Lessons learned from field demonstrations of model predictive control and reinforcement learning for residential and commercial HVAC: A review

• The paper reviews field work on model predictive control and reinforcement learning for heating, ventilation, and air conditioning.Source: Highlights · p. 2
• The review draws lessons from 206 tests in residential and commercial buildings.Source: Highlights · p. 2
• The review critically analyzes experiment protocols used in past field work.Source: Highlights · p. 2
• The paper recommends directions for future field work and algorithm development.Source: Highlights · p. 2
• The paper provides an open dataset with paper details and an experiment checklist.Source: Highlights · p. 2

1. Introduction

• MPC optimizes actions using predictive models, while RL can learn policies from real-world or simulation data and DeePC predicts responses directly from data without an explicit system model.These approaches may be useful where model development is costly or HVAC dynamics make mathematical modeling difficult.Source: §1 · p. 4
• Only 4% of 2,892 relevant MPC and RL papers focus on field demonstrations, which can reveal practical benefits and implementation challenges overlooked by simulations.The review notes that field work is constrained by costs, time, data access, and control authority in real buildings.Source: §1 · p. 4, Figure 1 · p. 5

1.1. Contributions of this paper

• The paper presents the first systematic review of all known peer-reviewed MPC and RL field demonstrations for residential and commercial HVAC.Its contributions also include an open-access dataset, critical analysis and guidance on experiment protocols, and future-work recommendations based on 104 studies and the authors’ experience conducting 21 field studies.Source: §1.1 · pp. 5–6

1.2. Organization of this paper

• The paper proceeds from prior reviews and methodology to demonstration demographics, protocol reliability, implementation issues, recommendations, and conclusions.Appendix B provides summary tables of the residential and commercial demonstrations.Source: §1.2 · p. 6

2.1. General HVAC control reviews

• Reviews cover occupant-centric HVAC control integrating real-time occupancy data and comfort feedback for energy efficiency and thermal comfort, alongside optimization-based strategies for performance and decision-making under uncertainty.Source: §2.1 · p. 7

2.2. Reviews of MPC for buildings

• Building-MPC reviews offer implementation frameworks and practical guidance,, while topical work examines commercial deployment, occupancy, thermal storage, renewables, and grid integration,,,,,.Broader reviews also cover energy flexibility, fault detection, and comparisons of MPC variants with RL –.Source: §2.2 · pp. 7–8

2.3. Reviews of RL for buildings

• RL reviews report limited real-building evidence: about 11% of 77 studies used real buildings, while 91% omitted historical states and 77% used value-based algorithms [59].Other reviews noted simulation-heavy evidence, insufficient standardization for comparisons, and gaps in safety guarantees and practical adoption,,.Source: §2.3 · pp. 8–9

2.4. Reviews of MPC and RL field demonstrations

• Field-focused reviews remain few: Sturzenegger, Blum, and Pergantis surveyed commercial or residential MPC demonstrations and identified implementation and cost challenges.Khabbazi et al. reviewed 56 commercial-building field tests and found lower savings in longer-duration and whole-building tests; Saloux found savings estimates vary with benchmarking method, underscoring the need for consistent assessment.Source: §2.4 · pp. 9–10

3.1. Search strategy

• The review searched peer-reviewed journal and conference publications for real-world residential or commercial MPC/RL studies, excluding conference papers with detailed journal versions and non-peer-reviewed reports.Searches used scholarly engines, publisher databases, and keywords concerning field demonstrations, buildings, MPC, RL, and HVAC.Source: §3.1 · p. 10
• The review organized study characteristics into four tables covering 24 residential and 80 commercial papers, including methods, buildings, HVAC systems, durations, objectives, benefits, and benchmarks.A full open-access table is available in.Source: §3.1 · p. 10
• Of 104 reviewed field studies, 45 prompted author outreach; 26 of 32 contacted researchers responded, an 81% response rate, while questions about six papers remained unresolved.For unresolved studies, the reviewers sought missing information in related publications and estimated some parameters themselves.Source: §3.1 · pp. 10–11

3.2. Savings evaluation

• Savings results are counted by field test rather than paper, with duration-weighted averages giving greater weight to longer tests and a focus on energy and cost savings.The review also considers reported benefits such as comfort, emissions, peak demand, and frequency regulation.Source: §3.2 · p. 11

3.3. Statistical methods

• Mann-Whitney U tests compared two-category distributions without assuming Gaussian data, but results were omitted for categories with fewer than 15 samples.The authors note that some comparisons therefore could not be tested rigorously.Source: §3.3 · p. 11

4. Demographics

• The review identified 24 residential papers covering 52 field tests and 80 commercial papers covering 154 tests; Fig. 2 charts cumulative counts from 2005–2025.Source: §4 · p. 11, Figure 2 · p. 12
• Commercial MPC began in 2005 and reached 69 papers and 134 tests, while residential MPC began in 2013 and reached 21 papers and 49 tests.Source: §4 · p. 11
• Commercial RL began in 2006 [73] and resumed gradual growth in 2016 to 11 papers and 20 tests; residential RL began in Leuven, with two papers each reporting one five-day test.Source: §4 · pp. 11–12

4.1. Geography

• Most papers came from North America and Europe, with the United States and Switzerland leading their continents at around 35% and 13%; Asia contributed 12 of 104 papers (12%), versus one each from Africa and South America and two from Oceania.Source: §4.1 · p. 12, Figure 3 · p. 13

4.2. Weather

• Climate-zone distributions count field tests rather than papers because some papers report tests across regions; ASHRAE Standard 90.1 categorizes zones by temperature and humidity.Source: §4.2 · p. 12
• Residential tests were concentrated in cold zones (32), followed by hot (11), mixed (6), and very hot (3); all 52 occurred in humid regions, none in marine or dry locations.Source: §4.2 · pp. 12–13
• Commercial tests were distributed across mixed (49), cold (47), hot (43), and very hot (13) zones; 119 were in humid areas, 17 marine, and 16 dry.The grouped temperature-zone totals were 56 extremely hot and hot, 49 mixed, and 47 cold.Source: §4.2 · p. 13

4.3.1. Type of test building

• 88% of residential tests (46 of 52) and 75% of commercial tests (115 of 154) were conducted at field sites rather than laboratories.Field sites offer more realistic test conditions, while laboratories offer greater reproducibility.Source: §4.3.1 · p. 14

4.3.2. Electricity rate structure

• Residential tests were nearly split between flat rates (27 of 52, or 52%) and dynamic rates (25 of 52, or 48%); commercial tests favored flat rates (121 of 154, or 79%).Rate structures influence the potential for cost savings by load shifting.Source: §4.3.2 · p. 14

4.3.3. Space conditioning mode(s)

• Residential tests focused on heating (27 of 52, 52%), cooling (22, 42%), or both (3, 6%); commercial tests focused on heating (66 of 154, 43%), cooling (71, 46%), or both (17, 11%).Some commercial studies reported only aggregate results when zones heated and cooled simultaneously, making the separate effects indistinguishable.Source: §4.3.3 · p. 15

4.3.4. Thermal mass of building

• 11 of 52 residential and 33 of 154 commercial tests lacked adequate thermal-mass information; reported residential cases were nearly even, while commercial cases more often had moderate-to-high thermal mass (88 versus 33).Thermal mass influences temperature stability and load-shifting effectiveness; building classification relied on available descriptions because reporting was not standardized.Source: §4.3.4 · pp. 15–16

4.3.5. Insulation level of building

• Insulation information was absent for 10 of 52 residential and 29 of 154 commercial tests; reported cases were nearly evenly split between moderate-to-high and low-to-moderate insulation.Moderate-to-high insulation was reported in 21 residential and 70 commercial tests, versus low-to-moderate insulation in 21 residential and 55 commercial tests.Source: §4.3.5 · p. 16

4.3.6. Control architecture

• High-level control was used in 39 of 52 residential tests (75%) and 109 commercial tests (71%), compared with low-level control in 13 (25%) and 45 (29%), respectively.High-level control adjusts supervisory setpoints, whereas low-level control directly actuates HVAC components.Source: §4.3.6 · p. 16

4.4.1. MPC implementation

• Two-thirds of MPC field papers used linear building-dynamics models, usually one- to three-state-per-zone thermal circuits; most used five- to 60-minute steps.Other approaches included autoregressive time-series models, nonlinear models, and simulation software; equipment curves were typically represented by low-degree polynomials.Source: §4.4.1 · pp. 16–17
• MPC forecasts commonly used autoregressive models or neural networks, sometimes supplemented by external weather or energy-price forecasts, with horizons of two to 24 hours.Source: §4.4.1 · p. 17
• MPC commonly minimized cumulative HVAC energy or energy cost, using soft penalties for constraint violations while enforcing equipment limits as hard constraints.Some studies also targeted peak demand or grid-service revenues; many solved convex programs, while nonconvex problems often yielded locally optimal solutions.Source: §4.4.1 · p. 17

4.4.2. RL implementation

• All reviewed RL field papers pre-trained agents offline in simulators before hardware deployment, using system models such as neural networks, EnergyPlus, or linear thermal circuits.This avoided exposing equipment to untrained agents or requiring long hardware training before convergence.Source: §4.4.2 · p. 17
• Nearly all RL papers used Q-learning or actor-critic methods, with indoor temperatures commonly represented as states and equipment settings as actions.Most aimed to reduce energy or cost, typically adding reward penalties for comfort or equipment-limit violations; some rewarded smoother actions.Source: §4.4.2 · pp. 17–18
• The most popular RL exploration methods were ϵ-greedy and Boltzmann approaches; a few studies added Gaussian noise to actions.Source: §4.4.2 · p. 18

4.5. Experiment protocol choices

• Simulation benchmarks were used in 7 (14%) of 51 residential tests and 47 (33%) of 153 commercial tests; one test in each group was not benchmarked.Most tests in both building types benchmarked against measurements.Source: §4.5 · p. 18
• Whole buildings were controlled in 38 (73%) of 52 residential tests, while subsets of building spaces were controlled in 110 (71%) of 154 commercial tests.Source: §4.5 · p. 18
• Tests lasting less than seven days accounted for 15 (29%) of 52 residential tests and 67 (46%) of 146 commercial tests; eight commercial tests did not report duration.Source: §4.5 · p. 18
• Tests lasting one month or more accounted for 20 (38%) of 52 residential tests and 39 (27%) of 146 commercial tests.Source: §4.5 · p. 18

5.1. Benchmarking method

• Simulation-based benchmarks showed substantially higher duration-weighted savings than measurement-based benchmarks in all four comparisons, but small samples precluded meaningful statistical conclusions.The authors identify model prediction errors comparable to typical 10–20% savings and manual tuning of model parameters as possible sources of overestimation.Source: Figure 7 · p. 19, §5.1 · p. 20
• Only 9% of reviewed papers reported confidence intervals for energy or cost savings estimates.Confidence intervals can characterize uncertainty in savings estimates; they appeared in 3 of 23 residential papers and 6 of 77 commercial papers, with some papers in each group not applicable.Source: §5.1 · pp. 20–21

5.2. Controlled portion of building

• In commercial buildings, subset control yielded 27% energy savings versus 13% for whole-building control, a statistically significant difference (U = 738, p = 0.004).Commercial cost savings were 26% versus 12%, but the sample sizes were too small for statistical testing; in residential buildings, energy savings were 25% for subsets versus 23% for whole buildings, and cost savings were 22% versus 17%. The authors attribute possible overestimation to uncounted heat transfer from adjacent zones.Source: §5.2 · p. 22

5.3. Test duration

• Residential tests shorter than one week reported 57% duration-weighted average cost savings, compared with 22% for longer tests.Short tests also showed more variability and generally higher savings estimates; the authors suggest transient thermal-mass effects can distort results for tests lasting only a few days, while their bias is small for month-long tests.Source: Figure 9 · p. 23, §5.3 · pp. 23–24

5.4. Energy and cost savings in filtered data

• After three protocol filters, duration-weighted average cost savings across 53 tests were 16% in residential and 13% in commercial buildings.The filters required measurement-based benchmarks, whole-building or large-subset control, and tests lasting at least seven days; the retained tests came from 9 of 24 residential papers and 21 of 80 commercial papers.Source: §5.4 · p. 24

6.1. Inconclusive analyses

• Analyses of electricity rates, building properties, control architecture, climate, and operating mode yielded no robust conclusions because filtered samples were small or contextual reporting was sparse and inconsistent.Complex electricity incentives and thermally massive buildings may favor advanced control, but the reviewed evidence could not establish these or other contextual effects reliably.Source: §6.1 · p. 24, §6.1.1 · p. 25, §6.1.2 · p. 26

6.2. Costs, labor, and other challenges

• Duration-weighted cost savings were 16% in residential buildings and 13% in commercial buildings, but only 13 of 104 papers reported deployment, operation, or maintenance costs.The paper organizes the limited cost evidence into instrumentation, data and communication, labor, and cost-benefit analysis.Source: §6.2 · p. 26

6.2.1. Instrumentation; 6.2.2. Data and networking

• Reported instrumentation and data-networking costs range from low-cost plug-and-play hardware to additional sensing infrastructure and weather-data subscriptions.Wang et al. described a $50 system; Sturzenegger et al. reported annual weather-data subscriptions of 100 to 600 CHF, while some manufacturer APIs were free for research but might require fees at scale.Source: §6.2.1 · pp. 26–27, §6.2.2 · p. 27

6.2.3. Labor; 6.2.4. Cost-benefit analysis

• Deployment and maintenance can require substantial labor, and reported payback or cost recovery does not eliminate concerns that engineering costs may outweigh savings.Pergantis et al. reported 190 engineer-days for residential MPC deployment; Blum et al. estimated 239 person-days for commercial MPC. Thorsteinsson et al. estimated €1,200 cost recovery over 10 years, while Ham et al. reported a 13-heating-month payback; Sturzenegger et al. cautioned that high deployment and engineering costs may outweigh savings.Source: §6.2.3 · p. 27, §6.2.4 · p. 27

6.3.1. Improve experiment protocols; 6.3.2. Focus on costs

• Future demonstrations should use measured baseline comparisons, control whole or well-insulated building areas or account for heat transfer, and run at least seven days, ideally a month or longer.Researchers should also report deployment, operation, and maintenance costs, distinguish one-time from recurring expenses, and prioritize reducing recurring costs through less sensing and communication and faster deployment and commissioning.Source: §6.3.1 · p. 28, §6.3.2 · p. 28

6.3.3. Report the context

• Field studies should consistently report electricity-rate structure, building characteristics, control architecture, and climate, using the provided checklist or comparable information.The checklist is an open-access spreadsheet [20], and the authors recommend attaching it to submissions and checking for it during editorial and peer review.Source: §6.3.3 · pp. 28–29

6.3.4. Expand the hardware scope; 6.3.5. Explore emerging objectives

• Future demonstrations should coordinate HVAC with other distributed energy resources and test objectives including peak shaving, wholesale-price arbitrage, and grid reliability services.The review notes that almost all field studies controlled HVAC alone, while few evaluated grid-interactive opportunities in real buildings.Source: §6.3.4 · p. 29, §6.3.5 · p. 29

6.4. Limitations and future literature reviews

• The review excludes non-peer-reviewed sources and focuses on MPC and RL; future reviews could broaden source and control-method coverage and revisit contextual hypotheses after reporting improves.Sparse and inconsistent contextual reporting limited this analysis; possible additions include open-loop optimal, fuzzy-logic, rule-based, robust, adaptive, hybrid, and DeePC methods.Source: §6.4 · pp. 29–30

7. Conclusion

• 71% of reviewed papers used protocols that may produce unreliable savings estimates; over the remaining 29%, duration-weighted average utility bill savings were 16% in residential and 13% in commercial buildings.These savings are potentially attractive, but few studies reported deployment, operating, or maintenance costs, labor, or related challenges.Source: §7 · p. 30
• The authors recommend that research characterize and reduce the costs of deploying, operating, and maintaining advanced HVAC control to help make its economic case and accelerate adoption.Few studies reported costs, labor, or related deployment, operation, and maintenance challenges.Source: §7 · p. 30
• The authors also recommend controlling HVAC for emerging objectives, potentially alongside other distributed energy resources, to improve the economic case for advanced control.Source: §7 · p. 30

Appendix A. Nomenclature

• The nomenclature expands abbreviations for building systems, control approaches, and energy resources.These include AHU (air handling unit), DeePC (data-enabled predictive control), DER (distributed energy resource), HP (heat pump), HVAC (heating, ventilation, and air-conditioning), MPC (model predictive control), OCC (occupant-centric control), PV (photovoltaics), RL (reinforcement learning), TABS (thermally-activate building system), VAV (variable air volume), and VRF (variable refrigerant flow).Source: Appendix A · p. 31

• Table B.1 summarizes field demonstrations of MPC and RL for HVAC in residential buildings.Source: Table B.1 · p. 32
• The residential-building overview table continues from a previous page.Source: Table B.1 · p. 33
• Table B.2 summarizes commercial-building demonstrations and continues across additional pages.Source: Table B.2 · pp. 34–39