| Under the combined influences of climate change and human activities, reservoir operations are increasingly challenged by the proliferation of multi-source heterogeneous data and weakened correlations within existing knowledge frameworks. As a result, traditional knowledge management models are proving inadequate for making refined, data-driven decisions. To address issues such as fragmented knowledge and semantic disconnects in reservoir operations, this study proposes a novel method for constructing a knowledge graph that integrates large language models (LLMs) with deep learning technologies. The first step involves developing a multidimensional knowledge framework that incorporates hydrological elements, engineering attributes, operational methods, constraints, and optimization models. A knowledge extraction framework for unstructured data is then designed, featuring dynamic encoding of textual semantic features. To identify entity boundaries within reservoir operation texts, the Bert-BiLSTM-CRF joint decoding approach is employed, with attention mechanisms enhancing the extraction accuracy of specialized entities such as operational methods and models. Next, a relation extraction strategy based on semantic role labeling and dependency parsing is introduced, complemented by conflict resolution rules for aligning entities across different literature sources. The resulting reservoir operation knowledge graph, built from core Chinese journal literature, includes 1,594 entities and 927 relationships. Entity recognition achieves precision, recall, and F1-scores of 97.38%, 97.94%, and 97.66%, respectively, surpassing traditional BiLSTM-CRF and BiLSTM-CNN models by 13.25% and 13.29% in F1-score. Case studies demonstrate the efficacy of the knowledge graph in visualizing topological relationships within reservoir operation knowledge. It supports applications such as knowledge reasoning and question answering, and serves as an extensible knowledge hub for the intelligent management of basin-scale reservoir systems. This methodology provides valuable insights for the development of digital twin systems in water resources management. |