Building Energy Consumption Prediction Integrating Stereo Photogrammetry and GIS-Based Urban Digital Twins with Machine Learning

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Buildings account for a substantial share of global energy consumption and greenhouse gas emissions, highlighting the need for accurate large-scale building energy assessment. However, in Türkiye, the limited availability of Energy Performance Certificates (EPCs) and detailed 3D architectural building models restricts comprehensive urban energy analyses. This study proposes a novel GIS and Urban Digital Twin (UDT)-based methodology integrating stereo photogrammetry and Machine Learning (ML) to predict annual building energy consumption and evaluate rooftop solar energy potential. Initially, EPCs were integrated with architectural and photogrammetric 3D building models, while building energy parameters were validated using the BEP-TR2 calculation methodology. Validated datasets were then employed to train and compare Random Forest, XGBoost, CatBoost, and Artificial Neural Network models. For buildings lacking detailed architectural models, geometric attributes were extracted from stereo photogrammetric aerial imagery to enable energy consumption prediction. XGBoost achieved the best predictive performance, yielding an R2 of 0.82 using architectural GML data and an R2 of 0.83 using photogrammetrically derived data. Subsequently, rooftop solar radiation potential was estimated and compared with predicted annual energy consumption to assess building self-sufficiency. Results revealed that 27% of buildings were fully self-sufficient, while 20% achieved 50–70% self-sufficiency. The proposed framework demonstrates the potential of integrating UDTs, GIS, stereo photogrammetry, and ML to support scalable urban energy planning, renewable energy integration, and sustainable decision-making.