Bayesian network for predicting energy consumption in schools in Florianópolis - Brazil

Abstract

It is important to study innovative approaches that consider real-world data to predict energy consumption, especially in existing buildings. This paper presents a data-driven model to predict energy consumption using Bayesian Networks. Monthly energy bills over three years were obtained from 90 public schools in Florianópolis, southern Brazil. Information such as floor-plan area, number of students, type of education, number of floors and occurrence of events were gathered for each building. The network output indicator was assessed using Energy Use Intensity based on floor-plan area or number of students. Three types of discretization methods and three network structures were tested, generating eighteen networks. A performance analysis comparing predicted as well as real Energy Use Intensity determined the Normalized Root Mean Square Error for each network and pointed out Equal Width Discretization as the best method and Naïve-Bayes as the most advantageous structure type. The discretization method had a high impact on the network performance. In addition, the Energy Use Intensity based on floor-plan area was more reliable than that based on the number of students.