A weighted ensemble DL framework for granular RPV inventory and energy yield analysis in Pakistan

Abstract

The global imperative to triple renewable energy by 2030 underscores the vital role of Rooftop Photovoltaics (RPV). Despite high solar irradiation, Pakistan's RPV potential remains underexplored due to the poor generalizability of conventional Deep Learning (DL) models across complex urban architectures. This study addresses this limitation by proposing a weighted ensemble DL framework for large-scale RPV delineation in Islamabad, Lahore, and Karachi. Multiple DL models were trained on region-specific, high-resolution Google Satellite imagery (0.3 m spatial resolution), with final predictions synthesized through a performance-based weighted majority voting scheme. For solar resource assessment, a 10 m Digital Surface Model (DSM) was generated from Sentinel-1 data to account for urban morphology. The ensemble model demonstrated superior robustness, achieving an F1-score of 0.92, 0.96 accuracy, and a Matthews Correlation Coefficient of 0.89. Applying this framework, we identified 21,586 installations in Islamabad, 65,304 in Lahore, and 35,710 in Karachi, representing the most granular assessment to date. Analysis of annual electricity yield and carbon mitigation revealed that Karachi leads with 602 GWh (0.37 Mt CO₂ reduction), followed by Lahore at 373 GWh (0.23 Mt CO₂), and Islamabad at 141 GWh (0.09 Mt CO₂). Total annual generation from identified modules reached 1117.94 GWh, offsetting 0.693 Mt CO₂. Ground validation in Lahore confirmed high model reliability (R² = 0.98, MAPE = 6.02%). This research provides a scalable, data-driven methodology for RPV mapping in diverse developing regions, offering essential intelligence for policymakers to accelerate Pakistan's transition to a sustainable energy future.