Volume 68

FaçadeRAG: A Retrieval-Augmented Generation Framework for Intelligent Façade Photovoltaic Planning in Urban Energy Scenarios Jixiang Cai , Zhuoyue Shen , Yi Zhang

Abstract

Building façades represent a vast yet underexploited vertical surface for urban distributed photovoltaics (PV) beyond rooftops. Advances in radiation simulation, 3D urban modeling, and deep learning-based semantic segmentation have enabled batch acquisition of street?level façade data, including irradiance, window-to-wall ratio (WWR), material classification, and obstruction conditions. However, these multi-source heterogeneous data remain as numerical matrices and categorical labels
that planners cannot readily translate into building?specific retrofit decisions. This paper presents FaçadeRAG, a retrieval-augmented generation (RAG) framework that automatically converts structured façade data into traceable, rule-compliant photovoltaic planning recommendations. The framework integrates a domain-specific knowledge base spanning policy documents, technical guides, academic literature, and engineering case studies with a hybrid retrieval pipeline combining metadata filtering, BM25 keyword matching, semantic vector search, reciprocal rank fusion (RRF), and cross-encoder reranking. A constraint-guided generation module employing GPT-4o produces structured decision outputs with explicit risk identification, regulatory compliance checking, and evidence traceability. Evaluated on 15 façade segments across three streets in Shenzhen, China, with blind assessment by six domain experts, FaçadeRAG achieved 91.4% factual accuracy, 87.3% rule compliance, 95.2% decision completeness, and 93.8% evidence hit rate, significantly outperforming standalone LLM and naive RAG baselines while approaching expert-level quality at approximately 90 times the generation speed. These results demonstrate that RAG-LLM architectures can bridge the persistent gap between urban energy data acquisition and actionable planning decisions, offering a reusable paradigm for AI?driven building energy scenario assessment in high?density built environments.

Keywords façade photovoltaic, retrieval-augmented generation, urban energy planning, building-integrated photovoltaics, large language model