pith. sign in

arxiv: 2607.01966 · v1 · pith:QMKJDT76new · submitted 2026-07-02 · 💻 cs.LG

Probabilistic Low-Voltage Peak Load Forecasting with Time Series Foundation Models Evaluated on Application-Oriented Metrics

Benedikt Kaas , Manuel Treutlein , Hannes Benedikt Gerber , Oliver Neumann , Cheewan Phatthanakhuha , Oliver Resch , Ralf Mikut , Veit Hagenmeyer This is my paper

Pith reviewed 2026-07-03 17:21 UTC · model grok-4.3

classification 💻 cs.LG
keywords low-voltage load forecastingtime series foundation modelsprobabilistic forecastingpeak load predictionapplication-oriented metricsgrid asset planningChronosTabPFN-TS
0
0 comments X

The pith

Time series foundation models outperform baselines on probabilistic low-voltage peak load forecasts and tie accuracy to grid cost-risk trade-offs via a new metric.

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

The paper evaluates short-term net load forecasts on 200 real-world low-voltage feeders, comparing three time series foundation models against six baselines. It finds that Chronos-2 in particular delivers better probabilistic forecasts and peak predictions. An ablation study shows these models maintain performance even when weather data is removed. The work introduces a new metric that directly connects forecast quality to the practical balance grid operators must strike between reducing asset costs and limiting failure risk. This evaluation focuses on metrics that matter for actual grid planning rather than generic error scores.

Core claim

Chronos-2 and the other foundation models achieve superior performance over the baselines across the 200 feeders, with the new application-oriented metric showing how improved peak prediction supports lower-cost asset decisions while controlling failure risk; the models also adapt to higher uncertainty when weather covariates are omitted.

What carries the argument

The novel application-oriented metric that converts peak-forecast skill into an explicit cost-reduction versus failure-risk trade-off for grid asset planning and operation.

If this is right

  • Operators can use Chronos-2 outputs to reduce asset over-provisioning while keeping failure risk within acceptable bounds.
  • Foundation models remain effective for low-voltage forecasting even when weather forecasts are unavailable or unreliable.
  • Peak-focused evaluation reveals advantages that standard error metrics miss.
  • The same models can support probabilistic planning without extensive manual feature engineering.

Where Pith is reading between the lines

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

  • The metric could be adapted to medium-voltage or transmission-level planning problems where similar cost-risk tensions exist.
  • Retraining the foundation models on more diverse feeder data might further improve generalization beyond the tested 200 sites.
  • The observed robustness to missing weather data suggests these models could integrate easily into systems with incomplete sensor coverage.

Load-bearing premise

The 200 real-world low-voltage feeders are representative of the wider population and the new metric correctly reflects the actual cost-risk trade-off that operators face.

What would settle it

Running the same models and metric on a substantially larger or differently distributed set of feeders and checking whether the implied cost-risk balance matches recorded planning costs and observed failure events.

Figures

Figures reproduced from arXiv: 2607.01966 by Benedikt Kaas, Cheewan Phatthanakhuha, Hannes Benedikt Gerber, Manuel Treutlein, Oliver Neumann, Oliver Resch, Ralf Mikut, Veit Hagenmeyer.

Figure 1
Figure 1. Figure 1: Visualization of the four steps to compute the application-oriented metric. The metric derives the parameters [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Excerpt of three LV load forecasts of feeder 𝐹141 from Chronos-2 with a forecast horizon of four days from May 1, 2024 to May 12, 2024 (UTC). Note the wider intervals produced by the univariate model and the model predicting significant feed-in on all days, while the weather-influenced forecast exhibits narrower intervals and is able to capture the time series better. best values for the consumer and produ… view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of the Winkler score for Chronos-2 over the forecast horizon. All forecasts over all LV feeders are aggregated by their forecast horizon. The vertical lines indicate full days from midnight UTC, the start of all forecasts. The Winkler score distribution barely differs between the configurations during the nights, while the score values during the day are often much higher for the univariate mo… view at source ↗
Figure 4
Figure 4. Figure 4: Frequency of six selected metadata columns with respect to all [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Active power with 15 minutes mean aggregation which is used as target data from the dataset used in our experiments [59]. The histogram over all values is supplemented with histograms for the mean, minimum and maximum value per LV feeder. The figure illustrates the diversity of the LV feeders. train test (seen) ignored test (unseen) Feeder ID Time 1 161 200 2023/04/01 2024/04/01 2025/04/01 [PITH_FULL_IMAG… view at source ↗
Figure 6
Figure 6. Figure 6: Train-test split used in the experiment for the models. While fully [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
read the original abstract

Low-voltage load forecasting is an important component in current and future energy systems with a high degree of electrification and decentralized generation. However, current forecasting methods require significant manual effort, often lack uncertainty estimation and proper peak prediction, and they are often not adequately evaluated in terms of grid requirements. In the present study, we provide an extensive evaluation of short-term net load forecasts of 200 real-world low-voltage feeders with a focus on the rapidly evolving time series foundation models. Our study compares Chronos-Bolt, Chronos-2 and TabPFN-TS to six baseline models and demonstrates superior performance, in particular for Chronos-2. An ablation study, in which weather covariates are omitted, shows that time series foundation models adapt to increased uncertainty, despite the importance of weather information. A novel application-oriented metric links the model's forecasting capabilities in peak prediction to the trade-off in grid asset planning and operation between cost reduction and minimizing the risk of failure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper evaluates short-term probabilistic net load forecasting on 200 real-world low-voltage feeders, comparing three time series foundation models (Chronos-Bolt, Chronos-2, TabPFN-TS) against six baselines. It reports superior performance for Chronos-2, presents an ablation study omitting weather covariates to show adaptation to uncertainty, and introduces a novel application-oriented metric intended to connect peak-prediction accuracy to the cost-risk trade-off in grid asset planning and operation.

Significance. If the empirical superiority and metric hold under scrutiny, the work would be significant for demonstrating practical utility of foundation models in energy systems with real feeder data and an application-specific evaluation. The ablation study and use of 200 feeders are strengths that support the central empirical claim.

major comments (2)
  1. [Abstract] Abstract: the novel application-oriented metric is presented as linking peak prediction to the explicit cost-risk trade-off faced by grid operators, yet no derivation, mapping to cost or failure-probability functions, or external validation against operator decisions is supplied. This is load-bearing for the claim of practical utility.
  2. [Results] Results / Discussion: the claim that the 200 feeders support generalization of Chronos-2 superiority requires explicit discussion of representativeness and sampling; without it, performance on this specific set does not establish broader applicability to the population of low-voltage feeders.
minor comments (2)
  1. [Methods] Methods: baseline implementations, exact probabilistic metrics (e.g., CRPS variants), statistical significance tests, and train/test splits should be described with sufficient detail for reproducibility.
  2. Ensure figure captions and table footnotes clearly define all abbreviations and units used in the application-oriented metric.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will revise the manuscript to incorporate additional details where needed.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the novel application-oriented metric is presented as linking peak prediction to the explicit cost-risk trade-off faced by grid operators, yet no derivation, mapping to cost or failure-probability functions, or external validation against operator decisions is supplied. This is load-bearing for the claim of practical utility.

    Authors: We agree that the presentation of the novel metric would be strengthened by an explicit derivation and mapping to cost and failure-probability functions. In the revised manuscript we will expand the relevant section (and update the abstract accordingly) to include the mathematical formulation of the metric, its connection to asset cost-risk trade-offs, and a clearer statement of its assumptions and limitations. No external validation against real operator decisions was performed, as the metric is intended as a proxy; we will note this explicitly. revision: yes

  2. Referee: [Results] Results / Discussion: the claim that the 200 feeders support generalization of Chronos-2 superiority requires explicit discussion of representativeness and sampling; without it, performance on this specific set does not establish broader applicability to the population of low-voltage feeders.

    Authors: We concur that an explicit discussion of dataset representativeness is required. In the revision we will add a paragraph in the Results/Discussion section describing the sampling procedure used to select the 200 feeders, key characteristics (geographic distribution, load types, voltage levels), and the limitations on generalizing Chronos-2 superiority beyond this sample to the full population of low-voltage feeders. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical comparison on held-out data

full rationale

The paper conducts an empirical model comparison of Chronos variants and TabPFN-TS against six baselines on 200 real-world low-voltage feeders, with an ablation and a novel application-oriented metric for peak prediction. No equations, derivations, or self-citations are load-bearing; performance claims rest on external held-out data and baselines rather than any reduction of predictions to fitted inputs or self-defined quantities. The novel metric is introduced without derivation from the paper's own parameters, satisfying the self-contained empirical benchmark criterion.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical superiority of Chronos-2 and the validity of the new metric; the main unverified premise is the representativeness of the 200-feeder sample.

axioms (1)
  • domain assumption The 200 low-voltage feeders constitute a representative sample for evaluating general forecasting performance and the new metric.
    All reported performance numbers and the metric's usefulness depend on this assumption.

pith-pipeline@v0.9.1-grok · 5728 in / 1265 out tokens · 47409 ms · 2026-07-03T17:21:38.563877+00:00 · methodology

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

71 extracted references · 59 canonical work pages · 11 internal anchors

  1. [1]

    Abdul Fatir Ansari, Caner Turkmen, Oleksandr Shchur, and Lorenzo Stella. 2024. Fast and accurate zero-shot forecasting with Chronos-Bolt and AutoGluon. AWS Artificial Intelligence. (Dec. 2024). Retrieved Apr. 24, 2026 from https://aws.am azon.com/de/blogs/machine-learning/fast-and-accurate-zero-shot-forecasti ng-with-chronos-bolt-and-autogluon/

    2024

  2. [2]

    Yvenn Amara-Ouali, Bachir Hamrouche, Guillaume Principato, and Yannig Goude. 2025. Quantifying the Uncertainty of Electric Vehicle Charging with Probabilistic Load Forecasting.World Electric Vehicle Journal, 16, 2, (Feb. 2025),

    2025

  3. [3]

    doi:10.3390/wevj16020088

    work page doi:10.3390/wevj16020088

  4. [4]

    Abdul Fatir Ansari et al. 2025. Chronos-2: From Univariate to Universal Fore- casting. (Oct. 2025). arXiv: 2510.15821[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Abdul Fatir Ansari et al. 2024. Chronos: Learning the Language of Time Series. (Nov. 2024). arXiv: 2403.07815[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    Jason Ansel et al. 2024. PyTorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In29th ACM Inter- national Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS ’24). ACM, (Apr. 2024). doi:10.1145/36206 65.3640366

    work page doi:10.1145/36206 2024

  7. [7]

    Sebastian Pineda Arango et al. 2025. ChronosX: Adapting Pretrained Time Series Models with Exogenous Variables. (Mar. 2025). arXiv: 2503.12107[cs]

    work page arXiv 2025

  8. [8]

    Marcel Arpogaus, Marcus Voss, Beate Sick, Mark Nigge-Uricher, and Oliver Dürr

  9. [9]

    2023), 4902–4911

    Short-Term Density Forecasting of Low-Voltage Load Using Bernstein- Polynomial Normalizing Flows.IEEE Transactions on Smart Grid, 14, 6, (Nov. 2023), 4902–4911. doi:10.1109/TSG.2023.3254890

    work page doi:10.1109/tsg.2023.3254890 2023

  10. [10]

    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. (May 2025). arXiv: 2505.23719 [cs]

    work page arXiv 2025

  11. [11]

    Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan

  12. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence, 47, 4, (Apr

    Foundation Models Defining a New Era in Vision: A Survey and Outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47, 4, (Apr. 2025), 2245–2264. doi:10.1109/TPAMI.2024.3506283

    work page doi:10.1109/tpami.2024.3506283 2025

  13. [13]

    Cristian Bodnar et al. 2025. A foundation model for the Earth system.Nature, 641, 8065, (May 2025), 1180–1187. doi:10.1038/s41586-025-09005-y

    work page doi:10.1038/s41586-025-09005-y 2025

  14. [14]

    On the Opportunities and Risks of Foundation Models

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, and Simran Arora. On the Opportunities and Risks of Foundation Models. (Aug. 2021). arXiv: 2108.07258[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  15. [15]

    Cakmak and Veit Hagenmeyer

    Huseyin K. Cakmak and Veit Hagenmeyer. 2022. Using Open Data for Modeling and Simulation of the All Electrical Society in eASiMOV. In2022 Open Source Modelling and Simulation of Energy Systems (OSMSES). 2022 Open Source Mod- elling and Simulation of Energy Systems (OSMSES). IEEE, Aachen, Germany, (Apr. 2022), 1–6. doi:10.1109/OSMSES54027.2022.9769145

    work page doi:10.1109/osmses54027.2022.9769145 2022

  16. [16]

    Zhaojing Cao, Can Wan, Zijun Zhang, Furong Li, and Yonghua Song. 2020. Hy- brid Ensemble Deep Learning for Deterministic and Probabilistic Low-Voltage Load Forecasting.IEEE Transactions on Power Systems, 35, 3, (May 2020), 1881–

    2020

  17. [17]

    doi:10.1109/TPWRS.2019.2946701

    work page doi:10.1109/tpwrs.2019.2946701 2019

  18. [18]

    Ching Chang, Wei-Yao Wang, Wen-Chih Peng, and Tien-Fu Chen. 2024. LLM4TS: Aligning Pre-Trained LLMs as Data-Efficient Time-Series Forecasters.ACM Transactions on Intelligent Systems and Technology, 16, 3, 1–20. arXiv: 2308.08469 [cs]. doi:10.1145/371920710.1145/3719207

    work page internal anchor Pith review doi:10.1145/371920710.1145/3719207 2024

  19. [19]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD ’16: The 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, San Francisco California USA, (Aug. 2016), 785–794. doi:10.1145/2939672.2939785

    work page doi:10.1145/2939672.2939785 2016

  20. [20]

    Ben Cohen et al. 2025. This Time is Different: An Observability Perspective on Time Series Foundation Models. (May 2025). arXiv: 2505.14766[cs]

    work page arXiv 2025

  21. [21]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder- only foundation model for time-series forecasting. (Apr. 2024). arXiv: 2310.10688 [cs]. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    Sarkar Snigdha Sarathi Das et al. 2025. Synapse: Adaptive Arbitration of Com- plementary Expertise in Time Series Foundational Models. (Nov. 2025). arXiv: 2511.05460[cs]

    work page arXiv 2025

  23. [23]

    Samuel Dooley, Gurnoor Singh Khurana, Chirag Mohapatra, Siddartha Naidu, and Colin White. 2023. ForecastPFN: Synthetically-Trained Zero-Shot Forecast- ing. (Nov. 2023). arXiv: 2311.01933[cs]

    work page arXiv 2023

  24. [24]

    Nguyen, Wesley M

    Vijay Ekambaram, Arindam Jati, Pankaj Dayama, Sumanta Mukherjee, Nam H. Nguyen, Wesley M. Gifford, Chandra Reddy, and Jayant Kalagnanam. 2024. Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series. InAdvances in Neural Information Pro- cessing Systems. A. Globerson, L. Mackey, D. Belgrave, A. F...

    work page doi:10.52202/079017-2359 2024

  25. [25]

    Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecasting. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. KDD ’23: The 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, Long Beach CA...

    work page doi:10.1145/3580305.3599533 2023

  26. [26]

    Epoch AI. 2026. Data on AI models. (Apr. 2026). Retrieved Apr. 24, 2026 from https://epoch.ai/data/ai-models

    2026

  27. [27]

    Anthony Faustine, Nuno Jardim Nunes, and Lucas Pereira. 2025. Efficiency Through Simplicity: MLP-Based Approach for Net-Load Forecasting With Un- certainty Estimates in Low-Voltage Distribution Networks.IEEE Transactions on Power Systems, 40, 1, (Jan. 2025), 46–56. doi:10.1109/TPWRS.2024.3400123

    work page doi:10.1109/tpwrs.2024.3400123 2025

  28. [28]

    Kun Feng, Shaocheng Lan, Yuchen Fang, Wenchao He, Lintao Ma, Xingyu Lu, and Kan Ren. 2025. Kairos: Towards Adaptive and Generalizable Time Series Foundation Models. (Sept. 2025). arXiv: 2509.25826[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    Shanghua Gao, Teddy Koker, Owen Queen, Thomas Hartvigsen, Theodoros Tsiligkaridis, and Marinka Zitnik. 2024. UniTS: A Unified Multi-Task Time Series Model. (Nov. 2024). arXiv: 2403.00131[cs]

    work page arXiv 2024

  30. [30]

    Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco. 2024. TimeGPT-1. (May 2024). arXiv: 2310.03589[cs]

    work page arXiv 2024

  31. [31]

    Ciaran Gilbert, Jethro Browell, and Bruce Stephen. 2023. Probabilistic load fore- casting for the low voltage network: Forecast fusion and daily peaks.Sustainable Energy, Grids and Networks, 34, (June 2023), 100998. doi:10.1016/j.segan.2023.10 0998

    work page doi:10.1016/j.segan.2023.10 2023

  32. [32]

    González-Sopeña, V

    J.M. González-Sopeña, V. Pakrashi, and B. Ghosh. 2021. An overview of per- formance evaluation metrics for short-term statistical wind power forecasting. Renewable and Sustainable Energy Reviews, 138, (Mar. 2021), 110515. doi:10.1016 /j.rser.2020.110515

    work page arXiv 2021

  33. [33]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. 2024. MOMENT: A Family of Open Time-series Foundation Models. (Oct. 2024). arXiv: 2402.03885[cs]

    work page arXiv 2024

  34. [34]

    Lars Graf, Thomas Ortner, Stanisław Woźniak, and Angeliki Pantazi. 2025. FlowState: Sampling Rate Invariant Time Series Forecasting. (Aug. 2025). arXiv: 2508.05287[cs]

    work page arXiv 2025

  35. [35]

    Stephen Haben, Siddharth Arora, Georgios Giasemidis, Marcus Voss, and Danica Vukadinovic Greetham. 2021. Review of Low Voltage Load Forecasting: Methods, Applications, and Recommendations.Applied Energy, 304, (Dec. 2021), 117798. doi:10.1016/j.apenergy.2021.117798

    work page doi:10.1016/j.apenergy.2021.117798 2021

  36. [36]

    Stephen Haben, Georgios Giasemidis, Florian Ziel, and Siddharth Arora. 2019. Short term load forecasting and the effect of temperature at the low voltage level.International Journal of Forecasting, 35, 4, (Oct. 2019), 1469–1484. doi:10.1 016/j.ijforecast.2018.10.007

    2019

  37. [37]

    Stephen Haben, Jonathan Ward, Danica Vukadinovic Greetham, Colin Singleton, and Peter Grindrod. 2014. A new error measure for forecasts of household- level, high resolution electrical energy consumption.International Journal of Forecasting, 30, 2, (Apr. 2014), 246–256. doi:10.1016/j.ijforecast.2013.08.002

    work page doi:10.1016/j.ijforecast.2013.08.002 2014

  38. [38]

    Hamann et al

    Hendrik F. Hamann et al. 2024. Foundation models for the electric power grid. Joule, 8, 12, (Dec. 2024), 3245–3258. doi:10.1016/j.joule.2024.11.002

    work page doi:10.1016/j.joule.2024.11.002 2024

  39. [39]

    Benedikt Heidrich, Matthias Hertel, Oliver Neumann, Veit Hagenmeyer, and Ralf Mikut. 2024. Using conditional Invertible Neural Networks to perform mid-term peak load forecasting.IET Smart Grid, 7, 4, (Apr. 2024), 460–472. doi:10.1049/stg2.12169

    work page doi:10.1049/stg2.12169 2024

  40. [40]

    Matthias Hertel, Sebastian Pütz, Jonathan Kolar, Benjamin Schäfer, Ralf Mikut, and Veit Hagenmeyer. 2026. A Benchmark for Electrical Load Forecasting across Grid Levels: Time-Series Transformers outperform established Methods. In 15th DACH+ Conference on Energy Informatics. Linz, Austria, (Sept. 2026). accepted

    2026

  41. [41]

    2013.Elektrische En- ergieversorgung: Erzeugung, Übertragung und Verteilung elektrischer Energie für Studium und Praxis

    Klaus Heuck, Klaus-Dieter Dettmann, and Detlef Schulz. 2013.Elektrische En- ergieversorgung: Erzeugung, Übertragung und Verteilung elektrischer Energie für Studium und Praxis. Springer Fachmedien Wiesbaden, Wiesbaden. doi:10.1007 /978-3-8348-2174-4

    2013

  42. [42]

    Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter

  43. [43]

    TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. (Sept. 2023). arXiv: 2207.01848[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  44. [44]

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. 2025. Accurate predictions on small data with a tabular foundation model.Nature, 637, 8045, (Jan. 2025), 319–326. doi:10.1038/s41586-024-08328-6

    work page doi:10.1038/s41586-024-08328-6 2025

  45. [45]

    Shi Bin Hoo, Samuel Müller, David Salinas, and Frank Hutter. 2025. From Tables to Time: How TabPFN-v2 Outperforms Specialized Time Series Forecasting Models. (May 2025). arXiv: 2501.02945[cs]

    work page arXiv 2025

  46. [46]

    Richard Yu

    Haowen Hou and F. Richard Yu. 2024. RWKV-TS: Beyond Traditional Recurrent Neural Network for Time Series Tasks. (Jan. 2024). arXiv: 2401.09093[cs]

    work page arXiv 2024

  47. [47]

    Ming Jin et al. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. (Jan. 2024). arXiv: 2310.01728[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [48]

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. 2024. Foundation Models for Time Series Anal- ysis: A Tutorial and Survey. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. KDD ’24: The 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, Barce...

    work page doi:10.1145/3637528.3671451 2024

  49. [49]

    Arik, Nicolas Loeff, and Tomas Pfister

    Bryan Lim, Sercan O. Arik, Nicolas Loeff, and Tomas Pfister. 2020. Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting. (Sept. 2020). arXiv: 1912.09363[stat]

    work page arXiv 2020

  50. [50]

    Chenghao Liu et al. 2025. Moirai 2.0: When Less Is More for Time Series Fore- casting. (Nov. 2025). arXiv: 2511.11698[cs]

    work page arXiv 2025

  51. [51]

    Xu Liu et al. 2024. Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts. (Oct. 2024). arXiv: 2410.10469[cs]

    work page arXiv 2024

  52. [52]

    Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2025. Sundial: A Family of Highly Capable Time Series Foundation Models. (May 2025). arXiv: 2502.00816[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  53. [53]

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024. Timer: Generative Pre-trained Transformers Are Large Time Series Models. (Oct. 2024). arXiv: 2402.02368[cs]

    work page arXiv 2024

  54. [54]

    Stefan Meisenbacher, Johannes Galenzowski, Kevin Förderer, Wolfgang Suess, Simon Waczowicz, Ralf Mikut, and Veit Hagenmeyer. 2025. Automation Level Taxonomy for Time Series Forecasting Services: Guideline for Real-World Smart Grid Applications. InEnergy Informatics. Vol. 15271. Bo Nørregaard Jørgensen, Zheng Grace Ma, Fransisco Danang Wijaya, Roni Irnawan...

    work page doi:10.1007/978-3-031-747 2025

  55. [55]

    Marcel Meyer, Sascha Kaltenpoth, Kevin Zalipski, Henrik Albers, and Oliver Müller. 2025. TS-Arena Technical Report – A Pre-registered Live Forecasting Platform. (Dec. 2025). arXiv: 2512.20761[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  56. [56]

    Marcel Meyer, David Zapata, Sascha Kaltenpoth, and Oliver Müller. 2025. Bench- marking Time Series Foundation Models for Short-Term Household Electricity Load Forecasting.IEEE Access, 13, 218141–218153. doi:10.1109/ACCESS.2025.36 48056

    work page doi:10.1109/access.2025.36 2025

  57. [57]

    Moreno-Munoz, J

    A. Moreno-Munoz, J. J. G. De La Rosa, R. Posadillo, and V. Pallares. 2008. Short term forecasting of solar radiation. In2008 IEEE International Symposium on Industrial Electronics. 2008 IEEE International Symposium on Industrial Elec- tronics (ISIE 2008). IEEE, Cambridge, UK, (June 2008), 1537–1541. doi:10.1109 /ISIE.2008.4676880

    work page arXiv 2008

  58. [58]

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. (Mar. 2023). arXiv: 2211.14730[cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  59. [59]

    Zarzalejo, Andreas Kazantzidis, and Stefan Wilbert

    Bijan Nouri, Yann Fabel, Niklas Blum, Dominik Schnaus, Luis F. Zarzalejo, Andreas Kazantzidis, and Stefan Wilbert. 2024. Ramp Rate Metric Suitable for Solar Forecasting.Solar RRL, 8, 24, (Dec. 2024), 2400468. doi:10.1002/solr.202400 468

    work page doi:10.1002/solr.202400 2024

  60. [60]

    Kashif Rasul et al. 2023. Lag-Llama: Towards Foundation Models for Time Series Forecasting. InR0-FoMo: Workshop on Robustness of Few-shot and Zero-shot Learning in Foundation Models at NeurIPS 2023. NeurIPS 2023

    2023

  61. [61]

    Johannes Schneider, Christian Meske, and Pauline Kuss. 2024. Foundation Models: A New Paradigm for Artificial Intelligence.Business & Information Systems Engineering, 66, 2, (Apr. 2024), 221–231. doi:10.1007/s12599-024-00851- 0

    work page doi:10.1007/s12599-024-00851- 2024

  62. [62]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2025. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. (Feb. 2025). arXiv: 2409.16040[cs]

    work page arXiv 2025

  63. [63]

    Shreyashi Shukla and Tao Hong. 2024. BigDEAL Challenge 2022: Forecasting peak timing of electricity demand.IET Smart Grid, 7, 4, 442–459. doi:10.1049/st g2.12162

    work page doi:10.1049/st 2024

  64. [64]

    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. (Feb. 3, 2026). arXiv: 2602.03521[eess]. 13

    work page arXiv 2026

  65. [65]

    Manuel Treutlein, Marc Schmidt, Roman Hahn, Matthias Hertel, Benedikt Hei- drich, Ralf Mikut, and Veit Hagenmeyer. 2025. Generating peak-aware pseudo- measurements for low-voltage feeders using metadata of distribution system operators.IET Smart Grid, 8, 1, (Jan. 2025), e12210. doi:10.1049/stg2.12210

    work page doi:10.1049/stg2.12210 2025

  66. [66]

    Xue Wang, Tian Zhou, Jinyang Gao, Bolin Ding, and Jingren Zhou. 2025. Output Scaling: YingLong-Delayed Chain of Thought in a Large Pretrained Time Series Forecasting Model. (May 2025). arXiv: 2506.11029[cs]

    work page arXiv 2025

  67. [67]

    Çakmak, and Veit Hagenmeyer

    Dorina Werling, Benedikt Heidrich, Hüseyin K. Çakmak, and Veit Hagenmeyer

  68. [68]

    InProceedings of the Thirteenth ACM International Conference on Future Energy Systems

    Towards line-restricted dispatchable feeders using probabilistic forecasts for PV-dominated low-voltage distribution grids. InProceedings of the Thirteenth ACM International Conference on Future Energy Systems. E-Energy ’22: The Thirteenth ACM International Conference on Future Energy Systems. ACM, Virtual Event, (June 2022), 395–400. doi:10.1145/3538637.3538868

    work page doi:10.1145/3538637.3538868 2022

  69. [69]

    Robert L. Winkler. 1972. A decision-theoretic approach to interval estimation. Journal of the American Statistical Association, 67, 337, 187–191. doi:10.1080/016 21459.1972.10481224

    work page doi:10.1080/016 1972

  70. [70]

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. 2024. Unified Training of Universal Time Series Forecasting Transformers. (May 2024). arXiv: 2402.02592[cs]

    work page arXiv 2024

  71. [71]

    arXiv preprint arXiv:2302.11939 , year=

    Tian Zhou, PeiSong Niu, Xue Wang, Liang Sun, and Rong Jin. 2023. One Fits All: Power General Time Series Analysis by Pretrained LM. (Oct. 2023). arXiv: 2302.11939[cs]. A Time Series Foundation Models A.1 Paradigm Change Foundation models are large-scale models trained on broad data frequently using self-supervision [11]. They are often characterized by em...

    work page arXiv 2023