TLDR: A new AI framework, SF-SPA, automates the assessment of solar PV potential on building facades using street-view images. It corrects perspective distortion, segments facade components with vision-language models, and uses LLMs for optimal PV layout, then simulates energy yield. Validated across 80 buildings, it achieves high accuracy (6.2% error) and efficiency (100 seconds/building), demonstrating that facades, especially on high-rises, offer significant untapped solar energy potential, often exceeding rooftop contributions.
A new automated framework, SF-SPA (Semantic Facade Solar-PV Assessment), has been developed to accurately evaluate the solar photovoltaic (PV) potential on building facades. This innovative approach addresses the long-standing challenge of assessing vertical solar energy generation in dense urban environments, which often goes overlooked in favor of rooftop installations.
Traditionally, assessing PV potential on building facades has been difficult due to complex geometries, perspective distortions in street-level imagery, and the need to understand various semantic components like windows, doors, and walls. Existing methods often simplify facades into uniform planes, leading to significant overestimations of installable capacity.
The SF-SPA framework tackles these issues through a four-stage pipeline that transforms a single street-view photograph into a quantitative PV deployment assessment. The process begins with geometric rectification, where perspective distortions from street-level images are corrected. This is achieved by identifying semantic keypoints, such as window corners, to create a true-to-scale, front-on view of the facade. This crucial step ensures that all subsequent measurements and analyses are geometrically accurate.
Following rectification, the framework employs a vision-language foundation model for zero-shot semantic segmentation. This means the model can accurately delineate various architectural elements—like walls, windows, doors, and balconies—at a pixel level without requiring extensive, domain-specific training data. By simply modifying a text prompt, the system can recognize new component types, making it highly adaptable to diverse architectural styles.
The segmented information is then fed into a Large Language Model (LLM) for PV layout reasoning. Guided by a structured prompt chain, the LLM acts like a human designer, applying codified design rules to identify optimal, non-overlapping rectangular regions for PV installation. It maximizes the size of each region, ensures mutual exclusivity with obstructions, and adheres to practical constraints such as minimum module dimensions and necessary clearances. This intelligent reasoning ensures that the proposed PV layouts are both efficient and code-compliant.
The final stage involves irradiance and energy simulation using the pvlib Python toolkit. This step converts the identified usable areas into annual energy yield estimates, considering geographical coordinates, orientation, and meteorological data. The framework can perform both ideal clear-sky simulations for an upper-bound estimate and more realistic simulations using Typical Meteorological Year (TMY) datasets, which account for factors like cloud cover and temperature. This provides a quantitative measure of the facade’s solar energy potential under specific climatic conditions.
Validation across 80 buildings in four countries demonstrated the framework’s robust performance, achieving a mean area estimation error of 6.2% compared to expert annotations. The automated assessment process takes approximately 100 seconds per building, offering a substantial gain in efficiency over manual methods. The study also highlighted that while the per-unit-area efficiency of facade PV might be lower than rooftop PV, the significantly larger available facade area, especially in high-rise buildings, means facades can contribute 1.1 to 2.5 times the total energy output of rooftop PV for such structures.
While promising, the SF-SPA framework has limitations, including its assumption of planar facades, the inability to account for inter-building shading in its current 2D approach, and the manual step required for metric scaling. Future work aims to enhance geometric fidelity using 3D models, automate the scaling process, and integrate economic analysis for a more holistic decision-support tool.
Also Read:
- New Benchmark Challenges AI’s Understanding of Physical Spaces
- ViTSP: A Hybrid AI Framework for Large-Scale Traveling Salesman Problems
This automated assessment capability holds significant implications for urban energy policy and economic planning, enabling rapid, city-scale screening of facade PV potential. For more details, you can refer to the full research paper: Solar PV Installation Potential Assessment on Building Facades Based on Vision and Language Foundation Models.
