Forecasting energy trends and peak usage at the University of Virginia

Forecasting energy trends, especially peak usage, is a valuable and necessary part of energy management. Accurate prediction allows for the control and alleviation of overuse during peak times with the implementation of energy efficiencies. Using hourly kilowatt data from over 200 buildings on the University of Virginia's campus this paper examines the most effective techniques for developing both individual building and overall grid use energy models with a specific focus on predetermining peak usage points. This paper proposes that the expectation-maximization algorithm within the state-space framework is the most effective method for smoothing missing values, a common occurrence when working with metered energy data. Next, this paper covers two separate methods for creating forecasting models. The first, a linear model, was found to be most successful in predicting campuswide energy usage, with the inclusion of features of temperature, humidity, school session, and three temporal variables. The second method, found to be most successful when forecasting short term individual building energy use, uses a seasonal autoregressive integrated moving average (SARIMA) model. Finally this paper delves into the intricacies involved in clustering buildings based on their energy usage trends rather than building use, and the implications that arise from such a process. The conclusions made in this paper can be rescaled and applied to larger energy systems outside the university setting.