Synthetic Electric Vehicle Charging Session Generation Using a Conditional Variational Autoencoder
Graeme Kelly, Emilio J. Palacios-Garcia, Barry P. Hayes
eess.SY · cs.LG
Quick verdict
A CVAE turns 157k real EV charging sessions into realistic synthetic data (TSTR 1.08–1.21), but lacks baselines and physical-consistency or privacy checks.
Summary
This paper presents a conditional variational autoencoder (CVAE) that generates synthetic individual EV charging sessions from real transaction-level data. Trained on roughly 157,000 Electric Nation sessions (CrowdCharge and GreenFlux; 602 charge points, 2017–2018), the model reconstructs six engineered features — session duration, charging duration, delivered energy, plug-in-to-charging gap, and cyclical time-of-week — conditioned on day-of-week and managed-charging status, using a Gaussian NLL reconstruction loss for heteroscedastic uncertainty and KL regularisation. Fidelity is assessed with per-feature Wasserstein distance, sliced WD and Spearman-correlation MAE against a real-vs-real baseline, plus TSTR/TRTR utility using Random Forest regressors. Reported Test sliced WD is 0.058 (baseline 0.017) and Spearman-MAE 0.015, indicating reasonable joint fidelity; TSTR/TRTR ratios of 1.075–1.207 show modest utility loss, larger for managed sessions. Key limitations: no comparison with simpler generative baselines (plain VAE, GAN, copula, resampling); the paper's own histograms show material moment mismatches (synthetic session-duration SD 7.95 vs real 11.27; gap SD 1.51 vs 2.51); heavy tails and charging-gap spikes are not captured; 46k sessions (~29%) were dropped for missing timestamps without bias analysis; and despite the privacy motivation and hosting-capacity ambition, neither privacy leakage nor physical consistency of generated sessions is evaluated. Code and data availability are stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 5
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A CVAE conditioned on day-of-week and managed-charging status can generate individual EV charging sessions that preserve key statistical properties of 157k+ real Electric Nation records.
- A Gaussian NLL reconstruction loss with feature-wise heteroscedastic variance yields joint fidelity close to the real-data floor (Test SWD 0.058 vs Base 0.017; Spearman-MAE 0.015 vs 0.007).
- Synthetic data retains downstream predictive utility, with RF TSTR/TRTR ratios of 1.075–1.207 across target features and 1.021–1.167 across day-of-week and managed-status cells.
- Managed charging scenarios are systematically harder for the generative model than unmanaged ones (higher TSTR/TRTR ratios, e.g., 1.167 vs 1.084 on Thursday).
- Session-level (rather than aggregate) synthetic generation enables selective addition of charging demand for hosting-capacity studies.
Strengths
Real large-scale data with open release
Trained and evaluated on over 157,000 transaction-level sessions from the public Electric Nation programme (CrowdCharge + GreenFlux, 602 charge points, 2017–2018); the authors state all code and data are available in a GitHub repository (ref [13]), supporting reproducibility.
Principled probabilistic formulation and feature engineering
Gaussian NLL reconstruction with feature-wise learned variances (heteroscedastic uncertainty), log(1+x) transforms for heavy-tailed durations/energy, cyclical sin/cos week-time encoding (Eqs. 1–2), one-hot day-of-week plus managed-status conditioning, and train-split-only standardisation to avoid leakage (Sec. II.A–II.C).
Well-calibrated fidelity evaluation design
Per-scenario utility analysis with actionable insight
Weaknesses
No generative baselines
Fidelity gaps visible in the paper's own numbers
No physical-consistency validation of generated sessions
Unexamined 29% data loss and untested privacy claim
Thin statistical and tuning reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 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 data was used to train the EV charging CVAE?
The UK Electric Nation trial data (CrowdCharge and GreenFlux datasets): over 157,000 residential charging sessions from 602 charge points spanning early 2017 to end-2018. After removing about 46k sessions lacking a charging-start timestamp and filtering extreme-duration outliers (under 5 minutes or over 10,000 minutes), roughly 111k sessions were split 70/15/15 for training, validation, and testing.
How realistic is the synthetic EV charging data produced by this model?
Reasonably faithful but imperfect. On the test split, the joint-distribution sliced Wasserstein distance is 0.058 versus a 0.017 real-vs-real baseline, and rank correlations are well preserved (Spearman MAE 0.015). However, synthetic standard deviations fall noticeably short for session duration (7.95 vs 11.27 hours) and the plug-in-to-charging gap (1.51 vs 2.51 hours), and heavy tails plus gap spikes are not captured. Random Forest models trained on synthetic data show about 7.5–20.7% higher error than models trained on real data (TSTR/TRTR 1.075–1.207).
Is the code and dataset from this paper publicly available?
Yes. The authors state that all code and data are available in a public GitHub repository (ucc-future-grid-lab/ev-cvae), and the underlying Electric Nation dataset is publicly released through National Grid UK's Electric Nation programme page.
Can these synthetic EV charging sessions be used for distribution network hosting-capacity studies?
The authors position session-level generation for that purpose, since individual sessions can be added on top of existing demand. However, this paper performs no network study, does not verify physical consistency of generated sessions (e.g., charging duration should not exceed session duration), and acknowledges heavy-tailed behaviour is not fully captured — so any planning use would still require such validation first.
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 2–4 were judged only from extracted caption and axis text (moments, legend labels); visual distribution shapes and loss-curve details could not be inspected directly.
- The per-feature reconstruction-loss weights and hyperparameter-sweep results are not in the text, so exact retraining depends on the linked GitHub repository (ucc-future-grid-lab/ev-cvae), which was not examined.
- The paper is IEEE conference-length; limited space may explain the absent baselines and uncertainty reporting, but the gaps remain for review purposes.
- Text extraction appears complete (Sections I–IV plus references), so truncation risk is low; the n=470 real-sample count in Fig. 3 is taken as printed.
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.