To address problems such as insufficient multi-source data integration, lacking system-level association analysis, and weak decision interpretability in current decision-making of river regulation schemes, an intelligent decision-making method for river regulation schemes via collaboration of knowledge graph and large language model was proposed. In this method, a knowledge graph for river regulation was constructed by integrating multi-source data such as river channels, sluice operating conditions, and dyke-cluster topology. A two-stage classification model was designed (Stage 1 modeled physical features and engineering ratio parameters based on random forests; Stage 2 extracted topological relationships from the knowledge graph and utilized the Node2Vec algorithm to generate 64-dimensional graph embedding features input into a logistic regression model for optimizing decision boundary). A generative evaluation engine based on DeepSeek-R1 was developed, which combined domain-knowledge-enhanced prompt learning to generate interpretable reasoning chains through multi-step reasoning for scheme feasibility assessment and improvement recommendation generation. The proposed method was validated through multi-label classification performance comparison and ablation experiments. Verification results of 228 river channels in Nansha District, Guangzhou indicate that this method can effectively alleviate the problems of insufficient multi-source data integration, missing system-level association analysis, and weak interpretability of decision-making process in traditional methods. Its F1 score (the harmonic mean of precision and recall) reaches 0.853, and the accuracy reaches 0.921, indicating that its overall performance is superior to that of the baseline models considered. Both the topological embedding and the domain-knowledge-enhanced multi-step reasoning mechanism can effectively improve method performance. The proposed method achieves the highest F1 score and the lowest Hamming loss, verifying the effectiveness of each key module.