Empowering Smart Meter Applications: GAN-Generated Synthetic Data for Water and Power Analysis

Aditya YADAV

Master's thesis

Abstract:

This thesis explores the application of Generative Adversarial Networks (GANs) in generating synthetic time-series data to support smart meter applications, particularly for water and power consumption analysis. The reliance on smart meters for applications like Non-Intrusive Load Monitoring (NILM) and Human Activity Recognition (HAR) present significant challenges due to privacy concerns, data scarcity, and the high cost of real-world data collection. To address these issues, this research investigates different GAN architectures, including TimeGAN, a modified TimeGAN variant, and DoppelGANger, analyzing their effectiveness in capturing the temporal dependencies and statistical properties of real smart meter data. The study systematically evaluates these models using discriminative and predictive scores, statistical similarity metrics, and visual inspection techniques. Key findings indicate that GAN-generated synthetic data approximates real-world patterns but still faces limitations, particularly in maintaining long-term dependencies and avoiding mode collapse. The modified TimeGAN introduces architectural enhancements, such as improved statistical alignment via variance loss terms and adjustments in GRU layers, to stabilize training and enhance the fidelity of generated data. However, persistent challenges in training stability, convergence, and sequence-level realism highlights areas for further refinement. The study concludes that while GAN-based approaches hold promise, future research should focus on hybrid architectures that integrate GANs with autoregressive models and Transformer-based time-series forecasting techniques to improve the predictive capabilities of synthetic data. Addressing high-frequency anomalies and optimizing GAN architectures for multi-variable smart meter datasets remains an open research challenge. These findings contribute to the broader discourse on synthetic data generation for smart grid applications, providing a foundation for more robust, privacy-preserving, and scalable solutions in energy and water resource management.

Abstract:

This thesis explores the application of Generative Adversarial Networks (GANs) in generating synthetic time-series data to support smart meter applications, particularly for water and power consumption analysis. The reliance on smart meters for applications like Non-Intrusive Load Monitoring (NILM) and Human Activity Recognition (HAR) present significant challenges due to privacy concerns, data scarcity, and the high cost of real-world data collection. To address these issues, this research investigates different GAN architectures, including TimeGAN, a modified TimeGAN variant, and DoppelGANger, analyzing their effectiveness in capturing the temporal dependencies and statistical properties of real smart meter data. The study systematically evaluates these models using discriminative and predictive scores, statistical similarity metrics, and visual inspection techniques. Key findings indicate that GAN-generated synthetic data approximates real-world patterns but still faces limitations, particularly in maintaining long-term dependencies and avoiding mode collapse. The modified TimeGAN introduces architectural enhancements, such as improved statistical alignment via variance loss terms and adjustments in GRU layers, to stabilize training and enhance the fidelity of generated data. However, persistent challenges in training stability, convergence, and sequence-level realism highlights areas for further refinement. The study concludes that while GAN-based approaches hold promise, future research should focus on hybrid architectures that integrate GANs with autoregressive models and Transformer-based time-series forecasting techniques to improve the predictive capabilities of synthetic data. Addressing high-frequency anomalies and optimizing GAN architectures for multi-variable smart meter datasets remains an open research challenge. These findings contribute to the broader discourse on synthetic data generation for smart grid applications, providing a foundation for more robust, privacy-preserving, and scalable solutions in energy and water resource management.

Keywords

generative adversarial networks, synthetic data, smart meters, time-series data, non-intrusive load monitoring, human activity recognition, energy consumption, water consumption, deep learning, machine learning

Language used: English

Date on which the thesis was submitted / produced: 11. 2. 2025

Thesis defence

• Supervisor: prof. Dr. Florian Wahl

UNIVERSITY OF SOUTH BOHEMIA IN ČESKÉ BUDĚJOVICE

Faculty of Science

Master programme / field:
Artificial Intelligence and Data Science / Artificial Intelligence and Data Science