the desired neural under test.
hparam_tuning = [
{
'model_name': 'NFED',
'hparams': [
{'depth': 1, 'kernel_size': 5, 'cnn_dim': 16,
'input_dim': None, 'hidden_dim': 256, 'dropout': 0.0},
{'depth': 2, 'kernel_size': 5, 'cnn_dim': 32,
'input_dim': None, 'hidden_dim': 64, 'dropout': 0.0},
]
},
{
'model_name': 'SAED',
'hparams': [
{'window_size': None, 'bidirectional': False, 'hidden_dim': 16},
{'window_size': None, 'bidirectional': False, 'hidden_dim': 16, 'num_heads': 2},
]
},
]
After the declaration of the hparam_tuning list the user should save the list as an HyperParameterTuning object:
hparam_tuning = HyperParameterTuning(hparam_tuning)
a) In order to run the experiments a NILMExperiments object should be defined as shown bellow.
experiment = NILMExperiments(
project_name='MyProject', // the project name
clean_project=True, // whether to delete the folders under the project name or not
devices=devices, // the device list
save_timeseries_results=False, // whether to save the network's output or not
experiment_categories=experiment_categories, // the experiment categories
experiment_volume=SupportedExperimentVolumes.LARGE_VOLUME,// the data volume
experiment_parameters=experiment_parameters, // the general experiment parameters
save_model=True, // whether to save model weights or not
export_plots=True,// whether to export result plots or not
)
b) This experiments object contains all the experiment APIs, which can be called as shown bellow.
experiment.run_benchmark(model_hparams=model_hparams)
experiment.run_cross_validation(model_hparams=model_hparams)
experiment.run_hyperparameter_tuning_cross_validation(hparam_tuning=hparam_tuning)
c) After an experiment is executed the corresponding statistical_report.xlsx and the plots are exported in the project_name/results/ directory.
d) If for some reason an experiment was interrupted and the report was not created the user can run the export_report API with the same settings as the desired experiment API. This way a new report will be created without re-running the experiment.
## experiment.run_benchmark(model_hparams=model_hparams)
experiment.export_report(model_hparams=model_hparams, experiment_type=SupportedNilmExperiments.BENCHMARK)
In order to execute the experiment run:
python3 set_experiment.pyThe results of each API are saved under the directory output/project-name/results/api-name. Inside that directory 3 folders exist (depending the user configurations); plots, results and saved_models. Plots directory contains all the produced graphs. Results contains the results folders for every appliance and model alongside the final statistical report in xlsx format. Saved_models contains all the weights of the models for every appliance and iteration.
The NILMTK[2] toolkit is used for reading the data. All the datasets that are compatible with NILMTK are supported, but the benchmark is constructed on end-uses from UK DALE[3], REDD[4] and REFIT[5]. It should be noted that the data have to be downloaded manually. In order to load the data, the files path_manager.py and datasource.py inside datasources/ directory should be modified accordingly.
This project is licensed under the MIT License - see the LICENSE file for details
- Symeonidis, N.; Nalmpantis, C.; Vrakas, D. A Benchmark Framework to Evaluate Energy Disaggregation Solutions. International 541 Conference on Engineering Applications of Neural Networks. Springer, 2019, pp. 19–30.
- Batra, N.; Kelly, J.; Parson, O.; Dutta, H.; Knottenbelt, W.; Rogers, A.; Singh, A.; Srivastava, M. NILMTK: an open source toolkit 525 for non-intrusive load monitoring. Proceedings of the 5th international conference on Future energy systems, 2014, pp. 265–276.
- Jack, K.; William, K. The UK-DALE dataset domestic appliance-level electricity demand and whole-house demand from five UK homes. Sci. Data 2015, 2, 150007.
- Kolter, J.Z.; Johnson, M.J. REDD: A public data set for energy disaggregation research. Workshop on data mining applications in sustainability (SIGKDD), San Diego, CA, 2011, Vol. 25, pp. 59–62.
- Firth, S.; Kane, T.; Dimitriou, V.; Hassan, T.; Fouchal, F.; Coleman, M.; Webb, L. REFIT Smart Home dataset, 2017. doi:10.17028/rd.lboro.2070091.v1.