In this article, you will learn how to use the Baseline Health Diagnostics feature within the Baseline Calculator, available in the M&V section of the platform.
Contents of this article:
Introduction
The Baseline Health Diagnostics feature allows you to evaluate the statistical reliability of the baselines created with the Baseline Calculator feature.
Requirements
In order to be able to use this feature, you need to have Advanced or Ultimate licenses. The same permissions that are required in order to use the Baseline Calculator feature.
You will also need to create a baseline in order to trigger its statistical evaluation.
How does it work?
Whenever you calculate a baseline, the platform automatically evaluates its statistical health. These diagnostics complement metrics such as R² and adjusted R² by checking whether the model has enough reliable data, uses meaningful variables and is likely to perform consistently.
The result is summarised by an overall baseline health status:
- OK: All completed checks have passed. No statistical quality issues were detected.
- Warning: At least one check identified an issue that may reduce the reliability of the baseline. Review the recommendations before using the model.
- Critical: At least one critical issue was detected. The baseline may not be reliable and should be reviewed before it is used for savings calculations or an M&V project.
A critical result takes priority over warnings. Some diagnostics may not be applicable to a particular model—for example, checks that require input variables cannot be performed when the model contains only a constant.
Reviewing the baseline diagnostics
In order to visualise the issues, you should click "View issues".
This will enable a right-click menu that will show you the Baseline health diagnostics section. It displays the result of each check. Every diagnostic card includes:
- The name of the check.
- A description of the issue detected.
- A recommended action to help improve the baseline.
- Additional technical information, where relevant.
Checks that have passed also include a short explanation confirming which model condition was validated.
After reviewing a warning or critical issue, update the relevant baseline configuration and select Calculate baseline again. The diagnostics will be repeated using the new configuration.
What is checked?
The platform performs up to 11 statistical checks:
1. Degrees of freedom
Checks whether the reference period contains enough data compared with the number of parameters used by the model.
If a warning appears, extend the reference period to include more data or reduce the number of selected variables.
2. Sample size ratio
Checks whether there are enough data points per input variable to obtain reliable estimates.
If a warning appears, extend the reference period or reduce the number of candidate variables.
3. Correlation
Checks whether input variables are too closely correlated with each other. Highly correlated variables make it difficult to determine the individual effect of each variable on energy consumption.
If a warning appears, remove one of the redundant variables, particularly when two variables represent similar weather or operational effects, and recalculate the baseline.
4. Error distribution
Checks whether the differences between actual and predicted consumption are evenly distributed. An unusual distribution can indicate outliers, abnormal operating periods or consumption patterns that the model does not capture.
If a warning appears, review the reference period for shutdowns, operational changes, outliers or other unrepresentative data. You can exclude these periods under Calculation options > Periods to discard.
5. No significant variables
Checks whether the calculated model contains meaningful input variables. A constant-only model uses the average consumption because none of the selected variables sufficiently explains the consumption pattern.
If a critical issue appears, review the candidate variables and consider adding more relevant energy drivers, such as weather, production, occupancy, calendar or operating-schedule data.
6. Model quality
Checks whether the selected variables explain energy consumption better than simply using average consumption.
If a critical issue appears, review whether the selected variables represent the site’s actual energy behaviour. Consider adding more informative variables or extending the reference period.
7. Search completeness
Checks whether the platform was able to analyse all relevant combinations of candidate variables within the available calculation time.
If a warning appears, a good model was found, but a better combination may exist. Reduce the number of candidate variables and prioritise those with a clear physical or operational relationship to energy consumption.
8. Data availability
Checks the completeness of the data used by the target and input variables. Missing readings from variables included in the model can reduce its accuracy or make the baseline unreliable.
Depending on the amount of missing data and whether the affected variable is used by the model, this check can return a warning or a critical result.
Check the affected devices or data sources for gaps, offline meters, sensor problems or integration issues. You can also select a reference period with more complete data.
9. Overfitting
Checks whether the model fits the reference-period data too closely or performs poorly when applied to verification data. An overfitted model may represent random variations instead of repeatable energy-consumption patterns.
This diagnostic considers the model’s R², the difference between R² and adjusted R² and, when verification data is available, its prediction error.
Depending on the severity of the issue, this check can return a warning or a critical result. Consider simplifying the model by reducing the number of variables or extending the reference period to include more representative data.
10. Linearity
Checks whether a linear model adequately represents the relationship between the input variables and energy consumption.
If a warning appears, the model may be missing non-linear effects, seasonal behaviour or different operating modes. Consider adding squared terms or interactions between relevant variables under Parameters interactions. Reviewing degree-day variables or separating different operating periods may also help.
11. Error variance
Checks whether the size of the model’s prediction errors remains consistent across different consumption levels.
If a warning appears, predictions may be less reliable during high- or low-consumption periods. Review the data for outliers, shutdowns, production changes or unusual operating conditions. You can also adjust the reference period, review the selected variables or discard unrepresentative periods.
Improving the baseline
A warning does not necessarily mean that the baseline cannot be used, but it identifies a condition that may affect its reliability. Critical issues require closer review because they can have a greater impact on predicted consumption and the resulting savings calculations.
When making changes, consider the physical and operational behaviour of the site. A statistically healthy baseline should also make sense from an energy-management perspective.
Repeat the calculation after each relevant change and compare the new health status, formula and performance metrics before selecting the baseline for your M&V project.
Pricing
This feature is included with Advanced or Ultimate licences.