ASHRAE Guideline 14 Regression Baselines: Change-Point Models Explained

Energy savings cannot be metered. You can meter what a building uses after a retrofit, but the thing you are trying to measure — what it would have used without the retrofit — never happens. ASHRAE Guideline 14 solves this by building a baseline model that predicts the counterfactual. Our earlier article covers the guideline's three approaches and the calibration statistics for simulation models. This one goes deeper into the approach used most often in practice: the whole-building regression baseline, and specifically the change-point models that sit underneath it.

The idea: energy as a function of weather

The largest driver of month-to-month variation in most buildings is outdoor temperature. Plot twelve or more months of utility consumption against average outdoor temperature and a pattern appears: a flat portion where only base loads operate (lighting, plug loads, hot water, fans), and a sloped portion where heating or cooling switches on. The temperature at which the line bends is the building's balance point, or change point.

Fit a model to that shape using the pre-retrofit data, then feed it the weather from the post-retrofit period. The model tells you what the old building would have used under the new weather. The difference between that prediction and the metered consumption is the avoided energy use — the savings. This is the method behind IPMVP Option C and most utility pay-for-performance programmes.

The change-point model family

These models were developed through ASHRAE research and packaged in the ASHRAE Inverse Modeling Toolkit. Each is named by its number of parameters:

• 1P (mean) — a constant. Suitable only for loads with no weather dependence, such as a sub-metered lighting circuit.
• 2P (linear) — a straight line with an intercept and slope. Works when a building is heating or cooling over the entire data range, such as a heating-only fuel meter in a cold climate.
• 3P heating or cooling — a flat base load plus one sloped leg beginning at a change point. The classic shape for a gas meter (heating) or an electric meter in a cooling-dominated building.
• 4P — two different slopes joined at a single change point, with no flat region. Common in buildings with variable-air-volume systems, where energy keeps changing on both sides of the balance point.
• 5P — a heating slope, a flat dead-band and a cooling slope, with two change points. The typical shape for an all-electric building that both heats and cools on the same meter.

The right model is the one that matches the physics of the meter, not the one with the best statistics. A 5P model fitted to a gas meter can produce a superb R² and an absurd cooling slope. Always check that slopes have the expected sign and that change points fall at plausible temperatures.

Temperature or degree-days?

Monthly models can use average outdoor temperature with fitted change points, or heating and cooling degree-days calculated to a chosen base. The two are closely related. Degree-days to an assumed base of 65°F or 18°C are convenient but impose a balance point the building may not have. Change-point models let the data find it. For daily and hourly data, temperature-based models with additional variables — occupancy, day type, time of week — generally perform better.

Weather matters in both directions. Use measured weather from a nearby station for the actual baseline and reporting periods. Use long-term typical weather only when you want to normalise savings to an average year.

Is the model good enough?

Guideline 14 judges a baseline model on how well it predicts the data it was fitted to, using two statistics:

• CV(RMSE) — the typical size of the model's errors relative to mean consumption, a measure of scatter.
• NMBE — the net bias. A regression fitted by least squares has near-zero bias on its own data by construction, so NMBE mostly catches errors in how the model was applied.

R² is widely reported, but on its own it is not a reliable test. A building with little weather sensitivity can have a low R² and still a very accurate baseline, because there is little variation to explain. CV(RMSE) tells you more about whether savings will be distinguishable from noise.

Savings uncertainty: the test that matters

The core question is not whether the model fits, but whether the savings are larger than the model's uncertainty. Guideline 14's whole-building compliance path requires the uncertainty in reported savings to be less than 50% of the annual savings, at a 68% confidence level. Three factors drive it:

• Size of the savings — a 25% saving is far easier to verify than a 5% one. Whole-building M&V is generally a poor choice for small measures, which is why the guideline points to retrofit-isolation methods for them.
• Model scatter — a high CV(RMSE) widens the uncertainty band.
• Amount of data — more baseline and reporting periods narrow it. With daily or hourly data, consecutive errors are usually autocorrelated — today's error predicts tomorrow's — and the uncertainty calculation must correct for it, or it will be badly understated.

Calculate the expected savings uncertainty before the project starts. If the projected savings are within the noise of the baseline, no amount of post-retrofit analysis will rescue the M&V plan.

Non-routine adjustments: where projects go wrong

A regression baseline assumes the building is otherwise unchanged. Real buildings are not. A new tenant, an extended operating schedule, a server room, a floor that goes vacant: each shifts consumption for reasons unrelated to the retrofit. These non-routine events must be identified and adjusted for, or they will be counted as savings (or losses).

The practical defence is documentation. Record occupancy, operating hours, floor area in use and significant equipment changes throughout both periods. Where a static factor changes, quantify its effect by engineering calculation, sub-metering or a calibrated simulation, and report the adjustment transparently. Unexplained steps in the post-retrofit data are the most common reason savings claims are challenged.

When regression is not the right tool

Regression baselines need a stable building, a meaningful baseline period (normally at least twelve months covering the full weather range) and savings large enough to see. Where those conditions fail — new construction with no baseline, deep renovations that change the building's character, or many overlapping measures that must be reported separately — Guideline 14's calibrated simulation approach is the better choice. That approach is covered in our Guideline 14 overview. For a single, well-defined measure, retrofit isolation with sub-metering is often simpler and more precise.

The takeaway

Change-point regression is the workhorse of whole-building M&V because it is simple, transparent and grounded in the physics of heating and cooling. Pick the model shape that matches the meter. Judge it by CV(RMSE) and savings uncertainty rather than R². Test the uncertainty before the retrofit, and record every non-routine change along the way. Do those four things and the savings you report will stand up to scrutiny from a client, a utility or an ASHRAE Standard 100 compliance review.

This article is general guidance and reflects information available at the time of writing. Acceptance criteria are defined by the published edition of ASHRAE Guideline 14 and by the programme, contract or M&V plan that applies to your project — always confirm the requirements that govern your savings claim.