Pith. sign in

REVIEW 5 major objections 5 minor 49 references

Integrating Physics-Based and Data-Driven Approaches for Probabilistic Building Energy Modeling

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper establishes that residual learning is the best-performing hybrid architecture for probabilistic building temperature forecasting and that conformalized quantile regression effectively calibrates the resulting intervals.

desk verdict A solid empirical benchmark of hybrid probabilistic building energy models; the residual approach's average superiority holds up, but the conformal calibration claim is overextended given the acknowledged distribution shift. read the letter →

arxiv 2507.17526 v1 pith:4DGLBVR2 submitted 2025-07-23 eess.SY cs.AIcs.LGcs.SY

classification eess.SYcs.AIcs.LGcs.SY
keywords buildingenergymodelinghybridquantileregressionconformalpredictionresiduallearningindoortemperatureuncertaintyquantificationphysics-informedmachine
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether hybrid models that marry a physics-based building simulator with data-driven quantile regression can deliver reliable probabilistic indoor-temperature forecasts, and which hybrid design is best. It evaluates five hybrid architectures—assistant, residual, surrogate, augmentation, and constrained—against purely data-driven baselines on a real occupied residential unit. Its central finding is that residual learning, where the machine-learning model predicts the difference between the EnergyPlus simulation and measured temperatures, performs best on average and is the only design that tracks physically intuitive behavior during out-of-distribution events such as prolonged window openings. The paper further finds that Conformalized Quantile Regression provides an effective post-hoc calibration of the quantile forecasts.

What carries the argument

The central object is the residual hybrid formulation, Eq. (2), in which the data-driven quantile model is trained on $\mathbf{y} - f_{EP}(\mathbf{x})$, the gap between measured indoor temperature and the EnergyPlus simulation, instead of on $\mathbf{y}$ directly. This turns the physics model into a prior that already explains most of the thermal dynamics, leaving the learner to model a smaller, smoother residual. The calibration machinery is Conformalized Quantile Regression, whose non-conformity score $\max\{f_{\alpha/2}(\mathbf{x})-y, y-f_{1-\alpha/2}(\mathbf{x})\}$ and empirical-quantile correction $\Delta_q$ widen the raw quantile outputs so that prediction intervals reach nominal coverage.

What would settle it

A season-split experiment would settle the calibration claim: calibrate on summer 2020 data only, test on winter 2021, and compare average coverage error with and without conformal correction; the paper's own Figure 12 shows the two distributions differ, so the burden is on showing the correction still achieves near-zero ACE.

Watch

Extended reading notes

Core claim

In the terms of the paper, predictive uncertainty in building thermodynamics is treated as primarily aleatoric, and all five hybrid strategies are cast into a quantile-regression framework with 99 quantiles. Across five rooms, the residual approach attains the lowest average pinball loss, with a quantile feedforward network reaching 0.33 against 0.37 for the pure data-driven QNN, and it also yields the best Winkler scores and among the best average coverage errors. In the out-of-distribution window-opening episodes, only the residual model captures the temperature drops, because it inherits the physics simulation's response while the other models do not. The ablation with Conformalized Quantile Regression shows that conformal calibration reduces average coverage error across all hybrid and data-driven QNN models, mainly by widening the prediction intervals.

Load-bearing premise

The conformal calibration step assumes the 20% of the 2020 training year used for calibration is exchangeable with the 2021 test year, an assumption the paper itself notes is partially violated by seasonality and changing occupant behavior.

Editorial extensions

If this is right

  • For probabilistic indoor-temperature modeling, a residual hybrid architecture appears preferable to assistant, augmentation, constrained, or surrogate designs, and best with a feedforward neural network as the residual learner.
  • The physical-plausibility advantage of residual learning in out-of-distribution conditions suggests that hybrid models can be more trustworthy than pure data-driven models when operating conditions shift.
  • Conformalized Quantile Regression is a viable post-training calibration step for building temperature quantiles; its main effect is to widen intervals to correct coverage.
  • The constrained approach's benefit is room-dependent: physics regularization helps bathrooms, where sensors are sparse, but adds little for well-sensed bedrooms.
  • Surrogate learning, which replaces the physics model entirely, performs worst among hybrids and should be avoided when the physics simulator is available at inference time.

Reading between the lines

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

  • A broader implication the paper leaves implicit is that residual learning may be the most sample-efficient way to inject a physics prior, because the learner only needs to model the residual rather than rediscover the dominant heat dynamics.
  • The exchangeability caveat suggests a natural extension: adaptive conformal procedures, or a calibration set spanning all seasons, would likely preserve the reported coverage gains under stronger distribution shift.
  • One could test whether the residual advantage persists when the physics model is deliberately miscalibrated, since the residual then contains more structured signal that the learner must absorb.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper presents a comparative evaluation of five hybrid physics-based/data-driven approaches (Assistant, Residual, Surrogate, Augmentation, Constrained) for probabilistic indoor temperature prediction in a real building, using quantile regression with three base models (QR, QNN, QRF) and 99 quantiles. A Conformalized Quantile Regression (CQR) calibration ablation is included for the QNN family. The two main findings are: (1) residual learning with a feedforward neural network performs best on average (PBL 0.33 vs 0.37 for the data-driven QNN) and is the only approach with physically intuitive out-of-distribution predictions; (2) CQR is an effective calibration procedure for indoor temperature quantiles. The evaluation is broad in scope (5 rooms, 3 base models, 5 hybrid variants, 3 metrics, conformal ablation, sensitivity analysis) and uses a real-world dataset, but the headline claims are weakened by a lack of statistical uncertainty quantification and by the paper's own acknowledgment that the CQR exchangeability assumption is only partially satisfied.

Significance. If the findings hold, the paper provides a useful benchmark for hybrid probabilistic building energy modeling, identifying residual learning as a promising architecture and demonstrating the practical value of conformal calibration in this domain. The study is methodical and broad in its design: five rooms, three base learners, five hybrid strategies, 99 quantiles, several evaluation metrics, plus a conformal ablation and a sensitivity analysis of the physics-constraint weight. The real-world test set and the explicit discussion of limitations (e.g., the exchangeability violation) are also strengths. However, the central conclusions rely on differences that are not tested for significance and on a calibration claim whose theoretical prerequisite is acknowledged to be violated, so the significance of the results is currently conditional on additional evidence.

major comments (5)
  1. [Section 4.1, Fig. 5] The headline result that Residual-QNN achieves an average PBL of 0.33 versus 0.37 for the data-driven QNN is reported without any uncertainty quantification, such as confidence intervals, bootstrap errors, or a significance test. With a single test year and a single building, the 0.04 difference may well be within noise. Please provide per-segment error bars, bootstrap intervals, or a statistical test to support the statement of a 'clear relative improvement of approximately 10%'.
  2. [Section 4.3, Eq. (14)] The second main finding states that CQR is effective, but Section 4.3 itself acknowledges that the exchangeability assumption is 'partially violate[d]' and Figure 12 shows clear empirical differences between the calibration and test distributions. Under such distribution shift, the correction factor Δ_q from Eq. (14) has no formal coverage guarantee; the ACE improvements shown in Figure 10 could therefore reflect this particular shift rather than a general calibration property. The suggested mitigation of using a full-year calibration set is not implemented. Please either implement that mitigation, employ an adaptive conformal method, or restrict the claim to settings where exchangeability approximately holds.
  3. [Section 4.1, Fig. 8] The claim that the Residual approach is the only model that produces physically intuitive out-of-distribution predictions is based on a single visual example of window-opening periods in Room 272, without any quantitative metric computed over those periods. Please provide quantitative evidence, for instance PBL or ACE restricted to the grey-shaded OOD periods, or a formal OOD evaluation, to support the abstract's strong statement.
  4. [Section 5] The conclusion states that the Residual approach 'consistently outperforms the others in terms of average pinball loss, average coverage error, and Winkler score.' This is internally inconsistent with Section 4.1, which reports that Constrained-QNN achieves the highest calibration accuracy (lowest ACE) among all models. Only the PBL and WKS claims are supported by the presented results. Please correct this contradiction or weaken the ACE claim.
  5. [Section 4.3] The conformal ablation is performed only for the QNN base model, yet the abstract makes a general claim about 'quantile predictions in case of indoor temperature modeling.' The effectiveness of CQR for QR and QRF is not demonstrated. Please either extend the ablation to the other base models or narrow the claim to the QNN family.
minor comments (5)
  1. [Section 4.1] There are label inconsistencies: 'bedroom 273' should be 'living room 273' (the rooms are defined as R273 living room, R272/R274 bedrooms, R275/R276 bathrooms), and 'bathrooms 275 and 275' should read '275 and 276.'
  2. [Section 3.2] The abbreviation 'QLR' appears where 'QR' is used elsewhere (e.g., in the description of Augmentation-QLR). Please use a single consistent abbreviation for Quantile Regression.
  3. [Section 3.2] The Constrained-QRF variant is implemented by adding the physics-based simulation as an additional target rather than by adding a physics term to the loss function. This difference should be acknowledged more prominently in the evaluation of the Constrained approach, as it means the 'Constrained' label does not represent the same mechanism across all base models.
  4. [Section 4.4] The caption of Figure 13 states 'two QNN hybrid approaches,' but the figure appears to show more than two models. Please clarify which models are plotted.
  5. [Section 5] The sentence 'In case of the Constrained approach, we demonstrate that the increasing proportion of physics-based information is only beneficial for rooms that are harder to model' is not fully supported by the sensitivity analysis in Section 4.2, which tests only the QNN variant; please make this limitation explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the results are an out-of-sample empirical benchmark, and the sole self-citation is not load-bearing.

full rationale

The paper is an empirical comparison, not a derivation that reduces to its inputs. The five hybrid architectures are defined by explicit equations (1)-(5), and all reported metrics (PBL, ACE, WKS) are evaluated on the 2021 test year after training and calibration on 2020 data. The Residual approach predicts y - f_EP, but its claimed advantage is measured out-of-sample on 2021, so the result is not equivalent to its training objective by construction. Conformalized Quantile Regression is implemented as a standard external method [39], with non-conformity scores computed on a held-out 20% calibration split and applied to test predictions; the ACE comparison is therefore an empirical claim, not a fitted parameter renamed as a prediction. The only self-citation is [27], which supplies the hybrid taxonomy, but the same taxonomy is re-specified in the equations of Section 2.1 and is not load-bearing for the conclusions. The paper's own caveat in Section 4.3 that seasonality and changing user behavior partially violate the exchangeability assumption, and the ECDF difference in Figure 12, are validity limitations rather than circularity; they do not make the conformal-prediction result true by definition. No equation is shown to equal its own output, and no uniqueness claim or ansatz is imported from the authors' prior work to force the preferred architecture.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical benchmark, so the ledger is mostly domain assumptions and hand-chosen hyperparameters. No new entities are introduced. The main burden is on the reliability of the EnergyPlus model and the exchangeability assumption for conformal prediction.

free parameters (3)
  • Physics regularization constant = 0.1
    The Constrained approach adds a physics-based loss term weighted by a constant chosen as 0.1 without a formal tuning protocol; the sensitivity analysis in Section 4.2 varies it from 0 to 5.
  • QNN architecture hyperparameters = 2 layers, 128 neurons
    Selected by grid search over two/three layers with 64/128 neurons; the chosen values are a design choice, though validated on a split.
  • QRF hyperparameters = 500 trees, min samples split 2, leaf 1
    Selected by grid search over 100 to 1000 trees; the exact settings affect quantile estimates.
assumptions (4)
  • domain assumption The EnergyPlus model, calibrated on 2020 training data, provides a physically reliable simulation of indoor temperatures, including qualitative behavior under conditions not present in training.
    The Residual approach's superiority and OOD robustness depend on this; the paper states the EP model 'remains an approximation of the true temperature' (Section 4.2).
  • domain assumption The calibration and test sets are approximately exchangeable for Conformalized Quantile Regression.
    Stated in Section 4.3 as an idealized assumption that is partially violated by seasonality and user behavior.
  • domain assumption Uncertainty in indoor temperature prediction is predominantly aleatoric, so quantile regression without explicit epistemic uncertainty modeling is sufficient.
    The authors justify this in Section 1 and use it to motivate the choice of quantile regression; if epistemic uncertainty is substantial, the comparison may change.
  • ad hoc to paper Time-lagged exogenous variables can be ignored for a fair comparison with EnergyPlus.
    Section 3.2: 'Time-lagged data is disregarded as model input to allow direct comparison to EnergyPlus.' This may handicap data-driven models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating Physics-Based and Data-Driven Approaches for Probabilistic Building Energy Modeling." pith.science (2026). https://pith.science/paper/4DGLBVR2

@misc{pith2026250717526,
  author       = {Pith},
  title        = {Pith review of: Integrating Physics-Based and Data-Driven Approaches for Probabilistic Building Energy Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DGLBVR2}},
  note         = {Machine review of arXiv:2507.17526}
}
read the original abstract

Building energy modeling is a key tool for optimizing the performance of building energy systems. Historically, a wide spectrum of methods has been explored -- ranging from conventional physics-based models to purely data-driven techniques. Recently, hybrid approaches that combine the strengths of both paradigms have gained attention. These include strategies such as learning surrogates for physics-based models, modeling residuals between simulated and observed data, fine-tuning surrogates with real-world measurements, using physics-based outputs as additional inputs for data-driven models, and integrating the physics-based output into the loss function the data-driven model. Despite this progress, two significant research gaps remain. First, most hybrid methods focus on deterministic modeling, often neglecting the inherent uncertainties caused by factors like weather fluctuations and occupant behavior. Second, there has been little systematic comparison within a probabilistic modeling framework. This study addresses these gaps by evaluating five representative hybrid approaches for probabilistic building energy modeling, focusing on quantile predictions of building thermodynamics in a real-world case study. Our results highlight two main findings. First, the performance of hybrid approaches varies across different building room types, but residual learning with a Feedforward Neural Network performs best on average. Notably, the residual approach is the only model that produces physically intuitive predictions when applied to out-of-distribution test data. Second, Quantile Conformal Prediction is an effective procedure for calibrating quantile predictions in case of indoor temperature modeling.

Figures

Figures reproduced from arXiv: 2507.17526 by the authors.

Figure 1
Figure 1. Overview of the five considered hybrid approaches with physics-based EnergyPlus and data-driven sub-models. The two main data sources include building documentation data and building sensor data of various groups. The green exogenous data arrow represents all sensor measurements besides indoor room temperature. While 𝑋 and 𝑌 indicate input features and target variables of the data-driven sub-model, 𝑌̂ denotes the fi… view at source ↗
Figure 2
Figure 2. Overview of hybrid framework consisting of physics-based EnergyPlus model and data-driven model. Leandro Von Krannichfeldt et al.: Preprint submitted to ... Page 4 of 19 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the EnergyPlus workflow with modules, adapted from [41] 2.3. Data-driven quantile models For the data-driven quantile models, we evaluate Quantile Regression (QR), Quantile Feedfoward Neural Networks (QNN) and Quantile Random Forest (QRF). QR is selected as a baseline model for its simplicity and linear structure. QNN is chosen for its effectiveness in regression tasks and computational efficiency compar… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: UMAR unit at Empa with two bedrooms, one living room and two bathrooms. Feature group Feature variables Datetime Season, week (weekday/weekend), daytime (morning/afternoon/evening/night) Weather Drybulb & dewpoint temperature, direct & diffuse solar radiation, rel. hum…
Figure 5
Figure 5. Figure 5: PBL across all 99 quantiles broken down by rooms, for pure data-driven and hybrid approaches. The grey-dotted line indicates the mean PBL across all rooms. For uncertainty estimation, it is often useful to consider prediction intervals, defined by upper and lower quant…
Figure 6
Figure 6. Figure 6: ACE for the 90% prediction interval broken down by rooms, for pure data-driven and hybrid approaches. The grey-dotted line indicates the mean ACE across all rooms. The orange-dotted line shows the zero axis Data-driven Assistant Residual Surrogate Augmentation Constrai…
Figure 7
Figure 7. Figure 7: WKS for the 90% prediction interval broken down by rooms, for pure data-driven and hybrid approaches. The grey-dotted line indicates the mean WKS across all rooms. One reason for the better performance of the Residual approach is its robustness to out-of-distribution s…
Figure 8
Figure 8. Figure 8: Visualization of the 5%- and 95%-quantile enclosing a 90% prediction interval for a selection of hybrid approaches using the QNN sub-model in the case of Room 272. The forecast of the Surrogate approach is omitted for better visibility. The grey shaded areas indicate t…
Figure 9
Figure 9. Figure 9: Sensitivity analysis of the physics regularization constant of the QNN physics-constrained hybrid approach. While 0 indicates a total loss function fully consisting of the data-driven loss, the regularization constants 0.1, 0.5, 1 and 5 represent a total loss with 10%,…
Figure 10
Figure 10. Figure 10: Comparison between ACE performance for the 90% prediction interval of the QNN hybrid approaches with and without conformal prediction. 4.4. Prediction evaluation by quantile and prediction interval The analysis in the preceding subsections were limited to either aggre…
Figure 11
Figure 11. Figure 11: Forecast comparison between standard and Conformalized Quantile Regression for the Residual-QNN in the case of Bedroom 272 [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: The Empirical Cumulative Distribution Function of the calibration and test set for the Temperature in Bedroom 272. purely data-driven QNN ( [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: PBL per quantile for two QNN hybrid approaches in the case of Bedroom 272. 5. Conclusion In this work, we introduce probabilistic methods to five established hybrid approaches in building energy modeling combining physics-based and data-driven modeling alongside a pur…
Figure 14
Figure 14. Figure 14: Reliability per prediction interval nominal confidence for two QNN hybrid approaches in the case of Bedroom 272 [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Purely data-driven QNN predictions of all 99 quantiles for selected time steps in March in the case of Bedroom 272. The median prediction is highlighted in red. during abrupt temperature increases with condensed predictive distribution, suggesting that further refinem…
Figure 16
Figure 16. Figure 16: Residual-QNN predictions of all 99 quantiles for selected time steps in March in the case of Bedroom 272. The median prediction is highlighted in red. CRediT authorship contribution statement Leandro Von Krannichfeldt: Conceptualization, Methodology, Software, Investi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 40 canonical work pages

  1. [27]

    Combining Physics-based and Data-driven Modeling for Building Energy Systems

    Krannichfeldt, L.V., Orehounig, K., Fink, O., 2024. Combining physics-based and data-driven modeling for building energy systems. URL: https://arxiv.org/abs/2411.01055, arXiv:2411.01055

  2. [1]

    Prediction of building power consumption using transfer learning-based reference building and simulation dataset

    Ahn, Y., Kim, B.S., 2022. Prediction of building power consumption using transfer learning-based reference building and simulation dataset. Energy and Buildings 258, 111717. URL:https://www.sciencedirect.com/science/article/pii/S037877882101001X, doi:10.1016/j.enbuild.2021.111717

  3. [2]

    DigitalTwinforHVACLoadandEnergyStoragebasedona Hybrid ML Model with CTA-2045 Controls Capability, in: 2022 IEEE Energy Conversion Congress and Exposition (ECCE), pp

    Alden,R.E.,Jones,E.S.,Poore,S.B.,Gong,H.,AlHadi,A.,Ionel,D.M.,2022. DigitalTwinforHVACLoadandEnergyStoragebasedona Hybrid ML Model with CTA-2045 Controls Capability, in: 2022 IEEE Energy Conversion Congress and Exposition (ECCE), pp. 1–5. URL: https://ieeexplore.ieee.org/document/9948141, doi:10.1109/ECCE50734.2022.9948141. iSSN: 2329-3748

  4. [3]

    Probabilistic indoor temperature forecasting: A new approach using bernstein- polynomial normalizing flows

    Arpogaus, M., Kempf, R., Baur, T., Schubert, G., 2025. Probabilistic indoor temperature forecasting: A new approach using bernstein- polynomial normalizing flows. Energy and Buildings 335, 115527. URL:https://www.sciencedirect.com/science/article/pii/ S0378778825002579, doi:10.1016/j.enbuild.2025.115527

  5. [4]

    Identifying suitable models for the heat dynamics of buildings

    Bacher, P., Madsen, H., 2011. Identifying suitable models for the heat dynamics of buildings. Energy and Buildings 43, 1511–

  6. [5]

    Pattern Recognition and Machine Learning

    Bishop, C.M., . Pattern Recognition and Machine Learning. Information science and statistics, Springer

  7. [6]

    Quantifying Uncertainty with Conformal Prediction for Heating and Cooling Load Forecasting in Building Performance Simulation

    Borrotti, M., 2024. Quantifying Uncertainty with Conformal Prediction for Heating and Cooling Load Forecasting in Building Performance Simulation. Energies 17, 4348. URL:https://www.mdpi.com/1996-1073/17/17/4348, doi:10.3390/en17174348. Leandro Von Krannichfeldt et al.:Preprint submitted to ... Page 16 of 19 Integrating Physics-Based and Data-Driven Appro...

  8. [7]

    Probabilistic electric load forecasting through Bayesian Mixture Density Networks

    Brusaferri, A., Matteucci, M., Spinelli, S., Vitali, A., 2022. Probabilistic electric load forecasting through Bayesian Mixture Density Networks. Applied Energy 309, 118341. URL: https://www.sciencedirect.com/science/article/pii/S0306261921015907, doi:10.1016/j.apenergy.2021.118341

Show all 49 references
  1. [8]

    Probabilistic modeling of the indoor climates of residential buildings using EnergyPlus

    Buechler, E., Pallin, S., Boudreaux, P., Stockdale, M., 2017. Probabilistic modeling of the indoor climates of residential buildings using EnergyPlus. Journal of Building Physics 41, 225–246. URL:https://journals.sagepub.com/doi/10.1177/1744259117701893, doi:10.1177/1744259117701893

  2. [9]

    A hybrid-model forecasting framework for reducing the building energy performance gap

    Chen, X., Guo, T., Kriegel, M., Geyer, P., . A hybrid-model forecasting framework for reducing the building energy performance gap. Ad- vancedEngineeringInformatics52,101627. URL: https://www.sciencedirect.com/science/article/pii/S1474034622000933, doi:10.1016/j.aei.2022.101627

  3. [10]

    Physics-informed neural networks for building thermal modeling and demand response control

    Chen, Y., Yang, Q., Chen, Z., Yan, C., Zeng, S., Dai, M., 2023. Physics-informed neural networks for building thermal modeling and demand response control. Building and Environment 234, 110149. URL:https://www.sciencedirect.com/science/article/pii/ S0360132323001762, doi:10.10...

  4. [11]

    Context-Aware Urban Energy Efficiency Optimization Using Hybrid Physical Models, in: Climate Change AI, Climate Change AI

    Choi, B., Nutkiewicz, A., Jain, R., 2020. Context-Aware Urban Energy Efficiency Optimization Using Hybrid Physical Models, in: Climate Change AI, Climate Change AI. pp. 1–7. URL:https://www.climatechange.ai/papers/neurips2020/24

  5. [12]

    Review of data-driven energy modelling techniques for building retrofit

    Deb, C., Schlueter, A., 2021. Review of data-driven energy modelling techniques for building retrofit. Renewable and Sustainable Energy Reviews 144, 110990. URL: https://linkinghub.elsevier.com/retrieve/pii/S1364032121002823, doi:10.1016/j. rser.2021.110990

  6. [13]

    CVXPY: A Python-embedded modeling language for convex optimization

    Diamond, S., Boyd, S., 2016. CVXPY: A Python-embedded modeling language for convex optimization. Journal of Machine Learning Research 17, 1–5

  7. [14]

    Machine Learning & Uncertainty Quantification: Application in Building Energy Consumption, in: 2022 Annual Reliability and Maintainability Symposium (RAMS), pp

    Fakour, F., Parhizkar, T., Ramezani, R., Mosleh, A., 2022. Machine Learning & Uncertainty Quantification: Application in Building Energy Consumption, in: 2022 Annual Reliability and Maintainability Symposium (RAMS), pp. 1–7. URL:https://ieeexplore.ieee.org/ document/9893988, d...

  8. [15]

    Modeling the Thermal Dynamics of Buildings: A Latent- Force-Model-BasedApproach

    Ghosh, S., Reece, S., Rogers, A., Roberts, S., Malibari, A., Jennings, N.R., 2015. Modeling the Thermal Dynamics of Buildings: A Latent- Force-Model-BasedApproach. ACMTransactionsonIntelligentSystemsandTechnology6,1–27. URL: https://dl.acm.org/doi/10. 1145/2629674, doi:10.1145/2629674

  9. [16]

    Adaptive conformal inference under distribution shift

    Gibbs, I., Candes, E., 2021. Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems 34, 1660–1672

  10. [17]

    Physics informed neural networks for control oriented thermal modeling of buildings

    Gokhale, G., Claessens, B., Develder, C., 2022. Physics informed neural networks for control oriented thermal modeling of buildings. AppliedEnergy314,118852. URL: https://www.sciencedirect.com/science/article/pii/S0306261922002884,doi: 10.1016/ j.apenergy.2022.118852

  11. [18]

    Uncertainty quantification and sensitivity analysis of energy consumption in substation buildings at the planning stage

    Guo, J., Wang, Z., Li, M., Jin, Y., 2023. Uncertainty quantification and sensitivity analysis of energy consumption in substation buildings at the planning stage. Journal of Building Performance Simulation 16, 327–345. URL:https://www.tandfonline.com/doi/full/10. 1080/19401493...

  12. [19]

    Quantile regression

    Hallock, K., Koenker, R., 2001. Quantile regression. Journal of Economic Perspectives 15, 143–156

  13. [20]

    Bayesian lstm for indoor temperature modeling

    Hannula, E., Häkkinen, A., Solonen, A., Uribe, F., de Wiljes, J., Roininen, L., 2025. Bayesian lstm for indoor temperature modeling. URL: https://arxiv.org/abs/2504.03350, arXiv:2504.03350

  14. [21]

    Gaussian process modeling for measurement and verification of building energy savings

    Heo, Y., Zavala, V.M., 2012. Gaussian process modeling for measurement and verification of building energy savings. Energy and Buildings 53,7–18. URL: https://www.sciencedirect.com/science/article/pii/S037877881200312X,doi: 10.1016/j.enbuild.2012. 06.024

  15. [22]

    Building simulation: Ten challenges

    Hong, T., Langevin, J., Sun, K., 2018. Building simulation: Ten challenges. Building Simulation 11, 871–898. URL:https://doi.org/ 10.1007/s12273-018-0444-x, doi:10.1007/s12273-018-0444-x

  16. [23]

    Hossain, M.M., Zhang, T., Ardakanian, O., 2019. Evaluating the Feasibility of Reusing Pre-trained Thermal Models in the Residential Sector, in: Proceedings of the 1st ACM International Workshop on Urban Building Energy Sensing, Controls, Big Data Analysis, and Visualization, A...

  17. [24]

    Buildings - Energy System

    International Energy Association, 2023. Buildings - Energy System. URL:https://www.iea.org/energy-system/buildings

  18. [25]

    quantile-forest: A python package for quantile regression forests

    Johnson, R.A., 2024. quantile-forest: A python package for quantile regression forests. Journal of Open Source Software 9, 5976. doi:10.21105/joss.05976

  19. [26]

    Benchmarking HVAC controller performance with a digital twin, in: Energy Proceedings, pp

    Khayatian, F., Cai, H., Bojarski, A., Heer, P., Bollinger, A., . Benchmarking HVAC controller performance with a digital twin, in: Energy Proceedings, pp. 1–6. doi:10.46855/energy-proceedings-10382

  20. [28]

    Distribution-Free Predictive Inference for Regression

    Lei, J., G’Sell, M., Rinaldo, A., Tibshirani, R.J., Wasserman, L., 2018. Distribution-Free Predictive Inference for Regression. Journal of the American Statistical Association 113, 1094–1111. URL:https://www.tandfonline.com/doi/full/10.1080/01621459.2017. 1307116, doi:10.1080/...

  21. [29]

    Energy-saving potential benchmarking method of office buildings based on probabilistic forecast

    Liu, C., Li, Y., Chen, H., Xing, L., Zhang, S., 2024. Energy-saving potential benchmarking method of office buildings based on probabilistic forecast. Journal of Building Engineering 95, 110282. URL:https://www.sciencedirect.com/science/article/pii/ S2352710224018503, doi:http...

  22. [30]

    A review of physics-informed machine learning for building energy modeling

    Ma, Z., Jiang, G., Hu, Y., Chen, J., 2025. A review of physics-informed machine learning for building energy modeling. Applied Energy 381, 125169. URL: https://www.sciencedirect.com/science/article/pii/S0306261924025534, doi:https://doi.org/10. 1016/j.apenergy.2024.125169

  23. [31]

    AhybridapproachtothermalbuildingmodellingusingacombinationofGaussianprocessesandgrey-box models

    MassaGray,F.,Schmidt,M.,2018. AhybridapproachtothermalbuildingmodellingusingacombinationofGaussianprocessesandgrey-box models. Energy and Buildings 165, 56–63. URL:https://www.sciencedirect.com/science/article/pii/S0378778817332942, Leandro Von Krannichfeldt et al.:Preprint su...

  24. [32]

    Quantile regression forests

    Meinshausen, N., 2006. Quantile regression forests. J. Mach. Learn. Res. 7, 983–999

  25. [33]

    Change-point multivariable quantile regression to explore effect of weather variables on building energy consumption and estimate base temperature range

    Meng, Q., Xiong, C., Mourshed, M., Wu, M., Ren, X., Wang, W., Li, Y., Song, H., 2020. Change-point multivariable quantile regression to explore effect of weather variables on building energy consumption and estimate base temperature range. Sustainable Cities and Society 53, 10...

  26. [34]

    Mohebi,P.,Wang,Z.,2024. Chance-constraintmpcapproachtooptimizebuildingenergysystemoperationconsideringpredictionuncertain- ties, in: Proceedings of ASim Conference 2024: 5th Asia Conference of IBPSA, IBPSA-Asia, Osaka, Japan. pp. 1265–1272. URL:https: //publications.ibpsa.org/...

  27. [35]

    Weighted aggregated ensemble model for energy demand management of buildings

    Pachauri, N., Ahn, C.W., 2023. Weighted aggregated ensemble model for energy demand management of buildings. Energy 263, 125853. URL: https://www.sciencedirect.com/science/article/pii/S0360544222027396, doi:10.1016/j.energy.2022.125853

  28. [36]

    Pytorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems 32

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S., 2019. Pytorch: An imp...

  29. [37]

    Demand response for residential building heating: Effective Monte Carlo Tree Search control based on physics-informed neural networks

    Pavirani, F., Gokhale, G., Claessens, B., Develder, C., 2024. Demand response for residential building heating: Effective Monte Carlo Tree Search control based on physics-informed neural networks. Energy and Buildings 311, 114161. URL:https://www.sciencedirect. com/science/art...

  30. [38]

    NEST – una plataforma para acelerar la innovación en edificios

    Richner, P., Heer, P., Largo, R., Marchesi, E., Zimmermann, M., 2018. NEST – una plataforma para acelerar la innovación en edificios. Informes de la Construcción 69, 222. doi:10.3989/id.55380

  31. [39]

    Conformalized quantile regression

    Romano, Y., Patterson, E., Candes, E., 2019. Conformalized quantile regression. Advances in neural information processing systems 32

  32. [40]

    Building consumption anomaly detection: A comparative study of two probabilistic approaches

    Stjelja, D., Kuzmanovski, V., Kosonen, R., Jokisalo, J., 2024. Building consumption anomaly detection: A comparative study of two probabilistic approaches. Energy and Buildings 313, 114249. URL: https://www.sciencedirect.com/science/article/pii/ S0378778824003657, doi:10.1016/...

  33. [41]

    Department of Energy, 2017

    U.S. Department of Energy, 2017. Energyplus™, version 00. URL:https://www.osti.gov//servlets/purl/1395882

  34. [42]

    Algorithmic Learning in a Random World

    Vovk, V., Gammerman, A., Shafer, G., 2005. Algorithmic Learning in a Random World. Springer-Verlag, New York. URL:http: //link.springer.com/10.1007/b106715, doi:10.1007/b106715

  35. [43]

    Wan,C.,Xu,Z.,Wang,Y.,Dong,Z.Y.,Wong,K.P.,2014.Ahybridapproachforprobabilisticforecastingofelectricityprice.IEEETransactions on Smart Grid 5, 463–470

  36. [44]

    Using Bayesian deep learning approaches for uncertainty-aware building energy surrogate models

    Westermann, P., Evins, R., 2021. Using Bayesian deep learning approaches for uncertainty-aware building energy surrogate models. Energy and AI 3, 100039. URL:https://www.sciencedirect.com/science/article/pii/S2666546820300392, doi:10.1016/j.egyai. 2020.100039

  37. [45]

    A decision-theoretic approach to interval estimation

    Winkler, R.L., 1972. A decision-theoretic approach to interval estimation. Journal of the American Statistical Association 67, 187–191

  38. [46]

    Conformal prediction interval for dynamic time-series, in: Meila, M., Zhang, T

    Xu, C., Xie, Y., 2021. Conformal prediction interval for dynamic time-series, in: Meila, M., Zhang, T. (Eds.), Proceedings of the 38th InternationalConferenceonMachineLearning,PMLR.pp.11559–11569. URL: https://proceedings.mlr.press/v139/xu21h.html

  39. [47]

    Xu, X., Hu, Y., Atamturktur, S., Chen, L., Wang, J., 2025. Systematic review on uncertainty quantification in machine learning-based buildingenergymodeling.RenewableandSustainableEnergyReviews218,115817.URL: https://www.sciencedirect.com/science/ article/pii/S1364032125004903,...

  40. [48]

    Stateoftheartreviewonmodelpredictivecontrol(MPC)inHeatingVentilationandAir-conditioning(HVAC) field

    Yao,Y.,Shekhar,D.K.,2021. Stateoftheartreviewonmodelpredictivecontrol(MPC)inHeatingVentilationandAir-conditioning(HVAC) field. Building and Environment 200, 107952. URL: https://linkinghub.elsevier.com/retrieve/pii/S0360132321003565, doi:10.1016/j.buildenv.2021.107952. Leandro...

  41. [1522]

    URL: https://www.sciencedirect.com/science/article/pii/S0378778811000491, doi:https://doi.org/10.1016/ j.enbuild.2011.02.005

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.