A machine learning approach for predicting office energy consumption in a Mediterranean region

The underlying causes of discrepancy between building energy modelling predictions and building energy usage and performance have been the hanging fruit of several studies over the past three decades. Many of the patterns governing this divergence relate to the integration of unrealistic input parameters of occupancy behaviour in building energy models as well as the lack of feedback to designers once a building is constructed and occupied. To address these gaps, this paper presents a machine learning framework to forecast an office energy consumption in a Mediterranean climate, while taking into consideration the occupants' behavioural patterns and weather conditions. Data was collected from an office building management system and was used to train and test the learning algorithm. Three key variables were selected as the most important predictors of electricity usage, namely time of the day, outdoor air dry-bulb temperature, and indoor office space temperature. Experiments were conducted and results revealed the importance and potential of a data-driven forecasting model in efficiently generating information about the patterns that govern energy demand and enabling designers to incorporate more realistic input parameters in energy models.