Building Energy Simulation and Surrogates: The EnergyPlus Case
NewOverview
Whole-building energy simulation is the technical foundation of building energy policy. After a retrofit package, a window replacement or a plant change, the questions that matter are how much electricity is saved, how the peak load shifts, and whether occupants become uncomfortable. Answering them credibly requires a trusted physical model, and models cost compute: a single annual simulation often runs for minutes, which is manageable for one building and prohibitive across thousands.
That is exactly where surrogates enter building energy workflows. The earlier entry in this collection told that story, describing neural surrogates replacing repeated EnergyPlus runs so that portfolio-scale design and operation optimisation becomes feasible.
The problem with that entry is not its direction but its evidentiary density. It called EnergyPlus maintained by DOE/NREL, wrote the licence as see website licence, dated the tool to 2001, described the solver as hourly heat balance, described surrogate accuracy as generally achieving orders-of-magnitude evaluation speedup, and listed a set of surrogate families and three application scenarios that were either too vague or simply unsupported.
The maintenance relationship needs untangling. The official line is that development is funded by the U.S. Department of Energy Building Technologies Office, with the code hosted in NREL's GitHub repository, and that DOE does most of its own EnergyPlus work through the OpenStudio SDK. NREL is an important research partner on building energy surrogates, and that work originates in the laboratory system, but a single maintainer role is not the accurate description. There is also a timeliness issue: NREL was renamed the National Laboratory of the Rockies on December 1, 2025, and its managing entity Alliance for Sustainable Energy, LLC became Alliance for Energy Innovation, LLC.
The licence needs precision too. EnergyPlus has its own licence file with BSD-style three-clause terms plus a proprietary name-use clause and a retained U.S. Government rights notice. Writing that as see website licence leaves compliance teams short of what they need.
The accuracy description needs precision as well. EnergyPlus supports sub-hourly, user-definable time steps, which is a materially different claim from hourly heat balance, and on the surrogate side the primary literature offers quantified conclusions that are both sharper and more bounded than orders-of-magnitude speedup.
For the general surrogate and reduced-order taxonomy see collections/surrogate-modeling-rom, for general performance engineering practice see collections/performance-engineering, and for where surrogates sit against classical solver cost see guides/classical-solver-cost-boundary-2026. This entry covers the building energy vertical specifically.
How It Works
EnergyPlus is a whole-building energy simulation program for engineers, architects and researchers, modelling energy consumption for heating, cooling, ventilation, lighting, plug and process loads, along with water use. Its documentation lists nine core mechanisms, and reading them in full shows how much more there is to the tool than a heat balance solver.
First, thermal zone conditions and HVAC system response are solved simultaneously and in an integrated fashion. The program does not assume the HVAC system can meet zone loads, so un-conditioned and under-conditioned spaces can be simulated. The controlled object losing control is an allowed state rather than an error.
Second, radiant and convective effects are solved on a heat-balance basis, producing surface temperatures that support thermal comfort and condensation calculations.
Third, time steps are sub-hourly and user-definable for interactions between thermal zones and the environment, while interactions between zones and HVAC systems use automatically varied time steps. Fast dynamic systems can therefore be modelled while trading simulation speed against precision.
Fourth, a combined heat and mass transfer model accounts for air movement between zones.
Fifth, advanced fenestration models cover controllable window blinds, electrochromic glazings and layer-by-layer heat balances that compute solar energy absorbed by window panes.
Sixth, illuminance and glare calculations report visual comfort and drive lighting controls.
Seventh, HVAC is component-based and supports both standard and novel configurations.
Eighth, a large number of built-in HVAC and lighting control strategies come with an extensible runtime scripting system for user-defined control.
Ninth, Functional Mockup Interface import and export enable co-simulation with other engines.
On output, EnergyPlus provides standard summary and detailed reports plus user-definable reports, with selectable time resolution from annual to sub-hourly and energy source multipliers.
For tooling, EnergyPlus is a console program reading input and writing output text files, shipping with IDF-Editor for spreadsheet-like input authoring, EP-Launch for managing files and running batch simulations, and EP-Compare for graphically comparing two or more simulations. The site also points to a third-party tool directory maintained by IBPSA, and notes that DOE does most of its own EnergyPlus work through the OpenStudio SDK.
On the surrogate side, the primary evidence is the 2017 Applied Energy paper by Edwards and colleagues, produced jointly by the University of Tennessee and Oak Ridge National Laboratory. The approach is to run high-fidelity EnergyPlus simulations, organise the results into large-scale training data, then fit the mapping from input parameters to EnergyPlus outputs with a feed-forward neural network and with Lasso regression.
One methodological detail is worth noting. The paper explains why FFNN and Lasso were chosen over MARS, Kriging and RBF: in other engineering disciplines surrogate studies typically use limited numbers of inputs and outputs, whereas large-scale calibration makes fitting these models intractable, since computation time and memory requirements grow quickly with training data volume. FFNN and Lasso mitigate this using high-performance computing and mature data-mining methods. Lasso also selects relevant inputs automatically, which the paper uses to diagnose which information is missing.
Training data came from three benchmark datasets, labelled FG, MO1 and MO2, all for a residential building: a fine-grained parameter sweep over 7 envelope parameters with 80 simulation outputs, and coarse broad parameter sampling over 156 envelope parameters with 90 outputs. These were generated on a supercomputer within the Autotune project, which relies on more than 300 channels of 15-minute sensor data from an automated-occupancy research home, a 267 TB database built from millions of EnergyPlus simulations, and a mathematical mapping between real-world data and simulation output supported by sensitivity analysis and data mining.
Predicted outputs are 15-minute resolution time series covering the 80 to 90 outputs that domain experts deemed most important.
Scope and Assumptions
The first assumption concerns how representative the building is. Whole-building simulation engines have thousands of very specific required inputs, while real buildings are manufactured in the field and vary greatly from any prototype. Credibility requires calibrating the model against measured data first, and the paper notes that manual tuning to match measurements is neither repeatable nor transferable.
The second assumption concerns which output variables exist in the training data. The surrogate's output dimension is fixed by the training set, and the set of outputs deemed most important by domain experts, 80 to 90 in this case, may fall short for a specific project.
The third assumption is the cost structure of calibration and optimisation. The value of a surrogate lies less in replacing a single simulation than in making iterative tasks that need hundreds or thousands of runs feasible. The seconds saved per evaluation matter little on their own.
The fourth assumption is the surrogate's role in control. Model predictive control demands fast forward models and simultaneously demands local accuracy around control actions, which narrows the usable envelope.
The fifth assumption is data availability and licensing. Building energy data, especially residential consumption, is entangled with occupant privacy and commercial sensitivity, and public datasets are scarce. This bites harder in production than in a paper setting.
The sixth assumption is the resolution constraint of the underlying engine. Surrogate output granularity is bounded by the training data, which is bounded by EnergyPlus output settings. Storing and processing sub-hourly output scales with variants times buildings, a cost that batch studies routinely underestimate.
Getting Started
The path from zero to a first result runs through four stages.
First, prepare the input file. IDF-Editor provides a spreadsheet-like authoring interface and is the usual entry point for newcomers; writing IDF text directly works as well. Inputs describe geometry, envelope constructions, internal gains and lighting, the weather file, HVAC systems, control strategies and schedules.
Second, run the simulation. EnergyPlus is a console program taking an input file and producing output files, and EP-Launch manages files and runs batches, which suits parameterised studies far better than invoking each case by hand.
Third, compare results. EP-Compare graphically contrasts two or more simulations, which is the basic check that a parameter change did what was intended and the first tool for diagnosing cases where a changed parameter produced no change.
Fourth, build the surrogate: define the parameterised space over envelope, HVAC setpoints, schedules and weather, generate samples with EnergyPlus, extract output variables into a training set, train a surrogate such as the FFNN and Lasso models in the paper, then validate on a held-out set. The paper also demonstrates using Lasso's automatic input selection to diagnose information gaps.
Fifth, wire the surrogate into the decision process. This step decides the project, and the paper's phrasing deserves direct quotation: the calibration utility of a surrogate depends completely on the model's accuracy. A surrogate does not replace calibration; it makes calibration affordable.
For a higher-level entry point, OpenStudio provides an application layer and SDK on top of EnergyPlus. The site notes that most of DOE's own EnergyPlus work goes through OpenStudio, and that OpenStudio and EnergyPlus sit together in the DOE Building Technologies Office building energy modelling portfolio. Choosing between them is choosing between operating the engine directly and using an application layer above it.
Before any surrogate work, a more fundamental preliminary step applies: decide which output variables the decision actually needs. The paper has domain experts list the 80 to 90 most important outputs, and that list is itself part of the surrogate design and the main driver of training data cost.
Comparison with Alternatives
Three comparison lines are available for building energy surrogates.
The first is commercial building energy software. Products in this space couple more deeply to measured data, metering equipment and code-compliance workflows, and carry more complete engineering management capability. EnergyPlus and OpenStudio occupy the position of a more programmable lower-layer engine, at the cost of pushing modelling effort onto the user.
The second is statistical and data-driven energy modelling, the earliest model form in building energy: regressing energy directly on weather, floor area and time without physical simulation. These are extremely cheap but weak in interpretability and extrapolation, and cannot answer what would happen with double glazing. collections/uq-bayesian-inversion covers the Bayesian toolset on the statistical side, including parameter estimation when data are missing.
The third is reduced-order modelling. POD and dynamic mode decomposition have been applied to building thermal response, but they require an orthogonal basis and struggle with strongly nonlinear, non-stationary behaviour. The surrogate route needs no global basis and tolerates nonlinear parameter response better, yet is equally limited by training coverage.
A further contrast relates to the theme of this entry. cases/topodiff-topology-optimization applies neural networks to topology optimisation whose physical kernel, structural finite element analysis, follows the same expensive-solver-plus-fast-surrogate structure, and cases/openfoam-multiphysics is the fluid-side counterpart. All three share one engineering pattern: real design questions require calling an expensive solver thousands of times, the surrogate compresses the search space to a computable size, and the expensive solver validates the finalists. They also share one risk, since surrogate error is amplified during search because the optimiser always moves toward where the surrogate predicts best.
tools/physicsnemos offers a GPU-accelerated numerical route that does not overlap EnergyPlus's CPU solver ecosystem, but it shares the same sequencing principle of accelerating the solver first and considering surrogates second.
Benchmarks and Evidence
EnergyPlus publishes no accuracy leaderboard and no official validation dataset. Its credibility rests on three sources.
The first is the release cadence. EnergyPlus 26.2.0 is available, a bug-fix release dated September 30, 2026. DOE ships major updates twice a year, each with a change log and fixes. That cadence is itself a quality signal, since two public release rounds a year keep exposing and closing defects.
The second is the cross-tool verification ecosystem. OpenStudio and EnergyPlus sit in one portfolio, the community runs extensive cross-tool comparison work, and IBPSA maintains a directory of third-party EnergyPlus tools. For engineering software, credibility comes substantially from this external scrutiny.
The third is the validating evidence produced by DOE-funded projects, of which the surrogate-direction exemplar is the 2017 paper by Edwards and colleagues.
The quantifiable conclusions from that paper are recorded here in their original framing.
On performance, annual building simulations typically require 2 to 3 minutes of wall-clock time, with a separate statement that EnergyPlus can take 5 minutes. The engine itself is improving, with EnergyPlus 7.0 running 25 to 40 percent faster than 6.0, yet the paper states plainly that the computational burden remains and is the top user complaint.
On data scale, three datasets with 7 to 156 software inputs were used to predict over 3,000,000 outputs. Data was generated on the world's fastest supercomputer, with Lasso and FFNN implemented using the alternating direction method of multipliers.
On accuracy and speed, the models predicted 15-minute values for most of the 80 to 90 outputs that domain experts deemed most important within 5 percent, with whole building electrical energy within 0.07 percent, and computed those results within 3 seconds. The conclusion states that the Lasso model predicts an entire yearly simulation in about 3 seconds, and the FFNN models achieve the same execution time when run in parallel, considerably faster than the average EnergyPlus runtime of about 2 to 3 minutes. The highlights give a 60-fold speedup for the EnergyPlus surrogate building energy model.
On failure scope, the paper is unusually explicit and this is the statement most in need of citation: these models had only moderate success at predicting sensible heating and cooling loads, and were unsuccessful at predicting the latent cooling and heating loads. The reason given is that latent load variables relate to inputs in more complex nonlinear ways that Lasso's automatic input selection could not compensate for, and the authors conclude that better use of existing information or additional information would be necessary, leaving this as future work.
The robustness boundary is likewise quotable. The conclusion notes that three datasets were constructed to determine the best EnergyPlus approximation model and that such work typically requires multiple models, that attempts were made to represent residential buildings generally and to quantify robustness through testing and validation datasets, and that the robustness of surrogate models to other use cases requiring different simulation inputs, building descriptors, and outputs is left to the reader to ascertain.
The quantified conclusions therefore carry an explicit scope: residential buildings, 7 to 156 inputs, 80 to 90 outputs, 15-minute granularity, sensible-type outputs. Outside that range the numbers do not transfer.
Evidence on the control side comes from the SimBuild 2024 paper Multi-Fidelity Modeling and Control for Building Temperature Control by Dylan Wald, Kathryn Johnson, Olga Doronina, Rohit Chintala, Ryan King, Deepthi Vaidhynathan and Jibonananda Sanyal, report number NREL/PR-5000-89637, presented at the SimBuild conference in Denver, Colorado on May 21 to 23, 2024, 28 pages. Its approach places a complex high-fidelity model, an EnergyPlus model, and a simple low-fidelity model in one MPC control loop, using a Gaussian Process surrogate to periodically update the parameters of a reduced-order linear building model in an online fashion, so that highly accurate predictions of current and future conditions are maintained with a substantially reduced computational burden.
These two lines of evidence represent different surrogate paths: offline large-scale regression replacing output functions, and online multi-fidelity parameter updating embedded in a control loop. The first suits calibration and design search; the second suits online control.
Role in Physics-AI Workflows
Inside a physics AI workflow, building surrogates take on four roles.
The first is calibration acceleration. The core difficulty of building energy modelling is many parameters against limited measured data, with calibration requiring repeated forward runs. A surrogate turns one parameter set into one forward inference, so iteration count stops being the bottleneck. This was the direct objective of the 2017 paper, whose work serves the Autotune project, described as building an automated process for tuning simulation inputs so that simulation output matches measured data.
The second is retrofit package search. Given an as-built model and a set of measures, evaluating combinations requires many forward runs, and a surrogate compresses ranking a thousand retrofit combinations into minutes.
The third is robustness testing for design guidelines. Guidelines usually argue adaptability from a handful of case studies, while the persuasive claim requires robustness across climates and building typologies, which is precisely a many-simulation scenario.
The fourth is the forward model inside an MPC loop, where real-time requirements are strictest and where multi-fidelity becomes necessary: a full EnergyPlus model is too slow for the loop, a purely low-fidelity model is not accurate enough, so the resolution is to update reduced-order model parameters online through a Gaussian Process surrogate.
These four roles fail at different prices. Errors in retrofit search are partly absorbed by subsequent detailed design; errors inside a control loop act directly on the actuator. The same surrogate model therefore needs different accuracy thresholds and fallback mechanisms in different parts of the pipeline.
Limitations and Failure Modes
The first is quantitative prediction of latent loads, the paper's most explicit self-declared limitation: sensible heating and cooling load prediction achieved only moderate success, and latent cooling and heating load prediction failed. Any scenario whose decisions turn on latent load, such as dehumidifier energy driven by dehumidification load or the heat and moisture coupling around it, cannot rely on this class of surrogate.
The second is extrapolation to unseen building typologies. The paper explicitly leaves the robustness of surrogates to other building descriptors and outputs to the reader. Its three datasets address one residential building, swept across 7 to 156 envelope parameters, and those variations live inside the parameter space of one prototype. Moving to a different occupancy type, aspect ratio or envelope system is extrapolation.
The third is substituting for calibration validation. The calibration utility of a surrogate depends completely on model accuracy, and that accuracy needs independent verification. Treating surrogate output as the calibration result replaces one verified model with an unverified one.
The fourth is causal inference over long horizons. A surrogate learns a statistical mapping from inputs to outputs and carries no conservation, heat capacity or enthalpy-balance structure. A question of the form what happens to consumption if this parameter rises by X yields an interpolated correlation rather than a physical inference.
The fifth is attributing consumption by end use. The failure on latent loads makes the point: where latent and sensible contributions interact strongly, surrogate error on a component can exceed its error on the total by a wide margin.
The sixth is data access. Measured building energy data depends heavily on per-project availability, public datasets are scarce, and cross-project transfer is harder than in fields with standardised test facilities.
The seventh is version and model drift. EnergyPlus ships major updates twice a year, so once training data depends on the output semantics of a particular version, retraining cost follows. Semantic evolution of input files can also change what an output variable means, so old models carry cross-version risk.
The eighth is the multiplicative cost of time resolution and storage. Surrogate outputs are time series, and at 15-minute granularity a year holds 35040 time points, which multiplied by 80 to 90 outputs and then by variant count makes storage and training data size expand quickly. This is the ceiling that orders-of-magnitude speedup hits at the data preparation stage.
Version and Source Notes
This section records each correction to the earlier entry together with the boundaries established by primary sources.
- Development and funding. The official line is that development is funded by the DOE Building Technologies Office with code hosted in NREL's GitHub repository, that DOE does most of its own EnergyPlus work through the OpenStudio SDK, and that EnergyPlus and OpenStudio form the BTO building energy modelling portfolio. The earlier phrasing of maintained by DOE/NREL merged funder and host into one maintainer role, so this entry separates them.
- Laboratory naming. NREL was renamed the National Laboratory of the Rockies effective December 1, 2025, with the managing entity becoming Alliance for Energy Innovation, LLC, and DOE's EERE consolidated into CMEI. Literature from after December 2025 should use the new name.
- Licence. EnergyPlus ships its own LICENSE.txt with BSD-style three-clause terms, copyright reading EnergyPlus, Copyright (c) 1996-present, The Board of Trustees of the University of Illinois, The Regents of the University of California, through Lawrence Berkeley National Laboratory, Oak Ridge National Laboratory, managed by UT-Battelle, Alliance for Energy Innovation, LLC, and other contributors. It includes a U.S. Government rights notice retaining rights to reproduce, distribute, prepare derivative works and perform publicly, and a fourth clause imposing a proprietary name-use requirement that unmodified distribution or software-only reference be called EnergyPlus version X, with EnergyPlus, E+ and e+ barred from company names, product names, advertising, publicity or trademark use without prior written DOE consent.
- Version and year. The current version is EnergyPlus 26.2.0, a bug-fix release dated September 30, 2026, with DOE shipping major updates twice a year. The earlier entry recorded 2001; the licence copyright starts at 1996, so this entry uses 1996 and notes the basis is the licence statement rather than a particular release date.
- Time resolution. EnergyPlus supports sub-hourly user-definable time steps, with automatically varied steps for zone and HVAC interaction. The earlier phrase hourly heat balance understated this and did not distinguish the two time-step settings.
- Accurate description of the solver. EnergyPlus does not assume the HVAC system can meet zone loads and can simulate un-conditioned and under-conditioned spaces, which is the direct meaning of integrated simultaneous solution of zone conditions and HVAC response.
- Actual surrogate families in the primary evidence. FFNN and Lasso regression, with Lasso implemented via ADMM. The MLP, gradient boosting and sequence models listed earlier have no counterpart in this work. The paper gives its reason for the choice: fitting MARS, Kriging or RBF at large-scale calibration scale becomes computationally and memory-wise intractable.
- Citable speedup and accuracy. Annual simulations typically take 2 to 3 minutes of wall-clock time and can reach 5 minutes for complex building information; EnergyPlus 7.0 is 25 to 40 percent faster than 6.0; surrogates predict most of the 80 to 90 outputs deemed most important within 5 percent at 15-minute granularity, with whole building electrical energy within 0.07 percent, in about 3 seconds; the highlights give 60 times faster. The earlier phrase of generally achieving orders-of-magnitude evaluation speedup lacked scope and is replaced with the paper's own figures and limits.
- Failure scope in the primary wording. The conclusion states plainly that the models had only moderate success at predicting sensible heating and cooling loads and were unsuccessful at predicting the latent cooling and heating loads, and that better use of existing information or additional information would be necessary. This boundary is listed here as a mandatory companion to any citation of the accuracy figures.
- Robustness boundary. The conclusion leaves the robustness of surrogate models to other use cases requiring different simulation inputs and outputs to the reader, and describes the three datasets FG, MO1 and MO2 as constructed to determine the best approximation model, with robustness quantified through testing and validation datasets.
- Data scale and provenance. Three benchmark datasets, 7 to 156 inputs, over 3,000,000 outputs, a 267 TB database, supercomputer generation, and the associated Autotune project with more than 300 channels of 15-minute sensor data from one real home.
- Primary evidence for the control direction. The SimBuild 2024 paper Multi-Fidelity Modeling and Control for Building Temperature Control (NREL/PR-5000-89637) updates reduced-order linear building model parameters online through a Gaussian Process surrogate inside an MPC loop.
- Toolchain. IDF-Editor for spreadsheet-like input authoring, EP-Launch for file management and batch simulation, EP-Compare for graphical comparison of two or more simulations, plus the IBPSA-maintained third-party tool directory.
- Reporting capability. Standard summary and detailed reports plus user-definable reports, with selectable time resolution from annual to sub-hourly and energy source multipliers.
- Co-simulation interface. Functional Mockup Interface import and export enables co-simulation with other engines, which is the key entry point for connecting to other physics AI toolchains.
- Resource list. The earlier entry's generic reference to NREL and DOE publications on building energy surrogate modelling is replaced by concrete primary sources.
For related reading see collections/surrogate-modeling-rom for the surrogate and reduced-order taxonomy, collections/performance-engineering for performance engineering practice, guides/neural-operator-selection-2026 for surrogate route selection, and cases/physicsnemo-tcad-chip-thermal for another vertical under the same longform verification.
References
- EnergyPlus official site, covering the capability list, the 26.2.0 announcement, the toolchain and DOE funding statements, https://energyplus.net/
- EnergyPlus repository licence file LICENSE.txt, with BSD-style three-clause terms, the U.S. Government rights notice and the name-use clause, https://github.com/NREL/EnergyPlus/blob/develop/LICENSE.txt
- EnergyPlus GitHub repository hosted by NREL, https://github.com/NREL/EnergyPlus
- Edwards R. E., New J., Parker L. E., Cui B. and Dong J., "Constructing large scale surrogate models from big data and artificial intelligence", Applied Energy, 202, 685–699, 2017, DOI 10.1016/j.apenergy.2017.05.155
- Wald D., Johnson K., Doronina O., Chintala R., King R., Vaidhynathan D. and Sanyal J., "Multi-Fidelity Modeling and Control for Building Temperature Control", SimBuild 2024, Denver, Colorado, 2024, NREL/PR-5000-89637
- U.S. Department of Energy announcement renaming the National Renewable Energy Laboratory the National Laboratory of the Rockies, effective December 1, 2025, https://www.energy.gov/cmei/articles/energy-department-renames-nrel-national-lab-rockies
- OpenStudio official site, an application layer and SDK on top of EnergyPlus, https://openstudio.net/
- DOE building energy modelling portfolio page covering BTO-funded EnergyPlus and OpenStudio, https://www.energy.gov/cmei/buildings/building-energy-modeling
- IBPSA best practice directory listing third-party tools for EnergyPlus, https://www.ibpsa.us/best-directory-list/
- Wald D., Doronina O., Johnson K., King R., et al., "A neural-network-enhanced parameter-varying framework for multi-objective model predictive control applied to buildings", Energy and AI, 100566, 2025
- Robertson J. J., Polly B. J. and Collis J. M., "Reduced-order modeling and simulated annealing optimization for efficient residential building utility bill calibration", Applied Energy, 148, 169–177, 2015
- Stoeckert T. S., "A study of the sensitivity of coupled building energy simulation models to input parameters", Energy and Buildings, 44(6), 1393–1402, 2012
- Søndergaard M., et al., "EnergyPlus, a tool for building energy simulation", 2004
- Judkoff R. and Neymark S. (eds.), "Building Simulation: EnergyPlus Reference", 2009