pith. sign in

arxiv: 2505.22873 · v2 · submitted 2025-05-28 · 💰 econ.GN · q-fin.EC· stat.ML

Forecasting Residential Heating and Electricity Demand with Scalable, High-Resolution, Open-Source Models

Stephen J. Lee , Cailinn Drouin This is my paper

Pith reviewed 2026-05-19 12:25 UTC · model grok-4.3

classification 💰 econ.GN q-fin.ECstat.ML
keywords residential heating demandelectricity demand forecastingprobabilistic deep learningbuilding-level modelshigh-resolution forecastingopen-source energy modelsResStock benchmarkdemand heterogeneity
0
0 comments X

The pith

Probabilistic models trained on gas-heated region data forecast building-level residential heating and non-heating electricity demand with substantially lower errors than ResStock.

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

The paper develops probabilistic deep learning models to forecast hourly residential heating demand and non-heating electricity demand at the individual building level. These models draw on building physical characteristics such as footprint area, height, nearby density, and land use patterns together with high-resolution weather data. Because they are trained in areas with mostly gas heating, they learn to isolate non-heating electricity uses like lighting, appliances, and cooling. A sympathetic reader would care because accurate high-resolution forecasts support better decisions in managing electricity grids and transitioning buildings away from fossil fuels. The work shows these models outperform the widely used ResStock tool in backtests by lowering RMSE 18.8 percent for heating and 27.6 percent for electricity while improving weighted interval scores by 59 percent.

Core claim

Our probabilistic deep learning models, leveraging multimodal building information and weather data, achieve RMSE scores that are 18.8 percent lower for heating demand and 27.6 percent lower for electricity demand, along with 59 percent improvement in weighted interval score, compared to estimates based on the ResStock model in building-level validation tests.

What carries the argument

Probabilistic deep learning models that process multimodal building-level data including footprints, heights, nearby density, land use, and high-resolution weather to separate and forecast heating and non-heating demands.

If this is right

  • Enables scalable high-resolution demand estimation across the residential sector.
  • Supports policymakers with granular insights into demand heterogeneity.
  • Facilitates more effective grid planning and decarbonization strategies for buildings.
  • Provides an open-source platform that can be applied broadly for climate goals.

Where Pith is reading between the lines

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

  • Applying the same separation logic in all-electric regions could reveal if heating electrification patterns differ from the training assumptions.
  • The approach might combine with satellite imagery for even finer spatial resolution in future updates.
  • Similar models could help forecast demand in other sectors like commercial buildings where data is sparse.

Load-bearing premise

The training data from a gas-heated region accurately capture non-heating electricity patterns without any leftover heating-related consumption affecting the separation.

What would settle it

A direct comparison of the model's predictions against measured heating and electricity demand in a region where electric heating is common would test whether the learned patterns truly isolate non-heating uses.

Figures

Figures reproduced from arXiv: 2505.22873 by Cailinn Drouin, Stephen J. Lee.

Figure 1
Figure 1. Figure 1: Seasonal heating and electricity demand backtests comparing MLP-ZIG estimates with historical consumer consump￾tion. The top row presents heating demand results for (a) winter and (b) summer, while the bottom row shows electricity demand results for (c) winter and (d) summer. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Boxplots of heating and electricity demand errors segmented by dataset coverage. The first row shows heating demand errors: (a) whole test set, (b) nonzero values only, and (c) zero values only. The second row presents analogous distributions for electricity demand in (d)-(f). 6. Discussion In this section, we evaluate our models’ performance relative to the research community-standard ResStock framework, … view at source ↗
Figure 3
Figure 3. Figure 3: Posterior predictive model checking results from our MLP-ZIG models for heating (a) and electricity (b) demand, comparing aggregated empirical cumulative dis￾tribution functions (CDFs) of model-estimated probabilities against those of the ideal uniform distribution. timizations. 6.1.1. Challenges in Modeling Heating Demand: Zero Consumption Periods A key challenge in modeling heating demand lies in accurat… view at source ↗
Figure 4
Figure 4. Figure 4: Feature importance analysis for heating (a) and electricity (b) demand models, computed via numerical gradient analysis. Higher values indicate features with greater influence on model predictions. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
read the original abstract

We present a novel framework for high-resolution forecasting of residential heating demand and non-heating electricity demand using probabilistic deep learning models. Because our models are trained on electricity consumption from a predominantly gas-heated region, the learned electricity demand patterns primarily reflect non-heating end uses such as lighting, appliances, and cooling. We focus specifically on providing hourly building-level electricity and heating demand forecasts for the residential sector. Leveraging multimodal building-level information -- including data on building footprint areas, heights, nearby building density, nearby building size, land use patterns, and high-resolution weather data -- and probabilistic modeling, our methods provide granular insights into demand heterogeneity. Validation at the building level underscores a step change improvement in performance relative to NREL's ResStock model, which has emerged as a research community standard for residential heating and electricity demand characterization. In building-level heating and electricity estimation backtests, our probabilistic models respectively achieve RMSE scores 18.8% and 27.6% lower than those based on ResStock, with probabilistic forecast quality measured via WIS improving by 59% for both applications. By offering an open-source, scalable, high-resolution platform for demand estimation and forecasting, this research advances the tools available for policymakers and grid planners, contributing to the broader effort to decarbonize the U.S. building stock and meeting climate objectives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a novel framework for high-resolution forecasting of residential heating demand and non-heating electricity demand using probabilistic deep learning models. Trained on electricity consumption from a predominantly gas-heated region, the models leverage multimodal building-level data (footprint areas, heights, density, land use) and high-resolution weather to generate hourly probabilistic forecasts. It claims a step-change improvement over NREL's ResStock benchmark, with building-level backtests showing 18.8% lower RMSE for heating, 27.6% lower RMSE for electricity, and 59% better WIS for both.

Significance. If the reported performance gains prove robust, the open-source and scalable nature of the platform would represent a meaningful advance over existing tools like ResStock for granular, probabilistic demand estimation. This could support more accurate policy analysis and grid planning for building decarbonization. The use of multimodal inputs and probabilistic modeling directly addresses demand heterogeneity, which is a noted limitation in coarser models.

major comments (2)
  1. Abstract: The abstract reports specific RMSE improvements (18.8% for heating, 27.6% for electricity) and WIS gains (59%) but supplies no information on model architecture, training procedure, data sources, cross-validation strategy, or statistical significance testing. This prevents assessment of whether the claimed gains are robust.
  2. Abstract: The central performance claims rest on the assumption that electricity consumption data from a 'predominantly gas-heated region' captures only non-heating end uses without residual heating-related electricity consumption. No verification, sensitivity analysis, or discussion of potential contamination (e.g., from electric resistance or heat pumps) is provided, which is load-bearing for the separation task and the direct comparison to ResStock.
minor comments (1)
  1. Abstract: The phrase 'multimodal building-level information' lists several variables but would benefit from explicit examples or a brief definition to improve immediate clarity for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments. We address each major comment below, indicating planned revisions to the manuscript where appropriate.

read point-by-point responses

  1. Referee: Abstract: The abstract reports specific RMSE improvements (18.8% for heating, 27.6% for electricity) and WIS gains (59%) but supplies no information on model architecture, training procedure, data sources, cross-validation strategy, or statistical significance testing. This prevents assessment of whether the claimed gains are robust.

    Authors: The abstract is intentionally concise to highlight the core contribution and results. The full manuscript details the probabilistic deep learning models (including architecture), training procedures on electricity consumption data, multimodal building and weather data sources, cross-validation approach in the building-level backtests, and statistical comparisons to ResStock. To facilitate assessment without requiring readers to consult the full text immediately, we will revise the abstract to include a brief summary of the methodological framework and validation strategy. revision: yes

  2. Referee: Abstract: The central performance claims rest on the assumption that electricity consumption data from a 'predominantly gas-heated region' captures only non-heating end uses without residual heating-related electricity consumption. No verification, sensitivity analysis, or discussion of potential contamination (e.g., from electric resistance or heat pumps) is provided, which is load-bearing for the separation task and the direct comparison to ResStock.

    Authors: The manuscript explicitly selects a predominantly gas-heated region to isolate non-heating electricity end uses and discusses this choice as central to the separation of heating and non-heating demands. We agree that the abstract itself provides limited elaboration on verification or sensitivity to residual electric heating. In the revised version, we will expand the relevant sections to include additional discussion of potential contamination risks, regional heating characteristics, and any robustness checks supporting the assumption. revision: yes

Circularity Check

0 steps flagged

No circularity; claims rest on external benchmark comparison

full rationale

The provided abstract presents empirical backtest results (RMSE and WIS improvements versus NREL's ResStock) as direct numerical comparisons to an independent external model. The training-data assumption about gas-heated regions is stated explicitly as a premise rather than derived from any equation, fitted parameter, or self-citation chain. No derivation steps, equations, or load-bearing self-references appear that would reduce the reported performance gains to tautological inputs by construction. The validation therefore remains self-contained against an outside standard.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities. The approach appears to rely on standard probabilistic deep learning techniques and publicly referenced data sources whose details are not provided.

pith-pipeline@v0.9.0 · 5750 in / 1140 out tokens · 46439 ms · 2026-05-19T12:25:18.030119+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages

  1. [1]

    The long-term strategy of the united states: Pathways to net-zero greenhouse gas emissions by 2050,

    United States Government, “The long-term strategy of the united states: Pathways to net-zero greenhouse gas emissions by 2050,” 2021

    work page 2050

  2. [2]

    Decarbonizing the US Economy by 2050: A National Blueprint for the Buildings Sector,

    J. Langevin, E. Wilson, C. Snyder, R. Narayanamurthy, J. Miller, K. Kaplan, M. Reiner, R. Risser, M. Mahoney, J. Geyer, et al., “Decarbonizing the US Economy by 2050: A National Blueprint for the Buildings Sector,” 2024

    work page 2050

  3. [3]

    Sources of greenhouse gas emissions,

    United States Environmental Protection Agency, “Sources of greenhouse gas emissions,” 2025

    work page 2025

  4. [4]

    2020 Residential U.S. Heat Pump Market Update,

    Atlas Buildings Hub, “2020 Residential U.S. Heat Pump Market Update,” 2023

    work page 2020

  5. [5]

    Electricity load implications of space heating decarbonization pathways,

    M. Waite and V . Modi, “Electricity load implications of space heating decarbonization pathways,” Joule, vol. 4, no. 2, pp. 376– 394, 2020

    work page 2020

  6. [6]

    End-use load profiles for the u.s. building stock: Methodology and results of model calibration, validation, and uncertainty quantification,

    U.S. Department of Energy, “End-use load profiles for the u.s. building stock: Methodology and results of model calibration, validation, and uncertainty quantification,” tech. rep., Department of Energy, 2022

    work page 2022

  7. [7]

    ResStock 2024.2 Re- lease: Technical Documentation,

    National Renewable Energy Laboratory, “ResStock 2024.2 Re- lease: Technical Documentation,” tech. rep., National Renewable Energy Laboratory, 2024

    work page 2024

  8. [8]

    Comstock reference documentation: 2024 release 1,

    National Renewable Energy Laboratory, “Comstock reference documentation: 2024 release 1,” tech. rep., National Renewable Energy Laboratory, 2024

    work page 2024

  9. [9]

    Cost-effective planning of decarbonized power-gas infrastructure to meet the challenges of heating electrification,

    R. Khorramfar, M. Santoni-Colvin, S. Amin, L. K. Norford, A. Botterud, and D. Mallapragada, “Cost-effective planning of decarbonized power-gas infrastructure to meet the challenges of heating electrification,” 2023

    work page 2023

  10. [10]

    Elec- trification impacts study part i: Bottom-up load forecasting and system-level electrification impacts cost estimates,

    California Public Utilities Commission, Energy Division, “Elec- trification impacts study part i: Bottom-up load forecasting and system-level electrification impacts cost estimates,” tech. rep., California Public Utilities Commission, San Francisco, Califor- nia, 2023

    work page 2023

  11. [11]

    Model america–data and models of every us building,

    J. New, M. Adams, A. Berres, B. Bass, and N. Clinton, “Model america–data and models of every us building,” tech. rep., Oak Ridge National Laboratory, 2021

    work page 2021

  12. [12]

    Sustainable cities,

    ASCR Discovery, “Sustainable cities,” ASCR Discovery, Jan. 2022

    work page 2022

  13. [13]

    Automatic building energy modeling (autobem),

    Oak Ridge National Laboratory, “Automatic building energy modeling (autobem),” 2024

    work page 2024

  14. [14]

    2023 long-term load forecasting workshop,

    Energy Systems Integration Group (ESIG), “2023 long-term load forecasting workshop,” 2023. Avail- able at:https://www.esig.energy/event/ 2023-long-term-load-forecasting-workshop/

    work page 2023

  15. [15]

    Degree days - energy explained,

    U.S. Energy Information Administration (EIA), “Degree days - energy explained,” n.d. Available at:https://www. eia.gov/energyexplained/units-and-calculators/ degree-days.php

    work page

  16. [16]

    Atlanta, GA: ASHRAE, 2021

    American Society of Heating, Refrigerating and Air- Conditioning Engineers, ASHRAE Handbook – Fundamentals. Atlanta, GA: ASHRAE, 2021

    work page 2021

  17. [17]

    A global model of hourly space heating and cooling demand at multiple spatial scales,

    I. Staffell, S. Pfenninger, and N. Johnson, “A global model of hourly space heating and cooling demand at multiple spatial scales,”Nature Energy, vol. 8, no. 12, pp. 1328–1344, 2023

    work page 2023

  18. [18]

    Building heat demand forecasting by training a common ma- chine learning model with physics-based simulator,

    L. Kannari, J. Kiljander, K. Piira, J. Piippo, and P. Koponen, “Building heat demand forecasting by training a common ma- chine learning model with physics-based simulator,”Forecasting, vol. 3, no. 2, pp. 290–302, 2021

    work page 2021

  19. [19]

    Improved day ahead heating demand forecasting by online correction methods,

    F. B ¨unning, P. Heer, R. S. Smith, and J. Lygeros, “Improved day ahead heating demand forecasting by online correction methods,” Energy and Buildings, vol. 211, p. 109821, 2020

    work page 2020

  20. [20]

    Machine-learning-based multi-step heat demand forecasting in a district heating system,

    P. Poto ˇcnik, P. ˇSkerl, and E. Govekar, “Machine-learning-based multi-step heat demand forecasting in a district heating system,” Energy and Buildings, vol. 233, p. 110673, 2021

    work page 2021

  21. [21]

    Short-term forecasting of heat demand of buildings for efficient and optimal energy man- agement based on integrated machine learning models,

    A. T. Eseye and M. Lehtonen, “Short-term forecasting of heat demand of buildings for efficient and optimal energy man- agement based on integrated machine learning models,” IEEE Transactions on Industrial Informatics, vol. 16, no. 12, pp. 7743– 15 7755, 2020

    work page 2020

  22. [22]

    Machine learning ap- proaches for estimating commercial building energy consump- tion,

    C. Robinson, B. Dilkina, J. Hubbs, W. Zhang, S. Guhathakurta, M. A. Brown, and R. M. Pendyala, “Machine learning ap- proaches for estimating commercial building energy consump- tion,”Applied Energy, vol. 208, pp. 889–904, 2017

    work page 2017

  23. [23]

    Predictive modeling for us commercial building energy use: A comparison of existing statistical and machine learning algorithms using cbecs micro- data,

    H. Deng, D. Fannon, and M. J. Eckelman, “Predictive modeling for us commercial building energy use: A comparison of existing statistical and machine learning algorithms using cbecs micro- data,”Energy and Buildings, vol. 163, pp. 34–43, 2018

    work page 2018

  24. [24]

    Machine learning for early stage building energy prediction: Increment and enrich- ment,

    M. M. Singh, S. Singaravel, and P. Geyer, “Machine learning for early stage building energy prediction: Increment and enrich- ment,”Applied Energy, vol. 304, p. 117787, 2021

    work page 2021

  25. [25]

    Building energy consumption prediction for residential build- ings using deep learning and other machine learning techniques,

    R. Olu-Ajayi, H. Alaka, I. Sulaimon, F. Sunmola, and S. Ajayi, “Building energy consumption prediction for residential build- ings using deep learning and other machine learning techniques,” Journal of Building Engineering, vol. 45, p. 103406, 2022

    work page 2022

  26. [26]

    Predicting energy consumption for residential buildings using ann through parametric modeling,

    E. Elbeltagi and H. Wefki, “Predicting energy consumption for residential buildings using ann through parametric modeling,” Energy Reports, vol. 7, pp. 2534–2545, 2021

    work page 2021

  27. [27]

    An ann-based fast building energy consumption prediction method for complex architectural form at the early design stage,

    Z. Li, J. Dai, H. Chen, and B. Lin, “An ann-based fast building energy consumption prediction method for complex architectural form at the early design stage,” vol. 12, pp. 665–681, 2019

    work page 2019

  28. [28]

    En- hancing hourly heat demand prediction through artificial neural networks: A national level case study,

    M. Zhang, M.-A. Millar, S. Chen, Y . Ren, Z. Yu, and J. Yu, “En- hancing hourly heat demand prediction through artificial neural networks: A national level case study,” Energy and AI, vol. 15, p. 100315, 2024

    work page 2024

  29. [29]

    A comparison of models for forecasting the residential natural gas demand of an urban area,

    P. Poto ˇcnik, J. ˇSilc, G. Papa, et al., “A comparison of models for forecasting the residential natural gas demand of an urban area,” Energy, vol. 167, pp. 511–522, 2019

    work page 2019

  30. [30]

    Natural gas consumption forecasting based on the variability of external meteorological factors using machine learning algorithms,

    W. Panek and T. Włodek, “Natural gas consumption forecasting based on the variability of external meteorological factors using machine learning algorithms,” Energies, vol. 15, no. 1, p. 348, 2022

    work page 2022

  31. [31]

    Data-driven ma- chine learning model performance of real annual natural gas con- sumption in residential buildings,

    M. Van Hove, M. Delghust, and J. Laverge, “Data-driven ma- chine learning model performance of real annual natural gas con- sumption in residential buildings,” in2022 Building Performance Analysis Conference and SimBuild, pp. 181–188, 2022

    work page 2022

  32. [32]

    A hy- brid hourly natural gas demand forecasting method based on the integration of wavelet transform and enhanced deep-rnn model,

    H. Su, E. Zio, J. Zhang, M. Xu, X. Li, and Z. Zhang, “A hy- brid hourly natural gas demand forecasting method based on the integration of wavelet transform and enhanced deep-rnn model,” Energy, vol. 178, pp. 585–597, 2019

    work page 2019

  33. [33]

    Natural gas consump- tion forecast with mars and cmars models for residential users,

    A. ¨Ozmen, Y . Yılmaz, and G.-W. Weber, “Natural gas consump- tion forecast with mars and cmars models for residential users,” Energy Economics, vol. 70, pp. 357–381, 2018

    work page 2018

  34. [34]

    Overture Maps Dataset,

    Overture Maps Foundation, “Overture Maps Dataset,” 2025

    work page 2025

  35. [35]

    World settlement footprint 3d-a first three-dimensional survey of the global building stock,

    T. Esch, E. Brzoska, S. Dech, B. Leutner, D. Palacios-Lopez, A. Metz-Marconcini, M. Marconcini, A. Roth, and J. Zeidler, “World settlement footprint 3d-a first three-dimensional survey of the global building stock,” Remote sensing of environment, vol. 270, p. 112877, 2022

    work page 2022

  36. [36]

    Global land use/land cover with sentinel 2 and deep learning,

    K. Karra, C. Kontgis, Z. Statman-Weil, J. C. Mazzariello, M. Mathis, and S. P. Brumby, “Global land use/land cover with sentinel 2 and deep learning,” in 2021 IEEE international geoscience and remote sensing symposium IGARSS, pp. 4704– 4707, IEEE, 2021

    work page 2021

  37. [37]

    Annual time series of global viirs nighttime lights derived from monthly averages: 2012 to 2019,

    C. D. Elvidge, M. Zhizhin, T. Ghosh, F.-C. Hsu, and J. Taneja, “Annual time series of global viirs nighttime lights derived from monthly averages: 2012 to 2019,”Remote Sensing, vol. 13, no. 5, p. 922, 2021

    work page 2012

  38. [38]

    Speedtest by Ookla Global Fixed and Mobile Net- work Performance Map Tiles,

    Ookla, “Speedtest by Ookla Global Fixed and Mobile Net- work Performance Map Tiles,” 2022.https://github.com/ teamookla/ookla-open-data

    work page 2022

  39. [39]

    Era5 hourly data on single levels from 1940 to present,

    C. D. S. Copernicus Climate Change Service, “Era5 hourly data on single levels from 1940 to present,” Copernicus Climate Change Service (C3S) Climate Data Store (CDS), 2023

    work page 1940

  40. [40]

    Thefuzz: Fuzzy string matching in python

    SeatGeek, “Thefuzz: Fuzzy string matching in python.”https: //github.com/seatgeek/thefuzz, 2021

    work page 2021

  41. [41]

    S. J. Lee, Multimodal Data Fusion for Estimating Electricity Access and Demand. PhD thesis, Massachusetts Institute of Tech- nology, 2023

    work page 2023

  42. [42]

    Weather downloader

    Oikolab, “Weather downloader.” Online, 2024. Available: https://weatherdownloader.oikolab.com/app

    work page 2024

  43. [43]

    Basic tutorial

    NREL, “Basic tutorial.” Online, 2024. Available: https://resstock.readthedocs.io/en/v3.3.0/basic_ tutorial/index.html

    work page 2024

  44. [44]

    2020 Public Use Microdata Areas (PUMA) Shapefile,

    IPUMS USA, “2020 Public Use Microdata Areas (PUMA) Shapefile,” 2024. Available at:https://usa.ipums.org/ usa/volii/pumas20.shtml. 16

    work page 2020