CInterno/Synthetic_Industrial_Dataset_For_Energy_Disaggregation_SIDED

Synthetic Industrial Dataset for Energy Disaggregation (SIDED)

Dataset Summary

The Synthetic Industrial Dataset for Energy Disaggregation (SIDED) is a novel, open-source dataset created to address the critical scarcity of high-quality data for Non-Intrusive Load Monitoring (NILM) in the industrial sector. Generated using a high-fidelity Digital Twin simulator that was carefully calibrated with real-world operational data, SIDED provides a rich and diverse benchmark for NILM research. It captures complex appliance behaviors, varying weather conditions, and diverse load profiles that are systematically different and more complex than those in residential settings.

This dataset was introduced in the paper "Industrial Energy Disaggregation with Digital Twin-generated Dataset and Efficient Data Augmentation," which also proposes the Appliance-Modulated Data Augmentation (AMDA) method—a computationally efficient technique designed to enhance NILM model generalization.

How to Use

The dataset is designed to be compatible with standard toolkits like NILMTK. For use with the Hugging Face datasets library, a loading script would be required. An example of how one might load a specific configuration is shown below:

from datasets import load_dataset

# Note: This requires a loading script in the repository.
# Replace 'YourUsername/SIDED' with the actual repository name on the Hub.
dataset = load_dataset("YourUsername/SIDED", name="Office_Offenbach")

# Explore the data
print(dataset)
print(dataset['train'][0])

Dataset Structure

SIDED provides nine full years of data, representing one year for each of the nine unique facility configurations.

Characteristic	Description
Total Configurations	9 (3 Facility Types × 3 Locations)
Facility Types	Office, Dealer, Logistics Center
Geographic Locations	Offenbach (Germany), Los Angeles (USA), Tokyo (Japan)
Appliances Tracked	5: Combined Heat and Power (CHP), Cooling Systems (CS), EV Supply Equipment (EVSE), Photovoltaic (PV), and Background Appliances (BA).
Duration	One complete calendar year per configuration.
Sampling Rate	$\tfrac{1}{60} \text{Hz}$ (one-minute intervals).
Total Data Points	525,600 per configuration.

Data Fields

Each configuration's time series file contains the following columns. All power values are in Watts (W).

• Time: Timestamp for the reading.
• Aggregate: Aggregate real power (W) measured at the main meter.
• EVSE: Electric Vehicle Supply Equipment real power (W).
• PV: Photovoltaic system real power (W). Production is represented by negative values.
• CS: Cooling Systems real power (W).
• CHP: Combined Heat and Power real power (W). Production is represented by negative values.
• BA: Background Appliances real power (W).
• TAmbient: Ambient temperature in °C.
• RadiationDiffuse: Diffuse sky radiation in W/m².
• RadiationDirect: Direct solar radiation in W/m².

Appliance Characteristics

The dataset reflects the complexity of industrial energy consumption by including several classes of appliances:

• Constantly-on Variable Appliances: Devices like the CHP that are always active but have variable consumption.
• Periodical: Appliances like EVSE and PV that show periodic behavior.
• Seasonal: Appliances whose consumption is influenced by weather and seasons, including PV, CHP, and CS.
• Consumers: Devices that consume energy (EVSE, CS, and BA).
• Producers: Devices that generate energy (PV and CHP).

How to Cite

If you use the SIDED dataset or the AMDA method in your research, please cite the following paper:

@misc{internò2025industrialenergydisaggregationdigital,
      title={Industrial Energy Disaggregation with Digital Twin-generated Dataset and Efficient Data Augmentation}, 
      author={Christian Internò and Andrea Castellani and Sebastian Schmitt and Fabio Stella and Barbara Hammer},
      year={2025},
      eprint={2506.20525},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={[https://arxiv.org/abs/2506.20525](https://arxiv.org/abs/2506.20525)}, 
}