Identifying the most influential parameters in predicting lighting energy consumption in office buildings using data-driven method

Predicting building energy consumption is a necessary step in energy management, energy saving and optimization. Recently, data-driven models have shown promising performance in predicting energy consumption in various building energy systems, such as lighting and HVAC. Many studies used data-driven models to predict building energy performance based on the correlation between building energy use and the input variable, such as building physical characteristics, building energy systems, geographical information, and occupancy variables. Yet, the application of data-driven methods in identifying the most significant input variables for forecasting the energy usage of particular building energy systems has not been explored thoroughly. Moreover, there is a lack of guidelines for using data-driven models and feature selection in predicting building energy consumption. This research proposes a data-driven predictive framework to determine the most significant features and the optimized number of input variables in lighting energy consumption prediction as significant end-use energy in office buildings. This study uses the Commercial Building Energy Consumption Survey (CBECS) dataset as the baseline. A fourstep research methodology is designed to develop the data-driven framework for identifying the significant inputs in predicting lighting energy consumption: (1) data preprocessing, (2) feature selection, (3) prediction model development, and (4) model evaluation. The results show that to predict the lighting energy consumption in US office buildings using CBECS, the Support Vector Machine (SVM) algorithm provides the best prediction performance with an R-square value of 78% while using only the twenty most influential parameters. In addition, building physical characteristics has the highest number of significant features that impact the lighting energy use.