Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Large Language Model-Empowered Interactive Load Forecasting

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 Interactive LLM agents let human insight beat automated load forecasting on two real datasets.

desk verdict A well-designed LLM-agent system for interactive load forecasting with a central accuracy claim that its own experiments do not yet support. read the letter →

arxiv 2505.16577 v1 pith:HBUSYITD submitted 2025-05-22 cs.LG

Yu Zuo , Dalin Qin , Yi Wang This is my paper
classification cs.LG
keywords largelanguagemodelsloadforecastingmulti-agentcollaborationinteractivemachinelearningBayesianoptimizationhuman-in-the-loopAutoMLpowersystems
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

Load forecasting models are usually static: once trained, operators cannot inject their knowledge of holidays, weather shocks, or operating conditions. This paper argues that large language models can change that by acting as a multi-agent bridge between a human operator and the forecasting pipeline. It proposes a five-agent framework for task coordination, data preparation, model search, training, and deployment, where users give natural-language guidance at each stage. The paper's central evidence is that its interactive Bayesian optimization, steered by user suggestions, beats plain Bayesian optimization on two real-world datasets, reaching lower mean absolute error in fewer trials. It also reports a full workflow costs about $0.75 in API tokens, arguing that interactive human-guided forecasting is practically affordable.

What carries the argument

The load-bearing mechanism is an interactive Bayesian optimization algorithm embedded in a five-agent LLM system. A human's natural-language instructions are converted by the Task Manager and Model Manager into external guidance that prunes the search space or adds preferences; the acquisition function then proposes the next batch of configurations conditioned on past trials and that guidance, while the Model Developer executes training and evaluation. The multi-agent layer uses subscription-based topic messaging, and each agent follows a profile, memory, workflow, and action prompt structure, with the Model Manager's think-then-act loop modeled on the ReAct paradigm. Together these pieces let the system keep the efficiency of Bayesian optimization while redirecting exploration in response to human judgment and contextual knowledge.

What would settle it

Run the interactive framework many times with operators of varied expertise, and compare each guided run against a vanilla Bayesian optimization baseline given the identical trial budget and random seed; the central claim is refuted if guided runs do not, on average, reach lower MAE than the baseline, or if an automated baseline with the same number of trials reaches the same best configuration without human guidance.

Watch

Extended reading notes

Core claim

The paper's central claim is that making load forecasting interactive, rather than a one-shot automated pipeline, lets domain knowledge improve both forecast accuracy and search efficiency. On GEFCom2014, the proposed interactive method reaches an MAE of 68.23 versus 73.41 for vanilla Bayesian optimization, and on the Guangdong dataset 1966.96 versus 2119.88; in both cases it reaches its best result in fewer trials. The claimed mechanism is not that the LLM forecasts load itself, but that it interprets user intent, coordinates specialist agents, and injects user-guided constraints into an otherwise standard Bayesian optimization loop. When the user suggests shifting trials toward under-explored model families or pruning unpromising hyperparameter ranges, the search escapes premature convergence to one model type, producing better final configurations.

Load-bearing premise

The load-bearing premise is that the experiment's human participant is a knowledgeable user whose 'proper' insights are representative of real operators—if that premise fails, the measured advantage could come from user skill rather than from the framework.

Editorial extensions

If this is right

  • Operators can monitor and steer data cleaning, feature construction, model search, and deployment through natural language, lowering the technical barrier the paper identifies as the main obstacle to practical use.
  • Users with no AI background still get the automated baseline through the framework's fallback mode, while knowledgeable users improve results by intervening, so one system serves both levels of expertise.
  • At about $0.75 per full GEFCom2014 workflow under GPT-4o pricing, the interactive overhead is small relative to the forecasting task itself, supporting the paper's feasibility argument.
  • Deployment-stage human adjustments can be substantial accuracy levers: scaling the forecast during Typhoon Saola cut MAPE from 6.99% to 3.28% on the Guangdong case.
  • The vanilla baseline's recorded failure mode, sticking with a lucky early Linear configuration until trial 150 on GEFCom2014, is exactly the premature convergence that user-guided exploration is designed to avoid.

Reading between the lines

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

  • The paper does not quantify how the reported advantage depends on user skill; one testable extension is to recruit operators at different expertise levels and compare guided runs against an automated baseline with the same trial budget.
  • The paper reports a single interactive trajectory chosen with a knowledgeable user, so the improvement is best read as a human-plus-system capability; repeating the run with several users would turn the point estimate into a distribution.
  • The same guidance-and-prune loop could transfer to other hierarchical AutoML problems where early luck with one model family causes premature convergence, though the paper does not claim that generality.
  • Because roughly two thirds of the cost is input tokens from accumulated conversation history, compressing the Task Manager's memory would likely reduce per-run cost substantially; the paper names this as future work but does not estimate the saving.
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.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a five-agent LLM-based framework (Task Manager, Preparation Assistant, Model Manager, Model Developer, Deployment Operator) that lets a non-expert user interact with a load-forecasting pipeline through natural language, including task preparation, an interactive Bayesian-optimization search over models/features/hyperparameters, and deployment-stage postprocessing. The main reported results are on GEFCom2014 and Guangdong datasets: the interactive method achieves lower MAE/MAPE than vanilla Bayesian optimization (e.g., GEFCom14 MAE 68.23 vs 73.41) and reaches its best result in fewer trials (Section IV-B, Table IV). Two postprocessing examples (Memorial Day and Typhoon Saola) illustrate user corrections. A token-consumption analysis reports a total cost of $0.751 for the GEFCom14 workflow. The abstract and conclusions state that interactive accuracy can be significantly improved when users provide proper insight.

Significance. The framework design is thoughtful: the separation of agent roles, the workflow-based prompts, the topic-based message routing, and the interactive batched BO loop are described in enough detail to be reproduced, and the token-cost analysis is a useful practical contribution. If the central claim were properly supported, this would be a valuable step toward human-in-the-loop AutoML for load forecasting. The weakness is not in the architecture but in the evidence: the reported advantage over vanilla BO is confounded with the framework's own default exploration policy and is based on a single interactive trajectory, so the paper's core claim is currently conditional on additional experiments.

major comments (3)
  1. [§IV-B1, Table IV; §III-D2] Table IV and the surrounding text in §IV-B1 compare 'vanilla Bayesian optimization' with the proposed interactive algorithm, but the proposed condition differs in several coupled ways: it uses batched interactive BO, LLM-agent-mediated planning, and user steering. Section III-D2 explicitly states that the Model Manager falls back to a default exploration/rebalancing strategy when the user offers no guidance, yet no experiment reports this user-disabled condition. Consequently, the observed gains on GEFCom14 (68.23 vs 73.41 MAE) and Guangdong (1966.96 vs 2119.88 MAE) cannot be attributed to human insight; they could come from the framework's default policy or from batched BO. The table also reports a single interactive run, while the baseline is averaged over five runs, and no significance test or error bar is given despite the word 'significantly' in the abstract. An ablation with user guidance disabled, plus repeated interactive runs, is needed before the causal claim can be accepted.
  2. [§IV-B1] The interactive user study is under-specified. §IV-B1 narrates one trajectory (Fig. 6b) but does not state the user's identity, AI expertise, instructions, whether they saw only the displayed visualizations, whether they were blinded to actual outcomes, or how many runs were performed. The abstract's qualifier 'proper insight' is circular unless the protocol specifies what counts as proper; otherwise any interaction that fails can be relabeled as improper post hoc. Please report a fixed instruction/protocol and multiple independent runs (with distribution over final MAE and trials), or clearly limit the claim to a demonstration that the interface can express a particular knowledgeable user's guidance.
  3. [§IV-B2] The postprocessing demonstrations in §IV-B2 are anecdotal and appear outcome-fitted: two hand-selected days are shown, and in both the user applies a 10% reduction selected with apparent knowledge of the actual outcome. Only one error improvement (Guangdong MAPE 6.99% to 3.28%) is reported; no alternative adjustment magnitudes, no automated postprocessing baseline, and no evaluation over a set of event days are provided. These examples establish that the Deployment Operator can express corrections, but they do not support the conclusion that human postprocessing improves accuracy in general. A small fixed-protocol evaluation over multiple event days (or a simulated-adjustment study) would be needed.
minor comments (5)
  1. [Table IV] The table header appears mangled ('TrialsAutomated Proposed'); please reformat it and clearly report trials-to-best for both the baseline and the proposed method separately.
  2. [Fig. 6b and Table IV] Fig. 6b is truncated at trial 190 with the caption explaining that no improvement occurs beyond that point, yet Table IV reports a trial count of 221 for GEFCom14; reconcile these numbers or clarify what 221 represents.
  3. [Fig. 7 captions] The order of series in the two subfigure captions is inconsistent ('Predicted, Actual and Adjusted Load' vs the legend order Actual, Predicted, Adjusted); standardize the wording.
  4. [Table I] The Load-Based Scaling expression uses \(\hat P_t \gtrless \theta\), leaving the direction of the threshold condition ambiguous; specify whether the adjustment applies when the prediction is above or below the threshold.
  5. [Eqs. (3) and (4)] The base loss \(\ell\) is used without definition; define it before presenting the weighted and asymmetric loss formulations.

Circularity Check

2 steps flagged · score 6.0 of 10

Post-processing demonstrations are hand-fitted; the main BO comparison is not circular, but the 'proper insight' claim is outcome-dependent.

  1. fitted input called prediction [Section IV-B2, Deployment postprocessing examples (Fig. 7)]
    "The user is aware of typical demand reductions on holidays, and may manually reduced the forecast after 6 a.m. by 10%, leading to improved alignment with actual load as shown in Fig.7a."

    The 10% correction factor is not derived from the model, from a pre-specified rule, or from an out-of-sample procedure. It is introduced by the user with knowledge of the event and is then scored against the same day's actual load. The improvement in MAPE therefore follows from choosing a correction that moves the forecast toward the target, rather than from a demonstrated predictive capability of the framework. The second example, where 'A user adjustment that reduced the forecast by 10% after 15:00' reduces MAPE from 6.99% to 3.28%, has the same structure: the correction is hand-fitted to the day being evaluated, so the claimed benefit is partially by construction.

  2. self definitional [Abstract and Section IV-B1]
    "Our experiments demonstrate that the interactive load forecasting accuracy can be significantly improved when users provide proper insight in key stages."

    The claim is conditional on 'proper insight', but the paper never defines properness independently of the outcome. The single reported interactive search trajectory is described as 'timely and proper guidance' only after it produced an improvement; an unsuccessful trajectory could be dismissed as improper without any fixed criterion. Thus the statement 'proper insight improves accuracy' is close to a tautology: successful guidance is, by post hoc labeling, the guidance that succeeds. This is compounded by the absence of a control condition running the same framework without user guidance, even though the Model Manager is said to have a default strategy when the user offers no guidance.

full rationale

The main quantitative comparison in Table IV, interactive Bayesian optimization against vanilla Bayesian optimization, is not circular: the baseline is a separate external algorithm, the search objective is validation MAE, and the user's guidance is based on intermediate visualizations rather than the final reported result. That part of the paper retains independent empirical content. However, the deployment-stage demonstrations in Section IV-B2 are hand-fitted: in both the Memorial Day and Typhoon Saola examples, a 10% downward correction is applied with knowledge of a special event and then evaluated against the same day's actual load, with no protocol explaining how the factor was chosen. This makes the displayed MAPE improvement partly a fitted input rather than an independent prediction. The abstract's qualifier 'proper insight' is also outcome-dependent: the paper does not define properness or fix an elicitation protocol, so any failed interaction could be labeled improper and excluded, and no no-guidance condition is run for the same framework. These issues make parts of the claimed benefit partially circular, while the central optimization comparison is not.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The paper is an engineering system paper, so the ledger is short on mathematical axioms and long on empirical assumptions. The only hand-fitted numbers are the 10% postprocessing adjustment and the ad hoc BO constraints chosen by the user. The core assumptions are that LLM agents reliably follow the multi-agent workflow, that the interactive user provides guidance that is actually beneficial, that the vanilla Bayesian optimizer is a fair baseline, and that the two datasets represent practical load forecasting. The five agents are introduced as system components with no independent validation beyond the paper's own experiments.

free parameters (2)
  • Postprocessing scaling factor lambda (10% reduction) = lambda = -0.10 (hand-specified, same value for both cases)
    In Section IV-B.2, the user applies a 10% downward adjustment after 6 a.m. on Memorial Day and after 15:00 during Typhoon Saola. The reported error reductions (for example, MAPE 6.99% to 3.28%) depend on this hand-chosen factor, which is not derived from any data-driven procedure.
  • User-selected BO constraints (CNN learning rate bound, calendar encoding exclusions for MLP) = learning rate less than 0.001; exclude 'none' and 'categorical' calendar encodings
    Section IV-B.1 describes these choices as user adjustments that accelerated convergence. They are ad hoc, based on the user's reading of intermediate plots, and are not specified as a reusable policy, so the improvement is not reproducible from the paper alone.
assumptions (4)
  • domain assumption LLM agents reliably follow the specified workflows and tool-calling protocols.
    The entire pipeline depends on GPT-4o consistently interpreting system prompts, emitting the right messages, and calling the right tools; the paper does not report failure rates, retries, or failure handling.
  • domain assumption The human participant provides 'proper' insights that improve the search.
    The claimed accuracy gain is conditional on user quality; the paper does not control for user expertise, does not use blinded users, and does not test with the non-expert operators the framework targets.
  • domain assumption The vanilla Bayesian optimization baseline is a fair and adequately resourced comparison.
    The baseline is averaged over five trials, while the proposed method appears to be a single run with user-directed early stopping; differing trial budgets and stopping rules may bias the comparison.
  • domain assumption The two datasets are representative of practical load forecasting tasks.
    GEFCom2014 and Guangdong data support the empirical claims; results could differ on other regions, seasons, or data qualities, and no generalization argument is made.
invented entities (1)
  • Specialized LLM agents (Task Manager, Preparation Assistant, Model Manager, Model Developer, Deployment Operator)
    purpose: Decompose the load forecasting pipeline into five roles to enable natural-language interaction, workflow coordination, and human guidance.
    These are proposed software components. The paper's experiments are the only evidence of their utility; no external, independent validation (for example, an ablation showing which agent is necessary) is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Model-Empowered Interactive Load Forecasting." pith.science (2026). https://pith.science/paper/HBUSYITD

@misc{pith2026250516577,
  author       = {Pith},
  title        = {Pith review of: Large Language Model-Empowered Interactive Load Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBUSYITD}},
  note         = {Machine review of arXiv:2505.16577}
}
read the original abstract

The growing complexity of power systems has made accurate load forecasting more important than ever. An increasing number of advanced load forecasting methods have been developed. However, the static design of current methods offers no mechanism for human-model interaction. As the primary users of forecasting models, system operators often find it difficult to understand and apply these advanced models, which typically requires expertise in artificial intelligence (AI). This also prevents them from incorporating their experience and real-world contextual understanding into the forecasting process. Recent breakthroughs in large language models (LLMs) offer a new opportunity to address this issue. By leveraging their natural language understanding and reasoning capabilities, we propose an LLM-based multi-agent collaboration framework to bridge the gap between human operators and forecasting models. A set of specialized agents is designed to perform different tasks in the forecasting workflow and collaborate via a dedicated communication mechanism. This framework embeds interactive mechanisms throughout the load forecasting pipeline, reducing the technical threshold for non-expert users and enabling the integration of human experience. Our experiments demonstrate that the interactive load forecasting accuracy can be significantly improved when users provide proper insight in key stages. Our cost analysis shows that the framework remains affordable, making it practical for real-world deployment.

Figures

Figures reproduced from arXiv: 2505.16577 by the authors.

Figure 1
Figure 1. The proposed multi-agent collaboration framework includes a Task Manager that handles user interaction and workflow control and other specialized [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Communication mechanism of the multi-agent framework, illustrating [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A simplified system prompt of the Task Manager with four core components: profile, memory, workflow, and actions. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Workflow of the task preparation stage. User interactions and system [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Example workflow of the model training and evaluation stage. The user’s input is first filtered and processed by the Task Manager, then forwarded [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Examples of the detailed search trajectories for the baseline and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Cumulative token cost analysis of agents in different stages of the [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 7
Figure 7. Figure 7: Users can postprocess the predicted load leveraging the situational [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages

  1. [1]

    Benchmarks and custom package for electrical load forecasting,

    Z. Wang, Q. Wen, C. Zhang, L. Sun, L. V on Krannichfeldt, and Y . Wang, “Benchmarks and custom package for electrical load forecasting,” arXiv e-prints, pp. arXiv–2307, 2023

    work page 2023

  2. [2]

    Towards human-guided machine learning,

    Y . Gil, J. Honaker, S. Gupta, Y . Ma, V . D’Orazio, D. Garijo, S. Gadewar, Q. Yang, and N. Jahanshad, “Towards human-guided machine learning,” in Proceedings of the 24th international conference on intelligent user interfaces, 2019, pp. 614–624

    work page 2019

  3. [3]

    Human-in-the-loop machine learn- ing: a state of the art,

    E. Mosqueira-Rey, E. Hern ´andez-Pereira, D. Alonso-R ´ıos, J. Bobes- Bascar´an, and ´A. Fern ´andez-Leal, “Human-in-the-loop machine learn- ing: a state of the art,” Artificial Intelligence Review, vol. 56, no. 4, pp. 3005–3054, 2023

    work page 2023

  4. [4]

    Galactica: A large language model for science,

    R. Taylor, M. Kardas, G. Cucurull, T. Scialom, A. Hartshorn, E. Saravia, A. Poulton, V . Kerkez, and R. Stojnic, “Galactica: A large language model for science,” arXiv preprint arXiv:2211.09085 , 2022

    arXiv 2022

  5. [5]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al. , “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information processing systems , vol. 35, pp. 24 824–24 837, 2022

    2022

  6. [6]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,” in 11th International Conference on Learning Representations, ICLR 2023 , 2023

    work page 2023

  7. [7]

    Toolformer: Language models can teach themselves to use tools,

    T. Schick, J. Dwivedi-Yu, R. Dess `ı, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools,” Advances in Neural Infor- mation Processing Systems , vol. 36, pp. 68 539–68 551, 2023

    2023

  8. [8]

    A survey on large language model based autonomous agents,

    L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y . Linet al., “A survey on large language model based autonomous agents,” Frontiers of Computer Science , vol. 18, no. 6, p. 186345, 2024

    work page 2024

Show all 36 references
  1. [9]

    Metagpt: Meta programming for a multi-agent collaborative framework,

    S. Hong, M. Zhuge, J. Chen, X. Zheng, Y . Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin et al., “Metagpt: Meta programming for a multi-agent collaborative framework,” in 12th International Conference on Learning Representations, ICLR 2024 , 2024

    2024

  2. [10]

    Exploring the capabilities and limitations of large language models in the electric energy sector,

    S. Majumder, L. Dong, F. Doudi, Y . Cai, C. Tian, D. Kalathil, K. Ding, A. A. Thatte, N. Li, and L. Xie, “Exploring the capabilities and limitations of large language models in the electric energy sector,” Joule, vol. 8, no. 6, pp. 1544–1549, 2024

    2024

  3. [11]

    Domain-specific large language models for fault diagnosis of heating, ventilation, and air conditioning systems by labeled-data-supervised fine-tuning,

    J. Zhang, C. Zhang, J. Lu, and Y . Zhao, “Domain-specific large language models for fault diagnosis of heating, ventilation, and air conditioning systems by labeled-data-supervised fine-tuning,” Applied Energy , vol. 377, p. 124378, 2025

    2025

  4. [12]

    Large language model interface for home energy management systems,

    F. Michelon, Y . Zhou, and T. Morstyn, “Large language model interface for home energy management systems,” arXiv preprint arXiv:2501.07919, 2025

    2025 arXiv

  5. [13]

    Rl2: Reinforce large language model to assist safe reinforcement learning for energy management of active distribution networks,

    X. Yang, C. Lin, H. Liu, and W. Wu, “Rl2: Reinforce large language model to assist safe reinforcement learning for energy management of active distribution networks,” arXiv preprint arXiv:2412.01303 , 2024

    2024 arXiv

  6. [14]

    On the potential of chatgpt to generate distribution systems for load flow studies using opendss,

    R. S. Bonadia, F. C. Trindade, W. Freitas, and B. Venkatesh, “On the potential of chatgpt to generate distribution systems for load flow studies using opendss,” IEEE Transactions on Power Systems , vol. 38, no. 6, pp. 5965–5968, 2023

    2023

  7. [15]

    Enhancing llms for power system simulations: A feedback-driven multi-agent framework,

    M. Jia, Z. Cui, and G. Hug, “Enhancing llms for power system simulations: A feedback-driven multi-agent framework,” arXiv preprint arXiv:2411.16707, 2024

    2024 arXiv

  8. [16]

    Real-time optimal power flow with linguistic stipulations: integrating gpt-agent and deep reinforcement learning,

    Z. Yan and Y . Xu, “Real-time optimal power flow with linguistic stipulations: integrating gpt-agent and deep reinforcement learning,” IEEE Transactions on Power Systems , vol. 39, no. 2, pp. 4747–4750, 2023

    2023

  9. [17]

    Large foundation models for power systems,

    C. Huang, S. Li, R. Liu, H. Wang, and Y . Chen, “Large foundation models for power systems,” in 2024 IEEE Power & Energy Society General Meeting (PESGM) . IEEE, 2024, pp. 1–5

    2024

  10. [18]

    Timegpt in load forecasting: A large time series model perspective,

    W. Liao, S. Wang, D. Yang, Z. Yang, J. Fang, C. Rehtanz, and F. Port´e-Agel, “Timegpt in load forecasting: A large time series model perspective,” Applied Energy, vol. 379, p. 124973, 2025

    2025

  11. [19]

    From news to forecast: Integrating event analysis in llm-based time series forecasting with reflection,

    X. Wang, M. Feng, J. Qiu, J. Gu, and J. Zhao, “From news to forecast: Integrating event analysis in llm-based time series forecasting with reflection,” Advances in Neural Information Processing Systems, vol. 37, pp. 58 118–58 153, 2024

    2024

  12. [20]

    Empower pre-trained large language models for building-level load forecasting,

    Y . Zhou and M. Wang, “Empower pre-trained large language models for building-level load forecasting,” IEEE Transactions on Power Systems , 2025

    2025

  13. [21]

    Buildprog: Program generation for testing ml-based building load forecasting mod- els via llm and prompt engineering,

    Y . Deng, D. Xie, R. Liang, J. Zeng, S. Tai, and D. Wang, “Buildprog: Program generation for testing ml-based building load forecasting mod- els via llm and prompt engineering,” in Proceedings of the 11th ACM International Conference on Systems for Energy-Efficient Buildings, ...

    2024

  14. [22]

    Automated data-driven building energy load prediction method based on generative pre-trained transformers (gpt),

    C. Zhang, J. Zhang, Y . Zhao, and J. Lu, “Automated data-driven building energy load prediction method based on generative pre-trained transformers (gpt),” Energy, p. 134824, 2025

    2025

  15. [23]

    Repower: An llm-driven autonomous platform for power system data-guided research,

    Y .-X. Liu, M. Jia, Y .-X. Zhang, J. Wang, G. He, S.-L. Zhong, and Z.-M. Dang, “Repower: An llm-driven autonomous platform for power system data-guided research,” Patterns, vol. 6, no. 4, 2025

    2025

  16. [24]

    Automl-agent: A multi-agent llm framework for full-pipeline automl,

    P. Trirat, W. Jeong, and S. J. Hwang, “Automl-agent: A multi-agent llm framework for full-pipeline automl,” arXiv preprint arXiv:2410.02958 , 2024

    2024 arXiv

  17. [25]

    Large language model agent: A survey on methodology, applications and challenges,

    J. Luo, W. Zhang, Y . Yuan, Y . Zhao, J. Yang, Y . Gu, B. Wu, B. Chen, Z. Qiao, Q. Long et al. , “Large language model agent: A survey on methodology, applications and challenges,” arXiv preprint arXiv:2503.21460, 2025

    2025 arXiv

  18. [26]

    Improving economic values of day-ahead load forecasts to real-time power system operations,

    Y . Wang and L. Wu, “Improving economic values of day-ahead load forecasts to real-time power system operations,” IET Generation, Trans- mission & Distribution , vol. 11, no. 17, pp. 4238–4247, 2017

    2017

  19. [27]

    Automl to date and beyond: Challenges and opportunities,

    S. K. Karmaker, M. M. Hassan, M. J. Smith, L. Xu, C. Zhai, and K. Veeramachaneni, “Automl to date and beyond: Challenges and opportunities,” Acm computing surveys (csur) , vol. 54, no. 8, pp. 1–36, 2021

    2021

  20. [28]

    Evolving ma- chine learning workflows through interactive automl,

    R. Barbudo, A. Ram ´ırez, and J. R. Romero, “Evolving ma- chine learning workflows through interactive automl,” arXiv preprint arXiv:2402.18505, 2024

    2024 arXiv

  21. [29]

    Tree-structured parzen estimator: Understanding its al- gorithm components and their roles for better empirical performance,

    S. Watanabe, “Tree-structured parzen estimator: Understanding its al- gorithm components and their roles for better empirical performance,” arXiv preprint arXiv:2304.11127 , 2023

    2023 arXiv

  22. [30]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford et al., “Gpt-4o system card,” arXiv preprint arXiv:2410.21276 , 2024

    2024 arXiv

  23. [31]

    Autogen: Enabling next-gen LLM applications via multi- agent conversations,

    Q. Wu, G. Bansal, J. Zhang, Y . Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang, “Autogen: Enabling next-gen LLM applications via multi- agent conversations,” in First Conference on Language Modeling, 2024. [Online]...

    2024

  24. [32]

    Optuna: A next- generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next- generation hyperparameter optimization framework,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , 2019, pp. 2623–2631

    2019

  25. [33]

    Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond,

    T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, and R. J. Hyndman, “Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond,” pp. 896–913, 2016

    2014

  26. [34]

    Electric load forecasting with recency effect: A big data approach,

    P. Wang, B. Liu, and T. Hong, “Electric load forecasting with recency effect: A big data approach,” International Journal of Forecasting , vol. 32, no. 3, pp. 585–597, 2016

    2016

  27. [35]

    A na ¨ıve multiple linear regression benchmark for short term load forecasting,

    T. Hong, P. Wang, and H. L. Willis, “A na ¨ıve multiple linear regression benchmark for short term load forecasting,” in 2011 IEEE power and energy society general meeting . IEEE, 2011, pp. 1–6

    2011

  28. [36]

    Deciding when to use a personalized model for load forecasting,

    D. Qin, Q. Wen, Z. Zhou, L. Sun, and Y . Wang, “Deciding when to use a personalized model for load forecasting,” IEEE Transactions on Smart Grid, 2024

    2024

Pith tools

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