The International Performance Measurement & Verification Protocol (IPMVP) Option C is suitable for assessing facility-wide consumption changes when tracking large reductions or complex electrification projects. This method relies on linear regression models to quantify energy use as a function of driving variables—most commonly weather, occupancy, or production. However, the strength of an Option C approach depends entirely on the reliability of the underlying model which is not solely determined by the coefficient of determination r2. A reliable model must be grounded in physics and thermodynamics. Adding additional variables to increase this coefficient may cause misleading results. 

The first step when developing models is to align the independent and dependent variables. Weather data (Heating Degree-Days (HDD) and Cooling Degree-Days (CDD)) must be synchronized with utility billing cycles. Failing to do so blurs the relationship between the two variables. When using monthly data, a minimum of 12 months of data is recommended to capture a full range of weather conditions. Using 24 to 36 months is ideal where facility behavior is stable, especially if additional variables (occupancy, production) are used. 

While short-term data (daily or weekly) can provide more granular insights and allow better segregation between occupied and unoccupied periods, practitioners should be wary of hourly models where correlations often weaken due to the slow thermal response of buildings.

The goal of any regression model is to produce an adjusted baseline with an acceptable level of uncertainty. Both ASHRAE Guideline 14 and the IPMVP consider a savings uncertainty of 50% or less acceptable. In other words, the reported savings need to be at least twice the aggregate model uncertainty to be statistically valid. 

The coefficient of determination r2 alone does not determine a model’s usefulness. While a value above 0.7 is preferred, r2only quantifies the strength of the relationship and can be low if there is little variation in the independent variables. A more critical metric for uncertainty and savings estimation is the Coefficient of Variation of the Root Mean Square Error (CRMSE) is it drives the overall uncertainty of the model and savings derived from the model. If the expected energy reduction is at least 20%, a CVRMSE below 0.2 is acceptable. Smaller savings require lower CVRMSE values. 

In addition to the r2 and CVRMSE values, the statistical significance of each variable must be verified using t-statistics. A magnitude greater than 2 indicates a 95% chance that the variable is statistically relevant; if it’s lower, the variable is likely contributing little and should be removed. One notable exception is the model intercept in heating-only applications where the balance point is accurate; in these cases, the intercept can remain even if its t-statistic is low. Finally, ensure the Net Mean Bias Error (NMBE) is negligible (less than 50 ppm) to confirm no systematic errors were introduced during data manipulation.

Because regression modeling is a powerful tool, it needs to be used with care to reduce errors. Synchronize all data to the utility billing cycle. Keep any models as simple as possible and include only variables that have physical or thermodynamic explanations while also being statistically relevant. Focus on reducing uncertainty by evaluating the CVRMSE as well as the r2 with an emphasis on the former. Avoid excessive data manipulation to reduce introducing bias error into the results. For additional information, see Uncertainty Assessment For IPMVP. 

Mark Stetz, P.E. PMVE CMVP | Sr. Energy Engineer
Iconergy CEG LLC
mstetz@iconergy.com / 303-882-4295