Peak-Aware Short-Term Load Forecasting Across Distribution Grid Aggregation Levels
Souhardya Chattopadhyay, Julian Oelhaf, Antonia Schoening, Jessica Deuschel, Bitan Bhattacharyya, Christian Bergler, Andreas Maier, Siming Bayer
cs.LG
Quick verdict
Chronos-2 wins peak-aware load forecasting at three grid levels; plausible, but GBT baselines look handicapped by fixed seasonal training windows.
Summary
This paper studies day-ahead short-term load forecasting at three distribution-grid aggregation levels — area codes (115 entities, Swiss CKW data), secondary substations (387) and LV feeders (485, UK Northern Powergrid) — arguing that average-error evaluation hides failures during high-demand (HD) periods that drive congestion risk. HD timestamps are defined via rolling normalization (demand >= 0.8 of the trailing 14-day 99th percentile of per-meter demand), and NMAE/MAPE are reported separately on all/HD/non-HD subsets for seasonal-average statistical baselines, LightGBM/XGBoost, and two Chronos foundation models, over ~359,000 daily forecasts from March 2024 to February 2025. Chronos-2 achieves the best HD metrics at every level (HD-NMAE 0.039/0.080/0.138; HD-MAPE 4.53/9.45/16.14%), Chronos-Bolt ranks second, and level-specific quantile selection (q0.65-q0.70) tunes the overall/HD trade-off; GPU inference (0.007-0.2 s per forecast) is deployment-ready. Main limitations: the gradient-boosted baselines use a fixed, hand-mapped seasonal training window while the foundation models consume rolling recent context — at AC level the 7-day-average baseline beats both GBTs overall — so the advertised 20-51% HD-NMAE reduction may partly reflect baseline misconfiguration; there is no significance testing, no HD-threshold sensitivity analysis, no deep-learning baselines, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
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 peak-aware evaluation framework that separates HD from non-HD error exposes failures hidden by average metrics, using a rolling 14-day p99 normalization as a scale-independent HD definition.
- Chronos-2 attains the best HD performance at every aggregation level (HD-NMAE 0.039/0.080/0.138 and HD-MAPE 4.53/9.45/16.14% at AC/SUB/LV), with Chronos-Bolt consistently second best.
- Relative to the gradient-boosted ML models, Chronos-2 reduces mean HD-NMAE by about 20-51% across levels while remaining best or near-best on overall metrics.
- Aggregation-specific quantile selection (q0.65-q0.70) plus sub-second GPU inference make probabilistic foundation models practically deployable for distribution-grid STLF.
Strengths
Operationally motivated, scale-independent peak-aware evaluation
Rolling 14-day p99 normalization defines HD timestamps without requiring per-entity capacity limits, and NMAE/MAPE are reported separately on all/HD/non-HD subsets (Sec. II-B, III-B), directly addressing the average-metric blind spot the paper identifies in the introduction.
Large, multi-level, full-year benchmark
About 359,000 day-ahead forecasts across 987 entities (115 AC, 387 SUB, 485 LV) from two countries, issued once daily at midnight over all four seasons with holiday and Open-Meteo weather covariates (Sec. II-C) — far beyond typical single-feeder STLF studies.
Probabilistic operating-point analysis
Deployment-relevant cost and robustness accounting
Weaknesses
Baseline fairness threatens the headline claim
No uncertainty quantification
HD definition not validated
Narrow baseline family
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- 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 peak-aware short-term load forecasting?
It evaluates load-forecasting models specifically during high-demand (peak) periods instead of only averaging errors over all hours. In this paper, a timestamp is labeled high-demand when per-meter demand exceeds 0.8 times the trailing 14-day 99th percentile for that grid entity, and forecast errors (NMAE and MAPE) are reported separately for peak and non-peak periods, because peak-period errors matter most for grid congestion, voltage control, and asset protection.
Which model performed best for distribution-grid load forecasting in this study?
The time-series foundation model Chronos-2 achieved the best high-demand accuracy at all three aggregation levels studied (area codes, secondary substations, and LV feeders), with Chronos-Bolt second at every level. It reduced high-demand NMAE by roughly 20-51% relative to LightGBM and XGBoost while staying competitive on overall accuracy, and its GPU inference takes about 0.2 seconds per day-ahead forecast (0.007 s for Chronos-Bolt).
What data were used in this load forecasting benchmark?
Open smart-meter data: the CKW Group dataset (Switzerland) aggregated per postal-code area (115 entities) and Northern Powergrid open data (UK) at secondary-substation (387) and LV-feeder (485) levels, covering November 2023 to February 2025 at 30-minute resolution. Calendar features, public holidays, and Open-Meteo weather variables were added, and about 359,000 day-ahead forecasts were evaluated from March 2024 to February 2025 across all seasons.
Is code available to reproduce this paper's results?
The paper does not state a public code release. The ingredients are open — the CKW and Northern Powergrid datasets and the Chronos, LightGBM, and XGBoost models are publicly available — but GBT hyperparameters and the validation set used for seasonal-window mapping and quantile selection are only partially described, so exact reproduction would require additional experimentation.
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 1-2 are present only as captions in the extracted text; visual evidence (HD-threshold illustration, seasonal peak patterns) could not be verified.
- Tables were extracted as plain text; cell values were transcribed carefully but formatting noise in +/- ranges cannot be fully excluded.
- No appendix or supplementary material appears in the extraction; additional hyperparameter or validation details may exist outside the provided text.
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.