An Improved Procedure for Developing a Calibrated Hourly Simulation Model of an Electrically Heated and Cooled Commercial Building

Abstract

With the increased use of building energy simulation programs, calibration of simulated data to measured data has been recognized as an important factor in substantiating how well the model fits a real building. Model calibration to measured monthly utility data has been utilized for many years. Recently, efforts have reported calibrated models at the hourly level. Most of the previous methods have relied on very simple comparisons including bar charts, monthly percent difference time-series graphs, and x-y scatter plots. A few advanced methods have been proposed as well which include carpet plots and comparative 3-D time-series plots. Unfortunately, at hourly levels of calibration, many of the traditional graphical calibration techniques become overwhelmed with data and suffer from data overlap. In order to improve upon previously established techniques, this thesis presents new calibration methods including temperature binned box-whisker-mean analysis to improve x-y scatter plots, 24-hour weather-daytype box-whisker-mean graphs to show hourly temperature-dependent energy use profiles, and 52-week box-whisker-mean plots to display long-term trends. In addition to the graphical calibration techniques, other methods are also used including indoor temperature calibration to improve thermostat schedules and architectural rendering as a means of verifying the building envelope dimensions and shading placement. Several statistical methods are also reviewed for their appropriateness including percent difference, mean bias error (MBE), and the coefficient of variation of the root mean squared error. Results are presented using a case study building located in Washington, D.C. In the case study building, nine months of hourly whole-building electricity data and site-specific weather data were measured and used with the DOE-2. 1D building simulation program to test the new techniques. Use of the new calibration procedures were able to produce a MBE of-0.7% and a CV(RMSE) of 23. 1 % which compare favorably with the most accurate hourly neural network models.