Research on PV power generation forecasting in zero-carbon residential buildings under multi-weather conditions based on a PV system large language model

Abstract Accurate photovoltaic power prediction underpins the management of carbon-neutral and smart buildings’ energy. This study proposes a knowledge-enhanced hybrid forecasting model for PV power generation in zero-carbon residential buildings to address the issue of model prediction accuracy degradation caused by the lack of expert knowledge and the complexity of weather variations. Firstly, based on Deepseek-R1, a Large Language Model (LLM) for PV power generation systems in zero-carbon residential buildings (PVELLM) is constructed by fine-tuning with an expert knowledge corpus, thereby reducing the dependency on data. A specialised question-answering process in the PV field (PV-QA) has been established. This process rapidly consolidates expert knowledge across various weather conditions to enhance the model’s dynamic response capability in complex weather scenarios. A method for encoding expert knowledge is proposed to initialise a Dynamic Bayesian structure, which is dynamically optimised with data-driven techniques, allowing for the rapid construction of PV power generation prediction models adaptable to different weather scenarios. The proposed model was validated in a zero-carbon residential building. The results show that the hybrid model improved the R2 by 2.19% to 5.2% on a one-month scale training set. Under various weather conditions, the model demonstrated higher accuracy, validating its excellent dynamic response capability to complex weather changes.