Building Energy Consumption Prediction: A Machine Learning Approach with Feature Selection

Building energy management system is required to achieve sustainable energy in the context of global environmental and climate change challenges. Energy consumption prediction is a typical procedure used to manage the energy consumption in buildings. By accurately predicting energy consumption, the effectiveness of various energy-saving measures can be assessed, and corresponding optimization strategies can be developed. This paper presents a machine learning-based energy consumption prediction in buildings. We evaluate nine machine learning models to predict energy consumption. There are 52 features characterizing the dataset, such as the building's energy rating, environmental impact, number of rooms, and lighting descriptions. These features are selected based on their potential relevance to energy consumption, and cover the physical properties of the building, energy usage characteristics and equipment parameters, thus providing the basis for a comprehensive analysis. By performing feature selection using mutual information, we can reduce the number of features into eight, thereby reducing the complexity of the models. The model evaluation is performed using the coefficient of determination, R2 and the Root Mean Square Error (RMSE) metrics. Simulation results show that selecting the features can reduce the complexity of the model while resulting in relatively acceptable performance.