Accurate evaluation of the performance of buildings participating in Demand Response (DR) programs is critical to the adoption and improvement of these programs. Typically, we calculate load sheds during DR events by comparing observed electric demand against counterfactual predictions made using statistical baseline models. Many baseline models exist and these models can produce different shed calculations. Moreover, modelers implementing the same baseline model can make different modeling implementation choices, which may affect shed estimates. In this work, using real data, we analyze the effect of different modeling implementation choices on shed predictions. We focused on five issues: weather data source, resolution of data, methods for determining when buildings are occupied, methods for aligning building data with temperature data, and methods for power outage filtering. Results indicate sensitivity to the weather data source and data filtration methods as well as an immediate potential for automation of methods to choose building occupied modes.

1 aAddy, Nathan, J.1 aMathieu, Johanna, L.1 aKiliccote, Sila1 aCallaway, Duncan, S. uhttps://gridintegration.lbl.gov/publications/understanding-effect-baseline03158nas a2200145 4500008003900000245008100039210006900120260002800189520262100217100002302838700002502861700002002886700002202906856008402928 2011 d00aUsing Whole-Building Electric Load Data in Continuous or Retro-Commissioning0 aUsing WholeBuilding Electric Load Data in Continuous or RetroCom aCincinnati, OHc08/20113 aWhole-building electric load data can often reveal problems with building equipment or operations. In this paper, we present methods for analyzing 15-minute-interval electric load data. These methods allow building operators, energy managers, and commissioning agents to better understand a building's electricity consumption over time and to compare it to other buildings, helping them to 'ask the right questions' to discover opportunities for electricity waste elimination, energy efficiency, peak load management, and demand response. For example: Does the building use too much energy at night, or on hot days, or in the early evening? Knowing the answer to questions like these can help with retro-commissioning or continuous commissioning.

The methods discussed here can also be used to assess how building energy performance varies with time. Comparing electric load before and after fixing equipment or changing operations can help verify that the fixes have the intended effect on energy consumption.

Analysis methods discussed in this paper include: ways to graphically represent electric load data; the definition of various parameters that characterize facility electricity loads; and a regression-based electricity load model that accounts for both time of week and outdoor air temperature. The methods are illustrated by applying them to data from commercial buildings. We demonstrate the ability to recognize changes in building operation, and to quantify changes in energy performance.

Some key findings are:

- Plotting time series electric load data is useful for understanding electricity consumption patterns and changes to those patterns, but results may be misleading if data from different time intervals are not weather-normalized.
- Parameter plots can highlight key features of electric load data and may be easier to interpret than plots of time series data themselves.
- A time-of-week indicator variable (as compared to time-of-day and day-of-week indicator variables) improves the accuracy of regression models of electric load.
- A piecewise linear and continuous outdoor air temperature dependence can be derived without the use of a change-point model (which would add complexity to the modeling algorithm) or assumptions about when structural changes occur (which could introduce inaccuracy).
- A model that includes time-of-week and temperature dependence can be used for weather normalization and can determine whether the building is unusually temperature-sensitive, which can indicate problems with HVAC operation.