Urban net-zero energy building pathways : from AI-driven consumption analytics to photovoltaic generation assessment

Abstract

Urban buildings account for approximately 40% of global energy consumption and 36% of carbon emissions, making them critical targets for achieving carbon neutrality goals. Net-Zero Energy Buildings (NZEB), which balance annual energy consumption with renewable energy generation, represent a promising target toward sustainable urban development. However, achieving NZEB at urban scale faces significant challenges due to complex inter-building interactions, computational limitations of traditional assessment methods, and the lack of integrated frameworks linking energy consumption analysis with renewable energy potential evaluation. This dissertation develops a comprehensive AI-driven framework for urban building energy analysis and photovoltaic (PV) potential assessment, providing scientific foundations for NZEB implementation in dense urban environments. The research establishes four interconnected methodological components using New York City's 1,084,465 buildings as the empirical testbed. First, an advanced Urban Building Energy Modeling (UBEM) framework integrates two-dimensional (2D)/three-dimensional (3D) geometric features with spatial proximity analysis based on Delaunay triangulation networks. Machine learning algorithms enhanced with Bayesian optimization demonstrate substantial performance improvements over default parameters, with Random Forest emerging as the optimal algorithm among six tested approaches. The framework successfully predicts both Total Energy Use and Energy Use Intensity at city scale, providing foundations for subsequent factor analysis and renewable energy assessment. Second, explainable AI techniques employing SHAP (SHapley Additive exPlanations) methodology reveal that 2D features contribute approximately 70% while 3D features contribute 30% to prediction accuracy. Inter-building shadow effects account for 4.5-10.5% of model importance, providing the systematic quantification of these previously unmeasured urban interaction effects in dense environments. Third, the Shadow-Attention Graph Neural Network (SAGNN) developed for comprehensive solar potential assessment achieves R² values ranging from 0.7295 to 0.9636, substantially outperforming traditional machine learning approaches while reducing computational time by 97.4% compared to physics-based simulations. The assessment reveals annual PV generation potential of 29,851.9 GWh from rooftops and 32,062.2 GWh from facades. However, compared to NYC's total energy consumption of 459,112.0 GWh, current PV potential remains insufficient for citywide NZEB achievement, though many areas demonstrate feasibility for NZEB. Fourth, the Multi-Layer Geo-Attention Graph Neural Network (MLGA-GNN) enables fine-grained facade analysis through vertical stratification, dividing building facades into 10 layers to capture height-dependent solar radiation variations. This innovative approach overcomes limitations of traditional methods that treat facades as homogeneous surfaces, potentially underestimating PV potential in high-radiation areas. Results reveal substantial vertical heterogeneity, with solar radiation intensity differences reaching up to 473.7 kWh/m² between top and bottom facade layers. Economic analysis indicates rooftop PV systems achieve 7-year payback periods with $71.98 billion net benefits over 25-year lifecycles, while facade systems face longer payback periods of 34.3 years under current conditions. This research contributes innovative methodologies bridging the gap between predictive accuracy and interpretable understanding in urban energy systems. The integrated framework advances from single-building analysis to city-scale assessment, from black-box predictions to explainable insights, and from isolated energy consumption modeling to comprehensive renewable energy integration evaluation. The findings provide essential scientific guidance for urban planners, policymakers, and researchers pursuing sustainable urban development through evidence-based NZEB implementation strategies.