Preliminary results of a methodological algorithm for the categorization of BIPV façades through the support of generative AI

Abstract

In recent years, the photovoltaic sector has continued growing, and in 2023 power generation from solar PV represented nearly 60% of the worldwide renewable energy produced. BIPV growth has also confirmed even if the market segmentation still faces difficulties to classify products, due to the wide range of technologies, designs, and materials. This study shows the process defined to categorize the BIPV products and prototypes through the analysis of 4,692 scientific papers indexed in Scopus, focused on BIPV façades, with the support of AI. The categorization is based on the one pro-posed by the International Energy Agency-Photovoltaic Power Systems Pro-gramme (IEA-PVPS). At first, the papers have been selected based on a keyword search. Then, fine textual information from both the Abstract and those sections where Methods and Results are described, have been automatically extracted. All the texts were arranged in a JavaScript Object Notation (JSON) structure, and corresponding embeddings were calculated via a Large Language Model (LLM) to obtain a Vector Database, the LLM then queried to retrieve the relevant documents from the vector database, according to a given query that can be useful for paper categorization. All the retrieved documents for each tested query were manually validated based on the coherence of the provided query. However, the method needs to be more refined in or-der to eliminate ambiguity in the definitions that impact on obtaining high values of consistency of the retrieved texts chunks.