Synthetic Time-Series Data Generation for Smart Grids Using 3D Autoencoder GAN

Given the growing significance of data-driven approaches in analysis and decision-making in smart grid, the availability of diverse and representative datasets is paramount. However, challenges such as privacy concerns, data size limitations, and data quality issues have constrained the usage of rea...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on industrial informatics Vol. 21; no. 7; pp. 5047 - 5058
Main Authors: Zhang, Guihai, Sikdar, Biplab
Format: Journal Article
Language:English
Published: Piscataway IEEE 01.07.2025
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects:
3D convolution
Autoencoders
Correlation
Data models
Data privacy
Datasets
Deep learning
Electric vehicles
Generative adversarial networks
generative adversarial networks (GANs)
Households
Load modeling
Multivariate analysis
Power demand
power grid data
Prediction models
Smart grid
Synthetic data
Three-dimensional displays
Time series
time series generation
Training
ISSN:1551-3203, 1941-0050
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Given the growing significance of data-driven approaches in analysis and decision-making in smart grid, the availability of diverse and representative datasets is paramount. However, challenges such as privacy concerns, data size limitations, and data quality issues have constrained the usage of real-world data. In this paper, we introduce the 3D Autoencoder Generative Adversarial Network (3DAE GAN) as a solution to generate high-resolution and multivariate synthetic time-series data capable of representing various real power consumption patterns across different households and driving data for Electric Vehicles (EVs). Beyond the conventional GAN structure, the incorporation of both the Autoencoder and 3D-convolution processes enables a more comprehensive extraction of patterns in data, thereby addressing limitations present in existing data generation methods. Evaluation results using the Pecan Street and Emobpy simulated EV dataset demonstrate that the proposed method generates synthetic data with higher similarity scores compared to existing approaches. Furthermore, downstream prediction tasks are conducted to establish the comparability between using the original data and the synthetic data, revealing no significant differences. Moreover, the risk of possible information leakage from synthetic data about original data is evaluated by performing membership inference attacks and population attacks on the prediction models that are trained with synthetic data. The robustness of the synthetic data are examined when facing FGSM attacks.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1551-3203
1941-0050
DOI:10.1109/TII.2025.3538115