Gaussian Process Baseline Regression Models in Industrial Facilities (ST-16-C008)

Due to energy becoming a prominent topic in the sciences, creating baseline energy models for buildings has become a major area of research. Baseline models help determine dependency on ambient temperature or other parameters, while also helping show if energy reduction is due to building retrofits, occupancy, or ambient temperature. Several different methods of creating a baseline models for commercial and residential buildings, however few attempts have been made to create baseline energy models in industrial facilities, with only simple methods being analyzed. Since industrial facilities account for 33% of the annual energy usage within the United States (EIA), energy usage and finding energy saving opportunities in industrial facilities needs to be analyzed. While most industrial facilities energy usage is strongly dependent on production, some can be very dependent on weather especially if the facility is being conditioned. Even in an industrial facilities major energy consumption is used for production the energy usage might still be temperature dependent due to equipment being temperature dependent (i.e. chillers, furnaces, boilers, etc.). To understand an industrial facility’s energy usage is to create a baseline energy model. Currently only change-point regression models have been commonly used for analysis in industrial facilities. An analysis of the effectiveness of using Gaussian process regression (GPR) to develop a baseline energy usage models in industrial facilities from utility bill data and hourly data is presented in this paper.  Gaussian process regression uses a covariance matrix of the input variables to construct the model compared to using a pre-defined relationship between the input-output variables. By using a covariance matrix Gaussian process regression is more flexible than traditional parametric regression models. Two case studies are presented: using utility bill data to create a Gaussian process regression model and a using hourly data to create a Gaussian process regression model. In all cases the baseline regression models gave a CV-RMSE of 20% or lower showing that hourly data or utility bill data is capable of producing accurate baseline models defined by ASHRAE Guideline 14.