中国水利

大语言模型（LLMs）是近年来人工智能领域的重大突破，依托Transformer架构与自注意力机制，在超大规模参数下涌现出接近人类的自然语言理解能力，为人类认知、思考、判断和决策提供辅助。当前大语言模型在垂直细分领域的应用已成为热点，特别是基于MOE融合架构的DeepSeek开源发布，为行业大模型应用提供了更为便捷的技术路径，进一步推动了相关研究热潮。“四预”是基于数字孪生水利建设的新型水利智能业务应用，具有专业性强、业务链条长、系统架构复杂等特点，功能完备，但在易用性方面仍有优化空间。基于大语言模型的理解和推理能力分析，首次提出了大模型智能交互L0至L3级分类体系，以意图识别和智能调用为切入点，研究其支撑“四预”平台的交互应用场景和实现技术路径，提出了通过优化“预设内容”和叠加具体问题增强大模型输出确定性的方法，并在通用大模型上进行测试，探索大模型智能调用“四预”平台专业模型的路径，为提升防洪“四预”的交互友好性提供了可行方案，同时也为大语言模型在水利智能业务中的深度应用提供参考。

Large language models(LLMs) have emerged as a significant breakthrough in artificial intelligence in recent years. Leveraging the Transformer architecture and self-attention mechanisms, these models exhibit near-human natural language understanding capabilities at an ultra-large scale, assisting human cognition,reasoning, judgment, and decision-making. Currently, the application of LLMs in specialized domains has become a focal point, especially with the open-source release of DeepSeek based on the Mixture of Experts(MOE) architecture, which offers a more accessible technical pathway for industry applications and further stimulates related research. The “four pres”(forecasting, early warning, pre-planning, and emergency response)in flood control represent a novel intelligent business application in water conservancy based on digital twin technology. This system is characterized by strong specialization, lengthy business chains, and complex system architecture. While functionally comprehensive, there remains room for improvement in usability. Based on an analysis of the understanding and reasoning capabilities of large language models, this study proposes, for the first time, a classification system for intelligent interaction with large models, ranging from L0 to L3 levels.Focusing on intent recognition and intelligent invocation, the research explores application scenarios and technical implementation paths that support the “four pres” platform. Methods to enhance the output certainty of large models are proposed by optimizing “preset content” and incorporating specific problem overlays,which are tested on general large models. The study also explores pathways for large models to intelligently invoke professional models within the “four pres” platform, providing feasible solutions to improve interactive friendliness. Additionally, this research offers valuable references for the deep application of large language models in intelligent water conservancy business.