Baselines and Weather Normalization

Key takeaways

• The baseline must cover the full range of conditions the equipment normally sees, and it cannot be rebuilt after the change is made.
• Cooling degree days (CDD) and outdoor temperature are the most common weather variables in HVAC baselines.
• ASHRAE Guideline 14 describes change-point regression models (three-, four- and five-parameter) that suit cooling loads.
• Savings can be reported as avoided energy (actual reporting conditions) or normalized savings (typical weather); the plan must say which.
• Like-for-like binning compares periods with matching weather and load; it is transparent but needs enough overlapping data.

Every savings figure is a comparison with a baseline (for the overall framework, see how HVAC energy savings are measured). If the baseline is too short, unrepresentative or poorly documented, no amount of later analysis can rescue the result. FEMP's M&V guidance makes the point bluntly: after a measure is installed, it is impossible to go back and re-evaluate the baseline, because the baseline no longer exists. This guide explains how baselines are set and how weather and load differences are removed, so you can judge whether a comparison is fair.

What a baseline has to include

A baseline is more than a number of kilowatt-hours. EVO's IPMVP principles say baseline documentation should include:

• The baseline period: often a year, though it can be any period that fits the project.
• Energy data (the dependent variable): utility bills for whole-facility methods, or field-metered interval data and spot measurements for equipment-level methods.
• Independent variables: the drivers of energy use measured over the same period, such as outdoor temperature, production, equipment speed or pressure.
• Static factors: conditions assumed to stay constant, such as occupancy type and density, run times, set points, ventilation levels, and any significant equipment problems or outages during the baseline.

EVO adds that the baseline period should represent a full normal operating cycle, from maximum to minimum energy use, and should sit immediately before the decision to make the change. Periods further back may not reflect conditions just before the retrofit. FEMP 5.0 recommends reviewing the previous 12 to 36 months of utility bills to spot anomalies and seasonal patterns, and measuring long enough to capture performance across the full range of operating conditions.

Weather is the biggest routine adjustment for cooling

FEMP calls outdoor air temperature the most common independent variable for many types of measures, often expressed as heating or cooling degree days. The U.S. Energy Information Administration (EIA) defines degree days against a 65°F base: a day with a high of 90°F and a low of 66°F has a mean of 78°F, which gives 78 − 65 = 13 cooling degree days. Sum the days in a billing period and you have a simple index of how much cooling the weather demanded.

The fixed 65°F base is a convention, not a law of physics. DOE's Uniform Methods Project protocol for whole-building billing analysis notes that 65°F was the industry standard for many years. It found that, in general, bases of 60°F for heating and 70°F for cooling usually fit better, and it describes a variable-base approach that lets the data choose the balance point. Commercial buildings with high internal heat gains often need cooling at lower outdoor temperatures than homes, which is one reason a variable or tested base is better than an assumed one.

Humidity matters as well, particularly on the Gulf Coast (see hotter summers and cooling load). Dehumidification is real cooling work (latent load) that dry-bulb degree days do not fully capture. Equipment-level methods can measure humidity directly, which is one advantage of metering at the unit.

Regression models: turning weather into an adjusted baseline

To adjust for weather, the analyst fits a model of baseline energy against the independent variables, then uses it to predict what the baseline equipment would have used under reporting-period conditions. ASHRAE Guideline 14-2023 lists recommended model forms for the whole-building approach:

A simple worked illustration

The numbers below are hypothetical and only show the arithmetic. They are not a result from any product or site.

1. Baseline metering on a cooling system gives a model of daily energy = 400 kWh + 35 kWh × CDD, fitted across a full range of baseline weather.
2. In a 30-day reporting month with 600 CDD, the model predicts the baseline system would have used 30 × 400 + 35 × 600 = 12,000 + 21,000 = 33,000 kWh. This is the adjusted baseline.
3. The meter shows the system actually used 30,500 kWh.
4. Avoided energy for the month = 33,000 − 30,500 = 2,500 kWh, or about 7.6% of the adjusted baseline.

Without the model, the analyst might have compared 30,500 kWh with a milder baseline month of, say, 26,000 kWh and concluded that energy use went up. The model removes the weather from the comparison. Its accuracy depends on how well it fits and on whether reporting-period conditions fall inside the range seen in the baseline. EVO requires the plan to report the range of independent variables over which the model is valid.

Avoided energy or normalized savings?

EVO describes three ways to put the two periods on the same footing:

• Project the baseline to reporting-period conditions. The result is avoided energy: what you did not use under the conditions that actually occurred. The example above works this way.
• Project the reporting period back to baseline conditions.
• Project both to standard conditions, such as a typical meteorological year (TMY). The result is normalized savings, which are useful for budgeting a "typical" year.

FEMP 5.0 explains that TMY datasets represent long-term typical weather at a location, built from many years of data. They describe seasonal and daily patterns rather than predicting a particular year. Neither approach is wrong, but a report should say which one it uses, because the two can differ noticeably in an unusually hot or mild year.

How good is the model?

EVO requires the plan to report each model's coefficients and statistical metrics: CV(RMSE), mean bias error, R-squared, t-statistics and similar. In plain terms:

• R-squared: the share of variation in energy use the model explains. Higher is better, but a high R-squared alone does not prove a good model.
• CV(RMSE): typical scatter of predictions around actual values, as a percentage. Lower is better.
• Mean bias error (or NMBE): whether the model runs consistently high or low. It should be close to zero.

For calibrated simulation models, FEMP 5.0 reproduces the Guideline 14-2023 tolerances: within ±5% bias and 15% CV(RMSE) on monthly data, and within ±10% and 30% on hourly data. Regression baselines should have their statistical targets stated in the M&V plan before data collection begins. FEMP also recommends building whole-facility models from whole-year data sets (12, 24, 36 or 48 months) so seasonal effects are not overstated.

The basic test is simple. If the model's normal scatter is about as large as the savings you expect, a single month's result cannot tell you much. That is the same reason FEMP warns that small savings can be "lost in the noise" in whole-facility data.

Like-for-like comparison: binning by conditions

An alternative to fitting one model is to compare only periods with matching conditions. With high-frequency data, such as one reading per minute, the analyst groups intervals into "bins" by outdoor temperature, humidity and building load. Baseline and post-change performance are then compared within each bin. If the post-change data show lower kW at the same outdoor temperature and the same cooling delivered, that difference is unlikely to be caused by weather.

This is the approach CryogenX4 describes for its field testing: aggregating statistically significant data points under like conditions (similar external weather and similar internal heat loads) in the baseline and post-treatment data sets, then comparing averages.

When you review a binned analysis, ask for the number of data points in each bin for both periods, how bins with too little overlap were treated, and how bin results were weighted into an annual figure.

Non-routine adjustments: the things that should not change

EVO defines static factors as conditions not expected to change, such as facility size, equipment design and operation, number of shifts, or occupant type and number. If they do change, they must be monitored and corrected with non-routine adjustments. Common HVAC examples include:

• Chilled-water or space temperature set points changed during the test
• New tenants, added shifts or extended hours
• Another efficiency measure installed on the same equipment, such as new VFDs, controls changes or coil cleaning
• Equipment failures, refrigerant leaks or extended outages

Reviewers should see every one of these in the report; how to read an M&V report explains what to look for. The simplest way to avoid disputes is to agree on a change log before testing starts. Facility staff record any change to the treated equipment or the spaces it serves, with dates.

Baseline checklist

1. Known faults repaired and documented before the baseline starts.
2. Baseline long enough to span the equipment's normal range of conditions.
3. Independent variables (temperature, humidity, load) measured on the same time stamps as energy.
4. Static factors written down: set points, schedules, occupancy, maintenance status.
5. Adjustment method chosen in advance: regression model or condition bins, avoided energy or normalized savings.
6. Statistical targets or minimum bin data stated in the plan.
7. Change log agreed with facility staff.

Where CryogenX4 fits

According to the company, CryogenX4 changes how efficiently treated equipment transfers heat, so its effect shows up as lower energy for the same cooling under the same conditions. That is what a like-for-like comparison is designed to detect. The company describes minute-level monitoring over a testing period of usually 3 to 9 months, with a baseline captured before the one-time treatment. CryogenX4 reports savings of up to 30%, and results vary by equipment condition. A well-built baseline is what makes your own result trustworthy. The full list of data points is in what a CryogenX4 pilot measures.

Next step

Pull 12 to 36 months of utility bills and any BAS trend data for the equipment you are considering, and note any changes to set points, schedules or occupancy over that time. That record will shape a sound baseline for any efficiency project. To plan a measured baseline for a CryogenX4 pilot, contact the team.

Frequently asked questions

How long should an HVAC baseline be?

Long enough to cover the equipment's normal range of operating conditions. For whole-facility regression, FEMP recommends at least 12 and preferably 24 or more months of data. Equipment-level monitoring with high-frequency data can be shorter, provided the baseline and reporting periods overlap in the conditions they cover.

What base temperature should cooling degree days use?

EIA uses 65°F as the U.S. standard. DOE's Uniform Methods Project found that a 70°F cooling base usually fits better than 65°F in its whole-building billing protocol, and recommends a variable base where possible, letting the data identify the balance point.

Can a baseline be adjusted after the fact?

Routine and non-routine adjustments are applied after the fact, but only if the underlying data were recorded. Once equipment has been changed, the original baseline conditions can no longer be measured, which is why FEMP stresses documenting the baseline thoroughly.

Is binning by conditions as valid as regression?

Both are recognized ways to compare periods under similar conditions. Binning is transparent and works well with minute-level data, but it needs enough overlapping data in each bin and a stated method for turning bin results into annual savings.

Sources

1. IPMVP Generally Accepted M&V Principles (October 2018) — Efficiency Valuation Organization (EVO)
2. M&V Guidelines: Measurement and Verification for Performance-Based Contracts, Version 5.0 (September 2024) — U.S. Department of Energy, Federal Energy Management Program
3. M&V Guidelines: Measurement and Verification for Performance-Based Contracts, Version 4.0 (November 2015) — U.S. Department of Energy, Federal Energy Management Program
4. Degree days — U.S. Energy Information Administration
5. Uniform Methods Project, Chapter 8: Whole-Building Retrofit with Consumption Data Analysis Evaluation Protocol (NREL/SR-7A40-68564, 2017) — U.S. Department of Energy / National Renewable Energy Laboratory
6. Guideline 14-2023: additional material (contents, retrofit-isolation metering and whole-building model forms) — ASHRAE Journal