LoaDiff: Conditional Generation of Electricity Consumption Time Series for Energy Analytics
Mariia Baranova, Adrien Petralia, Etienne Le Naour, Nathan Etourneau, Guillaume Hofmann, Themis Palpanas
cs.LG · cs.AI · eess.SP
Quick verdict
DiT diffusion model for year-long smart-meter data shows strong fidelity/utility across three datasets; evaluation broad but privacy evidence heuristic.
Summary
The paper introduces LoaDiff, a DDPM with a DiT-style Transformer denoiser that generates year-long, 30-minute-resolution household electricity load curves (17,520 steps), reshaped into 365 daily patches and conditioned on static attributes (appliance ownership via AdaLN) and dynamic covariates (calendar, temperature) at patch level, with classifier-free guidance. It is evaluated on the public Irish CER set (4,225 households), proprietary EDF 1 (2,083 homes), and the simulated SMACH-based EDF 2 set (20K customers) against eight generative baselines including EnergyDiff and TimeWeaver. LoaDiff achieves best or second-best FID and ACD on all datasets (CER: FID 0.0073, NNDR 0.0009), wins 16 of 18 appliance-detection TSTR configurations (average balanced accuracy 0.710/0.766 for ROCKET/TransApp), and its synthetic-only training matches or beats real-data forecasting on CER and EDF 1 (TSTR 0.954/0.550 vs TRTR 0.980/0.562 RMSE/MAE). Counterfactual ±5°C winter temperature shifts move load +10.6%/−49.1%, an asymmetry the authors themselves admit is under-characterized. Main limitations: privacy assessment relies solely on distance-based heuristics with no membership-inference attacks or formal guarantees; two of three datasets are proprietary and EDF 2 is itself simulated; the NNDR metric convention is not formally defined; only two temperature shifts are tested; ablations run on EDF 2 only. Code and a 50K-curve synthetic dataset are released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LoaDiff generates year-long sub-hourly load curves with best or second-best FID and ACD across all three datasets (Table I).
- Training on LoaDiff samples alone matches or beats real-data training for forecasting on CER and EDF 1, and wins 16 of 18 appliance-detection configurations (Tables II-III).
- LoaDiff shows the lowest NNDR-based memorization evidence on two of three datasets, supporting a favorable quality-privacy trade-off.
- Conditioning produces coherent counterfactual behavior: ±5°C winter temperature shifts change mean winter load by +10.6% and −49.1% respectively (Sec. V-C3).
- Daily tokenization, AdaLN conditioning, classifier-free guidance, and both covariate families each contribute to generation quality (Table IV).
Strengths
Long-horizon, structure-aware design with ablation backing
Year-long 17,520-step curves tokenized into 365 daily patches; Table IV shows weekly patches (patch_336steps) yield Disc1NN 0.5000 with FID up 2.7x and ACD up 6.5x — a mode-collapse signature the authors correctly reject — while the daily patch reference wins on FID/ACD.
Evaluation breadth beyond field norms
Four dimensions (fidelity/diversity, memorization, downstream utility, controllability) across 8 baselines including EnergyDiff and TimeWeaver, with TSTR/TR+STR protocols on two real tasks (PatchTST forecasting, ROCKET/TransApp appliance detection); the paper correctly notes most prior smart-meter synthesis work reports fidelity only.
Unusually good metric literacy
Consistent downstream wins plus concrete releases
Weaknesses
Privacy evaluation is heuristic-only and partly uninterpretable
Limited external validity: two of three datasets unusable by others
Downstream comparisons confound conditioning with generator quality
No inferential statistics
Hyperparameter opacity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
LoaDiff (EDF R&D + Université Paris Cité) is a conditional DDPM with a DiT-style Transformer denoiser for year-long, 30-minute smart-meter load curves (T = 17,520). Its four claims: (1) best or second-best fidelity (FID/ACD) on three datasets (Table I); (2) a favorable quality-vs-memorization trade-off (NNDR/NNPriv); (3) downstream utility for forecasting (PatchTST, Table II) and appliance detection (ROCKET/TransApp, Table III); (4) coherent counterfactual temperature response (Sec. V-C3). It positions itself against short-horizon, low-controllability prior work [13]-[16], [29], and releases code [17] plus a 50K-curve appliance-conditioned Zenodo dataset [18].
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is LoaDiff?
LoaDiff is a conditional denoising-diffusion model (DDPM with a DiT-style Transformer denoiser) from EDF R&D and Université Paris Cité that generates a full year of half-hourly household electricity consumption (17,520 time steps), conditioned on appliance ownership, calendar information, and outdoor temperature, for privacy-constrained energy analytics.
Is LoaDiff's code or data available?
Yes, partially: the source code is on GitHub (github.com/adrienpetralia/loadiff) and a synthetic dataset of 50,000 appliance-conditioned residential load curves is on Zenodo (DOI 10.5281/zenodo.22257867). However, the two EDF evaluation datasets (2,083 real homes and a 20K-customer SMACH simulation) are proprietary and not released.
Does LoaDiff protect household privacy?
Only heuristically. The authors measure distance-based memorization signals (NNDR and NNPriv) and report low evidence of copying training records, with the best NNDR on two of three datasets, but there is no membership-inference attack and no differential-privacy guarantee, so no formal privacy protection is claimed or proven.
Can the synthetic data replace real smart-meter data for machine learning?
Largely yes in the paper's tests: for two-day-ahead forecasting, models trained only on LoaDiff data matched or slightly beat real-data training on the Irish CER and EDF 1 datasets; for appliance detection it won 16 of 18 classifier-appliance settings, and adding its data to real training data improved results in 12 of 18 settings.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 appear only as extracted caption and axis text; visual judgments (t-SNE overlap in Fig. 3, profile shapes in Figs. 4-5) rely on the authors' prose rather than the rendered figures.
- Inline equations are garbled by PDF extraction in places (e.g., the patch interaction term and the AdaLN/output-head equations); their content was reconstructed from the surrounding prose — this is extraction noise, not truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.