What building research actually says about HVAC savings

1. Commissioning is a meaningful alternative, not a weak baseline

A 2020 study by Crowe and colleagues assembled records for 1,185 commissioning projects covering 1,482 North American buildings. The headline sample was not the denominator for every result. The combined existing building energy savings analysis contained 446 projects, with median whole building source energy savings of 6.4%. The middle half of those projects fell between 3.4% and 12.4%.

The researchers used project engineering records, checked consistency and combined data collected in 2009 and 2018. Source energy includes the upstream energy associated with delivered electricity and fuel. It is not the same measure as electricity at an HVAC meter or dollars on a utility bill.

Delivery context mattered. Median savings were 5% across 298 utility commissioning projects, 9% across 41 utility projects using monitoring based commissioning, and 14% across 107 projects outside utility programs. These groups differed in scope, budgets and review practices. They were not randomly assigned competing treatments.

What transfers: correcting operation deserves a place in the options considered for an existing building. A control proposal should explain the opportunity that remains after practical corrections to existing operation.

What does not: none of these medians is a guaranteed saving for your building, or proof that one delivery model will outperform another at the same site.

Read the original commissioning study, sections 2 and 3.1, Tables 1 and 2, and Figure 2.

2. Analytics results describe tools plus operational action

The 2020 Smart Energy Analytics Campaign report distinguishes energy information systems, which analyze meter data, from fault detection and diagnostics, which analyze system operation. Both can support a continuing commissioning process.

Its frequently quoted second year results came from different reporting subsets:

1. Energy information systems: median energy savings of 3%, from 10 organizations covering 1,333 buildings.
2. Fault detection and diagnostics: median energy savings of 9%, from 18 organizations covering 509 buildings.

The reported percentages are medians across organizations, not medians across all individual buildings. They compare the second year after installation with the baseline year before installation. They are whole building results, not isolated HVAC or AI savings. The fault detection category also includes implementations combining diagnostics and meter analytics.

Across the broader savings dataset, four organizations used interval data analysis, two used engineering calculations and 44 used annual energy use analysis. Weather normalization depended on the method and available reporting. The report did not normalize the supplied energy data for changes in equipment or occupancy, and other efficiency projects could contribute to the reductions.

What transfers: analytics needs an operating process. The report identifies data quality and taking action on findings as major barriers. Buying a dashboard is not the same intervention as investigating and correcting the problems it reveals.

What does not: the difference between these medians is not a controlled test showing that diagnostics saves three times as much at any given building. It is also not evidence for autonomous control: only two campaign participants had automated system optimization, too few for the report to publish that category’s costs or savings.

Read the original analytics report, sections 2, 3.1, 3.2 and 3.5, and Tables 4 and 5.

3. Predictive control can shift demand, but the tariff is part of the result

A 2024 paper by Zanetti and colleagues reports field tests of model predictive control serving the office areas on two floors of one Berkeley Lab building. The system used models of building response, weather forecasts and electricity price signals to plan HVAC operation.

During summer and fall tests in 2023, it shifted HVAC consumption away from expensive periods. However, the electricity price profiles were synthetic research scenarios. The resulting calculated costs were hypothetical, not a report of savings on the building’s actual utility bills.

The comparison used periods with the existing controller and periods with predictive control. Weekends and periods affected by abnormal operation, faults or data failures were excluded. The authors reported no occupant complaints, while also describing difficulty maintaining acceptable temperatures during unusually hot conditions near the end of the fall test. The paper explicitly says the results do not describe a full year of operation.

What transfers: when electricity prices vary by time, the timing of consumption can matter as well as its total. The useful question is whether your actual tariff and building response create an opportunity worth testing.

What does not: a short test in a mild climate, on this particular HVAC configuration, does not establish annual savings elsewhere. Model predictive control in this study is not a test of Energia AI’s learning approach.

Read the original control study, sections 2, 3.3, 4.1 and 5.

4. The control algorithm is only part of the implementation

The same field study estimated around four person months of work to update an existing research control implementation. About half went to integrating new data streams and replacing sources that were no longer available. The remainder included data quality monitoring, testing and changes to the interface with the building controls.

This is an effort estimate from one research implementation update, not a market price or a typical installation duration. The researchers also described continuing work caused by data interruptions, server restarts and software updates.

What transfers: ask how a proposal handles changing data sources and who maintains the connection between analysis and equipment. Those responsibilities belong in the economic comparison, alongside the potential operating benefit.

What does not: the study cannot establish what Energia AI or another commercial provider will charge, how long your integration will take or what annual support it will require.

Read the implementation discussion, section 4.2.

Use the research to choose the next question

If the problem is incorrect operation, commissioning is a credible candidate. If the problem is finding and following up on recurring faults, evaluate analytics together with the people who will act on it. If the opportunity depends on anticipating building response or changing electricity prices, predictive control may warrant a bounded evaluation.

These are interpretations of the studies, not a prescribed sequence or a prediction of savings. The evidence does not justify adding the percentages together, comparing them as a league table or treating any of them as an Energia AI result.

For the next stage, use our pilot decision guide to define a building evaluation, or compare the scope of HVAC optimization offerings.

Discuss your building

Discuss your building with Energia AI.

Prepared with AI assistance using the original publications linked above. Energia AI is the publisher. This page does not claim a human engineering review.