Solar Rooftop Potential Mapping in India using Deep Learning

This paper proposes a novel approach for the assessment of rooftop solar potential by applying the integration of computer vision and deep learning techniques. The process involves the implementation of a two-network system consisting of a U-Net based segmentation network for locating the roof boundary and an in-house Convolutional Neural Network (CNN) for identifying obstacles. Our solution introduces a new three-pipeline design: initial rooftop segmentation through our learned U-Net model, followed by our in-house CNN design for obstacle detection, and lastly optimized panel placement taking real-world installation restraints into consideration. The novelty stems from our hybrid approach combining feature extraction via deep learning with geometric optimization techniques, particularly in handling complex rooftop geometries and architectural obstacles. Performance evaluation on various rooftop datasets shows remarkable advancements compared to the traditional methods with 91% accuracy in boundary identification and a 82% improvement in computation time compared to the traditional methods. The system is proven to have remarkable effectiveness in practical applications with an 85% reduction in incorrect positive obstacle detection and a 60% gain in efficiency in memory usage for high resolution imagery processing [11]. This research helps in the creation of automated solar capacity calculation systems by bridging the gap between theoretical estimates of solar capacity and actual installation constraints.