Benchmarking Deep Generative Models for Synthetic Smart-Meter Data
Embedded in: NWO-KIC project SHARE (Synthetic Harmonized Access to Renewable Energy data), WP3, Synthetic Data Generation using ML/AI.
Supervisor: Roel Bouman (RU), Yuliya Shapovalova (RU), Jacco Heres (Alliander)
Duration: 6 months.
1. Context and motivation
The Dutch energy transition needs high-resolution smart-meter and load data for grid planning, congestion management, and flexibility optimization — but this data is locked away by the GDPR, commercial sensitivity, and legal uncertainty. Synthetic data offers a way out: a generative model that reproduces the statistical and temporal structure of real load profiles while containing no real customer records can be shared openly. Synthetic energy data in SHARE spans both time series (load and generation profiles) and network graphs (grid topology); this thesis addresses the *time-series* side. It targets an open gap: modern generators (VAEs, GANs, diffusion) are each reported on their own data with their own metrics, and no published one-to-one comparison of, e.g., a GAN versus a diffusion model on the same data under the same metrics exists. Producing that fair benchmark — and independently checking recent strong claims (e.g. near-perfect autocorrelation for diffusion/flow-matching models) — is an important contribution to the field.
(This proposal pairs with a companion thesis on generating synthetic load profiles; the two can run in parallel — the models built there can be inputs to the benchmark here — but each stands alone. The models under test can equally be existing, published generators.)
2. Objective and research questions
Central question:which deep generative models produce the most useful synthetic residential/LV load-profile data, and how should "useful" be measured? The research questions below can be addressed based on the interest of the student.
- RQ1 (Fidelity). How well do conditional VAEs, GANs, and diffusion models reproduce the marginal distributions and temporal structure (autocorrelation, seasonality, spectrum) of real smart-meter data — and do the strong autocorrelation results reported for diffusion/flow-matching models hold up under independent replication?
- RQ2 (Utility). Under a Train-on-Synthetic, Test-on-Real (TSTR) protocol, how well does synthetic data replace real data for short-term load forecasting / peak classification — and does fidelity predict utility?
- RQ3 (Benchmark). Can these fidelity and utility measures be consolidated into a fair, reproducible, model-agnostic benchmark for synthetic energy time series (built on OpenSynth), and what does a head-to-head comparison of published generators reveal?
3. Background: state of the art
Four model families dominate. **VAEs** (Faraday cVAE; GUIDE-VAE, which explicitly aims to reduce the characteristic over-smoothing of peaks) are stable and interpretable. **GANs** (TimeGAN, Conv1D-WGAN-GP) capture sharp detail but suffer training instability and distribution mismatch. **Diffusion / flow-matching** models are the current state of the art — EnergyDiff (Lin et al., 2025) and SmartMeterFM report the best temporal fidelity, but these families have not been compared head-to-head on identical data and metrics. Classical baselines (copula, Gaussian processes) remain essential floors. For evaluation, the community now has a shared reference: **OpenSynth** (LF Energy / Centre for Net Zero) hosts synthetic-meter datasets, published models, and a standard evaluation framework (fidelity and utility), which this thesis builds on rather than reinventing.
4. Proposed approach
We will reuse existing datasets and models. All target models (EnergyDiff, Faraday / GUIDE-VAE, TimeGAN, SmartMeterFM) and datasets are already published; the thesis integrates and fairly evaluates them, bootstrapping from OpenSynth and from a recent comparative-evaluation codebase. An explicit early milestone is to reproduce and sanity-check the reused evaluation, in particular how per-window profiles are concatenated for the autocorrelation test.
Data. Primary: Low Carbon London (5,566 households, half-hourly, open CC-BY; downloaded and verified). Secondary: GoiEner (25,559 Spanish supply points, hourly, open on Zenodo). Plus OpenSynth datasets for cross-checks. 1-minute data is de-emphasized: large open sources are scarce and 15–30 min suffices for the target use cases.
Models & harness. Benchmark three published models (one diffusion, one VAE, one GAN) plus a copula baseline under identical preprocessing, splits, and compute. Extend OpenSynth's evaluation into a model-agnostic harness scoring (1) distributional (KS, Wasserstein, MMD), (2) temporal (ACF, spectral density, seasonal match) across daily / weekly / monthly horizons, and (3) utility (TSTR). Report training stability, compute cost, and ease of reuse as first-class results.
5. Deliverables and student profile
Deliverables: 1) a fair, reproducible benchmark of VAE / GAN / diffusion models on real smart-meter data, including an independent check of recent autocorrelation claims; 2) an open evaluation harness built on OpenSynth; 3) a thesis and code/data release; (stretch) tests/comparisons upstreamed to OpenSynth.
Profile: MSc in Data Science / AI; solid Python + PyTorch and GPU deep learning; strong in empirical evaluation; comfortable reading and running published research code; interest in time series and applied ML. No prior energy-domain knowledge required.
Contact
Interested? Contact Dr. Yuliya Shapovalova, [yuliya.shapovalova@ru.nl](yuliya.shapovalova@ru.nl)
References
1. Lin, N., Palensky, P., Vergara, P. P. "EnergyDiff: Universal time-series energy data generation using diffusion models." *IEEE Trans. Smart Grid* (2025).
2. Chai, S., Chadney, G. "Faraday: synthetic smart meter generator for the smart grid." arXiv:2404.04314 (2024).
3. Bölat, K., Tindemans, S. "GUIDE-VAE: Advancing data generation with user information and pattern dictionaries." arXiv:2411.03936 (2024).
4. Yoon, J., Jarrett, D., Van der Schaar, M. "Time-series generative adversarial networks." *NeurIPS* 32 (2019).
5. Thorve, S., et al. "Fidelity and diversity metrics for validating hierarchical synthetic data." *IEEE Big Data* (2022).
6. Turowski, M., et al. "Generating synthetic energy time series: A review." *Renew. Sustain. Energy Rev.* 206 (2024): 114842.
7. Chai, S., et al. "Defining 'Good': Evaluation framework for synthetic smart meter data." arXiv:2407.11785 (2024) — the OpenSynth (LF Energy) framework; lfenergy.org/projects/opensynth.
8. "An electricity smart meter dataset of Spanish households (GoiEner)." *Scientific Data* 11 (2024). doi:10.1038/s41597-023-02846-0; data on Zenodo.
9. "A Comparative Evaluation of Synthetic Load Profile Generators Across Unseen Datasets." IEEE (TU/e) — candidate starting codebase.