AI-Based Reconstruction of Traffic Dynamics and Building Energy Use in Seoul, South Korea

Abstract

Urban environmental conditions arise from the tightly coupled interactions among human activity, built form, and local microclimate. However, two dominant drivers of urban climate and sustainability, traffic activity and building electricity use, remain poorly observed at fine spatial and temporal scales in most cities. Sparsely distributed traffic sensors and unpublicized building energy data have limited our ability to quantify carbon emissions, anthropogenic heat release, and intra-urban energy inequality, thereby constraining evidence-based urban planning and climate policy. This dissertation addresses these longstanding observational gaps by developing an artificial intelligence (AI)-based framework that reconstructs missing layers of urban activity and integrates them into a unified, high-resolution representation of the urban system. First, a graph-based deep learning framework reconstructed nine years of hourly link-level traffic volumes across Seoul's entire road network by utilizing mobile signal- based traffic counts and probe vehicle speeds. This reconstruction enables the first long-term, citywide, bottom-up estimation of CO2 emissions, revealing pronounced spatial concentration of emissions, congestion-induced temporal amplification, and the distinct yet complementary mitigation effects of electrification and traffic-flow improvement. Second, the dissertation develops a citywide, parcel-level framework for estimating building electricity use that integrates detailed building geometry, socioeconomic context, dynamic human activity derived from mobile-phone data, and dense in situ microclimate observations. Using explainable machine learning, the analysis demonstrates that urban functional intensity and building structure—captured by land price, floor area, and construction age— form the primary determinants of electricity intensity. At the same time, human activity and microclimate conditions act as conditional amplifiers rather than dominant drivers. Notably, the results reveal a systematic decoupling between pedestrian-level outdoor air temperature and electricity demand in dense commercial and business cores, where urban canyon shading suppresses observed temperatures. However, indoor cooling demand remains high due to intensive occupancy, internal heat gains, and continuous operation. By reconstructing traffic activity and building electricity use at unprecedented resolution, this dissertation moves beyond gap-filling as an end in itself. It shows how recovered urban activity layers fundamentally reshape our understanding of emission hotspots, energy demand drivers, and climate–energy interactions within cities. The proposed framework provides a scalable foundation for quantifying congestion costs, targeting demand-side energy interventions, assessing energy inequities, and supporting urban decarbonization and heat-mitigation strategies. More broadly, this work demonstrates how AI can serve as an urban data-recovery engine, enabling cities to transition from coarse averages to spatially targeted, evidence-based environmental governance amid intensifying climate stress.