Pith. sign in

REVIEW 3 major objections 6 minor 23 references

LLM picks building sensors by physical topology, beats zero-shot baselines

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-08 08:42 UTC pith:SFA24XGG

load-bearing objection Training-free zero-shot building IoT forecasting using LLM agents over Brick knowledge graphs to select exogenous variables; competitive with supervised baselines on three real buildings. the 3 major comments →

arxiv 2607.06349 v1 pith:SFA24XGG submitted 2026-07-07 cs.AI

TopoBrick: Agentic Topology Sampling of Exogenous Variables for Zero-Shot Building IoT Forecasting

Xiachong Lin , Du Yin , Arian Prabowo , Hao Xue , Wen Hu , Imran Razzak , Matthew Amos , Sam Behrens
show 1 more author
Flora D. Salim
This is my paper
classification cs.AI
keywords zero-shot forecastingbuilding knowledge graphexogenous variable selectionagentic reasoningBrick Schemabuilding IoTtime series foundation modelstopology-aware forecasting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

TopoBrick uses an LLM-based agentic topology sampler over building knowledge graphs to select target-specific exogenous variables for zero-shot building IoT forecasting. The framework constructs a compact building skeleton from the raw knowledge graph, reasons over target-centric topology context to identify physically and operationally relevant exogenous variables, verifies selections against graph evidence, and organizes selected variables by deployment-time availability (past-known sensor states vs. future-known calendar, schedule, and meteorological variables). The central claim is that topology-aware exogenous-variable selection provides a stronger inductive bias than either channel-independent zero-shot forecasting or per-building supervised training, as evidenced by best-or-second-best normalized MAE and MSE across three real-world buildings on two continents. The paper demonstrates that a frozen time-series foundation model, when supplied with topology-routed exogenous variables organized by availability, can match or exceed fully trained building-specific models without any building-specific training. The agentic sampler outperforms random, ontology-only, and fixed-hop selection strategies, particularly for HVAC and weather-driven sensing variables. However, ablation results reveal that future-known meteorological forecasts and operational schedules drive most of the performance gains, while topology-selected past-known sensor observations provide minimal benefit and sometimes degrade performance.

Core claim

The paper's central discovery is that a building knowledge graph can serve as an effective routing layer for exogenous-variable selection in zero-shot forecasting, enabling a frozen foundation model to leverage building-specific physical and operational context without training. The agentic topology sampler—reasoning over equipment flow, spatial containment, and control-loop structure—selects more informative exogenous variables than graph-distance heuristics or ontology-class matching. The availability-aware formulation, which separates past-known sensor states (masked over the prediction horizon) from future-known calendar, schedule, and meteorological variables, proves critical: future-已知

What carries the argument

The central mechanism is the agentic topology sampler, which operates on a compact building skeleton distilled from the raw knowledge graph. Given a target Point node, the sampler receives a target-centric topology context comprising three components: the target anchor (equipment, room, zone, or system to which the target is attached), the local topology (upstream equipment, downstream served entities, sibling equipment, spatial containers), and global context (building-level drivers such as weather points and operational schedules). The agent reasons over this context following three principles—physical relevance (identifying plausible driver categories for the target's physical quantity),

Load-bearing premise

The paper assumes that LLM-based agentic reasoning over building topology context can reliably identify physically relevant exogenous variables that improve forecasts, but this assumption is validated only indirectly through downstream forecasting metrics. The ablation reveals that topology-selected past-known sensor observations provide minimal benefit or even degrade performance, while future-known weather forecasts drive most gains—raising the question of whether the agent

What would settle it

If a simpler pipeline using the same future-known variables (weather forecasts, operational schedules, calendar features) but with random or fixed-hop past-known variable selection achieves forecasting performance comparable to TopoBrick, then the agentic topology sampler is not the load-bearing mechanism.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Building knowledge graphs could serve as routing layers for other building analytics tasks beyond forecasting, including virtual sensing, anomaly detection, fault localization, and sensor outage recovery.
  • The availability-aware formulation (past-known vs. future-known) could be adopted by existing time-series foundation models to improve deployment-faithful forecasting in cyber-physical systems more broadly.
  • The finding that future-known weather and schedules drive most gains suggests that improving weather forecast quality and schedule integration may yield more returns than refining topology-based sensor selection.
  • The framework's training-free nature makes it deployable across heterogeneous building portfolios without per-building data collection or model tuning, potentially lowering the barrier to scalable building analytics.

Where Pith is reading between the lines

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

  • The ablation results raise the possibility that the agentic topology sampler—the paper's central contribution—may not be the primary driver of performance improvements. If future-known weather forecasts and operational schedules account for most gains, a simpler pipeline that injects these variables without topology-based selection might achieve comparable results, which would narrow the contribut
  • A direct test of this concern would compare TopoBrick against a baseline that uses the same future-known variables (weather, schedules, calendar) but selects past-known exogenous variables randomly or via fixed-hop sampling rather than agentic topology reasoning. If performance differences are negligible, the agentic sampler's value is called into question.
  • The ontology-level results suggest that the framework's effectiveness is bounded by whether the true physical drivers of a target sensor are represented in the knowledge graph. For variables dominated by unobserved occupancy, stochastic equipment usage, or closed-loop control policies, topology-aware selection may have limited leverage regardless of sampling quality.
  • The reliance on an LLM (gpt-oss-20b) for agentic reasoning introduces a dependency on model quality and prompt design. The paper does not report sensitivity to LLM choice or reasoning failures, leaving open whether the framework's robustness holds across different language model backbones.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. TopoBrick proposes a training-free framework for zero-shot building IoT forecasting that uses an LLM-based agentic topology sampler over building knowledge graphs (Brick schema) to select target-specific exogenous variables. The selected variables are organized by deployment-time availability into past-known (sensor states, masked over the prediction horizon) and future-known (calendar, operational schedules, meteorological forecasts) groups, then fed to a frozen time-series foundation model. The framework is evaluated on three real-world buildings (LBNL59, BTS-B, BTS-C) across four prediction horizons, against naive baselines, fully trained building-specific models, and zero-shot foundation models (Chronos-2, Moirai, TimesFM). The main results show TopoBrick achieving best or second-best performance in the majority of settings, with ablations decomposing contributions by covariate group and sampler strategy.

Significance. The paper addresses a genuine gap: existing zero-shot time-series foundation models lack mechanisms for selecting which building points serve as exogenous variables, and building-specific supervised models do not scale across heterogeneous portfolios. The idea of using building KG topology as a routing layer for exogenous variable selection is well-motivated. Strengths include reproducible code (GitHub link provided), evaluation on three real buildings across two continents, a deployment-faithful past/future-known variable split, and a falsifiable ablation design that honestly reports where the method helps and where it does not. The per-ontology analysis (Table 3) is a valuable contribution that connects performance to physical interpretability.

major comments (3)
  1. §4.4, Figure 3: The ablation reveals that +PastObs—the primary output of the agentic topology sampler—provides minimal or negative benefit (+4.7% nMAE on BTS-B at H=24, +10.0% at H=48), while gains come primarily from +WeatherFcst and +OpSched. Since weather and schedules are building-level global variables that do not require target-specific topology reasoning to select, this raises a question about whether the agentic topology sampler—the paper's central novel contribution—is actually responsible for the performance gains. The paper should either (a) run an ablation where weather and schedules are added without the agentic sampler (e.g., using all available past-known sensors selected by random or k-hop) to isolate the sampler's contribution, or (b) explicitly reframe the central claim to acknowledge that the availability-aware formulation and weather injection are the primary drivers,
  2. §4.3, Figure 2: The sampler comparison uses oracle weather for all methods but does not decompose whether the agentic sampler's advantage over random/ontology/k-hop comes from selecting better past-known sensors or better future-known variables. Given that §4.4 shows past-known sensors provide minimal benefit, the sampler comparison may conflate the sampler's variable-selection quality with the universal benefit of weather. A decomposition of the sampler comparison by covariate group (past-known vs. future-known) would clarify whether the agentic sampler's advantage is actually in topology-aware sensor selection or simply in consistently including weather/schedule variables that simpler alternatives might miss. This is load-bearing for the claim that topology-aware sampling is more reliable than simpler alternatives.
  3. Table 2: No error bars, confidence intervals, or statistical significance tests are reported for any result. Several key comparisons involve small margins (e.g., TopoBrick nMAE 0.319 vs. PatchTST 0.319 on BTS-C at H=24; TopoBrick nMAE 0.476 vs. PatchTST 0.473 on BTS-C at H=72). Without variance estimates, it is unclear whether these differences are meaningful. At minimum, bootstrap confidence intervals or paired tests across the per-sensor results should be reported for the key comparisons.
minor comments (6)
  1. §3.1.3: The LLM model choice (gpt-oss-20b) is mentioned only in §4 settings. A brief note in the methodology about model choice and sensitivity would help readers assess generalizability, or alternatively a statement that the approach is model-agnostic.
  2. Table 3: The per-ontology results are extensive but the color coding (green/red) is not visible in the text rendering. Consider adding sign indicators (+/−) or ensuring the color encoding is accessible in both print and digital formats.
  3. §4.4, Figure 3: The BTS-C ablation panels (f, i, l) appear to have fewer bars than other buildings, and the text notes BTS-C 'does not seem to respond well to our ablation modules.' This is mentioned briefly but deserves a clearer explanation or cross-reference to §5 discussion.
  4. The self-citation to Lin et al. 2024 (Bitsa, Ref [6]) is relevant but should be clearly distinguished from the current contribution to avoid confusion about novelty. A sentence clarifying what is new versus what was established in prior work would help.
  5. §3.1.1: The skeleton construction removes Point leaves from the traversal graph, but the notation V_s = {u ∈ V ∖ V_pt | τ(u) ∈ {Equipment, Location}} excludes other structural node types (e.g., System). The text mentions 'systems' as structural nodes but the formal definition does not include them. Clarify whether System nodes are included or mapped to Equipment/Location.
  6. The ACM template artifacts (conference acronym, copyright year 2018, Woodstock NY) suggest this is a draft submission. Ensure formatting is corrected for the final version.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a thorough and constructive report. The referee raises three major points: (1) whether the agentic topology sampler—the paper's central contribution—actually drives performance gains, given that the ablation shows past-known sensors (the sampler's primary output) provide minimal benefit while weather and schedules drive most gains; (2) whether the sampler comparison conflates variable-selection quality with the universal benefit of weather; and (3) the absence of error bars or significance tests. We address each below. We agree with all three points and will revise accordingly.

read point-by-point responses

  1. Referee: §4.4, Figure 3: The ablation reveals that +PastObs—the primary output of the agentic topology sampler—provides minimal or negative benefit (+4.7% nMAE on BTS-B at H=24, +10.0% at H=48), while gains come primarily from +WeatherFcst and +OpSched. Since weather and schedules are building-level global variables that do not require target-specific topology reasoning to select, this raises a question about whether the agentic topology sampler—the paper's central novel contribution—is actually responsible for the performance gains. The paper should either (a) run an ablation where weather and schedules are added without the agentic sampler (e.g., using all available past-known sensors selected by random or k-hop) to isolate the sampler's contribution, or (b) explicitly reframe the central claim to acknowledge that the availability-aware formulation and weather injection are the primary drivers.

    Authors: The referee is correct that the current ablation does not isolate the agentic sampler's contribution from the universal benefit of weather and schedules. This is a genuine gap in the experimental design, and we will address it by running the requested ablation: weather forecasts and operational schedules will be added on top of alternative past-known sensor selection strategies (random, k-hop, same-ontology) to test whether the agentic sampler's past-known sensor selections provide additional benefit beyond what weather and schedules alone deliver. We will also add a condition where weather and schedules are used with no past-known sensors at all, establishing a clean baseline for the availability-aware formulation independent of any sampler. If the results show that the sampler's past-known selections do not add significant value beyond weather and schedules, we will explicitly reframe the central claim. Specifically, we will reposition the paper's contribution as: (i) the availability-aware past/future-known formulation, which is the mechanism that enables weather and schedule injection in a deployment-faithful manner, and (ii) the agentic topology sampler as a principled variable-selection method that is especially valuable for physically coupled HVAC variables (as supported by the per-ontology analysis in Table 3), even if its aggregate-level contribution is smaller than that of future-known exogenous variables. We agree that the current framing overstates the sampler's role in aggregate performance and will revise the abstract, introduction, and conclusion to accurately reflect the relative contributions. We will not claim that the sampler is the primary driver of aggregate gains unless the new ablation supports that claim. revision: yes

  2. Referee: §4.3, Figure 2: The sampler comparison uses oracle weather for all methods but does not decompose whether the agentic sampler's advantage over random/ontology/k-hop comes from selecting better past-known sensors or better future-known variables. Given that §4.4 shows past-known sensors provide minimal benefit, the sampler comparison may conflate the sampler's variable-selection quality with the universal benefit of weather. A decomposition of the sampler comparison by covariate group (past-known vs. future-known) would clarify whether the agentic sampler's advantage is actually in topology-aware sensor selection or simply in consistently including weather/schedule variables that simpler alternatives might miss. This is load-bearing for the claim that topology-aware sampling is more reliable than simpler alternatives.

    Authors: The referee is right that the sampler comparison in Figure 2 does not decompose the advantage by covariate group, and this conflation undermines the claim that topology-aware sampling is more reliable than simpler alternatives. We will run the requested decomposition. Specifically, we will report sampler comparisons (agentic vs. random vs. same-ontology vs. k-hop) separately for: (a) past-known sensors only (no weather, no schedules), (b) future-known variables only (weather + schedules, no past-known sensors), and (c) the full combination. This will reveal whether the agentic sampler's advantage in Figure 2 comes from better past-known sensor selection or from consistently including weather and schedule variables that simpler alternatives might miss. We note that in the current experimental design, all samplers receive oracle weather, so the future-known component should be identical across samplers—meaning any difference in the current Figure 2 should already reflect past-known selection quality. However, the referee's concern is valid because the current presentation does not make this explicit, and we have not verified that all samplers include the same future-known variables. We will add the decomposition and, if it reveals that the sampler's advantage is primarily in consistently including weather/schedules rather than in topology-aware sensor selection, we will revise the claim accordingly. The claim that 'topology-aware sampling is more reliable than random, ontology-only, or fixed-hop selection' will be qualified to specify the covariate group where the advantage holds. revision: yes

  3. Referee: Table 2: No error bars, confidence intervals, or statistical significance tests are reported for any result. Several key comparisons involve small margins (e.g., TopoBrick nMAE 0.319 vs. PatchTST 0.319 on BTS-C at H=24; TopoBrick nMAE 0.476 vs. PatchTST 0.473 on BTS-C at H=72). Without variance estimates, it is unclear whether these differences are meaningful. At minimum, bootstrap confidence intervals or paired tests across the per-sensor results should be reported for the key comparisons.

    Authors: The referee is correct. We will add bootstrap confidence intervals and paired tests across per-sensor results for the key comparisons in Table 2. Specifically, we will report 95% bootstrap confidence intervals for nMAE and nMSE for TopoBrick and the strongest baselines at each building/horizon setting, and we will add paired comparisons (e.g., paired t-tests or Wilcoxon signed-rank tests across per-sensor errors) for the closest-margin comparisons the referee identifies, including TopoBrick vs. PatchTST on BTS-C at H=24 and H=72. Where differences are not statistically significant, we will state this explicitly and avoid claiming superiority. We will also add a note in the results discussion acknowledging which comparisons are within noise. This is a straightforward addition that does not change the experimental pipeline, only the reporting. revision: yes

Circularity Check

0 steps flagged

No circularity found — derivation chain is self-contained with no fitted-input-as-prediction or self-citation load-bearing steps

full rationale

The paper's derivation chain proceeds: (1) construct building skeleton from raw KG (§3.1.1), (2) LLM agent selects exogenous variables based on topology and metadata (§3.1.3), (3) KG-grounded verifier checks actions (§3.1.4), (4) deterministic materialization into time-series inputs (§3.1.5), (5) frozen forecaster produces predictions (§3.2.3). No step reduces to its inputs by construction. The sampler explicitly 'does not use future observations or forecasting errors during selection' (§3.1), so there is no fitted-input-called-prediction pattern. The forecaster F_θ is frozen and not updated on target data. Self-citations (Lin et al. 2024/BitSA [6], Prabowo et al. 2024/BTS dataset [15], Prabowo et al. 2025/Brick classification [14]) are for datasets and related-work context, not for load-bearing mathematical theorems or ansätze. The ablation in §4.4 honestly reports that +PastObs provides minimal benefit while +WeatherFcst drives most gains — this is a contribution-attribution concern (correctness risk), not circularity. The evaluation uses external baselines on held-out test periods with standard train/val/test splits (Table 1). No circularity detected.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical entities, particles, forces, or dimensions are introduced. The framework uses existing building KGs, existing foundation models, and standard LLM agents.

free parameters (6)
  • LLM model choice (gpt-oss-20b) = gpt-oss-20b
    The choice of LLM for agentic reasoning is a free parameter that affects which variables are selected. Not tuned on forecasting performance but determines sampler behavior.
  • Historical context length S = 96 (24h at 15min granularity)
    Set by the authors; standard for building time series but a modeling choice.
  • Prediction horizons H = {24, 48, 72, 96}
    Chosen evaluation horizons; standard but selected by authors.
  • Observation coverage threshold (L2) = obs_rate < 0.4
    Hand-set threshold for removing unreliable sensors from the exogenous variable pool.
  • Window validity threshold (L4) = 0.95
    Hand-set threshold requiring 95% valid target observations in each forecast window.
  • Near-flat signal threshold (L4) = std_train < 0.1
    Hand-set threshold for discarding low-variance target sensors.
axioms (4)
  • domain assumption Building knowledge graphs (Brick schema) accurately represent the physical and operational relationships between equipment, locations, and sensors.
    The entire framework depends on the KG being a faithful representation of building topology. Invoked throughout §3.1 where the skeleton is constructed from KG relations.
  • ad hoc to paper An LLM agent can reason over rendered topology context to select physically relevant exogenous variables better than random, ontology-only, or fixed-hop methods.
    This is the core assumption of the agentic sampler (§3.1.3). It is validated only indirectly through downstream forecasting metrics, not through expert evaluation of selected variables.
  • domain assumption A frozen time-series foundation model can effectively leverage exogenous variables without fine-tuning.
    The zero-shot forecasting formulation (§3.2.3) assumes the frozen model F_θ can use exogenous inputs meaningfully. The ablation in §4.4 partially challenges this for past-known variables.
  • domain assumption Meteorological forecasts of sufficient quality are available at deployment time.
    The future-known exogenous variable formulation (§3.2.2) assumes weather forecasts are obtainable. The sensitivity analysis in §4.4 tests this but the main results use a meteorological forecaster whose details are not specified.

pith-pipeline@v1.1.0-glm · 22484 in / 4612 out tokens · 289576 ms · 2026-07-08T08:42:14.693880+00:00 · methodology

0 comments
read the original abstract

Building sensors are embedded in physical topology, spatial hierarchy, and operational context, yet existing forecasters often treat them as isolated time series or rely on fixed covariate sets. We present TopoBrick, a training-free framework for zero-shot building IoT (Internet-of-Things) forecasting. TopoBrick uses building knowledge graphs to construct a compact structural skeleton and employs an agentic topology sampler to select target-specific exogenous variables. The selected variables are organized by deployment-time availability, separating past-known sensor states from future-known calendar, schedule, and meteorological exogenous variables. Across three real-world buildings, TopoBrick outperforms strong zero-shot foundation-model baselines and remains competitive with fully trained building-specific models. Ablations show that topology-aware sampling is more reliable than random, ontology-only, or fixed-hop selection, especially for physically coupled HVAC and weather-driven sensing variables.

Figures

Figures reproduced from arXiv: 2607.06349 by Arian Prabowo, Du Yin, Flora D. Salim, Hao Xue, Imran Razzak, Matthew Amos, Sam Behrens, Wen Hu, Xiachong Lin.

Figure 1
Figure 1. Figure 1: The visualization of the TopoBrick pipeline. (1) Building skeleton construction (2) Agentic reasoning over target-centric [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the sampler comparison. The x [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The average raw MAE of availability-aware exoge [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis of model performance to mete [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 4 internal anchors

  1. [1]

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, et al. 2025. Chronos-2: From Univariate to Universal Forecasting.arXiv preprint arXiv:2510.15821(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebas- tian Pineda Arango, Shubham Kapoor, et al. 2024. Chronos: Learning the language of time series.arXiv preprint arXiv:2403.07815(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2023. A decoder-only foundation model for time-series forecasting.arXiv preprint arXiv:2310.10688 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    Furong Jia, Kevin Wang, Yixiang Zheng, Defu Cao, and Yan Liu. 2024. Gpt4mts: Prompt-based large language model for multimodal time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 23343–23351

    work page 2024

  5. [5]

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. 2024. Time-LLM: Time series forecasting by reprogramming large language models. In International Conference on Learning Representations (ICLR)

    work page 2024

  6. [6]

    Xiachong Lin, Arian Prabowo, Imran Razzak, Hao Xue, Matthew Amos, Sam Behrens, and Flora D Salim. 2024. Bitsa: Leveraging time series foundation model for building energy analytics. In2024 IEEE International Conference on Data Mining Workshops (ICDMW). IEEE, 891–894

    work page 2024

  7. [7]

    Chenghao Liu, Taha Aksu, Juncheng Liu, Xu Liu, Hanshu Yan, Quang Pham, Silvio Savarese, Doyen Sahoo, Caiming Xiong, and Junnan Li. 2025. Moirai 2.0: When less is more for time series forecasting.arXiv preprint arXiv:2511.11698 (2025)

    work page arXiv 2025

  8. [8]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. [n. d.]. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    work page

  9. [9]

    Na Luo, Zhe Wang, David Blum, Christopher Weyandt, Norman Bourassa, Mary Ann Piette, and Tianzhen Hong. 2022. A three-year dataset supporting research on building energy management and occupancy analytics.Scientific data9, 1 (2022), 156

    work page 2022

  10. [10]

    Ozan Baris Mulayim, Avia Anwar, Umut Mete Saka, Lazlo Paul, Anand Krishnan Prakash, Gabe Fierro, Marco Pritoni, and Mario Bergés. 2025. BuildingQA: A Benchmark for Natural Language Question Answering over Building Knowledge Graphs. InProceedings of the 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation. 65–75

    work page 2025

  11. [11]

    Ozan Baris Mulayim, Gabe Fierro, Mario Bergés, and Marco Pritoni. 2025. To- wards Zero-shot Question Answering in CPS-IoT: Large Language Models and Knowledge Graphs. InProceedings of the 2nd International Workshop on Founda- tion Models for Cyber-Physical Systems & Internet of Things. 7–12

    work page 2025

  12. [12]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730(2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  13. [13]

    Arian Prabowo, Kaixuan Chen, Hao Xue, Subbu Sethuvenkatraman, and Flora D Salim. 2023. Navigating Out-of-Distribution Electricity Load Forecasting during COVID-19: Benchmarking energy load forecasting models without and with continual learning. InProceedings of the 10th ACM international conference on systems for energy-efficient buildings, cities, and tr...

    work page 2023

  14. [14]

    Arian Prabowo, Xiachong Lin, Imran Razzak, Hao Xue, Matthew Amos, Stephen D White, and Flora D Salim. 2025. Brick-by-Brick: Cyber-Physical Building Data Classification Challenge. InCompanion Proceedings of the ACM on Web Conference

    work page 2025

  15. [15]

    Arian Prabowo, Xiachong Lin, Imran Razzak, Hao Xue, Emily W Yap, Matthew Amos, and Flora D Salim. 2024. Building Timeseries Dataset: Empowering Large- Scale Building Analytics.Advances in Neural Information Processing Systems37 (2024), 133180–133206

    work page 2024

  16. [16]

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and JUN ZHOU. [n. d.]. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    work page

  17. [17]

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Guo Qin, Haoran Zhang, Yong Liu, Yunzhong Qiu, Jianmin Wang, and Mingsheng Long. 2024. Timexer: Empowering transformers for time series forecasting with exogenous variables.Advances in Neural Information Processing Systems37 (2024), 469–498

    work page 2024

  18. [18]

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. 2024. Unified training of universal time series forecasting transformers. InForty-first International Conference on Machine Learning

    work page 2024

  19. [19]

    Zhijian Xu, Ailing Zeng, and Qiang Xu. 2024. FITS: Modeling time series with 10𝑘 parameters. InInternational Conference on Learning Representations, Vol. 2024. 26295–26318

    work page 2024

  20. [20]

    Zakia Zaman, Praveen Gauravaram, Sanjay Jha, and Wen Hu. 2025. Bloom-LLM: Privacy-Preserving Large Language Model for Load Forecasting. InProceedings of the 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation. 128–138

    work page 2025

  21. [21]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting?. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 11121–11128

    work page 2023

  22. [22]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond efficient transformer for long se- quence time-series forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115

    work page 2021

  23. [23]

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al . 2023. One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems36 (2023), 43322–43355. TopoBrick: Agentic Topology Sampling of Exogenous Variables for Zero-Shot Building IoT Forecasting Conference acronym ’XX, June 03–05, 2018, Woodstock, NY A Pr...

    work page 2023