Article Contents
ARTICLE   Open Access     Cite

Localized LLM selection for agentic building energy modeling worker routing

More Information
  • Author Bio: Yanping Yuan; ypyuan@swjtu.edu.cn
  • Corresponding author: ypyuan@swjtu.edu.cn
  • DownLoad: Full size image
    1. 7 LLM backends benchmarked on bilingual agentic BEM; top regional models reach 98–99% Chinese success, near the 97% baseline.

      Accuracy alone is insufficient—cost, latency, and success-adjusted cost drive deployment selection.

      DeepSeek-V4-Pro offers the best cost–accuracy balance (~$119/M Chinese routing tasks).

      Success-adjusted cost spans 3.6×–21.2×; low per-call burden misleads when success is low.

      Low per-call inference burden can mislead when success rates are low; end-to-end validation needed.

  • Building Energy Modeling (BEM) is moving from expert-operated software toward Large Language Model (LLM)-based agentic systems that decompose user requests into specialized agents and executable tool calls. The LLM backend becomes part of the energy-modeling infrastructure, shaping routing reliability, latency, provider cost, and the inference burden incurred before a simulation even runs. Yet model choice is rarely quantified for localized agentic BEM, particularly when different language instructions must preserve canonical tool names, blackboard state, file paths, and engineering parameters. This study evaluates seven LLM backends, including a small local model and several regional models, using a bilingual worker-routing benchmark of 100 manually checked tasks per language covering prototype generation, quantitative modification, object insertion, simulation, output retrieval, comparative analysis, state-aware model selection, and missing-information detection. The strongest regional models achieved Chinese perfect-success rates of 98-99%, comparable to 97% for the baseline. Cost and latency analysis show that accuracy alone is insufficient for deployment selection. DeepSeek-V4-Pro provided the strongest cost–accuracy trade-off in this benchmark, requiring about US $119 per million Chinese worker-routing tasks and US $122 per million successful tasks. The corresponding success-adjusted cost was about 3.6× higher for the overseas baseline, 4.6× higher for MiniMax, 17.5× higher for Kimi, and 21.2× higher for GLM. A relative EnergyIndex further shows that low single-call inference burden can mislead when success rates are low. These results demonstrate that localized LLM selection is an operational design variable for reliable, cost-effective agentic BEM, and motivate end-to-end validation on full OpenStudio/EnergyPlus workflows.
  • 加载中
  • [1] González-Torres M., Pérez-Lombard L., Coronel JF., et al. (2022). A review on buildings energy information: Trends, end-uses, fuels and drivers. Energy Rep. 8:626−637. DOI:10.1016/j.egyr.2021.11.280

    View in Article CrossRef Google Scholar

    [2] Wan H. and Meyer R. (2025). A machine learning method of modern urban building energy modeling: a case study of Chicago. Sustain. Cities Soc. 134:106886. DOI:10.1016/j.scs.2025.106886

    View in Article CrossRef Google Scholar

    [3] Li X. and Wen J. (2014). Review of building energy modeling for control and operation. Renew. Sustain. Energy Rev. 37:517−537. DOI:10.1016/j.rser.2014.05.056

    View in Article CrossRef Google Scholar

    [4] Reinhart C.F. and Davila C.C. (2016). Urban building energy modeling: A review of a nascent field. Build. Environ. 97:196−202. DOI:10.1016/j.buildenv.2015.12.001

    View in Article CrossRef Google Scholar

    [5] Brackney L., Parker A., Macumber D., et al. (2018). Building energy modeling with OpenStudio: A practical guide for students and professionals. Cham Springer. DOI:10.1007/978-3-319-77809-9.

    View in Article Google Scholar

    [6] Ang Y.Q., Berzolla Z.M. and Reinhart C.F. (2020). From concept to application: A review of use cases in urban building energy modeling. Appl. Energy 279:115738. DOI:10.1016/j.apenergy.2020.115738

    View in Article CrossRef Google Scholar

    [7] Wan H., Zhang J., Chen Y., et al. (2025). Exploring Gen-AI applications in building research and industry: A review. Build. Simul. 18:1251−1273. DOI:10.1007/s12273-025-1279-x

    View in Article CrossRef Google Scholar

    [8] Jiang G., Ma Z., Zhang L., et al. (2024). EPlus-LLM: A large language model-based computing platform for automated building energy modeling. Appl. Energy 367:123431. DOI:10.1016/j.apenergy.2024.123431

    View in Article CrossRef Google Scholar

    [9] Lu J., Tian X., Zhang C., et al. Evaluation of large language models (LLMs) on the mastery of knowledge and skills in the heating, ventilation and air conditioning (HVAC) industry. Energy Built Environ. 6:875–892. DOI:10.1016/j.enbenv.2024.03.010.

    View in Article Google Scholar

    [10] Xu W., Wan H., Goel S., et al. (2025). Development of a dynamic multi-agent network for building energy modeling: A case study towards scalable and autonomous energy modeling. Energy Build. 372:116712. DOI:10.1016/j.enbuild.2025.116712

    View in Article CrossRef Google Scholar

    [11] Lyu J., Lai D., Zhang H., et al. (2026). LLM-powered agent for occupant-centric indoor environment control: Design, implementation, and evaluation. Energy Built Environ. DOI:10.1016/j.enbenv.2026.04.001.

    View in Article Google Scholar

    [12] Yao S., Zhao J., Yu D., et al. (2023). ReAct: Synergizing reasoning and acting in language models. Int. Conf. Learn. Rep. DOI:10.48550/arXiv.2210.03629.

    View in Article Google Scholar

    [13] Schick T., Dwivedi-Yu J., Dessì R., et al. (2023). Toolformer: Language models can teach themselves to use tools. Adv. Neural Inf. Proces. Syst. 37:68539−68551. DOI:10.5555/3666122.3669119

    View in Article CrossRef Google Scholar

    [14] Patil S.G., Zhang T., Wang X., et al. (2024). Gorilla: Large language model connected with massive APIs. Adv. Neural Inf. Proces. Syst 37:126544−12656. DOI:10.52202/079017-4020

    View in Article CrossRef Google Scholar

    [15] Qin Y., Liang S., Ye Y., et al. (2023). ToolLLM: Facilitating large language models to master 16000+ real-world APIs. arXiv preprint. DOI:10.48550/arXiv.2307.16789.

    View in Article Google Scholar

    [16] Li D., Liu M., Li J., et al. (2026). Exploring the applications of large language models in carbon management and assessment. Innov. Energy 3:100135. DOI:10.59717/j.xinn-energy.2026.100135

    View in Article CrossRef Google Scholar

    [17] Kulkarni M., Mazzia V., Gaspers J., et al. (2025). MASSIVE-Agents: A benchmark for multilingual function-calling in 52 languages. EMNLP 2025:20193−20215. DOI:10.18653/v1/2025.findings-emnlp.1099

    View in Article CrossRef Google Scholar

    [18] Luo Z., Kutralingam T.P., Okoani O.N., et al. (2026). Lost in Execution: On the multilingual robustness of tool calling in large language models. arXiv preprint. DOI:10.48550/arXiv.2601.05366.

    View in Article Google Scholar

    [19] Tang X., Peng H., Chen G., et al. (2026). Agent Systems with Harness Engineering. https://openreview.net/pdf?id=nM5tDHrQsx.

    View in Article Google Scholar

    [20] Niu C., Zhang W., Li J., et al. (2026). TokenPowerBench: Benchmarking the power consumption of LLM inference. AAAI 40:32582−32590. DOI:10.48550/arXiv.2512.03024

    View in Article CrossRef Google Scholar

    [21] Luccioni A.S., Jernite Y. and Strubell E. (2024). Power hungry processing: Watts driving the cost of AI deployment? ACM Fairn. Account. Trans. DOI:10.1145/3630106.3658542.

    View in Article Google Scholar

    [22] Patterson D., Gonzalez J., Le Q., et al. (2021). Carbon emissions and large neural network training. arXiv preprint. DOI:10.48550/arXiv.2104.10350.

    View in Article Google Scholar

    [23] Xia S., Yang Y. and Poon J.P.H. (2025). Ensuring a carbon-neutral future for artificial intelligence. Innov. Energy 2:100071. DOI:10.59717/j.xinn-energy.2024.100071

    View in Article CrossRef Google Scholar

  • Cite this article:

    Wan H. and Yuan Y. (2026). Localized LLM selection for agentic building energy modeling worker routing. The Innovation Energy 3:100194. https://doi.org/10.59717/j.xinn-energy.2026.100194
    Wan H. and Yuan Y. (2026). Localized LLM selection for agentic building energy modeling worker routing. The Innovation Energy 3:100194. https://doi.org/10.59717/j.xinn-energy.2026.100194

Welcome!

To request copyright permission to republish or share portions of our works, please visit Copyright Clearance Center's (CCC) Marketplace website at marketplace.copyright.com.

Figures(3)     Tables(3)

Supplementary Information

Share

  • Share the QR code with wechat scanning code to friends and circle of friends.

Article Metrics

Article views(366) PDF downloads(252)

Relative Articles

Cited by

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint