文章摘要
人工智能技术(AI)在水沙数学模型中应用
Applications of artificial intelligence in mathematical models of sediment transport
投稿时间:2025-07-29  修订日期:2026-02-25
DOI:
中文关键词: 水沙数学模型  人工智能  数据同化  数据驱动神经网络  物理信息神经网络
英文关键词: mathematical model of sediment transport  Artificial Intelligence  data assimilation  data-driven neural network  physics-informed neural network
基金项目:国家自然科学基金重大项目(52595701);十四五国家研发计划项目 (2022YFC3201803)
作者单位邮编
方红卫* 清华大学水利系 100084
张文俊 清华大学水利水电工程系 
摘要点击次数: 25
全文下载次数: 0
中文摘要:
      水沙数学模型以物理方程为基础,是研究水沙运动规律、支撑流域治理与水利水电工程建设的重要工具。传统数值方法受限于计算效率,难以满足数字孪生等新场景对实时、高效预测与智能决策的需求。针对这一问题,本文围绕数据同化、数据驱动神经网络与物理驱动神经网络三条技术路线,构建“输入边界与参数、控制方程和输出结果”的分析框架,系统梳理人工智能技术在水沙数学模型中的研究进展。结果表明:(1)数据同化方法有效抑制初始和边界条件误差传播,提升预报的稳定性与精度;(2)数据驱动神经网络模型能显著缩短计算时长,适用于短时间临时预报等场景;(3)物理驱动神经网络在数据稀缺条件下具备更好的物理一致性,但在高维、多尺度问题中面临收敛与精度受限的挑战。基于此,本文提出面向工程应用的融合思路:以物理—数据混合驱动为核心,结合数据同化实现在线校正,以期兼顾计算效率与精度,为数字孪生水沙系统的实时更新与智能决策提供参考。
英文摘要:
      Mathematical models of sediment transport, grounded in physical governing equations, are essential tools for understanding sediment transport processes and supporting basin management and hydraulic engineering. However, conventional numerical methods are often constrained by computational efficiency and thus struggle to meet the demands of emerging scenarios—such as digital twins—for real-time, high-efficiency prediction and intelligent decision-making. To address this issue, this study reviews three technical routes—data assimilation, data-driven neural networks, and physics-driven neural networks—within an analytical framework that spans input boundaries and parameters, governing equations, and model outputs. The results show that: (1) data assimilation effectively suppresses the propagation of errors in initial and boundary conditions, thereby improving the stability and accuracy of forecasts; (2) data-driven neural networks can markedly reduce computation time and are suitable for short-term, rapid prediction tasks; and (3) physics-driven neural networks provide improved physical consistency under data-scarce conditions, yet still face challenges related to convergence and accuracy in high-dimensional, multi-scale problems. On this basis, a fusion-oriented approach for engineering applications is proposed, which centers on hybrid physical–data modeling combined with online correction through data assimilation, with the aim of balancing efficiency and accuracy and providing a useful reference for real-time updating and intelligent decision-making in digital-twin sediment–water systems.
  查看/发表评论  下载PDF阅读器
关闭