Retrieval-Guided Photovoltaic Inventory Estimation from Satellite Imagery for Distribution Grid Planning Journal: Applied Energy
Muhao Guo, Lihao Mai, Erik Blasch, Jafarali Parol, Turki Rakan, Yang Weng · 2026
Read Retrieval-Guided Photovoltaic Inventory Estimation from Satellite Imagery for Distribution Grid Planning Journal: Applied Energy by Muhao Guo, Lihao Mai, Erik Blasch, Jafarali Parol, Turki Rakan, Yang Weng (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-22856.txt.
Introduction
Guo and co-authors estimate rooftop solar panel counts from satellite images for grid planning. Standard vision models use fixed features and falter when regions differ. Roof materials, city layouts, and imaging conditions all shift across areas. Their Solar-RAG framework retrieves similar rooftop scenes with verified labels. A vision-language model then reasons by comparing the target with those examples. Case studies at feeder level show fewer errors in voltage and hosting capacity checks. The method beats plain vision models and standalone vision-language models.
Catalog record
| CREATOR | Muhao Guo, Lihao Mai, Erik Blasch, Jafarali Parol, Turki Rakan, Yang Weng |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 69762 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.22856
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Retrieval-Guided Photovoltaic Inventory Estimation from Satellite Imagery for Distribution Grid Planning Journal: Applied Energy Muhao Guo, Lihao Mai, Erik Blasch, Jafarali Parol, Turki Rakan, Yang Weng Abstract The rapid expansion of distributed rooftop photovoltaic (PV) systems introduces increasing uncertainty in distribution grid planning, hosting capacity assessment, and voltage regulation. Reliable estimation of rooftop PV deployment from satellite imagery is therefore essential for accurate modeling of distributed generation at feeder and service-territory scales. However, conventional computer vision approaches rely on fixed learned representations and globally averaged visual correlations. This makes them sensitive to geographic distribution shifts caused by differences in roof materials, urban morphology, and imaging conditions across regions. To address these challenges, this paper proposes Solar Retrieval-Augmented Generation (Solar-RAG), a context-grounded framework for photovoltaic assessment that integrates similarity-based image retrieval with multimodal vision-language reasoning. Instead of producing predictions solely from internal model parameters, the proposed approach retrieves visually similar rooftop scenes with verified annotations and performs comparative reasoning against these examples during inference. This retrieval-guided mechanism provides geographically contextualized references that improve robustness under heterogeneous urban environments without requiring model retraining. Experiments on a globally distributed satellite imagery dataset covering six continents demonstrate that the proposed framework achieves 95.4% accuracy for PV presence detection, 86.2% accuracy for panel quantity estimation, and 80.2% accuracy for spatial localization. And, the method outperform both conventional deep vision models and standalone vision-language models. Furthermore, feeder-level case studies show that improved PV inventory estimation reduces errors in voltage deviation analysis and hosting capacity assessment. The results demonstrate that the proposed method provides a scalable and geographically robust approach for monitoring distributed PV deployment. This enables more reliable integration of remote sensing data into distribution grid planning and distributed energy resource management. 1 Introduction The global transition toward low-carbon energy systems has positioned solar photovoltaic (PV) technology as one of the fastest-growing sources of electricity generation [ 13 , 36 ] . Continuous reductions in PV costs, advances in manufacturing, and supportive policy frameworks have enabled rapid deployment across residential, commercial, and utility-scale applications worldwide [ 19 , 18 ] . As installed capacity moves toward terawatt-scale levels, distributed rooftop PV systems are increasingly changing the operation and planning of modern distribution grids [ 15 ] . Unlike centralized monitored generation assets, rooftop PV installations are often deployed by individual customers and may not be fully recorded in utility databases. This results in significant uncertainty in feeder-level PV inventories, which directly affects distribution system modeling, hosting capacity assessment, voltage regulation, protection coordination, and infrastructure investment decisions [ 37 , 30 ] . Therefore, reliable estimation of distributed PV deployment across service territories has become a requirement for modern grid planning and distributed energy resource (DER) integration studies [ 3 , 23 ] . For PV inventory estimation, existing monitoring approaches struggle to scale with the rapid and geographically heterogeneous growth of rooftop PV installations. For example, conventional methods rely on manual inspection [ 8 ] , field surveys, and specialized diagnostic equipment. But they are costly, labor-intensive, and difficult to apply across large service territories containing millions of buildings. As distributed PV continues to expand across diverse urban environments, these approaches can not maintain accurate PV deployment records. To overcome these limitations, remote sensing techniques using aerial and satellite imagery were introduced for large-scale PV monitoring [ 35 ] . By high-resolution imagery, such approaches enable automated identification of rooftop PV installations across wide geographic areas. This provides an opportunity to construct scalable PV deployment inventories that support distribution grid modeling and planning [ 29 ] . Recent advances in computer vision have significantly improved the capability of automated PV detection from satellite imagery. Deep learning models such as convolutional neural networks (CNNs) and vision transformers (ViTs) have demonstrated strong performance in solar panel detection [ 24 ] , rooftop segmentation [ 33 ] , and anomaly identification [ 12 , 7 ] . However, these methods often rely on fixed learned representations derived from training datasets that may not capture the diversity of rooftop structures and environmental conditions across different geographic regions. As a result, their performance can degrade under domain shifts caused by variations in roof materials, architectural styles, imaging resolution, weather conditions [ 5 ] , and urban morphology [ 42 , 34 ] . Addressing such variability typically requires retraining models with region-specific labeled datasets. But this increases the costs for obtaining data and limits scalability for large-scale PV monitoring applications. In parallel, recent developments in generative artificial intelligence have introduced multimodal large language models and vision–language models (VLMs) capable of semantic reasoning over visual content [ 40 , 39 ] . These models demonstrate strong zero-shot and few-shot generalization abilities and can produce structured explanatory outputs [ 10 , 9 , 11 ] , making them attractive for infrastructure monitoring tasks such as satellite-based assessment of distributed energy resources. However, inference in these models still relies primarily on internally learned representations and globally averaged visual priors. When applied to remote sensing scenarios with significant geographic variability, rooftop structures, architectural styles, and environmental conditions can differ substantially across regions, making photovoltaic identification highly dependent on locally relevant visual patterns. Without explicit access to region-specific reference information, VLM-based predictions may exhibit hallucination or misinterpretation of visual features [ 17 ] . Consequently, reliable photovoltaic inventory estimation across geographically diverse regions requires an inference framework capable of incorporating contextual visual evidence and dynamically referencing comparable rooftop cases observed under similar geographic and environmental conditions. To address this challenge, this paper proposes Solar Retrieval-Augmented Generation (Solar-RAG), a context-calibrated framework for photovoltaic assessment from satellite imagery. Unlike conventional pipelines that rely solely on parametric prediction, Solar-RAG performs retrieval-guided visual reasoning by integrating similarity-based rooftop retrieval with multimodal vision–language inference. As illustrated in Fig. 1 , the proposed architecture first identifies visually comparable rooftop scenes from a globally distributed database using visual embedding similarity [ 27 , 6 ] . These retrieved reference examples, together with their structured annotations, are then incorporated into the inference process to enable comparative reasoning over rooftop characteristics. This mechanism allows PV presence, panel quantity, and spatial distribution to be estimated relative to geographically relevant reference cases. Such estimations improve robustness under heterogeneous urban environments and enable scalable PV inventory estimation for distribution system analysis [ 38 , 21 ] . Unlike conventional pipelines that treat PV detection as a static classification or segmentation task [ 41 ] , the proposed Solar-RAG performs retrieval-guided and context-aware assessment of rooftop PV installations. The framework evaluates each query image relative to retrieved reference scenes to determine photovoltaic presence, estimate panel quantity, and localize installations within the scene, generating structured outputs suitable for downstream distribution system modeling and techno-economic analysis [ 20 ] . By decoupling knowledge expansion from fixed model parameters and shifting adaptation to a dynamically extendable retrieval layer, Solar-RAG mitigates domain shift, reduces dependence on repeated retraining, and improves interpretability. These capabilities enable scalable PV inventory estimation under geographically heterogeneous conditions, supporting more reliable integration of satellite-derived information into distribution grid planning and DER management [ 32 ] . The rest of this paper is organized as follows. Section 2 formulates the PV assessment problem. Section 3 presents the proposed methodology. Section