StairLLM-Net: A Large Language Model based Approach for Photovoltaic Power Forecasting

With the expansion of large-scale grid-connected distributed photovoltaic (PV) systems, accurate photovoltaic power forecasting has become increasingly crucial. Current methods often require extensive hyperparameter tuning and demonstrate limited zero-shot capability. This study introduces StairLLM-Net, an innovative Large Language Model (LLM) based approach that utilizes LLM as a feature extractor and implements feature decoding through segmented blocks. The proposed model effectively integrates time-series data with linguistic embeddings using a patch-based methodology and incorporates a specifically designed input format to leverage the inherent reasoning capabilities of LLMs, thereby enhancing predictive accuracy at individual PV stations. The StairLLM-Net exhibits remarkable robustness and can be directly applied to previously unseen PV stations without requiring fine-tuning. Extensive experiments on real-world station datasets demonstrate the model’s exceptional cross-dataset generalization capability.