Semantic-aware robust PV forecasting based on a localized enhanced LLM framework

High-quality photovoltaic (PV) power forecasting is crucial for the operation of modern power systems. However, conventional methods rely heavily on extensive historical data, leading to a significant decline in forecasting accuracy in data-scarce scenarios such as newly built plants or extreme weather conditions. To compensate for this lack of numerical data, incorporating supplementary textual and semantic information offers a promising approach. Current PV forecasting methods have yet to fully exploit the potential of this approach. Recently, Large Language Models (LLMs) have demonstrated strong generalization in time-series forecasting. Furthermore, their multimodal reasoning abilities present an opportunity to bridge the inherent gap between numerical data and natural language. Motivated by these insights, we propose a localized enhanced LLM framework for robust day-ahead PV forecasting. By incorporating a semantic self-expansion mechanism and specialized embedders into a fine-tuning architecture, the framework fuses time-series data with textual information, allowing the LLM to adaptively capture forecasting-relevant features, which effectively mitigates the accuracy decline caused by data scarcity. Empirical evaluations across geographically distinct regions in China demonstrate that our method consistently outperforms state-of-the-art models like Autoformer, PatchTST, GPT4TS, and Time-LLM, highlighting the learning capabilities and robustness of our method under data scarcity.