Pith. sign in

REVIEW 3 major objections 5 minor 44 references

LLMs for Agentic Home Energy Management

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A tool-calling LLM agent can schedule home appliances at near-optimal cost, but only if a deterministic validator checks every committed schedule before actuation.

desk verdict Worth a serious referee, but the undefined 'realized generation' could make the headline PV savings in-sample. read the letter →

arxiv 2607.04569 v2 pith:5VOW2UXS submitted 2026-07-06 eess.SY cs.SY

Sokipriala Jonah , Queen Moses , Abiola Babatunde , Michael Ajao-Olarinoye , Daniel Bammeke This is my paper · ORCID
classification eess.SYcs.SY
keywords homeenergymanagementlargelanguagemodelsagenticAIloadschedulingdynamictariffsphotovoltaicself-consumptionconstraintsafetyMILPbenchmark
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

The paper tries to establish that a single large-language-model agent, acting through structured tools, can turn plain-English requests into cost-optimal schedules for a washing machine, dishwasher, and EV charger on a half-hourly dynamic tariff, and that this holds across multiple model backends. On ordinary tariff days, native function calling yields a median cost gap of 0% for most backends, and in a seven-day rolling deployment the agents capture 96.7–98.0% of the savings available between an off-peak timer and an exact optimization benchmark. The same tests show that low cost is not safety: under deadline, power-cap, and infeasibility conflicts, some models commit invalid or fabricated schedules, and models sometimes explain a deadline correctly while scheduling it too late. The paper therefore concludes that LLMs are viable as an orchestration layer only when an independent deterministic feasibility validator rejects or repairs invalid schedules before actuation.

What carries the argument

The central object is a flat, single-agent tool-calling ReAct loop: a model iterates between reasoning and calls to structured tools — price retrieval, weather forecasting, PV generation prediction, cheapest-window search, appliance scheduling, and infeasibility reporting — until it commits a schedule or reports infeasibility. The load-bearing mechanism is native function calling, which restricts actions to validated tool names and arguments, compared directly with text-parsed actions. The decisive safety mechanism is a model-independent deterministic feasibility validator that re-checks appliance duration, horizon limits, deadlines, calendar conditions, power caps, and duplicate or missing

What would settle it

Re-run the seven-day deployment on real smart-meter data with measured PV generation and varied occupancy, keeping the same agent, prompts, and mixed-integer linear program oracle; if the timer-to-oracle savings share falls materially below the reported 96.7–98.0%, the headline result depends on the generation model rather than the LLM's scheduling ability.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that a flat, tool-calling ReAct agent with native function calling closes the multi-appliance coordination gap that earlier hierarchical LLM schedulers left open, and that the remaining differences between models show up under constraint conflict rather than on benign cost-minimization days. The evidence is a five-backend, 1,100+ run benchmark against a mixed-integer linear program, a constraint-conflict stress suite, a weather-aware PV co-optimization arm, and a seven-day rolling deployment in which the three commercial agents capture 96.7–98.0% of the timer-to-oracle savings. Because some models fabricate schedules for infeasible requests and comm

Load-bearing premise

The load-bearing premise is that the 'realized generation' used to score weather-aware and seven-day results is genuine measured PV output and that one archived week of simulated requests represents real households; if realized generation is instead the same forecast model used for planning, the savings claims become partly self-scoring.

Editorial extensions

If this is right

  • On ordinary tariff days, at least four of five backends commit every appliance with a 0% median cost gap under native function calling, so structured tool use, not model scale, closes the coordination gap.
  • Text-parsed actions degrade reliability for every model, so systems that parse free-form LLM output will be less dependable than systems built on native function calling.
  • Constraint violations can be cheaper than feasible schedules — one infeasible case produced a mean cost gap of −152.9% — so cost alone is a misleading success metric.
  • Weather-aware scheduling should target net cost, not self-consumption: it saves about £0.139/day on overcast days but can raise cost on sunny days with negative import prices.
  • A deterministic pre-actuation validator is mandatory because correct explanations can conceal invalid committed schedules, and prompt guidance can worsen weaker models.

Reading between the lines

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

  • A consequence the paper leaves implicit: the 96.7–98.0% savings share is an upper-bound estimate if the 'realized generation' used for scoring is the same PV model used for planning rather than measured meter output; measured PV would separate scheduling skill from model bias.
  • The one-week, one-household, simulated-request design suggests the headline ratios are relative-policy evidence, not annual savings; a multi-season deployment with real occupants is the natural next test.
  • The flat agent's success on three appliances leaves open the point at which complexity demands hierarchy: adding batteries, heating, or cooling introduces inter-temporal state that the current validator does not cover.
  • A concrete extension would be to apply the same validator-gated architecture to real smart-meter data and compare measured-PV savings capture, which would also test whether conservative PV forecasting remains the safer default outside the tested week.
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 paper proposes a single flat ReAct agent with native function calling and six tools (price, weather, PV, cheapest-window, scheduling, infeasibility) for scheduling a washing machine, dishwasher, and EV charger under half-hourly Octopus Agile prices, weather forecasts, PV generation estimates, and household base load. Five LLM backends are evaluated against an extended MILP ground truth across 12 tariff days, two action interfaces, 13 constraint-conflict scenarios, weather-aware PV co-optimization, and a seven-day rolling deployment. Headline claims are that native function calling yields high scheduling success and near-optimal cost on ordinary tariff days, that constraint conflicts expose safety gaps, that weather-aware scheduling gives regime-dependent cost effects, and that the three commercial agents capture 96.7–98.0% of the timer-to-oracle savings over the evaluated week, provided every committed schedule is checked by an independent deterministic feasibility validator.

Significance. If the results hold, this is a useful and unusually careful benchmark for LLM-based residential load orchestration. The design separates model, interface, and scenario effects; fixes prompts, tools, model identifiers, and archived inputs; uses temperature 0 with R=3 repeats; and compares against an independently implemented MILP. The paper also provides trace-level failure modes (fabricated schedules, non-commits, reasoning-to-action mismatches) and a concrete safety recommendation (model-independent feasibility validation before actuation), which is well supported by the conflict experiments. The main load-bearing concern is that the headline savings are scored against a quantity called "realized generation" that is never defined; the limitations section (Sec. 6.3) states that all experiments use archived data and simulated requests, with no mention of metered PV output. If "realized" means Eq. (9) evaluated on archived weather, then the weather-aware cost savings and the 96.7–98.0% savings capture are in-sample scores with respect to an unvalidated PV model, and the manuscript's most prominent quantitative claims are conditional on parameters (P_stc, PR, NOCT, α) that are never v

major comments (3)
  1. [Sec. 5.4–5.5, Eq. (9), Tables 6–7, Sec. 6.3] The paper repeatedly scores results against "realized generation" and "realized data" (e.g., Table 6 caption: "Realized daily net cost"; Table 7: "Seven-day realized net cost"), but this quantity is never defined. The only PV model in the paper is Eq. (9), which is also the model embedded in the agent's PV prediction tool. Sec. 6.3 states that experiments use archived data and simulated requests, and nowhere is metered PV output mentioned. If "realized generation" is Eq. (9) evaluated on archived weather, then the overcast-day benefit (£0.139/day), the sunny-day penalty (£0.146/day), and the 96.7–98.0% timer-to-oracle savings are measured against the same model the agent is optimizing, not against physical PV output. The LLM-vs-MILP comparison remains fair, but the headline savings are then conditional on unvalidated PV parameters. The authors must state explicitly whether "realized" mea
  2. [Sec. 5.1, Tables 4, 6, 7] All quantitative comparisons are reported as point estimates without confidence intervals, statistical tests, or per-run distributions. Each cell has only 36–45 runs, and the seven-day deployment has 3×7×3=63 model-day runs. The £0.139/day overcast benefit and the 96.7–98.0% savings figures could be within run-to-run or day-to-day variation. The paper does report median and maximum cost gaps, but the maximum gap of 122.86% for Qwen-3 shows that medians hide heavy tails; expected cost or a distributional summary is needed. Given that the contribution statement (Sec. 1.1, C1) promises "statistical analysis," the absence of any uncertainty quantification or significance testing is a gap that should be addressed before the comparative claims can be taken as strong.
  3. [Sec. 5.2, Table 4] Cost gaps and optimality rates are computed over successful runs only, as noted in the Table 4 footnote. This is appropriate when reported alongside success rates, but the text states "Median cost gaps are 0% for all function-calling models except Llama-3.3" and "GPT-4o-mini and Gemini attain the exact MILP optimum in every run." For the text-parsed interface, success rates fall to 0.50–0.94, so a median gap of 0% over successful runs can coexist with half the runs failing to commit a schedule. The narrative should make the joint distribution of success and cost explicit in the summary statements, especially for Claude's text-parsed configuration, where a fluent confirmation without any tool call could be misinterpreted as a successful low-cost schedule. This is a presentation issue, but it affects how the headline "near-optimal" claim is read.
minor comments (5)
  1. [Throughout] There are several duplicated references: Ref. [16] and Ref. [43] are the same work (Raghavan and Giridhar), and Sec. 5.3 says "Appendix Appendix B" and Sec. C says "Appendix Appendix A." Please clean these up.
  2. [Eq. (7)] The notation A_ddl is used in constraint (7) but is not defined in the text. Please define it explicitly, e.g., as the subset of appliances subject to calendar-deadline constraints.
  3. [Fig. 3] The heatmaps in Fig. 3 contain empty-looking cells that are presumably 0.00 but are rendered as blanks/dots. A legend or explicit note about zero entries would make the failure-rate visualization unambiguous.
  4. [Sec. 5.5.1] The sentence "The evaluation includes both weekdays and a weekend" is slightly ambiguous for a seven-day period; state explicitly how many weekdays and weekend days are included and how they are distributed across the week.
  5. [Sec. 5.3] For scenario S4, the definition of "correct infeasibility reporting" is clear, but the text says "A non-committing S4 run that neither fabricates nor reports infeasibility is scored as incorrect." This scoring rule is fine, but it should be stated before the results are presented, not only in the figure caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MILP-anchored benchmark, fixed champion selection, and no load-bearing self-citations; the undefined 'realized generation' is a validation gap, not a circular step.

full rationale

The paper's load-bearing claims are anchored to an independently implemented MILP (Sec. 4) rather than to quantities fitted from LLM outputs. The MILP defines the net-cost objective and constraints from first principles, and all agent schedules are scored against this oracle (Sec. 5). The champion-selection rule was fixed before the later experiments and selects GPT-4o-mini based on earlier success/optimality and constraint-compliance criteria, not on the outcome being predicted, so it does not create a fitted-input-called-prediction loop. No load-bearing self-citation is present: the only author-linked artifacts are the GitHub/live-demo links, and the prior-work comparison [15] is external. The seven-day savings metric (Eq. 10) is a descriptive ratio of measured costs, not a derived prediction. The one legitimate concern is the undefined term 'realized generation' in Sec. 5.4 and Table 6: if it means Eq. (9) evaluated on archived weather rather than metered PV output, the weather-aware cost benefits are in-sample with respect to the PV model. However, the paper never states this equivalence, the LLM-vs-MILP comparison remains independent of that model, and this is an external-validity/validation gap, not a circular derivation. Therefore no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No constants are fitted to the headline results. The central claims rest on the MILP benchmark, archived tariff/weather data, and the assumed PV model. The PV constants are the main input assumptions; the safety conclusions are robust to them but the weather-aware cost comparisons are not.

free parameters (4)
  • PV system size P_stc = 4.0 kWp (assumed; not measured)
    Used in Eq. (9) to produce PV generation estimates; directly sets the scale of solar output in Experiments 3 and 4 and is not validated against a real installation.
  • PV performance ratio PR = 0.80
    Multiplicative efficiency factor in Eq. (9); affects realized generation and net-cost comparisons.
  • Module NOCT = 45 °C
    Temperature model constant in Eq. (9); affects generation estimates under sunny conditions.
  • Temperature coefficient α = 0.004 K⁻¹
    Standard crystalline-silicon coefficient used in Eq. (9); minor but part of the generation model.
assumptions (5)
  • domain assumption MILP in Sec. 4.2–4.3 with CBC solver provides exact ground-truth schedules.
    Optimality gaps and savings are measured against this MILP; if the formulation misstates household constraints or cost, all headline comparisons shift.
  • domain assumption Household demand profile, appliance parameters, and export price F=5p/kWh are representative.
    They are assumed inputs, not fitted or measured; results may differ for other households.
  • domain assumption Eq. (9) PV model is an adequate proxy for realized PV output.
    Scoring uses "realized generation" without evidence of measured output; the model is the same family as the forecast model.
  • domain assumption Temperature-0 decoding plus R=3 repetitions approximates deterministic model behaviour.
    Provider-side nondeterminism and version changes are acknowledged but not fully controlled.
  • domain assumption The selected 12 tariff days, 13 conflict scenarios, and one archived week are representative enough for model ranking.
    Authors state one week is not enough for monthly or annual savings; ranking on other scenarios may change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLMs for Agentic Home Energy Management." pith.science (2026). https://pith.science/paper/5VOW2UXS

@misc{pith2026260704569,
  author       = {Pith},
  title        = {Pith review of: LLMs for Agentic Home Energy Management},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5VOW2UXS}},
  note         = {Machine review of arXiv:2607.04569}
}
read the original abstract

Home Energy Management Systems (HEMS) can reduce residential electricity costs, but many require users to express everyday preferences as technical constraints. This paper presents a tool-calling ReAct agent that converts natural-language requests into schedules for multiple household appliances using half-hourly Octopus Agile prices, weather forecasts, photovoltaic generation estimates, and household demand data. Five large language model backends are evaluated against a mixed-integer linear programming benchmark across dynamic tariff conditions, constraint conflicts, weather-aware scheduling, and a seven-day rolling deployment. Native function calling achieves high scheduling success and near-optimal cost on ordinary tariff days, whereas text-parsed actions reduce reliability. Constraint-conflict testing shows that low cost does not guarantee safe or feasible behaviour. Claude Sonnet 4.6 performs best in power-cap and infeasibility scenarios, while Qwen-3 achieves higher overall constraint compliance than GPT-4o-mini. The evaluation also identifies fabricated schedules, failed commitments, and reasoning-to-action failures in which models explain a deadline correctly but commit an invalid schedule. Weather-aware scheduling reduces cost and increases solar self-consumption under overcast conditions, but provides limited or adverse economic value under some dynamic-price regimes. Across the evaluated seven-day period, the agents capture 96.7-98.0% of the savings available between an off-peak timer and the MILP oracle and outperform the rule-based policies. The results support LLM-based HEMS orchestration, provided that every committed schedule is checked by an independent deterministic feasibility validator before actuation. Code and a live demonstration are available at https://github.com/sokistar24/ecohome-experiments and https://www.ecohomeagent.com/.

Figures

Figures reproduced from arXiv: 2607.04569 by the authors.

Figure 1
Figure 1. System architecture. A LangGraph ReAct agent reasons over a tool layer spanning live external APIs, including the Octopus Energy Agile tariff and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Cost, latency, and optimality trade-off across the three function-calling [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 2
Figure 2. Optimality, latency, and inference-cost trade-o [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: It does so by shifting daytime-flexible loads into the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 3
Figure 3. Figure 3: Failure rates by scenario family and model under baseline and guided prompts. For S4, failure is one minus the correct infeasibility- reporting rate. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 3
Figure 3. Figure 3: Constraint-conflict failure rates by scenario family and model under the baseline prompt and the guided prompt. Cell values show failure rate. For [PITH_FULL_IMAGE:figures/full_fig_p010_3.png]
Figure 4
Figure 4. Figure 4: Illustrative overcast day (2026-04-26) showing how weather-aware and price-only scheduling place the same three appliances differently. Top: half [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Realized net cost under PV forecast error. Over-forecasting is more [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 5
Figure 5. Figure 5: Sensitivity of weather-aware scheduling to forecast error. The accurate [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Cumulative realized net cost over seven archived days. Agent trajec [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 6
Figure 6. Figure 6: Cumulative realized net cost over the seven archived deployment [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Web interface of the EcoHome Energy Agent. The side panel shows the live Agile price profile and household usage and solar patterns. Example [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Thermostat scheduling interaction. The agent retrieves Agile prices, identifies the cheapest and most expensive hours, quantifies the per-kWh saving, [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 4 linked inside Pith

  1. [1]

    Hevia-Koch, B

    P. Hevia-Koch, B. Wanner, R. Kuwahata, et al., Electric- ity grids and secure energy transitions, IEA: Paris, Frence (2023)

    2023

  2. [2]

    Cabot, M

    C. Cabot, M. Villavicencio, The demand-side flexibility in liberalised power market: A review of current market design and objectives, Renewable and Sustainable Energy Reviews 201 (2024) 114643

    2024

  3. [3]

    URLhttps://www.iea.org/energy-system/ energy-efficiency-and-demand/ demand-response

    International Energy Agency, Demand response, ac- cessed: 2026-07-05 (2023). URLhttps://www.iea.org/energy-system/ energy-efficiency-and-demand/ demand-response

    2026

  4. [4]

    H. Duan, S. Chen, J. Song, Characterizing regional build- ing energy consumption under joint climatic and socioe- conomic impacts, Energy 245 (2022) 123290

    2022

  5. [5]

    Felez, J

    R. Felez, J. Felez, Advanced energy management for res- idential buildings optimizing costs and efficiency through thermal energy storage and predictive control, Applied Sciences 15 (2) (2025) 880

    2025

  6. [6]

    Golmohamadi, S

    H. Golmohamadi, S. Golestan, R. Sinha, B. Bak-Jensen, Demand-side flexibility in power systems, structure, op- portunities, and objectives: A review for residential sec- tor, Energies 17 (18) (2024) 4670

    2024

  7. [7]

    D’Ettorre, M

    F. D’Ettorre, M. Banaei, R. Ebrahimy, S. A. Pourmousavi, E. Blomgren, J. Kowalski, Z. Bohdanowicz, B. Łopaciuk- Gonczaryk, C. Biele, H. Madsen, Exploiting demand-side flexibility: State-of-the-art, open issues and social per- spective, Renewable and Sustainable Energy Reviews 165 (2022) 112605. 12 Table B.8: Constraint-conflict scenario matrix. ID Family ...

    2022

  8. [8]

    Shimoda, M

    Y . Shimoda, M. Sugiyama, R. Nishimoto, T. Momonoki, Evaluating decarbonization scenarios and energy man- agement requirement for the residential sector in japan through bottom-up simulations of energy end-use demand in 2050, Applied Energy 303 (2021) 117510

Show all 44 references
  1. [9]

    Khafiso, C

    T. Khafiso, C. Aigbavboa, S. A. Adekunle, Barriers to the adoption of energy management systems in residen- tial buildings, Facilities 42 (15-16) (2024) 107–125

    2024

  2. [10]

    Michelon, Y

    F. Michelon, Y . Zhou, T. Morstyn, Large language model interface for home energy management systems, in: Pro- ceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems, 2025, pp. 590– 602

    2025

  3. [11]

    F. F. Xu, Y . Song, B. Li, Y . Tang, K. Jain, M. Bao, Z. Z. Wang, X. Zhou, Z. Guo, M. Cao, et al., Theagentcom- pany: Benchmarking llm agents on consequential real world tasks, 2024, URL https://arxiv. org/abs/2412.14161 (2024)

    2024 arXiv

  4. [12]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, Y . Cao, React: Synergizing reasoning and acting in lan- guage models, arXiv preprint arXiv:2210.03629 (2022)

    2022 arXiv

  5. [13]

    Aksitov, S

    R. Aksitov, S. Miryoosefi, Z. Li, D. Li, S. Babayan, K. Kopparapu, Z. Fisher, R. Guo, S. Prakash, P. Srini- vasan, et al., Rest meets react: Self-improvement for multi-step reasoning llm agent, 2023, URL https://arxiv. org/abs/2312.10003. 13 Figure E.8: Interface details. Left...

    2023 arXiv

  6. [14]

    Jung, Multi-agent home energy management assistant (hema), SoftwareX 34 (2026) 102633

    W. Jung, Multi-agent home energy management assistant (hema), SoftwareX 34 (2026) 102633

    2026

  7. [15]

    El Makroum, S

    R. El Makroum, S. Zwickl-Bernhard, L. Kranzl, Agentic ai home energy management system: A large language model framework for residential load scheduling, Results in Engineering (2026) 109857

    2026

  8. [17]

    Welch, E

    S. Welch, E. Obonyo, A. M. Memari, A review of the pre- vious and current challenges of passive house retrofits, Building and Environment 245 (2023) 110938.doi: 10.1016/j.buildenv.2023.110938

    2023

  9. [18]

    F. Wise, A. Gillich, P. Palmer, Retrofit information chal- lenges and potential solutions: Perspectives of house- holds, retrofit professionals and local policy makers in the united kingdom, Energy Research & Social Science 119 (2025) 103866.doi:10.1016/j.erss.2024.103866

    2025

  10. [19]

    Ibrahim, F

    M. Ibrahim, F. Harkouss, P. Biwole, F. Fardoun, S. Ould- boukhitine, Building retrofitting towards net zero energy: A review, Energy and Buildings 322 (2024) 114707.doi: 10.1016/j.enbuild.2024.114707

    2024

  11. [20]

    C. A. Faulkner, R. Lutes, S. Huang, W. Zuo, D. Vrabie, Simulation-based assessment of ASHRAE guideline 36, considering energy performance, indoor air quality, and control stability, Building and Environment 240 (2023) 110371.doi:10.1016/j.buildenv.2023.110371

    2023

  12. [21]

    Drgo ˇna, J

    J. Drgo ˇna, J. Arroyo, I. Cupeiro Figueroa, D. Blum, K. Arendt, D. Kim, E. P. Ollé, J. Oravec, M. Wetter, D. L. Vrabie, L. Helsen, All you need to know about model predictive control for buildings, Annual Reviews in Con- trol 50 (2020) 190–232.doi:10.1016/j.arcontrol. 2020.09.001

    2020 doi

  13. [22]

    Michailidis, I

    P. Michailidis, I. Michailidis, F. Minelli, H. H. Coban, E. Kosmatopoulos, Model predictive control for smart buildings: Applications and innovations in energy man- agement, Buildings 15 (18) (2025) 3298

    2025

  14. [23]

    Y . Zong, G. M. Böning, R. M. Santos, S. You, J. Hu, X. Han, Challenges of implementing economic model predictive control strategy for buildings interacting with smart energy systems, Applied Thermal Engineering 114 (2017) 1476–1486

    2017

  15. [24]

    Z. Nagy, G. Henze, S. Dey, J. Arroyo, L. Helsen, X. Zhang, B. Chen, K. Amasyali, K. Kurte, A. Zamzam, et al., Ten questions concerning reinforcement learning for building energy management, Building and Environ- ment 241 (2023) 110435.doi:10.1016/j.buildenv. 2023.110435

    2023

  16. [25]

    Zhang, S

    H. Zhang, S. Seal, D. Wu, F. Bouffard, B. Boulet, Build- ing energy management with reinforcement learning and model predictive control: A survey, IEEE Access 10 (2022) 27853–27862

    2022

  17. [26]

    L. Yu, S. Qin, M. Zhang, C. Shen, T. Jiang, X. Guan, A re- view of deep reinforcement learning for smart building en- ergy management, IEEE Internet of Things Journal 8 (15) (2021) 12046–12063. 14

    2021

  18. [27]

    H. Li, L. Zhang, H. Zhou, T. Hong, Mcp-enabled agen- tic ai workflow for building energy modelling: framework and use cases, Journal of Building Performance Simula- tion (2026) 1–27

    2026

  19. [28]

    Sapkota, K

    R. Sapkota, K. I. Roumeliotis, M. Karkee, Ai agents vs. agentic ai: A conceptual taxonomy, applications and chal- lenges, Information Fusion (2025) 103599

    2025

  20. [29]

    Hosseini, H

    S. Hosseini, H. Seilani, The role of agentic ai in shap- ing a smart future: A systematic review, Array 26 (2025) 100399

    2025

  21. [30]

    Majumder, L

    S. Majumder, L. Dong, F. Doudi, Y . Cai, C. Tian, D. Kalathil, K. Ding, A. A. Thatte, N. Li, L. Xie, Ex- ploring the capabilities and limitations of large language models in the electric energy sector, Joule 8 (6) (2024) 1544–1549

    2024

  22. [31]

    Zhang, R

    H. Zhang, R. Zhang, W. Zhang, D. Niyato, Y . Wen, C. Miao, Advancing generative artificial intelligence and large language models for demand side management with internet of electric vehicles, IEEE Internet of Things Jour- nal (2026)

    2026

  23. [32]

    Z. Niu, J. Li, Q. Ai, J. Jiang, Q. Yang, H. Zhou, Ev charging system considering power dispatching based on multi-agent llms and cgan, IEEE Transactions on Intelli- gent Transportation Systems (2025)

    2025

  24. [33]

    H. Fan, M. Li, J. Cui, Z. Zhang, W. Run, D. Liu, Spa- tiotemporal prediction of electric vehicle charging load based on large language models, in: 2026 IEEE PES In- ternational Meeting (PES IM), IEEE, 2026, pp. 1–5

    2026

  25. [34]

    S. Chen, X. Liang, Y . Liu, X. Li, X. Jin, Z. Du, Cus- tomized large-scale model for human-ai collaborative op- eration and maintenance management of building energy systems, Applied Energy 393 (2025) 126169

    2025

  26. [35]

    Mirshekali, M

    H. Mirshekali, M. R. Shadi, F. G. Ladani, H. R. Shaker, A review of large language models for energy systems: Ap- plications, challenges, and future prospects, IEEE Access (2025)

    2025

  27. [36]

    L. Shu, A. Yeganeh, D. Zhao, Large language models for building energy retrofit decision-making: Technical and sociotechnical evaluations, Buildings 15 (22) (2025) 4081

    2025

  28. [37]

    Sawada, M

    T. Sawada, M. Mizuno, T. Hasegawa, K. Yokoyama, M. Kono, Office-in-the-loop: an investigation into agen- tic ai for advanced building hvac control systems, Data- Centric Engineering 6 (2025) e31

    2025

  29. [38]

    Jiang, Z

    G. Jiang, Z. Ma, L. Zhang, J. Chen, EPlus-LLM: A large language model-based computing platform for automated building energy modeling, Applied Energy 367 (2024) 123431.doi:10.1016/j.apenergy.2024.123431

    2024

  30. [39]

    Ren, S.-M

    C. Ren, S.-M. Hong, R. Tang, Agenticontrol-mpc: Auto- mated model predictive control for buildings using large language model-based multi-agent system, Advanced En- gineering Informatics 76 (2026) 105062

    2026

  31. [40]

    Giudici, A

    M. Giudici, A. Sironi, I. Villa, S. Scherini, F. Garzotto, Generating homeassistant automations using an llm-based chatbot, arXiv preprint arXiv:2505.02802 (2025)

    2025 arXiv

  32. [41]

    Li, Y .-K

    T.-C. Li, Y .-K. Liu, Y .-C. Tsai, Ai-driven smart home en- ergy optimization: integrating ai agents with iot for adap- tive decision-making, in: International Conference on Ap- plied System Innovation (ICASI 2025), V ol. 2025, IET, 2025, pp. 225–230

    2025

  33. [42]

    N. V . Gkalinikis, C. Nalmpantis, D. Vrakas, S. Chatzi- georgiou, C. Athanasiadis, D. Doukas, Rhea: Residential home energy advisor, in: 2025 10th International Confer- ence on Smart and Sustainable Technologies (SpliTech), IEEE, 2025, pp. 1–6

    2025

  34. [43]

    V . S. Raghavan, A. Giridhar, Cost-optimal residential en- ergy scheduling via a zero-shot llm agent and model pre- dictive control, Journal of Building Performance Simula- tion (2025) 1–19

    2025

  35. [44]

    Accessed: 2026-07-05 (2026)

    Octopus Energy, Agile octopus and the octopus energy API, product AGILE-24-10-01; tariffE-1R-AGILE-24- 10-01-C; region C. Accessed: 2026-07-05 (2026). URLhttps://developer.octopus.energy/

    2026

  36. [45]

    Zippenfenig, Open-meteo

    P. Zippenfenig, Open-meteo. com weather api, Zenodo (2023). 15

    2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.