Data-driven predictive models for residential building energy use based on the segregation of heating and cooling days

Abstract Data-driven models can estimate the buildings’ energy consumption using machine learning algorithms. This approach works based on the correlation between energy consumption and various inputs such as weather data, occupancy schedules, heating, air conditioning, and physical properties of buildings. Seasonal changes affect buildings’ energy use. Hence, the required data for data-driven models (DDMs) during the heating and cooling days could be different. Selecting the most impactful inputs can help to choose the type and quantity of sensors for deployment that improve the model’s accuracy and minimize the costs. This paper performs feature selection for heating, cooling, hot water, and ventilation loads in residential buildings under the mixed-humid climate zone. Filter method, wrapper backward elimination, wrapper recursive feature elimination, Lasso regression, linear regression, and Extreme Gradient Boosting (XGBoost) regression are adopted for heating and cooling days, separately. We use twenty-five outputs from a computer model, and the results show that the key features for a DDM are different for heating and cooling days, and XGBoost provides the most accurate forecast. The findings of this paper are useful for selecting proper models, sensors, and inputs for model-predictive control systems during the heating and cooling seasons.