Probabilistic Forecasting for Building Energy Systems using Time-Series Foundation Models

Ankush Chakrabarty; Christopher R. Laughman; Francois Germain; Gordon Wichern; Jing Liu; Navid Azizan; Toshiaki Koike-Akino; Ye Wang; Young Jin Park

REVIEW 3 major objections 5 minor 1 cited by

Probabilistic Forecasting for Building Energy Systems using Time-Series Foundation Models

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning a time-series foundation model (Chronos) with low-rank adapters yields more accurate and robust probabilistic forecasts for building energy signals than deep forecasters trained from scratch, even with limited data and lower…

desk verdict Solid empirical paper with a genuinely useful LoRA fine-tuning recipe for building energy forecasting; the missing from-scratch ablation weakens the foundation-model framing but not the practical results. read the letter →

arxiv 2506.00630 v1 pith:JANCC4YF submitted 2025-05-31 cs.LG

Young Jin Park , Francois Germain , Jing Liu , Ye Wang , Toshiaki Koike-Akino , Gordon Wichern , Navid Azizan , Christopher R. Laughman

show 1 more author

Ankush Chakrabarty

This is my paper

classification cs.LG

keywords time-seriesfoundationmodelsprobabilisticforecastingbuildingenergylow-rankadaptationparameter-efficientfine-tuningChronostransferlearningoccupancy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Buildings need probabilistic forecasts of occupancy, CO2, plug loads, and HVAC use to run control algorithms, but target-building data is often scarce. This paper claims that time-series foundation models—large transformers pre-trained on diverse time series—can fill that gap, provided they are fine-tuned rather than used zero-shot. Using real data from a commercial net-zero building, the authors show that fine-tuning Chronos, either fully or through low-rank adapters (LoRA), cuts forecast error roughly in half relative to zero-shot and beats deep models trained from scratch such as temporal fusion transformers. The result matters because LoRA keeps that accuracy at roughly a third of the training cost, and the fine-tuned model degrades gracefully with less than two months of data and transfers to zones it has never seen.

What carries the argument

The central objects are time-series foundation models, specifically Chronos, a transformer with roughly 200 million parameters pre-trained on diverse time series by tokenizing values and training with cross-entropy, and LoRA, which inserts small trainable low-rank matrices into the query and value projections of each attention layer so only a fraction of the weights are updated. The pre-training supplies a prior over temporal patterns—daily and weekly periodicity, energy and occupancy shapes—that regularizes fine-tuning, while LoRA preserves that prior while adapting to the target building. The argument also depends on the 24-hour context, 6-hour horizon setup and on the finding that 3–5 days of context is needed for zero-shot forecasts to identify daily and weekly periodicity.

What would settle it

Train a randomly initialized Chronos of identical architecture on the same building fine-tuning data with the same LoRA protocol and learning schedule; if its MASE and unseen-zone results match the pre-trained model's, the central foundation-model claim would be falsified.

Watch

Extended reading notes

Core claim

The paper's central empirical claim is that a pre-trained probabilistic time-series foundation model, after fine-tuning, is the best available forecaster for occupant-centric building signals in data-constrained settings. Specifically, Chronos fine-tuned with LoRA at rank 4 reaches a mean absolute scaled error of 0.18 on occupancy forecasts versus 0.20 for the temporal fusion transformer, with similar gains on CO2, lighting, and HVAC signals. Full fine-tuning and LoRA both improve on zero-shot TSFM forecasts by more than 50% on several metrics; LoRA matches full fine-tuning while training about 2.3 times faster, and the fine-tuned model degrades more gracefully than TFT or DeepAR with less than two months of training data.

Load-bearing premise

The load-bearing premise is that the pre-training stage, not the transformer architecture or the fine-tuning procedure, is what gives fine-tuned TSFMs their limited-data robustness and zone-transfer accuracy; the paper never tests this by fine-tuning a randomly initialized Chronos from scratch.

Editorial extensions

If this is right

A single fine-tuned TSFM can be deployed across a building's zones and signals instead of training a separate deep model per zone, since the reported zone-transfer results show the same adapted model generalizes to unseen floors.
LoRA fine-tuning makes on-site adaptation feasible on modest hardware: about 0.6 seconds per iteration, 33% fewer FLOPs, and 2.3x faster training than full fine-tuning.
A newly commissioned building, with only weeks of sensor history, can still get probabilistic forecasts whose mean absolute scaled error is comparable to models trained on months of data.
Zero-shot use is viable only with sufficiently long context (3–5 workdays), so practitioners without fine-tuning resources should lengthen the context rather than rely on 24-hour windows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper credits limited-data robustness and zone generalization to the pre-training prior, but it does not ablate a randomly initialized Chronos trained with the same protocol; if a from-scratch model matched the results, the advantage would be architectural or optimization-driven, not a foundation-model effect.
Because Chronos is trained with cross-entropy over tokenized values rather than quantile loss, its distributional metrics (MSIS and wQL) could presumably be improved further by fine-tuning directly against those scoring rules; that is a natural but untested extension.
The univariate setup leaves correlations among occupancy, CO2, lighting, and HVAC unused; extending the framework to multivariate forecasting could improve uncertainty estimates in control loops, though current TSFMs are mainly univariate.
If the pre-training prior is what matters, then smaller TSFMs or shorter fine-tuning schedules at even lower LoRA ranks may reach the same accuracy, which would lower deployment cost further.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

First thing to know: this is a competent empirical paper that delivers a useful recipe. Fine-tune Chronos with LoRA rank 4 for building energy forecasting and you get accuracy comparable to full fine-tuning at about 2.3x lower training cost, with robust performance even when you have less than two months of data. The tables support that. The reader's CONDITIONAL verdict is about right.

What's new: the comparison of zero-shot vs full fine-tuning vs LoRA for Chronos on real building signals (occupancy, CO2, plug loads, HVAC), across eight zones and four seasons, with careful metric design (constant naive scaling, rolling-window evaluation over 3,840 test steps). The limited-data ablation and the unseen-zone transfer tests are genuinely useful additions over earlier TSFM evaluations like Mulayim et al. The paper is honest about zero-shot performance being suboptimal with short contexts.

Soft spots, in proportion. The main one is exactly what the stress-test note says: Sections 4.5 and 4.6 attribute the limited-data robustness and cross-zone generalization to the pre-training prior, but there is no ablation that fine-tunes a randomly re-initialized Chronos under the same protocol. Without that control, the 'foundation model' framing is not actually tested — the supported claim is that a 200M-parameter transformer with this fine-tuning recipe works well on this building. That is a real gap, and it is load-bearing for the title's "time-series foundation models" language, though not for the practical recipe. I would push the authors to add that ablation. Second, the generalization claim rests on a single building (SUSTIE); the paper itself acknowledges this in the conclusions, so it's a scoping issue rather than an oversight. Third, there are no significance tests, and several comparisons fall within one standard deviation (e.g., LoRA vs FullFT on HVAC point estimates); a bootstrap interval or paired test would tighten the claims. Minor: code and data are proprietary, so reproducibility is limited to the protocol description.

Overall, the empirical core is sound and the paper would benefit from a serious referee. The missing from-scratch ablation is addressable and the authors should be asked to add it. If they do, the paper makes a solid contribution to the building energy forecasting literature.

Recommendation: send to peer review. I'd cite the LoRA rank recommendation if I were working on building forecasting.

Referee Report

3 major / 5 minor

Summary. This paper evaluates three publicly available time-series foundation models (Moirai, TimesFM, Chronos) for probabilistic forecasting of four building energy signals (occupancy, CO2, lighting power, and HVAC energy) measured in an eight-zone commercial net-zero building. The protocol fixes a 24-hour context and 6-hour horizon, uses rolling evaluation over 3,840 test steps per season, and reports MASE, RMSSE, wQL, and MSIS against DeepAR, N-BEATS, and TFT. The main empirical findings are that zero-shot TSFM performance is suboptimal with a short context, that fine-tuning Chronos (both full and LoRA) improves accuracy substantially over its zero-shot mode and over the deep baselines, that LoRA rank 4 matches full fine-tuning at lower computational cost, and that fine-tuned Chronos transfers to unseen zones and degrades more gracefully than TFT under limited training data.

Significance. If the results hold, the paper provides a useful engineering case study and a practical recipe (LoRA fine-tuning of Chronos) for data-constrained building forecasting. The evaluation protocol is careful: constant naive scaling factors, rolling windows over long test periods, four seasons, eight zones, and training-size ablations are all strengths. The main weakness is that the paper's central causal attribution — that gains come from the pre-training prior rather than from the architecture or optimization recipe — is not tested, so the 'foundation model' framing outruns the evidence. Nevertheless, the empirical comparisons, if properly qualified, would be a valuable addition to the building energy forecasting literature.

major comments (3)

[§4.5, §4.6, and Conclusion bullet 3] The paper attributes its limited-data robustness and unseen-zone generalization to pre-training, stating that 'the pre-training stage of TSFMs enables models to leverage general patterns' and 'the pre-training prior of Chronos helps regularize the training.' However, no experiment trains a randomly initialized Chronos with the same architecture and the same FullFT/LoRA protocol. Because the compared baselines (N-BEATS, DeepAR, TFT) differ from Chronos in architecture, parameter count, tokenization, loss, and training protocol, their underperformance cannot be attributed to the lack of a pre-trained prior. Add a control experiment that re-trains Chronos from random initialization under identical fine-tuning conditions (same 1,000 iterations, context/horizon, data splits, and LoRA placement). Without this ablation, the supported claim is the narrower one that a particular large transformer with this fine-tuning recipe works well on SUSTIE.
[Table 1 and Abstract] The abstract claims that fine-tuned TSFMs 'consistently outperform' state-of-the-art deep forecasting models, but several reported differences are within one standard deviation and no paired significance tests are reported. For example, on HVAC, TFT MASE is 0.24±0.18 vs. Chronos+FullFT 0.20±0.14 and Chronos+PEFT 0.21±0.14; the wQL values are 0.48±0.36 vs. 0.47±0.35 and 0.46±0.33. Similarly, in Table 3 the Occ wQL for TFT (0.432) is essentially tied with Chronos+FullFT (0.435). Report paired statistical tests across zones and seasons (e.g., Diebold-Mariano or permutation tests) or temper the wording to 'often outperforms' and identify the cases where the improvement is not significant.
[§4.2] The context-length analysis is reported only for the occupancy signal, with the statement 'other channels exhibit similar trends' unsupported by any figure or table in the main text or appendix. Since the paper uses this analysis to explain the zero-shot fine-tuning gap, at least one additional signal should be shown in the main text or the claim should be explicitly marked as a conjecture. Also, the sentence 'Longer context windows were infeasible due to GPU memory constraints' is confusing given that a 20-day window (1,920 steps) is reported; clarify the actual memory limitation (e.g., what length was attempted and failed).

minor comments (5)

[Keywords] The keyword line contains a stray colon after 'Keywords:' and an extra comma in the list; please format the keyword list consistently.
[§2.2] The statement that Chronos 'natively handles irregular sampling and missing data' is not supported by the cited reference and appears inaccurate; Chronos operates on tokenized, regularly sampled series. Please correct or qualify this claim.
[Tables 3 and 4] Tables 3 and 4 have identical captions that both say 'on unseen zone'; specify the target floor (3rd vs. 4th floor) in each caption to avoid ambiguity.
[Equation (3)] The fine-tuning dataset DFT is defined as a set of pairs for i=C,...,T−H, but the index range is not fully explicit about whether the context window and forecast window are non-overlapping; please state this explicitly in the text.
[Appendix Figures A.5–A.8] The y-axis ranges differ across signals in the training-size ablation figures, making cross-signal comparisons difficult; consider using consistent scales or adding a note explaining the range choices.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported numbers are held-out test results against external baselines, and no metric or claim reduces to its own input by construction.

full rationale

This is an empirical evaluation paper, not a derivation: every reported number is a held-out test result compared against external baselines (N-BEATS, DeepAR, TFT) and against the same TSFMs in zero-shot mode, and no metric is defined in terms of a fitted parameter. The fine-tuning protocols (FullFT, LoRA) are standard procedures applied to pre-trained weights, and the predictions are evaluated on test periods hidden at training time. No prediction is the fit itself, no parameter is renamed as a prediction, and no load-bearing argument reduces to a self-citation. The stated interpretation that the pre-training prior of Chronos helps regularize fine-tuning (Sections 4.5 and 4.6) is not isolated by a from-scratch Chronos ablation, so the attribution to pre-training versus architecture is not fully controlled; however, this is a missing control or identifiability concern, not a circularity, because the claim is an interpretation of held-out results rather than an equation equivalent to its inputs. Self-citations (e.g., Park et al. 2022, Chakrabarty et al. 2024, Azizan 2020) appear only as contextual background and do not support the central empirical claims. No circular step can be exhibited, so the honest finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on protocol choices (context/horizon sizes, LoRA rank, iteration count, training-window size) and domain assumptions about stationarity, univariate sufficiency, workday-only scope, and pre-training relevance, rather than on fitted physical constants or new postulated entities. The least-supported load-bearing assumption is that the pre-training prior, not the architecture or fine-tuning procedure, drives the limited-data and generalization gains, since no from-scratch ablation is run.

free parameters (5)

LoRA rank r = 4 (default); 16 and 64 also tested
Chosen by hand; the paper recommends rank 4 or 16 for compute savings. Results are stable across ranks (Table 2), so this choice does not drive the conclusions.
Fine-tuning iterations = 1000
Fixed for all fine-tuning runs with no convergence analysis or early stopping reported (Section 4.1).
Context window length C = 96 steps (24 hours)
Main results use a 24-hour context motivated by building MPC literature; longer contexts (1 to 20 days) are tested only for zero-shot inference in Section 4.2.
Forecast horizon H = 24 steps (6 hours)
Fixed at 6 hours following the 6 to 8 hour MPC horizon convention; no horizon sensitivity analysis is reported.
Fine-tuning (training) data length = 3 months preceding each test period
Main setup, ablated over 1 week to 9 months in Section 4.5. The 'limited data' finding is defined relative to this choice.

assumptions (5)

domain assumption Chronos's pre-training corpus is distributionally relevant to building energy signals, so transfer learning from it provides the robustness and generalization gains.
Invoked in Section 4.1 (building signals are 'not completely unrelated' to pre-training datasets such as SpanishEnergyAndWeather) and Sections 4.5 to 4.6, where gains are credited to the pre-training prior; never tested by fine-tuning a randomly initialized Chronos.
domain assumption The 3-month training window preceding each test period is stationary and representative of the 40-workday held-out test period.
The train/test protocol in Section 4 assumes no concept drift or operational changes within each season; every fine-tuned result depends on this.
domain assumption Univariate per-signal forecasting is adequate, and cross-signal correlation can be ignored without harming the claims.
Stated in Section 2.1 with citations for the common finding that multivariate forecasting often underperforms univariate.
domain assumption Workday-only evaluation is appropriate, and excluding weekends does not bias the assessment of workday forecasting.
Stated in Section 3.1: weekend signals are excluded because they are 'trivial to forecast'; all claims are scoped to office workdays.
standard math The scaled metrics (MASE, RMSSE, wQL, MSIS) with a constant naive scaling factor provide a fair default comparison across signals and models.
Defined in Section 3.2; the constant scaling factor is one of several defensible normalizations and affects how much weight easy sub-series receive.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Probabilistic Forecasting for Building Energy Systems using Time-Series Foundation Models." pith.science (2026). https://pith.science/paper/JANCC4YF

@misc{pith2026250600630,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Forecasting for Building Energy Systems using Time-Series Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JANCC4YF}},
  note         = {Machine review of arXiv:2506.00630}
}

read the original abstract

Decision-making in building energy systems critically depends on the predictive accuracy of relevant time-series models. In scenarios lacking extensive data from a target building, foundation models (FMs) represent a promising technology that can leverage prior knowledge from vast and diverse pre-training datasets to construct accurate probabilistic predictors for use in decision-making tools. This paper investigates the applicability and fine-tuning strategies of time-series foundation models (TSFMs) in building energy forecasting. We analyze both full fine-tuning and parameter-efficient fine-tuning approaches, particularly low-rank adaptation (LoRA), by using real-world data from a commercial net-zero energy building to capture signals such as room occupancy, carbon emissions, plug loads, and HVAC energy consumption. Our analysis reveals that the zero-shot predictive performance of TSFMs is generally suboptimal. To address this shortcoming, we demonstrate that employing either full fine-tuning or parameter-efficient fine-tuning significantly enhances forecasting accuracy, even with limited historical data. Notably, fine-tuning with low-rank adaptation (LoRA) substantially reduces computational costs without sacrificing accuracy. Furthermore, fine-tuned TSFMs consistently outperform state-of-the-art deep forecasting models (e.g., temporal fusion transformers) in accuracy, robustness, and generalization across varying building zones and seasonal conditions. These results underline the efficacy of TSFMs for practical, data-constrained building energy management systems, enabling improved decision-making in pursuit of energy efficiency and sustainability.

Figures

Figures reproduced from arXiv: 2506.00630 by the authors.

**Figure 1.** Forecasting accuracies of the zero-shot in-context inference with TSFM models across different context window sizes [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗

**Figure 2.** Forecasting accuracy of the benchmark TFT model and the proposed TSFM approach using PEFT across different [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

**Figure 3.** Forecasting accuracy of the benchmark TFT model and the proposed TSFM approach using PEFT across different [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Forecasting accuracies of the benchmark models and the proposed TSFM approach using PEFT across different [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

ThermoForce: A Physics-Structured Interventional World Model for Building HVAC Control
eess.SY 2026-07 conditional novelty 6.5 of 10

Separating a frozen TSFM free response from a monotone forced-response operator yields control-valid HVAC world models that beat observational and covariate baselines on intervention effects and closed-loop comfort.

Reference graph

Works this paper leans on

17 extracted references · 8 canonical work pages · cited by 1 Pith paper

[3]

22 Bourdeau, M., qiang Zhai, X., Nefzaoui, E., Guo, X., Chatellier, P.,

Net zero by 2050: A roadmap for the global energy sector. 22 Bourdeau, M., qiang Zhai, X., Nefzaoui, E., Guo, X., Chatellier, P.,

work page
[5]

Assessing building control performance using physics-based simulation models and deep generative networks, in: 2024 IEEE Conference on Control Technology and Applications (CCTA), IEEE. pp. 547–554. Cox, S.J., Kim, D., Cho, H., Mago, P.,

work page 2024
[7]

A Worrying Analysis of Probabilistic Time-series Models for Sales Forecasting

A worrying analysis of probabilistic time-series models for sales forecasting.arXiv:2011.10715. Khalil, M., McGough, A.S., Pourmirza, Z., Pazhoohesh, M., Walker, S.,

work page Pith review arXiv 2011
[12]

A large-scale ensemble learning framework for demand forecasting, in: 2022 IEEE International Conference on Data Mining (ICDM), IEEE. pp. 378–387. Pérez-Lombard, L., Ortiz, J., Pout, C.,

work page 2022
[13]

Bayesian forecasting with deep generative disturbance models in stochastic MPC for building energy sys- tems, in: 2024 IEEE Conference on Control Technology and Applications (CCTA), pp. 414–419. doi:10.1109/CCTA60707.2024.10666537. Sun, L., Hu, Z., Mae, M., Imaizumi, T.,

arXiv 2024
[14]

Transformers: State-of-the-art natural language processing, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for Computational Linguistics. pp. 38–45. Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., et al.,

work page 2020
[15]

arXiv preprint arXiv:2402.02592

Unified training of universal time series forecasting transformers. arXiv preprint arXiv:2402.02592 . Xing, Z., Pan, Y., Yang, Y., Yuan, X., Liang, Y., et al.,

arXiv
[16]

Developing data-driven models to predict bems energy consumption for demand response systems, in: Modern Advances in Applied Intelligence, IEA/AIE 2014, Springer. pp. 188–197. Zeng, Y., Lee, K.,

work page 2014

Show all 17 references

[2014]

arXiv preprint arXiv:1412.6980

Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 . Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., et al.,

arXiv
[2017]

arXiv preprint arXiv:1711.05101

Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 . Min, Y., Ahn, K., Azizan, N.,

arXiv
[2018]

arXiv preprint arXiv:1810.04805

BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 . Dong, X., Luo, Y., Yuan, S., Tian, Z., Zhang, L., et al.,

arXiv
[2019]

arXiv preprint arXiv:1905.10437

N-BEATS: Neural basis expansion analysis for interpretable time series forecasting. arXiv preprint arXiv:1905.10437 . Park, Y.J., Kim, D., Odermatt, F., Lee, J., Kim, K.M.,

1905 arXiv
[2020]

Advances in Neural Information Processing Systems 33, 1877–1901

Language models are few-shot learners. Advances in Neural Information Processing Systems 33, 1877–1901. Chakrabarty, A., Vanfretti, L., Tang, W.T., Paulson, J.A., Zhan, S., et al.,

1901
[2021]

arXiv preprint arXiv:2108.07258

On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 . Botman, L., Lago, J., Fu, X., Chia, K., Wolf, J., et al.,

arXiv
[2022]

One-pass learning via bridging orthogonal gradient descent and recursive least-squares, in: 2022 IEEE 61st Conference on Decision and Control (CDC), IEEE. pp. 4720–4725. 24 Mohebi, P., Li, S., Wang, Z.,

2022
[2023]

arXiv preprint arXiv:2310.17513

The expressive power of low-rank adaptation. arXiv preprint arXiv:2310.17513 . Zhang, C., Zhang, J., Zhao, Y., Lu, J., 2025a. Automated data-driven building energy load prediction method based on generative pre-trained transformers (GPT). Energy 318, 134824. Zhang, X., Glaws, ...

arXiv
[2024]

arXiv preprint arXiv:2403.07815

Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815 . Arroyo, J., Manna, C., Spiessens, F., Helsen, L.,

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

22 Bourdeau, M., qiang Zhai, X., Nefzaoui, E., Guo, X., Chatellier, P.,

Net zero by 2050: A roadmap for the global energy sector. 22 Bourdeau, M., qiang Zhai, X., Nefzaoui, E., Guo, X., Chatellier, P.,

work page

[2] [5]

Assessing building control performance using physics-based simulation models and deep generative networks, in: 2024 IEEE Conference on Control Technology and Applications (CCTA), IEEE. pp. 547–554. Cox, S.J., Kim, D., Cho, H., Mago, P.,

work page 2024

[3] [7]

A Worrying Analysis of Probabilistic Time-series Models for Sales Forecasting

A worrying analysis of probabilistic time-series models for sales forecasting.arXiv:2011.10715. Khalil, M., McGough, A.S., Pourmirza, Z., Pazhoohesh, M., Walker, S.,

work page Pith review arXiv 2011

[4] [12]

A large-scale ensemble learning framework for demand forecasting, in: 2022 IEEE International Conference on Data Mining (ICDM), IEEE. pp. 378–387. Pérez-Lombard, L., Ortiz, J., Pout, C.,

work page 2022

[5] [13]

Bayesian forecasting with deep generative disturbance models in stochastic MPC for building energy sys- tems, in: 2024 IEEE Conference on Control Technology and Applications (CCTA), pp. 414–419. doi:10.1109/CCTA60707.2024.10666537. Sun, L., Hu, Z., Mae, M., Imaizumi, T.,

arXiv 2024

[6] [14]

Transformers: State-of-the-art natural language processing, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for Computational Linguistics. pp. 38–45. Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., et al.,

work page 2020

[7] [15]

arXiv preprint arXiv:2402.02592

Unified training of universal time series forecasting transformers. arXiv preprint arXiv:2402.02592 . Xing, Z., Pan, Y., Yang, Y., Yuan, X., Liang, Y., et al.,

arXiv

[8] [16]

Developing data-driven models to predict bems energy consumption for demand response systems, in: Modern Advances in Applied Intelligence, IEA/AIE 2014, Springer. pp. 188–197. Zeng, Y., Lee, K.,

work page 2014