The water levels associated with mountain floods exhibit rapid fluctuations within small watersheds, necessitating extensive data on various factors influencing such disasters to facilitate real-time forecasting. This study investigates the application of Long Short-Term Memory (LSTM) networks for mountain flood forecasting, designing a watershed-internal Knowledge Graph (KG) and Large Language Model (LLM) that encompass watershed relationships and internal information structures. We have developed a hydrological KG for the Qixi Reservoir and Qiaodongcun forecasting points located in Zhejiang Province, China, to systematically organize water conservancy data, identify significant disaster-related factors, optimize the input hydrological data, and determine the most effective combination of input data for forecasting water levels. Additionally, we have implemented Recurrent Neural Networks (RNN) and Gated Recurrent Units (GRU) for comparative analysis with LSTM. The findings indicate that the LSTM model, when integrated with the watershed-internal KG and LLM, can effectively incorporate critical elements influencing water level changes, the accuracy of the LLM-KG-LSTM model is enhanced by 3% compared to the standard LSTM model, and the LSTM series outperforms both RNN and GRU models, Our method will guide future research from the perspective of focusing on forecasting algorithms to the perspective of focusing on the relationship between multi-dimensional disaster data and algorithm parallelism.