Abstract
As the energy sector transitions toward sustainable grids, the implementation of dynamic Time-of-Use (dToU) tariffs becomes critical. However, identifying households suitable for these tariffs requires granular consumption data, posing significant privacy risks. This paper benchmarks four synthetic data generation methods—Wasserstein GAN (WGAN), CTGAN, Diffusion Models (DDPM), and Gaussian noise—to evaluate their effectiveness in replacing sensitive smart meter data. Our findings reveal a distinct privacy-utility tradeoff: while Diffusion Models achieve the highest classification utility (up to 88% F1-score), CTGAN provides the most robust defense against reconstruction attacks.
Introduction
The transition to sustainable energy systems relies heavily on demand-side management, specifically dynamic Time-of-Use (dToU) tariffs. These tariffs incentivize consumers to shift their electricity usage to off-peak hours. Identifying "responsive" households typically requires analyzing high-resolution smart meter data, which can inadvertently reveal intimate details of a resident's daily life. Synthetic data offers a promising path forward by mirroring the statistical properties of real datasets without exposing individual identities. However, a fundamental tension exists: higher realism often leads to increased vulnerability to privacy attacks. This study systematically explores this tradeoff within the context of tariff suitability classification.
Experimental Framework and Methods
The researchers evaluated four generative approaches under two distinct training regimes: semi-synthetic (augmenting real data) and full-synthetic (replacing real data entirely).
The following models were benchmarked:
- Wasserstein-GP GAN (WGAN): A standard GAN architecture enhanced with entropy and label-balance terms to mitigate mode collapse.
- Diffusion Models (DDPM): Adapted for tabular data using a dual-head design to simultaneously predict Gaussian noise and class labels.
- CTGAN: A specialized GAN for tabular data utilizing conditional sampling to address imbalanced discrete columns.
- Gaussian Noise Augmentation: A baseline method where random noise is added to real features for regularization.
Evaluation metrics focused on three pillars: Distribution Fidelity (statistical similarity), Utility (performance of downstream ML models on real test data), and Privacy Robustness (resistance to Membership Inference and Reconstruction attacks).
Dataset and Feature Engineering
The study utilized data from the UK Power Networks’ Low Carbon London project, comprising 167 million half-hourly readings from over 5,500 households. To move beyond raw time-series, the authors engineered 24 behavioral features, such as peak/off-peak usage ratios and consumption autocorrelation. Since explicit "responsiveness" labels were unavailable, an unsupervised approach using Principal Component Analysis (PCA) was employed to identify the top 25% of responsive households as the target class.
Key Insights and Results
The experimental results highlighted several critical findings regarding the nature of synthetic data in smart grid applications.
Distribution Fidelity and Utility
A visual inspection of the synthetic data distributions via t-SNE plots reveals varying degrees of overlap with the original data.
- Utility Gains: Surprisingly, models trained on Full-Synthetic data often outperformed those trained on real data. While the real-data baseline achieved a Macro-F1 of 67.5%, Diffusion-based synthetic data reached 88.2%. This suggests that generative models can act as "super-regularizers," denoising the data and capturing the underlying signal more effectively than the raw, noisy household data.
- Fidelity: CTGAN in the semi-synthetic regime produced the most statistically accurate distributions, achieving the lowest Kullback-Leibler (KL) divergence (0.16).
The Privacy-Utility Tradeoff
The study identified a significant "Reconstruction Gap." While all models showed similar resistance to Membership Inference Attacks (MIA), their performance against reconstruction attacks varied wildly:
- CTGAN was the standout for privacy, achieving a Privacy Risk Score (PRS) of 0.16. It successfully obscured fine-grained patterns that could be used to reconstruct original data points.
- Diffusion and Noise-based models, despite their high utility, leaked significant information, with PRS values near 0.98, comparable to raw data.
Conclusion and Future Work
This research demonstrates that while Diffusion Models are superior for maximizing predictive utility, CTGAN is the preferred choice for public data releases where privacy is paramount. Future work should focus on integrating formal privacy guarantees, such as Differential Privacy (DP-SGD), and exploring time-aware architectures like TimeGAN to generate high-fidelity raw time-series profiles.
Summary updated