Abstract: The traditional agricultural irrigation load exhibits concentration and disorder, significantly exacerbating the imbalance between supply and demand and increasing operational costs in power systems. To enhance the economic efficiency of agricultural parks and the absorption capacity of photovoltaic resources, this study proposes a multi-agricultural park optimization scheduling model based on scenario generation using a large language model (LLM) and bi-level optimization. By characterizing the agricultural irrigation load and considering the energy demands of crop growth, the model forecasts the temporal analysis capabilities of the LLM for short-term photovoltaic power generation. Additionally, the knowledge inference capabilities of the LLM are employed to construct an agricultural water use knowledge graph. The rich semantic relationships within the knowledge graph assist the LLM in reasoning and prediction, generating more realistic scenarios for agricultural power supply, irrigation, and water storage. The bi-level optimization scheduling model, based on the generated scenarios, aims to optimize the economic operation of the parks by coupling the power supply, irrigation, and water storage across multiple agricultural parks. Finally, simulations demonstrate that the proposed method can be adopted to significantly enhance the stability of the power system operations in agricultural parks and effectively reduce operational costs.