Full text available at ADF & TransApp.

Adrien Petralia, Philippe Charpentier, and Themis Palpanas. ADF & TransApp: A Transformer-Based Framework for Appliance Detection Using Smart Meter Consumption Series. Proceedings of the VLDB Endowment (PVLDB), 17(3): 553 - 562, 2023. doi:10.14778/3632093.363211

We propose the Appliance Detection Framework (ADF) to detect the presence of appliances in households, using real-world consumption series, which are sampled at a very low frequency, and are long and variable-length. ADF addresses these challenges by operating at individual subsequences of each consumption series, instead of each series in its entirety. The framework can be used with any time series classifier designed to predict probabilities.

We propose TransApp, a Transformer-based time series classifier, which can first be pretrained in a self-supervised manner to enhance its ability on appliances detection tasks. This way, TransApp can significantly improve its accuracy.

The proposed architecture lies in combination of a strong embedding block made of dilated convolutional layers followed by a Transformer encoder using Diagonally Masked Self-Attention (DMSA).

Self-supervised pretraining. The use of a self-supervised pretraining of a Transformer architecture on an auxiliary task has been used in the past to boost the model performance on downstream tasks. This process is inspired by the mask-based pretraining of vision transformer and requires only the input consumption series without any appliance information label. It results in a reconstruction objective of a corrupted (masked) time series fed to the model input.

Supervised pretraining. The supervised training results in a simple binary classification process using labeled time series.

We provide a jupyter-notebook example to use our Appliance Detection Framework combined with our TransApp classifier on the CER data : experiments/TransAppExample.ipynb.

In addition, to reproduce papers experiments, use the following guidelines.

Pretraining TransApp in a self-supervised way using non labeled data :

sh LaunchTransAppPretraining.sh

Use our Appliance Detection Framework combined with TransApp to detect appliance in consumption time series :

sh LaunchTransAppClassif.sh

Use our Appliance Detection Framework combined with ConvNet, ResNet or InceptionTime to detect appliance in consumption time series :

sh LaunchModelsClassif.sh

Please refer to this Github ApplianceDetectionBenchmark to reproduce the experiments, where an extensive evaluation of different time series classifiers have been conducted, inluding on the datasets used in this study.

Python version : >= Python 3.7

Overall, the required python packages are listed as follows:

• numpy
• pandas
• scikit-learn
• imbalanced-learn
• torch==1.8.1
• torchinfo
• scipy
• sktime
• matplotlib

Use pip to install all the required libraries listed in the requirements.txt file.

pip install -r requirements.txt

The data used in this project comes from two sources:

• CER smart meter dataset from the ISSDA archive.
• Private smart meter dataset provide by EDF (Electricité De France).

You may find more information on how to access the datasets in the data folder.

• Adrien Petralia, EDF R&D, Université Paris Cité
• Philippe Charpentier, EDF R&D
• Themis Palpanas, Université Paris Cité, IUF

We would like to thanks Paul Boniol for the valuable discussions on this project. Work supported by EDF R&D and ANRT French program.