Prediction of building energy use in an urban case study using data driven approaches

Data driven models are widely used to perform prediction of energy consumption at building level and are of increasing importance as a complementary tool to traditional energy simulation approaches. A limited number of studies have tried to address the challenge of providing energy related information for large numbers of buildings within an urban area using data-driven models as a first and rapid step before an accurate simulation approach. This paper investigates the potential for data mining and machine learning approaches to provide estimations of the annual energy consumption at individual building level in a large urban context. The study is conducted for different building categories: Residential, commercial and mixed-use buildings, using data from the city of Geneva, Switzerland. The research methodology, replicable for similar city datasets, was developed using available online data from two different databases of approximately 88,000 buildings. A comparative analysis was undertaken in order to target the best predictive model for a given building class according to the Mean Absolute Percentage Error (MAPE). K-means clustering was employed in order to test prediction enhancement of the selected models. The final MAPE results are in the range of20-30% for all building categories.