Predicting solar potentials using machine learning: a data-driven approach to map solar energy potential in the urban fabric. The case of Bologna

Background. The EU’s Green Deal and EPBD recast require rapid urban PV uptake—especially on rooftops—yet dense historic fabrics hinder city-scale assessment. Objective. This dissertation develops a reproducible, open-data, machine-learning workflow to predict technical rooftop solar potential (annual in-plane irradiation, kWh·m⁻²·yr⁻¹) and deliver planning-grade maps for Bologna (Italy). Methods. Building footprints (OSM), terrain-based sky proxies (SVF, multi-azimuth shadow index, local relief), roof morphometry (slope; circular aspect via sin/cos), 50 m land-cover context (WorldCover, Urban Atlas, tree cover), and city-scale climate baselines are harmonized (EPSG:25832). Physics-based annual irradiation labels are computed on a stratified sample (~1,525 roofs). Gradient-boosted trees (LightGBM) are trained with grouped cross-validation (district × morphology cluster), supported by model-agnostic explainability (SHAP). Transferability is tested using Leave-One-Zone-Out (LOZO) and Leave-One-Cluster-Out (LOCO). Results. Cross-validated performance reaches R² ≈ 0.63 ± 0.11, RMSE ≈ 25 kWh·m⁻²·yr⁻¹, and MAPE ≈ 0.9%; calibration on hold-outs yields slope ≈ 0.954 and intercept ≈ 82. Under LOZO/LOCO, R² ≈ 0.47/0.42, confirming usable transfer with guardrails. City-wide inference covers 48,688 roofs; the 75th percentile (Q3 ≈ 1,782 kWh·m⁻²·yr⁻¹) supports a stable public threshold of ≥1,800 kWh·m⁻²·yr⁻¹ for Tier-1 shortlisting. Explainability confirms physics-consistent drivers, with southness (cosine aspect) and SVF dominating while slope, shadow index, and local relief moderate outcomes. Contributions and implications. The work delivers a transparent open-data pipeline, dual zoning for evaluation and fair targeting, interpretable ML attribution, and decision-ready products (maps, uncertainty cues, shortlist rules) to support municipal screening and REC formation in heritage-sensitive contexts, with portability to peer European cities.