Deep Learning-Based Commercial Building Energy Consumption Prediction Model Considering Occupant Density Using Mobile Network Big Data: A Case Study of Seoul

Abstract

This study proposes a multidimensional modelling framework to estimate annual energy consumption in commercial buildings across Seoul by integrating physical, environmental, economic, and behavioural variables. While previous research has predominantly focused on structural and environmental factors, this study addresses a critical gap by incorporating high-resolution spatial data and temporally disaggregated floating population variables derived from mobile network data, segmented into three time periods: late night, daytime, and evening. This enables a more refined representation of occupant dynamics in dense urban settings. Separate regression models for annual electricity and gas/heating consumption per unit floor area provide a contextual and scalable analytical approach. In the next phase of the research, a non-linear modelling technique based on a Multi-Layer Perceptron (MLP) will be implemented, along with SHAP (SHapley Additive exPlanations) analysis to enhance model interpretability. These advanced methods aim to capture complex interdependencies among variables and improve predictive performance. Ultimately, the proposed framework offers a robust foundation for data-driven energy policies and smart urban planning strategies in high-density metropolitan areas.