Ultra-short-term photovoltaic power prediction based on reprogrammed large language models

Ultra-short-term photovoltaic (PV) power forecasting is critical for power grid stability, yet existing deep learning models suffer from two major bottlenecks: insufficient robustness under complex weather conditions and poor cross-station generalization. To address this, we propose a reprogrammed large language model framework, SolarTime-LLM. This framework innovatively designs a Multi-feature Gating Fusion Reprogramming (MFGFR) module to capture multi-scale dynamics during high-volatility weather and integrates a domain-specific Simplified Prompt-as-Prefix (SPaP) to achieve Pareto optimality (Pareto, 1964) between prediction performance and inference costs within the LLM reprogramming paradigm. Comprehensive experiments on three real-world PV sites across different climate zones demonstrate that SolarTime-LLM significantly outperforms baseline models in all intra-site tests, maintaining top accuracy even under limited-data conditions; notably, on the data-limited site (approx. 11 months), its winter RMSE was reduced by 9.17% compared to the next-best model. Critically, in rigorous cross-station zero-shot transfer tests, where most baselines failed due to generalization collapse, SolarTime-LLM still achieved exceptional accuracy (up to 92.95%), nearly matching locally-trained performance. Furthermore, this study reveals the efficiency mechanism of SPaP from an attention mechanism perspective. SolarTime-LLM provides a high-accuracy, robust, and rapid cold-start solution for new PV stations, holding significant value for enhancing renewable energy integration.