The Application of Time Series Methodology to Improve Short-Term Forecasts of Building Power Demand

• Activity Number: 241
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Jeremy Coyle*+ and Jason Trager and Jade Benjamin-Chung and Paul Wright
• Companies: University of California, Berkeley and University of California, Berkeley and University of California, Berkeley and University of California, Berkeley
• Keywords: time series ; buildings ; energy

Abstract

Forecasts of short-term energy use in buildings (baselines) are useful for a variety of applications related to attempts to reduce energy demand in buildings both overall and on days when high demand is expected. To date, efforts to develop forecasting methodologies for this application have focused on explanatory methods, which attempt to predict use with covariates. In general, these efforts have ignored accepted best practices for dealing with the challenges of time series data, including autocorrelation, nonstationarity, and seasonality. The application of classical time series approaches (e.g. seasonal decomposition, ARIMA, and exponential smoothing) to hourly power data from buildings represents an unexplored topic. Using hourly demand data from 20 buildings on the UC Berkeley campus and temperature data from a nearby weather station, we will explore the application of a number of time series methodologies to this domain. Using the framework of out-of-sample forecast evaluation, we will compare the predictive performance of such methods to the methods currently used in the building baseline literature.