Pith. sign in

REVIEW 4 major objections 7 minor 28 references

This paper claims that an HVAC controller trained in a digital twin and fine-tuned on a few days of biased sensor data can match or beat PID, OSLN, and DDQN on comfort and setpoint stability, and can transfer between similar real offices in

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 · deepseek-v4-flash

2026-08-01 22:38 UTC pith:USBYOZKA

load-bearing objection Worth peer review on the strength of its real-building transfer results, but the marquee control margins over OSLN/PID/DDQN are measured in a one-step-validated surrogate, so they should not be taken at face value. the 4 major comments →

arxiv 2607.15663 v1 pith:USBYOZKA submitted 2026-07-17 eess.SY cs.SY

Adaptive Model-Based Transfer Learning for Dynamic HVAC Control

Quang-Thang Le , Kevin Wijaya , Hsin-Yi Lai , Che-Kai Liu , Ching-Chun Huang This is my paper
classification eess.SY cs.SY
keywords HVAC controltransfer learningdigital twinsetpoint controlphysics-informed losstemporal difference learningonline adaptationbuilding energy management
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.

The paper claims that an HVAC controller can be trained almost entirely in a simulated twin of a building, then transferred to real offices with only days of biased sensor data, and still hold room temperature within about 0.05-0.2°C of target while producing smoother setpoint changes than PID, OSLN, and DDQN. The core idea is dynamic pretraining: the agent actively controls the virtual building, so it encounters diverse and rare conditions that static datasets miss, then fine-tunes online in the target. Two additions carry the performance: physics-rule losses keep predictions physically consistent, and a long-term-aware setpoint selector uses the predicted long-term stable temperature rather than only the short-term forecast to avoid oscillation. If right, the practical payoff is that new buildings can be brought under stable autonomous control in about a week rather than after months of data collection, and similar buildings can share controllers directly.

Core claim

The paper proposes an adaptive model-based transfer learning framework for dynamic HVAC control. The agent is dynamically pretrained in an EnergyPlus digital twin of the target building — actively deciding setpoints and observing thermal responses rather than learning from fixed schedules — then transferred to the real building, fine-tuned on a small amount of biased sensor data, and updated online during operation. Two mechanisms are claimed to be essential: embedding physics rules as constraint losses (gradient constraint and residual constraint) keeps predictions physically consistent, and a long-term-aware setpoint selection strategy scores candidate setpoints by the predicted long-term

What carries the argument

The dynamic model is an encoder-decoder network with two prediction branches: the short-term regression branch (STRB) predicts next-step temperature S'_{t+Δ}, and the temperature difference branch (TDB) predicts the gap D'_{t+Δ} between the long-term stable temperature and the short-term prediction, trained against a temporal-difference-style bootstrap target Y = γ( F_TDB(·) + F_STRB(·) − S_{t+Δ} ) computed with a frozen target network. The control loop scores each candidate setpoint via Score = |S_t − T_exp| − α|S'_long − T_exp| − β|S'_long − S_t|, then fuses the best-scoring setpoint with an expert-defined schedule using a time-varying confidence driven by the model's prediction error. Phy

Load-bearing premise

The headline control comparisons against PID, OSLN, and DDQN are evaluated inside a learned neural-network surrogate of Room 301, not in real rooms — if that surrogate does not reproduce true closed-loop HVAC dynamics, especially oscillation and long-horizon behavior, the reported DAE/DARV margins and the tuned hyperparameters (α=1.0, β=0.9, γ=0.5) may not transfer to real buildings.

What would settle it

Run the identical agent and the three baselines for two weeks in the real Room 701 or 702, logging actual VAV temperatures and setpoint changes every 18 minutes, and compare daily average error (DAE) and daily average regret (DARV). If the proposed method does not stay below OSLN's real-room DAE (around 0.17°C) while keeping comparable or lower DARV, the central claim collapses. A quicker check: verify on real sensor data that the sign of the gradient of predicted temperature with respect to the AHU setpoint matches the +1 label assumed by the gradient constraint loss.

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

If this is right

  • If the reported margins hold, similar buildings no longer need their own digital twins: a stable controller from one room can be fine-tuned on about five days of biased data and reach below 0.15°C error within a week.
  • The long-term-aware setpoint selection (using predicted stable temperature rather than next-step error) is claimed to cut setpoint regret from 0.498 (OSLN) to 0.373 and reduce energy waste from oscillation without an explicit energy term in the objective.
  • Online adaptation lets the model follow seasonal and occupancy drift, so the controller does not need manual retuning after transfer.
  • The 18-minute control cycle is presented as the best balance: 6- and 12-minute cycles oscillate, and the 24-minute cycle loses accuracy, with DAE rising from 0.07 to 0.10.
  • Ablations show that the physics constraint losses reduce prediction error when training data is biased, and that removing TDB, online learning, or confidence fusion degrades both accuracy and smoothness.

Where Pith is reading between the lines

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

  • The decisive experiment is still missing: a head-to-head real-room comparison of the same policies rather than a surrogate. Until that runs, the 0.032°C vs 0.172°C margin over OSLN should be read as a simulated result, not a field result.
  • The TD-style bootstrap for long-term temperature (TDB with a target network) is a transferable idea: any slow thermal process with sparse actuation could use the same trick to learn a 'final stable value' head without waiting for physical convergence.
  • The confidence-based fusion with the expert schedule acts as a safety net; a natural extension is to replace the fixed expert schedule with a learnable or constraint-based limit on setpoint rate, which could remove the need for a domain expert.
  • The transfer claim is explicitly for similar buildings (same layout and sensor set); testing on more diverse geometries and climates would map the boundary of the 'one-week deployment' promise.

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

4 major / 7 minor

Summary. The paper proposes an adaptive model-based transfer learning framework for AHU setpoint control in HVAC systems. A neural dynamic model with a short-term regression branch and a long-term temperature-difference branch is pretrained by interacting with an EnergyPlus digital twin, transferred to target buildings, fine-tuned with limited biased real data, and updated online. The controller selects setpoints by scoring candidate actions using a predicted long-term stable temperature and a smoothness penalty, then fuses the winner with an expert schedule through a confidence mechanism. Physics-based gradient and residual constraints are added to the training loss. Experiments compare the method against OSLN, PID, and DDQN in a learned surrogate environment built from Room 301 data, provide ablation studies, and report transfer results between real offices (Room 301 as source; Rooms 401/402 as targets) and from virtual to real rooms (701/702). The paper claims superior occupant comfort, lower setpoint regret, smoother transitions, and deployment within one week.

Significance. If the main claims hold, the framework is practically valuable: it addresses real-data scarcity, avoids unsafe exploration, and demonstrates week-scale deployment between similar buildings. The real-building transfer evidence (Table III, Fig. 13) and the comparison against transfer-learning baselines FDTDM and DAF are genuine strengths, as are the detailed architectural and hyperparameter disclosures in Appendix A. The long-term-aware setpoint selection idea is interesting and the ablation studies are informative. However, the headline superiority over OSLN, PID, and DDQN rests on a learned surrogate that is validated only for one-step-ahead prediction, while hyperparameters are tuned on that same surrogate. The real transfer experiments do not include these baselines. Thus the practical significance of the central control claim is not yet established, although the direction is promising.

major comments (4)
  1. [Sec. V.A, Fig. 7, Tables IV–V] The central comparison against OSLN, PID, and DDQN is evaluated entirely inside a learned 'environmental simulation model' built from Room 301 data. The only validation reported is one-step-ahead test error (0.0347°C); no multi-step or closed-loop validation is given. Since DARV and setpoint oscillation are inherently multi-step dynamic phenomena, the surrogate may not reproduce them. The data were collected under OSLN control, so the surrogate's action distribution is centered on the very baseline it is used to outperform. To support the headline claim, either validate the surrogate's closed-loop rollouts against real Room 301 trajectories (e.g., compare predicted vs actual temperature responses under identical setpoint series) or report real-world comparisons against at least OSLN and PID.
  2. [Sec. V.E.1, Table IV] The hyperparameters α, β, and γ are selected by ablating on the same surrogate that produces the headline results. The final choice (α=1.0, β=0.9, γ=0.5) may therefore be overfit to surrogate artifacts. No error bars, repeated trials, or statistical tests are reported for the DAE/DARV values in Fig. 7, Tables IV, or Table V. Because the surrogate is not validated in closed loop, the reported margins could change materially in real buildings. Please report sensitivity of the real transfer results (Table III) to these hyperparameters, or at least explicitly acknowledge this limitation and avoid presenting the tuned values as globally optimal.
  3. [Eqs. (14), (23)] The long-term stability target Y is computed from the network's own F_TDB and F_STRB predictions, and S'_long = S'_{t+Δ} + D'_{t+Δ} is never compared with measured long-term temperatures. This is legitimate TD-style bootstrapping, but it means the 'long-term stability' claim is internally consistent rather than externally validated. The authors should evaluate D' against actual future temperatures (e.g., using recorded room temperatures following held setpoints), or temper the claims about 'long-term stability' and 'final stable temperature' accordingly.
  4. [Sec. V.D.2, Table III] The real-building transfer experiments compare only against FDTDM and DAF; OSLN, PID, and DDQN are not run in real rooms. Consequently, the conclusion 'Compared with OSLN, PID, and DDQN, our method consistently delivers superior occupant comfort, stable control, and smooth setpoint transitions' is not directly supported by real data. The authors should either deploy the key baselines in the real transfer setting or restrict the conclusion to comparisons inside the surrogate and against the transfer-learning baselines.
minor comments (7)
  1. [Eq. (7)] The first displayed expression for S_long appears malformed/duplicated; please fix the typesetting. Also, the accompanying text says a larger γ emphasizes temperatures closer to the present, but a larger γ actually assigns more weight to more distant future temperatures; the wording is misleading.
  2. [Algorithm 1] The equation references in Algorithm 1 are off: Loss_GC is Eq. (18), Loss_RC is Eq. (19), and Loss_recons is Eq. (21), not (15)/(17)/(18).
  3. [Table III and Sec. V.D.2] The baseline name is written 'FDTDM' in the text and 'FATDM' in Table III. Please harmonize the notation.
  4. [Table I] The relation label for 'VAV on/off' is +1, which is counterintuitive if VAV on corresponds to active cooling (which should reduce room temperature). Please clarify the encoding or verify the sign, since the gradient constraint loss enforces these labels as physical rules.
  5. [Eq. (24)] The first term |S_t − T_exp| is constant across all candidate setpoints a and therefore does not affect the arg max in Eq. (25). If this is intentional, please say so; otherwise it may be a leftover from the OSLN formulation.
  6. [Eq. (26)] The confidence fusion uses parameters β and σ, while Eq. (24) also uses α and β with different meanings. Consider renaming the fusion parameters to avoid ambiguity.
  7. [Sec. V.C and Fig. 11] Energy use is quantified via CDH/HDH areas and setpoint-curve areas rather than measured energy consumption, even though power sensors are listed in Table II. Please label these as proxies or indirect indicators, not measured energy savings.

Circularity Check

0 steps flagged

No significant circularity: the long-term TDB target is explicit temporal-difference bootstrapping, and the central control and transfer claims are evaluated on an independently fitted surrogate and on real-building deployments.

full rationale

I walked the derivation chain and found no step where a claimed prediction or first-principles result reduces by construction to its own input. The most potentially self-referential point is the long-term branch (Sec. IV-A, Eqs. 7-15): the target Y for D'_t+Δ is computed from the model's own F_TDB and F_STRB predictions after replacing the unavailable future temperature S_t+2Δ with the short-term model output. However, the paper explicitly states this is temporal-difference bootstrapping ('This formulation is inspired by temporal difference learning, where predictions are updated through bootstrapping using future estimated states'), and the short-term branch F_STRB is independently supervised against measured S_t+Δ via Loss_STRB in Eq. 6. Thus the TDB loss enforces a Bellman-style consistency condition anchored to real sensor data, not a fit of a parameter to the same quantity it later predicts as novel. The pre-trained/transferred model is also evaluated on real buildings (Table III, Fig. 12-13), which provides external support for the transfer-learning claim. The control-comparison headline margins (Fig. 7, Tables IV-V) are obtained inside a learned environmental simulation model built from Room 301 data (Sec. V-A), which is a validation-risk concern—especially because that surrogate was trained on OSLN-controlled data—but it is not a circularity: the proposed controller and baselines are all evaluated in the same fitted environment, and winning in that environment is an empirical outcome, not a definitional consequence. The only self-citation is to the authors' prior OSLN work [6], used as a baseline and as the source of the confidence-fusion heuristic (Eq. 26). This is a design component, not a load-bearing proof or uniqueness theorem, and the central transfer-learning and physics-embedding contributions do not rest on it. I therefore find no circular step and assign a score of 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The framework rests on two fitted environments (EnergyPlus digital twin and Room-301 neural surrogate), three control hyperparameters tuned by ablation on the surrogate, and a self-referential long-term target. No new physical entities are introduced; the TDB and long-term temperature are internal model constructs.

free parameters (6)
  • alpha (long-term target weight in scoring function) = 1.0
    Tuned by ablation on the surrogate environment (Table IV); balances target alignment with long-term stability in Eq. (24).
  • beta (change penalty in scoring function) = 0.9
    Tuned by ablation on the surrogate environment (Table IV); controls how strongly S'_long differs from S_t is penalized.
  • gamma (long-term decay) = 0.5
    Tuned by ablation on the surrogate environment (Table IV); defines the recursive long-term target horizon in Eq. (7) and Eq. (14).
  • beta and sigma in confidence fusion (Eq. 26) = not reported
    Tunable parameters controlling the exponential confidence weight and temporal smoothing; values are not reported.
  • control cycle Delta = 18 minutes
    Selected from a 6/12/18/24-minute comparison (Fig. 10) on the surrogate environment; not independently justified from real-building response times.
  • environmental simulation model weights = not reported (test error ≈0.0347°C)
    The learned neural surrogate of Room 301 used as the evaluation environment for all headline control comparisons; its weights are a fitted model whose closed-loop validity is assumed.
axioms (5)
  • domain assumption The Room-301 neural model is a valid closed-loop surrogate for real HVAC dynamics
    All control-method comparisons and energy analyses (Sec. V.A-C) are run inside this learned environment; only open-loop test error is reported, not closed-loop fidelity.
  • domain assumption The EnergyPlus digital twin with 0.4-1.6°C calibration error is a useful source domain
    Dynamic pretraining occurs in the digital twin (Sec. IV.C.1, Fig. 5); calibration errors are larger in perimeter zones, and transfer quality depends on this assumption.
  • ad hoc to paper The relation labels in Table I encode correct physical signs
    The gradient constraint loss (Eq. 18) enforces these signs, but the labels are asserted without validation; some signs, such as VAV on/off = +1 and power supplied = +1, are physically questionable.
  • ad hoc to paper The exponentially weighted S_long is the correct long-term stability objective
    Eq. (7) defines long-term temperature as an exponentially decaying average with no explicit derivation from occupant comfort or energy physics.
  • standard math Temporal-difference bootstrapping with a target network converges to a stable F_TDB
    Eq. (13)-(14) use the model's own predictions to form the target Y; this is standard TD practice, but no convergence proof is given for this specific architecture and nonstationary setting.

pith-pipeline@v1.3.0-alltime-deepseek · 23777 in / 11711 out tokens · 110844 ms · 2026-08-01T22:38:28.666802+00:00 · methodology

0 comments
read the original abstract

In this paper, we aim to automate the adjustment of air handling unit (AHU) setpoints within heating, ventilation, and air conditioning (HVAC) systems to maintain indoor temperatures at user-specified levels. A key challenge lies in obtaining sufficient high-quality sensor data from real buildings. To address this, we explore transfer learning and leverage simulation software to generate training data. We propose an adaptive model-based transfer learning approach for dynamic HVAC control, where the agent directly controls the source domain under conditions identical to the target domain. This eliminates the need for extensive target-specific knowledge to define data generation schedules and reduces the risk of collecting irrelevant samples, while also providing greater flexibility during learning. At the control level, we enhance performance through physics rule embedding, which ensures physical consistency, and long-term-aware setpoint selection strategy, which mitigates abrupt setpoint changes. Finally, to accelerate and stabilize deployment in new buildings, we enable knowledge transfer directly between similar real-world buildings, reducing the need to construct virtual source domains repeatedly.

Figures

Figures reproduced from arXiv: 2607.15663 by Che-Kai Liu, Ching-Chun Huang, Hsin-Yi Lai, Kevin Wijaya, Quang-Thang Le.

Figure 1
Figure 1. Figure 1: AHU temperature setting schedule designed by experts. Other than [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The HVAC system schematic. remain fixed and cannot adapt online or dynamically update control strategies on its own, which limits their applicability in dynamic environments. Overall, while existing model-based transfer learning meth￾ods improve prediction or control in specific buildings, they mostly rely on static simulation datasets with fixed setpoint schedules, which may not capture extreme conditions… view at source ↗
Figure 3
Figure 3. Figure 3: The architecture of the dynamic model. the short-term temperature St+∆. This measure is designed to gauge setpoint stability, offering an approximate estimation of subsequent indoor temperature changes if the setpoint At remains unchanged at time t + ∆. In essence, a smaller value of D′ t+∆ associated with a given setpoint indicates greater stability in the environmental conditions. Before entering the dyn… view at source ↗
Figure 4
Figure 4. Figure 4: Zone regionalization of the target building. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: The whole picture of pretrain scenario. our control agent. While information related to the target building remains relatively stable, weather plays a critical role in predicting indoor conditions. We incorporate local historical climate data from Taichung City, Taiwan, obtained from the Central Weather Administration. Together with geographical factors such as latitude, longitude, and sea level, we separa… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of our method and other control methods. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of four HVAC control methods during working hours. The red shaded area represents Cooling Degree Hours (CDH), and the blue shaded [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The results of controlling the virtual environment of Room 701. (From left to right, top to bottom, these are the first, second, third, sixth, seventh, and [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of temperature control performance for 6, 12, 18, and 24-minute control cycles. The 18-minute cycle achieves the best balance between [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of transferred models and models trained from scratch. [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The transfer between real-world offices. Room 301 serves as the [PITH_FULL_IMAGE:figures/full_fig_p016_13.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

28 extracted references · 2 canonical work pages

  1. [1]

    On performance and tuning of pid controllers in hvac systems,

    G. Geng and G. Geary, “On performance and tuning of pid controllers in hvac systems,” inProceedings of IEEE International Conference on Control and Applications, 1993, pp. 819–824 vol.2

    1993

  2. [2]

    Deep reinforcement learning for building HV AC control,

    T. Wei, Y . Wang, and Q. Zhu, “Deep reinforcement learning for building HV AC control,” inProceedings of the 54th Annual Design Automation Conference, DAC 2017, Austin, TX, USA, June 18-22, 2017. ACM, 2017, pp. 22:1–22:6. [Online]. Available: https://doi.org/10.1145/3061639.3062224

    arXiv 2017

  3. [3]

    Deep reinforcement learning for smart home energy management,

    L. Yu, W. Xie, D. Xie, Y . Zou, D. Zhang, Z. Sun, L. Zhang, Y . Zhang, and T. Jiang, “Deep reinforcement learning for smart home energy management,”IEEE Internet Things J., vol. 7, no. 4, pp. 2751–2762,

  4. [4]

    An hybrid model-free reinforcement learning approach for hvac control,

    F. M. Solinas, A. Bellagarda, E. Macii, E. Patti, and L. Bottaccioli, “An hybrid model-free reinforcement learning approach for hvac control,” in2021 IEEE International Conference on Environment and Electrical Engineering and 2021 IEEE Industrial and Commercial Power Systems Europe (EEEIC / I&CPS Europe), 2021, pp. 1–6

    2021

  5. [5]

    Hvac room temperature prediction control based on neural network model,

    S. Li, S. Ren, and X. Wang, “Hvac room temperature prediction control based on neural network model,” in2013 Fifth International Conference on Measuring Technology and Mechatronics Automation, 2013, pp. 606– 609

    2013

  6. [6]

    Online self-learning for smart HV AC control,

    T. Chao, M. Nguyen, C. Huang, C. Liang, and C. Chung, “Online self-learning for smart HV AC control,” in2019 IEEE International Conference on Systems, Man and Cybernetics, SMC 2019, Bari, Italy, October 6-9, 2019. IEEE, 2019, pp. 4324–4330. [Online]. Available: https://doi.org/10.1109/SMC.2019.8914027

    arXiv 2019

  7. [7]

    Effectiveness of neural networks and transfer learning for indoor air-temperature forecasting,

    A. Bellagarda, S. Cesari, A. Aliberti, F. Ugliotti, L. Bottaccioli, E. Macii, and E. Patti, “Effectiveness of neural networks and transfer learning for indoor air-temperature forecasting,”Automation in Construction, vol. 140, p. 104314, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S092658052200187X

    2022

  8. [8]

    Cosmos world foundation model platform for physical AI,

    N. Agarwal, A. Ali, M. Bala, Y . Balaji, E. Barker, T. Cai, P. Chattopadhyay, Y . Chen, Y . Cui, Y . Ding, D. Dworakowski, J. Fan, M. Fenzi, F. Ferroni, S. Fidler, D. Fox, S. Ge, Y . Ge, J. Gu, S. Gururani, E. He, J. Huang, J. S. Huffman, P. Jannaty, J. Jin, S. W. Kim, G. Kl ´ar, G. Lam, S. Lan, L. Leal-Taix ´e, A. Li, Z. Li, C. Lin, T. Lin, H. Ling, M. L...

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.03575 2025

  9. [9]

    Cosmos-drive-dreams: Scalable synthetic driving data generation with world foundation models,

    X. Ren, Y . Lu, T. Cao, R. Gao, S. Huang, A. Sabour, T. Shen, T. Pfaff, J. Z. Wu, R. Chen, S. W. Kim, J. Gao, L. Leal-Taix´e, M. Chen, S. Fidler, and H. Ling, “Cosmos-drive-dreams: Scalable synthetic driving data generation with world foundation models,”CoRR, vol. abs/2506.09042,

    Pith/arXiv arXiv

  10. [10]

    Cosmos-transfer1: Conditional world generation with adaptive multimodal control,

    H. A. Alhaija, J. M. ´Alvarez, M. Bala, T. Cai, T. Cao, L. Cha, J. Chen, M. Chen, F. Ferroni, S. Fidler, D. Fox, Y . Ge, J. Gu, A. Hassani, M. Isaev, P. Jannaty, S. Lan, T. Lasser, H. Ling, M. Liu, X. Liu, Y . Lu, A. Luo, Q. Ma, H. Mao, F. Ramos, X. Ren, T. Shen, X. Sun, S. Tang, T. Wang, J. Z. Wu, J. Xu, S. Xu, K. Xie, Y . Ye, X. Yang, X. Zeng, and Y . Z...

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.14492 2025

  11. [11]

    Model predictive control for HV AC systems - A review,

    R. Kwadzogah, M. Zhou, and S. Li, “Model predictive control for HV AC systems - A review,” in2013 IEEE International Conference on Automation Science and Engineering, CASE 2013, Madison, WI, USA, August 17-20, 2013. IEEE, 2013, pp. 442–447. [Online]. Available: https://doi.org/10.1109/CoASE.2013.6654072

    arXiv 2013

  12. [12]

    Model predictive control for the operation of building cooling systems,

    Y . Ma, F. Borrelli, B. Hencey, B. Coffey, S. C. Bengea, and P. Haves, “Model predictive control for the operation of building cooling systems,” IEEE Trans. Control. Syst. Technol., vol. 20, no. 3, pp. 796–803, 2012. [Online]. Available: https://doi.org/10.1109/TCST.2011.2124461 [13]Model-Based Hierarchical Optimal Control Design for HVAC Systems, ser. Dy...

    arXiv 2012

  13. [14]

    Modeling and optimization of complex building energy systems with deep neural networks,

    Y . Chen, Y . Shi, and B. Zhang, “Modeling and optimization of complex building energy systems with deep neural networks,” in51st Asilomar Conference on Signals, Systems, and Computers, ACSSC 2017, Pacific Grove, CA, USA, October 29 - November 1, 2017, M. B. Matthews, Ed. IEEE, 2017, pp. 1368–1373. [Online]. Available: https://doi.org/10.1109/ACSSC.2017.8335578

    arXiv 2017

  14. [15]

    Towards optimal control of air handling units using deep reinforcement learning and recurrent neural network,

    Z. Zou, X. Yu, and S. Ergan, “Towards optimal control of air handling units using deep reinforcement learning and recurrent neural network,”Building and Environment, vol. 168, p. 106535, 2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0360132319307474

    2020

  15. [16]

    Forecasting customers’ re- sponse to incentives during peak periods: A transfer learning approach,

    L. Cai, H. Wen, J. Gu, J. Ma, and Z. Jin, “Forecasting customers’ re- sponse to incentives during peak periods: A transfer learning approach,” International Transactions on Electrical Energy Systems, vol. 30, no. 7, p. e12251, 2020, e12251 ITEES-19-1069.R1. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/2050-7038.12251

    arXiv 2020

  16. [17]

    Study on data transfer in meteorological forecast of small and medium-sized cities and its application in zhaoqing city,

    T. Huang and F. Jiao, “Study on data transfer in meteorological forecast of small and medium-sized cities and its application in zhaoqing city,” Comput. J., vol. 63, no. 7, pp. 1076–1083, 2020. [Online]. Available: https://doi.org/10.1093/comjnl/bxz087

    work page doi:10.1093/comjnl/bxz087 2020

  17. [18]

    Domain adaptation for time series forecasting via attention sharing,

    X. Jin, Y . Park, D. C. Maddix, H. Wang, and Y . Wang, “Domain adaptation for time series forecasting via attention sharing,” in International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, ser. Proceedings of Machine Learning Research, K. Chaudhuri, S. Jegelka, L. Song, C. Szepesv ´ari, G. Niu, and S. Sabato, Eds., ...

    2022

  18. [19]

    The building adapter: Towards quickly applying building analytics at scale,

    D. Hong, H. Wang, J. Ortiz, and K. Whitehouse, “The building adapter: Towards quickly applying building analytics at scale,” in Proceedings of the 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments, BuildSys 2015, Seoul, South Korea, November 4-5, 2015, D. E. Culler, Y . Agarwal, and R. Mangharam, Eds. ACM, 2015, ...

    arXiv 2015

  19. [20]

    Transfer learning for smart buildings: A critical review of algorithms, applications, and future perspectives,

    G. Pinto, Z. Wang, A. Roy, T. Hong, and A. Capozzoli, “Transfer learning for smart buildings: A critical review of algorithms, applications, and future perspectives,”Advances in Applied Energy, vol. 5, p. 100084, 2022. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S2666792422000026 17

    2022

  20. [21]

    Fast adaptation of thermal dynamics model for predictive control of HV AC and natural ventilation using transfer learning with deep neural networks,

    Y . Chen, Y . Zheng, and H. Samuelson, “Fast adaptation of thermal dynamics model for predictive control of HV AC and natural ventilation using transfer learning with deep neural networks,” in 2020 American Control Conference, ACC 2020, Denver, CO, USA, July 1-3, 2020. IEEE, 2020, pp. 2345–2350. [Online]. Available: https://doi.org/10.23919/ACC45564.2020.9147321

    arXiv 2020

  21. [22]

    Development of an ann-based building energy model for information-poor buildings using transfer learning,

    A. Li, F. Xiao, C. Fan, and M. Hu, “Development of an ann-based building energy model for information-poor buildings using transfer learning,”Building Simulation, vol. 14, no. 1, pp. 89–101, Feb 2021. [Online]. Available: https://doi.org/10.1007/s12273-020-0711-5

    work page doi:10.1007/s12273-020-0711-5 2021

  22. [23]

    One for many: Transfer learning for building HV AC control,

    S. Xu, Y . Wang, Y . Wang, Z. O’Neill, and Q. Zhu, “One for many: Transfer learning for building HV AC control,” inBuildSys ’20: The 7th ACM International Conference on Systems for Energy- Efficient Buildings, Cities, and Transportation, Virtual Event, Japan, November 18-20, 2020. ACM, 2020, pp. 230–239. [Online]. Available: https://doi.org/10.1145/340830...

    arXiv 2020

  23. [24]

    Deep transfer learning for thermal dynamics modeling in smart buildings,

    Z. Jiang and Y . M. Lee, “Deep transfer learning for thermal dynamics modeling in smart buildings,” in2019 IEEE International Conference on Big Data (IEEE BigData), Los Angeles, CA, USA, December 9-12, 2019, C. K. Baru, J. Huan, L. Khan, X. Hu, R. Ak, Y . Tian, R. S. Barga, C. Zaniolo, K. Lee, and Y . F. Ye, Eds. IEEE, 2019, pp. 2033–2037. [Online]. Avail...

    arXiv 2019

  24. [25]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. A. Riedmiller, A. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep reinforcement learning,”Nat., vol. 518, no. 7540, pp. 529–533,

  25. [26]

    virtualpulse/EnergyPlus-7-1-0/DataSets at master · joshwentz/virtualpulse

    Joshwentz, “virtualpulse/EnergyPlus-7-1-0/DataSets at master · joshwentz/virtualpulse.” [Online]. Available: https://github.com/ joshwentz/virtualpulse/tree/master/EnergyPlus-7-1-0/DataSets APPENDIXA HYPERPARAMETERS FORMODEL ANDTRAINING SETTINGS We provide a table to summarize the hyperparameters used in our work, including both model and training setting...

  26. [2015]

    Available: https://doi.org/10.1038/nature14236

    [Online]. Available: https://doi.org/10.1038/nature14236

    doi:10.1038/nature14236

  27. [2020]

    Available: https://doi.org/10.1109/JIOT.2019.2957289

    [Online]. Available: https://doi.org/10.1109/JIOT.2019.2957289

    arXiv 2019

  28. [2025]

    Available: https://doi.org/10.48550/arXiv.2506.09042

    [Online]. Available: https://doi.org/10.48550/arXiv.2506.09042

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.09042