Distributed photovoltaic (DPV) systems are vital for promoting renewable energy and achieving energy independence. Accurate DPV power forecasting enhances system scheduling and reduces energy losses. However, intermittent solar radiation and the scattered nature of DPV systems complicate accurate forecasting. To overcome these challenges, this paper applies a large language model-based forecasting framework Time-LLM, which pioneers the integration of frequency-domain analysis and language-model reasoning for distributed photovoltaic (DPV) power forecast. Time-LLM first employs Fast Fourier Transform (FFT) to extract periodic characteristics from historical power data, integrating these insights into a natural language-formatted prompt. Subsequently, numerical time-series data are segmented into shorter patches, linearly mapped, and then aligned with textual prompts using a multi-head attention mechanism. The frozen pre-trained Qwen2.5-3B model processes the aligned embeddings, leveraging its powerful pattern recognition and reasoning abilities. Finally, a linear projection layer translates the large language model (LLM)'s high-dimensional outputs into accurate power forecasts. Experimental results demonstrate that Time-LLM significantly outperforms conventional methods in both short-term and long-term forecasting scenarios. The study highlights the innovative potential of LLMs in renewable energy forecasting, offering a scalable and efficient forecasting solution without dependence on costly external data.