Pith. sign in

REVIEW 3 major objections 6 minor 13 references

Accurate passive forecasts of building temperature are not enough for HVAC control; the model must separately predict the causal effect of each control action.

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 · grok-4.5

2026-07-11 22:50 UTC pith:AYAMTDU5

load-bearing objection Clean demonstration that passive thermal skill can invert control effects, plus a simple frozen-TSFM + monotone forced operator that works well on EnergyPlus and BOPTEST. the 3 major comments →

arxiv 2607.03942 v1 pith:AYAMTDU5 submitted 2026-07-04 eess.SY cs.SY

ThermoForce: A Physics-Structured Interventional World Model for Building HVAC Control

Yifan Wang This is my paper
classification eess.SY cs.SY
keywords Building HVAC controlModel predictive controlTime-series foundation modelsPhysics-informed machine learningInterventional world modelCounterfactual predictionForced response
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.

Model predictive control for building heating and cooling needs a thermal model that can answer a causal question: what will indoor temperature, energy, and comfort be if a particular control action is applied. Time-series foundation models can forecast how a building would evolve passively with strong skill, yet the paper shows that this skill does not guarantee a correct response to interventions. An observational grey-box model that is best at passive forecasting can predict cooling with the wrong sign, and simply feeding control and weather as covariates into a foundation model does not fix the problem. ThermoForce keeps a foundation model frozen as a free-response prior and learns a small, physics-structured forced-response operator that is monotone in the control input by construction. The two pieces compose into a counterfactual world model that recovers intervention effects from one to three days of excitation. On paired EnergyPlus heating and cooling tests it achieves the lowest intervention-effect error and correct sign where the baselines fail; inside MPC on BOPTEST it cuts thermal discomfort by 33–84 percent while also cutting energy, with only 195 trainable parameters and CPU-only computation.

Core claim

Factual forecasting accuracy is not sufficient for HVAC control. A thermal model is control-ready only when it predicts the causal effect of control actions, which requires structural separation of passive free evolution from forced intervention response. ThermoForce implements that separation: a frozen time-series foundation model supplies the free response, a compact monotone physics operator supplies the forced response, and their sum answers counterfactual control queries with correct sign and low effect error where observational, covariate, and distillation models fail.

What carries the argument

Free/forced superposition: total predicted temperature equals the frozen foundation-model free response plus a forced-response operator whose state evolves with a stable pole, sign-definite gain, and bounded state modulation, guaranteeing that the marginal effect of control remains monotone by construction.

Load-bearing premise

Zone temperature can be split additively into a control-free free response and a forced response from a low-order monotone operator; if real multi-zone, nonlinear, or occupancy-coupled dynamics make free and forced effects non-additive, the composition is no longer control-valid.

What would settle it

On a building with multi-zone coupling or strong occupancy-driven loads, measure whether ThermoForce’s predicted intervention effect (and its sign) still matches the measured temperature difference under a true paired control probe; a systematic sign error or large effect RMSE would falsify the free/forced claim.

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

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. The paper argues that high factual forecasting accuracy is not sufficient for HVAC model predictive control, because a control-ready thermal model must answer counterfactual queries about the effect of planned actuation. It shows that an observational grey-box model with the best passive accuracy can invert the sign of cooling effects, and that feeding control/weather covariates to a time-series foundation model (TSFM) does not reliably fix intervention response. ThermoForce keeps a TSFM frozen as a passive free-response prior and learns a compact, physics-structured forced-response operator that is monotone in the control input by construction (Eqs. 1, 6–11 and the Proposition). The operator is identified from one to three days of excitation via a two-stage procedure that addresses feedback confounding. On paired EnergyPlus heating/cooling interventions ThermoForce reports the lowest intervention-effect RMSE and high sign accuracy where observational RC, covariate-TSFM, and distillation baselines fail (Table 4); on REFIT it preserves or improves few-shot factual skill (Tables 5–6); and in BOPTEST MPC it reduces thermal discomfort by 33–84% while reducing energy across three two-week windows (Table 7), with 195 trainable parameters and CPU-only computation.

Significance. If the free/forced separation is accepted as control-valid, the work is a clear and useful reframing for foundation models in building control: passive prediction and forced intervention response should be structurally separated rather than compressed into a single forecaster or distilled surrogate. Strengths include an explicit monotone-by-construction operator with a stated Proposition, a two-stage identification procedure aimed at thermostat confounding, complementary evaluation (REFIT factual, EnergyPlus paired interventions with cell-level Wilcoxon tests, BOPTEST closed-loop KPIs), and practical efficiency (frozen backbone, 195 trainable parameters, no GPU). The contrast with ThermoStill and the demonstration that best passive accuracy can still invert cooling sign are particularly valuable for the community. The result is of direct interest to building MPC and to the broader use of TSFMs as plant models.

major comments (3)
  1. §4.1–4.2, Eqs. (1), (4)–(7), and the Proposition: the load-bearing premise is additive free/forced superposition motivated by a one-step LTI approximation. The Proposition guarantees only that the forced channel has fixed sign and bounded gain; it does not guarantee that the additive composition matches multi-zone, nonlinear, or occupancy-coupled dynamics (e.g., HVAC-induced airflow changing effective resistances). All interventional evidence (EnergyPlus Table 4; BOPTEST Table 7) is generated inside simulators that approximately obey RC-like additivity and from which the operator is identified; REFIT supplies only factual checks. The central control-readiness claim therefore needs either (i) an independent stress test or bound on non-additivity outside this simulator class, or (ii) a clearly scoped limitation that absolute counterfactual trajectories (and thus MPC action ranking) may rem
  2. §5 Intervention suite and Table 4: interventional validity is aggregated over only six cells (2 climates × 3 windows), with Wilcoxon tests at n=6. That is a thin sample for the paper’s strongest claim (lowest effect RMSE and correct sign where baselines fail). The manuscript should either expand the intervention suite (more climates, building types, or multi-zone cases) or temper the generality of the ranking and report per-cell effect RMSE ranges for all methods, not only the distillation surrogate.
  3. §4.3 and the claim of identification from one to three days of control excitation: the two-stage estimator and paired-probe residual (Eq. 12) are well motivated, but the paper does not quantify how sensitive the recovered gain and subsequent MPC ranking are to the HVAC-off threshold u_off, the amount of open-loop excitation, or residual thermostat confounding when true paired probes are unavailable. A short sensitivity or failure-mode analysis on these identification choices would make the few-shot control-readiness claim more credible.
minor comments (6)
  1. Figure 1 is still a placeholder (“to be drawn”); the framework overview is central and should be completed before acceptance.
  2. Notation: the composite forecast is written both as Eq. (1) and Eq. (6); keep a single numbering and consistent use of T̂ free / F throughout.
  3. Several figure captions and the BOPTEST plot (Fig. 7) still use “ForceCast” while the paper title and text use ThermoForce; unify naming.
  4. Table 4 omits factual RMSE for the distillation surrogate; either report it or state explicitly that it is not applicable so the passive-vs-control comparison is complete.
  5. §8 / Table 7: state the MPC horizon, sampling time, and energy–discomfort weights so the closed-loop comparison is reproducible.
  6. Minor typos and style: “HV AC” spacing in the title block, “bitwise-identical” for simulations, and occasional missing spaces before citations.

Circularity Check

0 steps flagged

No significant circularity: free/forced split is an LTI-motivated ansatz, monotone authority is an architectural guarantee, and intervention claims are scored on held-out paired EnergyPlus and external BOPTEST KPIs.

full rationale

The paper’s load-bearing modeling step is the free/forced superposition (Eqs. 1, 4–7) motivated by a one-step LTI decomposition, plus a forced operator whose pole, sign-definite gain, and bounded modulation make control authority monotone by construction (Eqs. 8–11 and the Proposition). That Proposition is not an empirical “prediction” smuggled from a fit; it is an architectural guarantee stated as such. Operator parameters (and the two-stage / paired-difference identification) are fit from short excitation, but the central claims—lowest intervention-effect RMSE and correct sign (Table 4), few-shot factual skill (Tables 5–6), and closed-loop discomfort/energy gains (Table 7)—are evaluated on held-out EnergyPlus paired cells and the external BOPTEST benchmark under validation-safe selection, not on quantities defined by the fitted weights. There is no self-definitional loop equating a fitted parameter to the reported effect, no load-bearing self-citation or uniqueness theorem imported from the same author, and no renaming of a known result as a derivation. Weaknesses of the additive free/forced premise are modeling/correctness risks, not circularity. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The central claim rests on a standard LTI free/forced decomposition, domain assumptions about HVAC actuation sign and short-horizon stability, a frozen pretrained TSFM as free prior, and a small set of hand-fixed or fitted operator parameters (γ, pole bounds, gain, MLP). No new physical particles or forces are invented; the invented entity is the ThermoForce composite world-model architecture itself. Free parameters are few and mostly constrained by construction; the main scientific bet is that additive free/forced superposition plus monotone gain transfers from the LTI sketch to the simulators and, by implication, real buildings.

free parameters (6)
  • modulation coefficient γ = 0.75
    Fixed a priori at 0.75 to bound state-dependent gain modulation m(x)∈(1−γ,1+γ); sensitivity is shown but the default is a design choice the effect RMSE depends on.
  • forced-operator pole bounds (a_lo, a_hi) and unconstrained ˜a
    Stable pole a is parameterized via logistic map into a box; bounds and learned ˜a set free-response memory of the forced state.
  • sign-definite gain g via softplus(˜g) = learned per building/mode
    Magnitude of control authority is learned from HVAC-on residuals or paired differences; sign is fixed by heating/cooling mode.
  • MLP weights in m(x) (195 trainable parameters total for operator) = 195 params, Adam 300–400 epochs
    State-modulated gain is fit on target-building excitation; core few-shot capacity of the method.
  • HVAC-off threshold u_off and free RC coefficients (a,c,d,e) in stage-1 ID
    Two-stage identification uses idle samples to pin free dynamics before estimating g; threshold and free coeffs are data-dependent design choices under box constraints.
  • MPC energy–discomfort weights and comfort bounds
    Closed-loop KPIs depend on the economic objective used to rank trajectories; not fully enumerated as a fixed public config in the text.
axioms (6)
  • domain assumption Short-window zone dynamics admit an LTI free/forced additive decomposition T = T_free + F with F driven only by u (Eqs. 4–6).
    §4.1 motivates the architecture from a linear one-step model; superposition is assumed to remain useful for counterfactual MPC queries.
  • domain assumption Control authority is monotone and sign-definite in u for heating (g>0) or cooling (g<0), with bounded state modulation excluding u from m(x).
    Proposition and Eqs. 7–11 encode physical HVAC directionality by construction.
  • domain assumption A frozen univariate TSFM (Chronos-Bolt) applied to control-removed free series is a valid passive prior for free response.
    §4.1–4.2 never update θ; factual skill of the backbone is taken as given from pretraining and REFIT/EnergyPlus context.
  • domain assumption Paired probe episodes that share weather/occupancy but differ in control isolate causal control effects for identification (Eq. 12).
    §4.3 two-stage and paired-difference estimator; assumes confounders cancel in D = T^A − T^B.
  • standard math Standard forecast metrics, Wilcoxon cell-level tests, and BOPTEST KPIs are appropriate measures of control readiness.
    §5–8 evaluation protocol; ordinary statistical and benchmark practice.
  • ad hoc to paper One to three days of control excitation suffice to identify the forced operator for deployment.
    Stated throughout abstract and §4.3/§9; supported by data-efficiency plots but is a paper-specific operational claim.
invented entities (1)
  • ThermoForce free/forced interventional world model (frozen TSFM + monotone forced operator) no independent evidence
    purpose: Answer counterfactual HVAC control queries and serve as the plant model inside MPC without fine-tuning the foundation model.
    The composite architecture and monotone operator are the paper’s proposed object; independent evidence is the EnergyPlus effect metrics and BOPTEST closed-loop results, not an external prior measurement of the same entity.

pith-pipeline@v1.1.0-grok45 · 16616 in / 4254 out tokens · 36384 ms · 2026-07-11T22:50:22.314028+00:00 · methodology

0 comments
read the original abstract

Model predictive control (MPC) of building HVAC systems needs thermal models that answer a causal question: what indoor temperature, energy use, and comfort will result if a control action is applied? Time-series foundation models (TSFMs) can forecast passive building trajectories with strong zero-shot skill, but high factual accuracy does not imply valid response to control interventions. We show that an observational grey-box model with the best passive accuracy predicts cooling effects with the wrong sign, and that adding control and weather covariates to a TSFM does not fix intervention response. We introduce ThermoForce, a control-ready interventional thermal world model that keeps a TSFM frozen as a passive free-response prior and learns a compact, physics-structured forced-response operator for the causal effect of HVAC actuation. The operator is monotone in the control input by construction, is identified from one to three days of control excitation, and composes with the free response into a counterfactual-capable world model. Across paired EnergyPlus heating and cooling interventions, ThermoForce attains the lowest intervention-effect error and correct effect sign where covariate-TSFM, observational grey-box, and distillation baselines fail. Embedded in MPC on the BOPTEST benchmark, it reduces thermal discomfort by 33--84\% relative to the native controller across three two-week windows while simultaneously reducing energy, using a frozen backbone, 195 trainable parameters, and CPU-only computation. ThermoForce reframes foundation models for building control: passive prediction and forced intervention response must be structurally separated for a model to be control-ready.

Figures

Figures reproduced from arXiv: 2607.03942 by Yifan Wang.

Figure 1
Figure 1. Figure 1: Overview of the ThermoForce interventional world model. The frozen foundation [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Factual accuracy does not imply control validity. Each marker is a method [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Few-shot data efficiency on REFIT. ThermoForce is below the zero-shot TSFM [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Free/forced superposition on a REFIT window. The control-free component [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Counterfactual capability. ThermoForce produces distinct temperature trajec [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Closed-loop energy-comfort tradeoff on BOPTEST. Each line connects the na [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Closed-loop operation on the BOPTEST hydronic heat-pump peak window. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity to the modulation coefficient [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The free/forced gain persists across frozen backbone scales (tiny, mini, small). [PITH_FULL_IMAGE:figures/full_fig_p022_9.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

13 extracted references · 1 linked inside Pith

  1. [1]

    Afram, F

    A. Afram, F. Janabi-Sharifi, Theory and applications of hvac control systems: A review of model predictive control (mpc), Building and En- vironment 72 (2014) 343–355

    2014

  2. [2]

    Drgoňa, J

    J. Drgoňa, 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 Control 50 (2020) 190–232. 19 0.0 0.2 0.4 0.6 0.8 modulation coefficient γ 0.30 0.35 0.40 0.45 0.50held-out effect RMSE (K) Sensitivity to gai...

    2020

  3. [3]

    A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. Pineda Arango, S. Kapoor, J. Zschieg- ner, D. C. Maddix, H. Wang, M. W. Mahoney, K. Torkkola, A. Gor- don Wilson, M. Bohlke-Schneider, Y. Wang, Chronos: Learning the language of time series, Transactions on Machine Learning Research (2024)

    2024

  4. [4]

    A. Das, W. Kong, R. Sen, Y. Zhou, A decoder-only foundation model for time-series forecasting, Proceedings of the 41st International Conference on Machine Learning (ICML) (2024)

    2024

  5. [5]

    Y.-J. Park, F. Germain, J. Liu, Y. Wang, T. Koike-Akino, G. Wich- 20 tiny 9M mini 21M small 48M frozen backbone (parameter count) 0.0 0.1 0.2 0.3 0.4 0.5median 6 h RMSE (K) Free/forced gain persists across backbones TSFM raw ForceCast Figure 9: The free/forced gain persists across frozen backbone scales (tiny, mini, small). ern, N. Azizan, C. Laughman, A....

    Pith/arXiv arXiv 2025

  6. [6]

    O. B. Mulayim, P. Quan, L. Han, X. Ouyang, D. Hong, M. Bergés, M. Srivastava, Can time-series foundation models perform building en- ergy management tasks?, Data-Centric Engineering (2025)

    2025

  7. [7]

    Drgoňa, A

    J. Drgoňa, A. R. Tuor, V. Chandan, D. L. Vrabie, Physics-constrained deep learning of multi-zone building thermal dynamics, Energy and Buildings 243 (2021) 110992

    2021

  8. [8]

    T.Xiao, F.You, Buildingthermalmodelingandmodelpredictivecontrol with physically consistent deep learning for decarbonization and energy optimization, Applied Energy 342 (2023) 121165

    2023

  9. [9]

    Zheng, D

    W. Zheng, D. Wang, Z. Wang, Economic model predictive control for building hvac system: A comparative analysis of model-based and data- driven approaches using the boptest framework, Applied Energy 374 (2024) 124079

    2024

  10. [10]

    Drgoňa, A

    J. Drgoňa, A. Tuor, E. Skomski, S. Vasisht, D. Vrabie, Deep learning explicit differentiable predictive control laws for buildings, in: IFAC- 21 PapersOnLine (7th IFAC Conference on Nonlinear Model Predictive Control), Vol. 54, Elsevier, 2021, pp. 14–19

    2021

  11. [11]

    Liang, Y

    R. Liang, Y. Deng, Y. Liu, D. Zhao, O. B. Mulayim, I. Taniguchi, D. Wang, Thermostill: Distilling time series foundation model into ther- mal dynamics model for hvac model predictive control, in: Proceedings of the 17th ACM International Conference on Future and Sustainable Energy Systems (e-Energy), 2026, pp. 1–12

    2026

  12. [12]

    D. B. Crawley, L. K. Lawrie, F. C. Winkelmann, W. Buhl, Y. Huang, C. O. Pedersen, R. K. Strand, R. J. Liesen, D. E. Fisher, M. J. Witte, J. Glazer, Energyplus: creating a new-generation building energy simu- lation program, Energy and Buildings 33 (4) (2001) 319–331

    2001

  13. [13]

    D. Blum, J. Arroyo, S. Huang, J. Drgoňa, F. Jorissen, H. T. Walnum, Y. Chen, K. Benne, D. Vrabie, M. Wetter, L. Helsen, Building optimiza- tion testing framework (boptest) for simulation-based benchmarking of control strategies in buildings, Journal of Building Performance Simula- tion 14 (5) (2021) 586–610. 22

    2021