Pith. sign in

REVIEW 4 major objections 5 minor 57 references

A standard Transformer yields the lowest 96-hour load-forecast error on all three grid-level datasets, with 6.6–10.7% lower error than the best non-Transformer method.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 22:31 UTC pith:Y47EG4PA

load-bearing objection Solid multi-grid load forecasting benchmark; Transformer margins are plausible but unquantified against seed/HPO variance, so treat the headline percentages as provisional. the 4 major comments →

arxiv 2607.15705 v1 pith:Y47EG4PA submitted 2026-07-17 cs.LG

A Benchmark for Electrical Load Forecasting Across Grid Levels: Time-Series Transformers Outperform Established Methods

Matthias Hertel , Sebastian P\"utz , Jonathan Kolar , Benjamin Sch\"afer , Ralf Mikut , Veit Hagenmeyer This is my paper
classification cs.LG
keywords electrical load forecastingTransformerbenchmarkhyperparameter optimizationshort-term forecastingtime-series foundation modelsChronos-2grid levels
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to settle a practical question: for short-term electrical load forecasting, are the many Transformer variants proposed in the literature genuinely better than the standard Transformer, the established classical and deep-learning methods, and a pretrained time-series foundation model? It builds a benchmark with three public datasets spanning a transmission-system control area, 200 low-voltage feeders, and 287 individual clients, with a 96-hour horizon, hourly forecasts, covariates, and monthly retraining. The main finding is that Transformer-based models rank at the top on every dataset: a standard encoder-decoder Transformer reduces forecast error by 8.3%, 6.6%, and 10.7% compared with the best non-Transformer method on the transmission, feeder, and client data, respectively. The paper also shows that a flexible architecture (YAformer) that assembles proposed modifications—patching, CNN and LSTM layers, max pooling, sparse attention—does not beat the plain Transformer, suggesting these modifications are unnecessary for load forecasting. The zero-shot foundation model Chronos-2 performs competitively on the feeder and client data but misses special events such as holidays on the transmission data.

Core claim

On the paper's own terms, the discovery is that forecasting performance on electrical load data is not helped by the architectural machinery added to Transformers over the years. After hyperparameter optimization with 100 configurations per method, the standard encoder-decoder Transformer achieves the lowest mean absolute error on the transmission-system and feeder datasets and the second-lowest on the client dataset, behind Chronos-2 whose training data included that dataset. Across the three datasets the best Transformer beats the best non-Transformer model by 6.6–10.7%. The intentionally flexible YAformer, which can express the standard Transformer as a special case but searches a much la

What carries the argument

YAformer — 'yet another Transformer' — is the paper's central instrument: an encoder-decoder Transformer with optional patching, convolutional layers, LSTM layers, max pooling, and LogSparse attention, each controlled by a hyperparameter. The standard Transformer is a special case of YAformer with all modifications off, so comparing the two under the same hyperparameter-optimization budget isolates the effect of the architectural additions. The other load-bearing components are the three datasets (TSO control area, low-voltage feeders, individual clients), the rolling 96-hour evaluation with hourly retraining, and the shared 100-trial Bayesian optimization used for every method.

Load-bearing premise

The load-bearing premise is that giving every model the same 100-trial hyperparameter optimization yields a fair comparison; for models with much larger search spaces, like YAformer, those 100 trials may not find the best configuration, and the paper explicitly notes there is no guarantee of a global optimum.

What would settle it

Rerun the benchmark with a substantially larger HPO budget (e.g., 500–1,000 trials) for YAformer, TFT, and the standard Transformer on the same three datasets, and check whether the plain Transformer still achieves the lowest error; if YAformer or another tuned model overtakes it, the paper's conclusion about architectural modifications being unnecessary would fail. A second check would replace the idealized reanalysis weather data with lagged or forecast weather to see whether the reported gaps persist under operational data availability.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For a new load-forecasting project, a plain Transformer with covariates appears to be a stronger default starting point than specialized Transformer variants or classical methods.
  • Transformer-based models occupy the top positions across all three grid levels, so non-Transformer baselines alone are insufficient for benchmarking new load-forecasting methods.
  • Zero-shot foundation-model forecasting is viable at feeder and client level, but special events (holidays, long weekends, school holidays) remain a failure mode on aggregated TSO data; adding binary features cuts but does not close the gap.
  • Long input contexts, calendar and weather covariates, and continuous monthly retraining each materially improve accuracy; models that omit them are likely to underperform.
  • In the runtime regime studied, dataset-specific models have inference below 100 ms, while Chronos-2 is slower but avoids training time, making the choice application-dependent.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the paper's conclusion would give YAformer and TFT substantially larger hyperparameter-optimization budgets and check whether the plain Transformer still wins; the paper itself notes that 100 trials do not guarantee a global optimum.
  • The benchmark uses reanalysis weather data that is not available in real time, so the reported error reductions may be smaller in operational settings unless live forecasts of equivalent quality are available.
  • Combining the short-horizon strength of Chronos-2 with the long-horizon strength of the Transformer is a natural extension that could lower overall error further, since the two models are complementary across horizons.
  • The same benchmark structure could be adapted to probabilistic forecasting, which would address the volatility and uncertainty that the paper identifies as especially relevant for feeder- and client-level data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents a benchmark for short-term (96-hour) electrical load forecasting across three grid levels: a TSO control area (TransnetBW), low-voltage feeders (FeederBW), and individual clients (Electricity-287). Ten methods are compared under a common protocol: Bayesian HPO with 100 trials per model, monthly retraining on the two longer datasets, and rolling test evaluation. The central claim is that Transformer-based models achieve the lowest forecast errors on all three datasets, with the best Transformer reducing error by 6.6–10.7% relative to the best non-Transformer method. A new flexible architecture, YAformer, is introduced and compared with a standard Transformer; the standard Transformer performs better, which the authors attribute to the larger HPO search space of YAformer. The paper also evaluates the Chronos-2 foundation model and finds it competitive on the feeder and client datasets but weak on TSO special events. Ablations highlight the importance of covariates, long contexts, and continuous retraining. The code, preprocessed data, and model predictions are released publicly.

Significance. If the central claim is robust, the benchmark would be a useful reference for practitioners: it spans three aggregation levels, uses public datasets, applies a consistent evaluation protocol, and includes state-of-the-art methods plus a strong zero-shot foundation model. The public release of data and predictions is a concrete strength that enables future comparisons. The ablation studies on covariates, context length, and retraining provide actionable guidance. However, the headline claim rests on point estimates from a single HPO run and single training seed per model, with no variance quantification, and the per-series Diebold–Mariano tests show a mixed significance pattern that is not fully consistent with the 'consistently outperform' phrasing. The paper is transparent about these limitations, but they are load-bearing for the main conclusion and require additional experimental support before the claim can be accepted as stated.

major comments (4)
  1. [§3.3, §6, Table 5] The central result is based on one HPO run and one training run per model. The reported improvements (6.6–10.7%) are point estimates with no confidence intervals, and the paper itself states in §6 that 'there is no guarantee that the Transformer configuration... is the global optimum.' Equal trial counts (100) do not equalize optimization difficulty, since YAformer searches a strictly larger space and other models have different search-space shapes. Please provide multiple seeds for the final configurations (at least for the top models), report variance measures (e.g., std/CI across seeds), or perform a sensitivity analysis around the HPO budget. Without this, the ranking of Transformer versus CNN/LSTM/N-HITS could be an artifact of HPO luck or seed variability.
  2. [§4.4, Figure 5] The Diebold–Mariano tests are applied to loss differentials from overlapping forecast origins: forecasts are generated hourly for a 96-step horizon, so errors at nearby origins are highly autocorrelated. The manuscript does not describe any autocorrelation-robust covariance estimator (e.g., Newey–West) or specify an effective horizon for the overlapping window, which can overstate significance. Please use non-overlapping origins (e.g., daily or weekly spacing), apply HAC standard errors, or otherwise justify the DM-test implementation for strongly overlapping multi-step forecasts.
  3. [§5.1, Figure 5] The claim that Transformers 'consistently outperform' established methods is weakened by the paper's own per-series DM results. On FeederBW, the Transformer is significantly better than Chronos-2 on 46% of series, while Chronos-2 is significantly better on 20% (a tie on the rest). On Electricity-287, the Transformer is significantly better on 39% of series and Chronos-2 on 14%, despite Chronos-2 having a lower aggregate nMAE. These results indicate the ranking is not uniform across series. Please clarify whether the headline claim refers to aggregate mean error or per-series superiority, and consider reporting paired bootstrap confidence intervals across series to quantify the aggregate advantage.
  4. [§4.2, §6] The use of perfect reanalysis weather data (ERA5) and short-term weather forecasts is acknowledged as a limitation in §6, but it deserves more emphasis as a scoping condition. The absolute error levels and the relative ranking could change under real-time weather forecasts, especially for the TSO and feeder datasets where weather covariates are reported to have a large effect. Since the paper makes practical recommendations, please either evaluate with forecast-available weather (or a noise-perturbed version) or state explicitly that the benchmark measures relative model performance under oracle weather, and temper the practical conclusions accordingly.
minor comments (5)
  1. [§5.2, Figure 6] The sentence 'so that it is barely visible in Figure 2' should refer to Figure 6, not Figure 2.
  2. [Table 4] The row for Chronos-2 on Electricity-287 appears to have misaligned columns ('42.2 407.07.9212.95'); please check the formatting. Similar alignment issues may exist in other rows.
  3. [§5.3] The 'Static' ablation is described only as 'trained once per year.' Please specify the exact training window and validation procedure for the static model to make the comparison reproducible.
  4. [§3.3] The phrase 'one validation per epoch on TransnetBW and ten validations per epoch on the other datasets' is ambiguous. Does 'validation' mean a validation step (batch) or a full pass over the validation set? Please clarify.
  5. [§4.1] For Electricity-287, the validation period is described as 15.3.2014–7.8.2014, which overlaps the training set ending 14.3.2014. The boundary is clear, but a sentence explaining that the validation set is used only for early stopping/HPO and is distinct from the test set would avoid confusion.

Circularity Check

0 steps flagged

No material circularity: the benchmark rests on external public data and empirical HPO comparisons; self-citations are present but not load-bearing.

full rationale

The paper's central claim—that Transformer-based models reduce 96-hour load-forecast error by 6.6–10.7% relative to the best non-Transformer method—is an empirical benchmark result, not a derivation. The three datasets are public (ENTSO-E, FeederBW [45], UCI Electricity-287) and models are evaluated on held-out test periods with defined metrics (Section 4). No fitted parameter is relabeled as a prediction: HPO is performed on separate validation periods (Section 3.3, Section 4.1) and test errors are reported after model selection. The standard Transformer vs. YAformer comparison is not circular either. Section 3.3 notes 'the standard Transformer is a special case of YAformer with patch size 1, zero CNN and LSTM layers, pooling width 1 and full attention,' which would, if HPO were perfect, predict YAformer at least as good as the Transformer. The paper explicitly disclaims this inference in Section 6: 'there is no guarantee that the Transformer configuration (or the configuration of any other model) after HPO is the global optimum.' Thus the observed YAformer deficit is presented as an empirical HPO-search-space effect, not a forced identity. Self-citations exist—the FeederBW dataset comes from the authors' companion paper [45], and the selection of Chronos-2 cites [7,22,32], two of which are by the same group—but they are not load-bearing: the comparison methods and datasets are externally documented, Chronos-2's quality is also justified by the external GIFT-Eval benchmark [1], and the core Transformer vs. non-Transformer ranking does not reduce to these citations. The paper also flags its own limitations, e.g., 'we have used weather data that is not available at the forecast time in practice, such as reanalysis data,' and notes for Electricity-287 that 'the training set ... is contained in the pretraining data of Chronos-2, so the evaluation on this dataset is not zero-shot.' These are honesty checks rather than circular reasoning. Remaining concerns about HPO trial-count fairness and single-run point estimates are correctness/robustness risks, not circularity.

Axiom & Free-Parameter Ledger

13 free parameters · 5 axioms · 0 invented entities

The paper is an empirical benchmark; the free parameters are model hyperparameters tuned via HPO and reported in Table 5. The main domain assumptions concern representativeness of datasets, availability of perfect weather/load data, and fairness of the HPO budget. No new physical or mathematical entities are introduced.

free parameters (13)
  • HPO trial budget = 100 configurations per model
    Number of hyperparameter configurations evaluated per model; larger budgets could change relative performance, so this is a free choice affecting the benchmark conclusion.
  • Forecast horizon H = 96 hours
    Fixed by task definition; results may vary with horizon.
  • Context length search space = [168, 336, 720]
    Only these context lengths are considered for scratch-trained models; other choices might affect ranking.
  • Transformer context length = 720 (all datasets)
    Selected via HPO on validation; direct impact on Transformer performance.
  • Transformer learning rate = 0.00028 / 0.00017 / 0.00033
    HPO-selected per dataset (TransnetBW / FeederBW / Electricity-287).
  • Transformer number of layers = 1 / 3 / 4
    HPO-selected.
  • Transformer number of heads = 1 / 4 / 4
    HPO-selected.
  • Transformer dropout = 0.17 / 0.10 / 0.07
    HPO-selected.
  • YAformer context length = 336 / 720 / 720
    HPO-selected.
  • YAformer learning rate = 0.00014 / 0.00038 / 0.00056
    HPO-selected.
  • YAformer patch size = 1 (all datasets)
    HPO-selected; patch size 1 means no patching.
  • YAformer LSTM layers = 0 / 1 / 0
    HPO-selected.
  • YAformer attention type = full / sparse / full
    HPO-selected.
axioms (5)
  • domain assumption The three datasets are representative of TSO, low-voltage feeder, and individual-consumer load forecasting.
    The paper selects these datasets to represent grid levels; if they are not representative, conclusions may not generalize.
  • domain assumption Weather reanalysis data (ERA5) and immediately available past loads are appropriate proxies for real-time operational information.
    The authors acknowledge that reanalysis is not available at forecast time and load data may be delayed; this is load-bearing for absolute error claims.
  • domain assumption The HPO with 100 configurations provides a fair comparison across methods.
    Uneven tuning budgets can alter rankings; the same budget is used for all models, but sufficiency is assumed.
  • standard math Diebold–Mariano tests on per-series loss differentials are valid for these data.
    Standard statistical test; assumes stationary loss differentials and appropriate handling of overlapping forecasts.
  • domain assumption The evaluation metrics (MAE, nMAE) capture relevant aspects of forecast quality.
    Point forecast accuracy is the focus; probabilistic metrics are excluded.

pith-pipeline@v1.3.0-alltime-deepseek · 21161 in / 14348 out tokens · 109738 ms · 2026-08-01T22:31:55.741993+00:00 · methodology

0 comments
read the original abstract

Accurate load forecasting at multiple grid levels is essential for future smart grids, ranging from aggregated control area forecasts for balancing supply and demand to forecasts of individual end-consumer loads for demand-side management and energy management systems. We present a comprehensive benchmark for load forecasting across grid levels, comprising three datasets that represent a transmission system operator control area, low-voltage grid feeders, and individual end consumers. We evaluate ten methods for short-term load forecasting and find that Transformer-based approaches consistently outperform established methods, reducing forecast error by 6.6-10.7 %. To analyze the impact of architectural design, we introduce YAformer, a flexible Transformer architecture that integrates modifications from prior work and is optimized via hyperparameter optimization. However, the standard Transformer achieves superior performance, suggesting that these architectural modifications are not required for accurate load forecasting. We further evaluate the Transformer-based time-series foundation model Chronos-2, which demonstrates competitive zero-shot performance on two datasets but fails to accurately capture special events in the TSO data. Detailed analyses reveal model-specific strengths and weaknesses, and ablation studies highlight the importance of long input contexts, covariates and continuous retraining - aspects that are often overlooked in the time-series forecasting literature.

Figures

Figures reproduced from arXiv: 2607.15705 by Benjamin Sch\"afer, Jonathan Kolar, Matthias Hertel, Ralf Mikut, Sebastian P\"utz, Veit Hagenmeyer.

Figure 1
Figure 1. Figure 1: Overview of the YAformer architecture based on an encoder-decoder [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Benchmarking results on the three datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the forecast error of the Transformer and Chronos-2 over time. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Exemplary forecasts of the Transformer and Chronos-2 on the TransnetBW data, showing cases where Chronos-2 exhibits large errors. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Statistical test results on the three datasets. A Diebold-Mariano test is performed per time series for each pair of models. The fraction of time series on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Annual total runtimes for twelve monthly model trainings and 8760 inferences on the three datasets. The training time is shown as dashed bar. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation results on test data for models without weather data ( [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Taha Aksu, Gerald Woo, Juncheng Liu, Xu Liu, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. 2024. GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation. doi:10.48550/arXiv.2410.10393

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.10393 2024

  2. [2]

    Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael Bohlke-Schneider

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael B...

    doi:10.48550/arxiv.2510 2025

  3. [3]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebas- tian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. 2024. Chronos: Learning the L...

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.07815 2024

  4. [4]

    Andreas Auer, Patrick Podest, Daniel Klotz, Sebastian Böck, Günter Klambauer, and Sepp Hochreiter. 2025. TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning. doi:10.48550/arXiv.2505.23719

    doi:10.48550/arxiv.2505.23719 2025

  5. [5]

    Lukas Biewald. 2020. Experiment Tracking with Weights and Biases. wandb.com

    2020

  6. [6]

    Olivares, Boris N

    Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. 2023. NHITS: Neural Hierarchi- cal Interpolation for Time Series Forecasting.Proceedings of the AAAI Conference on Artificial Intelligence37, 6 (June 2023), 6989–6997. doi:10.1609/aaai.v37i6.25854

    doi:10.1609/aaai.v37i6.25854 2023

  7. [7]

    Wei Soon Cheong, Lian Lian Jiang, and Jamie Ng Suat Ling. 2026. Assessing Electricity Demand Forecasting with Exogenous Data in Time Series Foundation Models. doi:10.48550/ARXIV.2602.05390

    work page doi:10.48550/arxiv.2602.05390 2026

  8. [8]

    Copernicus Climate Change Service. 2020. Climate and energy indicators for Europe from 1979 to present derived from reanalysis. doi:10.24381/CDS.4BD77450

    work page doi:10.24381/cds.4bd77450 2020

  9. [9]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder-only foundation model for time-series forecasting. doi:10.48550/arXiv.2310.10688

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.10688 2024

  10. [10]

    Diebold and Roberto S

    Francis X. Diebold and Roberto S. Mariano. 1995. Comparing Predictive Accuracy. Journal of Business & Economic Statistics13, 3 (1995), 253–263. jstor:1392185 doi:10.2307/1392185

    doi:10.2307/1392185 1995

  11. [11]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2020. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InInterna- tional Conference on Learning Representations. https:...

    2020

  12. [12]

    Patrick Emami, Abhijeet Sahu, and Peter Graf. 2023. BuildingsBench: A Large- Scale Dataset of 900K Buildings and Benchmark for Short-Term Load Forecasting. In37th Conference on Neural Information Processing Systems (NeurIPS 2023) Track on Datasets and Benchmarks. doi:10.48550/arXiv.2307.00142

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.00142 2023

  13. [13]

    ENTSO-E. 2026. Transparency platform. https://transparency.entsoe.eu

    2026

  14. [14]

    Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco. 2023. TimeGPT-1. doi:10.48550/ARXIV.2310.03589

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.03589 2023

  15. [15]

    Elena Giacomazzi, Felix Haag, and Konstantin Hopf. 2023. Short-Term Elec- tricity Load Forecasting Using the Temporal Fusion Transformer: Effect of Grid Hierarchies and Data Sources. InProceedings of the 14th ACM Interna- tional Conference on Future Energy Systems. ACM, Orlando FL USA, 353–360. doi:10.1145/3575813.3597345 11

    arXiv 2023

  16. [16]

    Webb, Rob Hyn- dman, and Pablo Montero-Manso

    Rakshitha Wathsadini Godahewa, Christoph Bergmeir, Geoffrey I. Webb, Rob Hyn- dman, and Pablo Montero-Manso. 2021. Monash Time Series Forecasting Archive. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track. https://openreview.net/forum?id=wEc1mgAjU-

    2021

  17. [17]

    Miha Grabner, Yi Wang, Qingsong Wen, Boštjan Blažič, and Vitomir Štruc. 2023. A Global Modeling Framework for Load Forecasting in Distribution Networks. IEEE Transactions on Smart Grid14, 6 (Nov. 2023), 4927–4941. doi:10.1109/TSG. 2023.3264525

    arXiv 2023

  18. [18]

    Stephen Haben, Siddharth Arora, Georgios Giasemidis, Marcus Voss, and Danica Vukadinović Greetham. 2021. Review of low voltage load forecasting: Methods, applications, and recommendations.Applied Energy304 (Dec. 2021), 117798. doi:10.1016/j.apenergy.2021.117798

    arXiv 2021

  19. [19]

    2023.Core Concepts and Methods in Load Forecasting: With Applications in Distribution Networks

    Stephen Haben, Marcus Voss, and William Holderbaum. 2023.Core Concepts and Methods in Load Forecasting: With Applications in Distribution Networks. Springer International Publishing, Cham. doi:10.1007/978-3-031-27852-5

    work page doi:10.1007/978-3-031-27852-5 2023

  20. [20]

    Matthias Hertel, Lara Ambrosius, Manuel Treutlein, Ralf Mikut, and Veit Ha- genmeyer. 2025. A comparison of local, cluster-specific and global Transformer models for forecasting electrical loads of individual buildings and substations. In 2025 IEEE Kiel PowerTech. IEEE, Kiel, Germany, 1–8. doi:10.1109/PowerTech59965. 2025.11180482

    arXiv 2025

  21. [21]

    Matthias Hertel, Maximilian Beichter, Benedikt Heidrich, Oliver Neumann, Ben- jamin Schäfer, Ralf Mikut, and Veit Hagenmeyer. 2023. Transformer training strategies for forecasting multiple load time series.Energy Informatics6, 1 (Oct. 2023), 20. doi:10.1186/s42162-023-00278-z

    doi:10.1186/s42162-023-00278-z 2023

  22. [22]

    Matthias Hertel, Alexandra Nikoltchovska, Sebastian Pütz, Benjamin Schäfer, Ralf Mikut, and Veit Hagenmeyer. 2026. Explainable Load Forecasting with Covariate-Informed Time Series Foundation Models. InProceedings of the 17th ACM International Conference on Future and Sustainable Energy Systems. ACM, Banff , Alberta , Canada, 612–626. doi:10.1145/3744255.3811724

    arXiv 2026

  23. [23]

    Hansika Hewamalage, Klaus Ackermann, and Christoph Bergmeir. 2023. Forecast evaluation for data scientists: common pitfalls and best practices.Data Mining and Knowledge Discovery37, 2 (March 2023), 788–832. doi:10.1007/s10618-022- 00894-5

    doi:10.1007/s10618-022- 2023

  24. [24]

    Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. 2023. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. doi:10.48550/arXiv.2207.01848

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.01848 2023

  25. [25]

    Tao Hong. 2014. Energy Forecasting: Past, Present, and Future.Foresight: The International Journal of Applied Forecasting32 (2014), 43–48

    2014

  26. [26]

    Tao Hong, Pierre Pinson, Yi Wang, Rafał Weron, Dazhi Yang, and Hamidreza Zareipour. 2020. Energy Forecasting: A Review and Outlook.IEEE Open Access Journal of Power and Energy7 (2020), 376–388. doi:10.1109/OAJPE.2020.3029979

    arXiv 2020

  27. [27]

    Shi Bin Hoo, Samuel Müller, David Salinas, and Frank Hutter. 2025. From Tables to Time: How TabPFN-v2 Outperforms Specialized Time Series Forecasting Models. doi:10.48550/arXiv.2501.02945

    doi:10.48550/arxiv.2501.02945 2025

  28. [28]

    Rob J Hyndman and Yeasmin Khandakar. 2008. Automatic time series forecasting: the forecast package for R.Journal of Statistical Software27, 3 (2008), 1–22. doi:10.18637/jss.v027.i03

    doi:10.18637/jss.v027.i03 2008

  29. [29]

    Benedikt Kaas, Manuel Treutlein, Hannes Benedikt Gerber, Oliver Neumann, Cheewan Phatthanakhuha, Oliver Resch, Ralf Mikut, and Veit Hagenmeyer. 2026. Probabilistic Low-Voltage Peak Load Forecasting with Time Series Foundation Models Evaluated on Application-Oriented Metrics. doi:10.48550/ARXIV.2607. 01966

    doi:10.48550/arxiv.2607 2026

  30. [30]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: a highly efficient gradient boosting decision tree. InProceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 3149–

    2017

  31. [31]

    Max Kleinebrahm, Jonathan Berrisch, Philipp Eiser, Wolf Fichtner, Veit Hagen- meyer, Matthias Hertel, Nils Koster, Sebastian Lerch, Ralf Mikut, Jan Pries- mann, Melanie Schienle, Benjamin Schaefer, Jann Weinand, and Florian Ziel

  32. [32]

    Alexander Kreusel, Matthias Hertel, Moritz Noskiewicz, Frederik Zahn, Heiko Maaß, Ralf Mikut, and Veit Hagenmeyer. 2026. Covariates Are the Key to Accurate Probabilistic Building Energy Forecasting with Time Series Foundation Models. InProceedings of the 2026 ACM Sustainability Week. ACM, Banff , Alberta , Canada, 172–182. doi:10.1145/3765611.3815507

    arXiv 2026

  33. [33]

    Shiyang Li, Xiaoyong Jin, Yao Xuan, Xiyou Zhou, Wenhu Chen, Yu-Xiang Wang, and Xifeng Yan. 2019. Enhancing the Locality and Breaking the Memory Bottle- neck of Transformer on Time Series Forecasting. InAdvances in Neural Informa- tion Processing Systems, Vol. 32. Curran Associates, Inc. doi:doi/10.5555/3454287. 3454758

    doi:10.5555/3454287 2019

  34. [34]

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. 2024. Foundation Models for Time Series Analysis: A Tutorial and Survey. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6555–6565. doi:10.1145/3637528.3671451

    arXiv 2024

  35. [35]

    Arik, Nicolas Loeff, and Tomas Pfister

    Bryan Lim, Sercan O. Arik, Nicolas Loeff, and Tomas Pfister. 2021. Temporal Fusion Transformers for interpretable multi-horizon time series forecasting.International Journal of Forecasting37, 4 (Oct. 2021), 1748–1764. doi:10.1016/j.ijforecast.2021. 03.012

    doi:10.1016/j.ijforecast.2021 2021

  36. [36]

    Chenghao Liu, Taha Aksu, Juncheng Liu, Xu Liu, Hanshu Yan, Quang Pham, Silvio Savarese, Doyen Sahoo, Caiming Xiong, and Junnan Li. 2026. Moirai 2.0: When Less Is More for Time Series Forecasting. doi:10.48550/arXiv.2511.11698

    doi:10.48550/arxiv.2511.11698 2026

  37. [37]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. doi:10.48550/arXiv.1711.05101

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.05101 2019

  38. [38]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. InThe Eleventh International Conference on Learning Representations. https: //openreview.net/forum?id=Jbdc0vTOcol

    2023

  39. [39]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners.OpenAI (2019). https://cdn.openai.com/better-language-models/language_models_are_ unsupervised_multitask_learners.pdf

    2019

  40. [40]

    Ramchurn, Perukrishnen Vytelingum, Alex Rogers, and Nicholas R

    Sarvapali D. Ramchurn, Perukrishnen Vytelingum, Alex Rogers, and Nicholas R. Jennings. 2012. Putting the ’smarts’ into the smart grid: a grand challenge for artificial intelligence.Commun. ACM55, 4 (April 2012), 86–97. doi:10.1145/ 2133806.2133825

    arXiv 2012

  41. [41]

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Hena Ghonia, Rishika Bhag- watkar, Arian Khorasani, Mohammad Javad Darvishi Bayazi, George Adamopou- los, Roland Riachi, Nadhir Hassen, Marin Biloš, Sahil Garg, Anderson Schneider, Nicolas Chapados, Alexandre Drouin, Valentina Zantedeschi, Yuriy Nevmyvaka, and Irina Rish. 2023. Lag-Llama: Towards Foundati...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.08278 2023

  42. [42]

    Fátima Rodrigues and Artur Trindade. 2018. Load forecasting through functional clustering and ensemble learning.Knowledge and Information Systems57, 1 (Oct. 2018), 229–244. doi:10.1007/s10115-018-1169-y

    work page doi:10.1007/s10115-018-1169-y 2018

  43. [43]

    Frederik vom Scheidt, Hana Medinová, Nicole Ludwig, Bent Richter, Philipp Staudt, and Christof Weinhardt. 2020. Data analytics in the electricity sector – A quantitative and qualitative literature review.Energy and AI1 (Aug. 2020), 100009. doi:10.1016/j.egyai.2020.100009

    arXiv 2020

  44. [44]

    Oleksandr Shchur, Abdul Fatir Ansari, Caner Turkmen, Lorenzo Stella, Nick Erickson, Pablo Guerron, Michael Bohlke-Schneider, and Yuyang Wang. 2025. fev-bench: A Realistic Benchmark for Time Series Forecasting. doi:10.48550/ ARXIV.2509.26468

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.26468 2025

  45. [45]

    Manuel Treutlein, Pascal Bothe, Marc Schmidt, Roman Hahn, Oliver Neumann, Ralf Mikut, and Veit Hagenmeyer. 2026. Real-world energy data of 200 feeders from low-voltage grids with metadata in Germany over two years. doi:10.48550/ arXiv.2602.03521

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.03521 2026

  46. [46]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InAdvances in Neural Information Processing Systems, Vol. 30. Curran Associates, Inc., 5998–6008. https://dl.acm.org/doi/10.5555/3295222.3295349

    arXiv 2017

  47. [47]

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. 2024. Unified Training of Universal Time Series Forecasting Transformers. doi:10.48550/arXiv.2402.02592

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.02592 2024

  48. [48]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. InAdvances in Neural Information Processing Systems, Vol. 34. Curran Associates, Inc., 22419–22430. https://dl.acm.org/doi/10.5555/3540261.3541978

    arXiv 2021

  49. [49]

    Neo Wu, Bradley Green, Xue Ben, and Shawn O’Banion. 2020. Deep Transformer Models for Time Series Forecasting: The Influenza Prevalence Case. doi:10.48550/ arXiv.2001.08317

    Pith/arXiv arXiv doi:10.48550/arxiv.2001.08317 2020

  50. [50]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are Transformers Effective for Time Series Forecasting?Proceedings of the AAAI Conference on Artificial Intelligence37, 9 (June 2023), 11121–11128. doi:10.1609/aaai.v37i9.26317

    doi:10.1609/aaai.v37i9.26317 2023

  51. [51]

    Guangqi Zhang, Chuyuan Wei, Changfeng Jing, and Yanxue Wang. 2022. Short- Term Electrical Load Forecasting Based on Time Augmented Transformer.In- ternational Journal of Computational Intelligence Systems15, 1 (Aug. 2022), 67. doi:10.1007/s44196-022-00128-y

    work page doi:10.1007/s44196-022-00128-y 2022

  52. [52]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting.Proceedings of the AAAI Conference on Artificial Intelligence35, 12 (May 2021), 11106–11115. doi:10.1609/aaai.v35i12.17325

    doi:10.1609/aaai.v35i12.17325 2021

  53. [53]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin

  54. [54]

    Çakmak and Veit Hagenmeyer

    Hüseyin K. Çakmak and Veit Hagenmeyer. 2022. Using Open Data for Mod- eling and Simulation of the All Electrical Society in eASiMOV. In2022 Open Source Modelling and Simulation of Energy Systems (OSMSES). 1–6. doi:10.1109/ OSMSES54027.2022.9769145 12 Table 3. Number of trainable parameters per model on the three datasets. Model TransnetBW FeederBW Electri...

    arXiv 2022

  55. [2022]

    InProceedings of the 39th International Conference on Machine Learning

    FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting. InProceedings of the 39th International Conference on Machine Learning. PMLR, 27268–27286. https://proceedings.mlr.press/v162/zhou22g.html

  56. [2026]

    Energy-Arena: A Dynamic Benchmark for Operational Energy Forecasting

    Energy-Arena: A Dynamic Benchmark for Operational Energy Forecasting. doi:10.48550/arXiv.2604.24705

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.24705

  57. [3157]

    https://dl.acm.org/doi/10.5555/3294996.3295074

    arXiv