Modeling office building occupancy in hourly data-driven and detailed energy simulation programs.

INTRODUCTION

Most simple linear regression (SLR) models assume the energy use as a function of outdoor temperature only, and can be used effectively at monthly or daily time scales, and thus do not require the use of a separate variable for occupancy. However, when a data-driven model is used at an hourly time scale, including separate variables for the driving factors becomes more effective, and thus the multiple linear regression (MLR) models become better candidates (Katipamula et al. 1994, 1995a, 1995b, and 1998, Reddy and Claridge 1994), and Reddy et al. 1998). Abushakra (2001) showed the advantage of including four driving variables in the hourly modeling of the energy use: (1) outdoor temperature, (2) outdoor specific humidity potential, (3) lighting and receptacles, and (4) occupancy. Very little published work has dealt with the occupancy factor as a variable in building energy use. In the forward modeling approach, as implemented in BLAST (U.S.Army 1979), DOE-2 (DOE 1981), and most recently EnergyPlus (DOE 2005), the total number of people occupying the building is estimated, and then multiplied by a diversity factor profile. In data-driven modeling, the occupancy factor is often ignored, or implicitly considered by grouping the daily data in different daytypes such as "weekday/weekend" groups which are modeled separately (IPMVP 2001). In university buildings it is common to identify "weekday", "weekend", "semester break", and "holiday" day-types (Thamilseran and Haberl 1995, and Dhar et al. 1994 and 1998). In order to develop one model that accounts for all periods, i.e., occupied and unoccupied, on an hourly time scale, a dummy variable (regressor) can be used. The dummy variable is often used in a simplified way; for instance, having a value of 1 between 8:00 AM and 5:00 PM, and 0 between 5:00 PM and 8:00 AM, for an office building. One of the DOE2.1 E manuals shows two occupancy schedules: one for weekends / holidays (0 for all 24 hours) and one for weekdays (0 from 10:00 PM to 8:00 AM, 1 from 9:00 AM to 6:00 PM, with the exception of a "dip" of 0.8 for the lunch break, and a slope from 1 down to 0 between 7:00 PM and 9:00 PM (0.5, 0.1, and 0.1 respectively, accounting for people working overtime) (Winkelmann et al. 1993).

In this paper, the effect of using different alternatives to account for the occupancy variable in data-driven modeling of building energy use is investigated, and the resulting uncertainty in the predictions is presented. Five different options to account for occupancy in data-driven models of building energy use are considered. All five are basically methods for generating fractions between 0 and 1 to represent the fractional level of occupancy during the day. These are multipliers (i.e., diversity factors, or hourly density) when multiplied by the estimated maximum occupancy would give the occupancy level at a specific hour of the day.

PREVIOUS WORK ON OCCUPANCY

The U.S. Department of Energy (DOE) Residential Buildings Program and the National Renewable Energy Laboratory (NREL) developed the Building America Research Benchmark in consultation with the Building America industry teams (NREL 2004). A series of user profiles intended to represent the behavior of a "standard" set of occupants, was created for use in conjunction with the benchmark. Average house and detailed (living room vs. bedroom) occupancy profiles (fractions between 0 and 1) were developed for weekdays and weekends, based on the basic ASHRAE occupancy schedule combined with engineering judgment.

Keith and Krarti (1999) developed a methodology for simplified prediction of the peak occupancy rate from readily available information, specifically the average occupancy rate and number of rooms within an office building. They developed a multiple linear regression model of peak occupancy rate as a function of average occupancy rate, number of rooms, and other variables that are combinations of these two variables. This model is based on occupancy data from 195 occupancy sensors in three buildings in Boulder, CO.

…