PDF(1993 KB)
PDF(1993 KB)
PDF(1993 KB)
大模型与水文模型协同应用关键技术研究及展望
Key Technologies and Prospects for Synergistic Application of Large Language Models and Hydrological Models
将大模型与水文模型结合,可提高水文模型的适应性与拓展性,降低模型使用门槛,加速水文科学的智能化进程。在搜集整理水文模型资料并分类阐述其特点的基础上,指出水文模型存在可移植性差、对数据质量敏感、极端事件预测能力弱等不足;明确其智慧化需融合知识辅助决策、优化交互方式以增强多模型协同解决复杂问题的需求;结合当前大模型的研究应用现状,指出水文模型与大模型具备优势互补特性。提出了大模型与水文模型单向耦合和双向耦合的协同方式;结合协同方式,基于自然下垫面与城市下垫面的湖泊入流复合水文过程,构建大模型与水文模型耦合的思路框架。这种耦合不仅以智能接口形式降低了专业水文模型的应用门槛,更通过动态参数优化与协同计算机制,提升了复杂水文过程模拟的效率和应急决策的精准性。
[Objective] Traditional hydrological models are mostly designed for specific scenarios, representing simplifications of complex physical processes or fitting of historical data. They require extensive modeling and driving data, and have high technical threshold for use, which largely limits their widespread application. Large language models, with their powerful understanding and generative capabilities, demonstrate superiority in the framework design and interaction processes of complex systems. Integrating large language models with hydrological model systems can provide an intelligent engine for the data collection, model construction and calibration, and simulation result analysis required by hydrological models. This enhances the adaptability and extensibility of hydrological models and lowers the model usage threshold. Therefore, this study explores the synergistic application paths for large language models and hydrological models and provides a prospect for future research priorities in their synergy. It aims to provide support and technical reference for the synergistic development of the two models, reduce the usage threshold of hydrological models, and accelerate the intelligent transformation of hydrological science. [Methods] Starting from hydrological models, based on the collection and organization of current hydrological model information and the classified elaboration of their characteristics, this study identifies problems such as poor portability, sensitivity to data quality, and weak predictive capability for extreme events. Additionally, it clarifies three requirements for their intelligent transformation: integrating knowledge-assisted decision-making, optimizing interaction methods to lower the usage threshold, and enhancing multi-model synergy to address complex problems. Combining the current research and application status of large language models, this study reviews the technical trends of mainstream large language models and analyzes the challenges in their practical application, including a lack of domain-specific knowledge, weak business reasoning capability, and difficulties in collaborating with professional tools. Finally, the study points out that hydrological models and large language models have complementary advantages, making their coupling inevitable. An application example is provided, focusing on the composite hydrological process of lake inflow from natural and urban underlying surfaces. [Results] This study proposes synergistic approaches for unidirectional and bidirectional coupling between large language models and hydrological models. This enables the effective utilization of large language models across various application scenarios of hydrological models, provides support and technical reference for their synergistic development, and contributes to lowering the usage threshold of hydrological models and accelerating the intelligent transformation of hydrological science. Combining the synergistic approaches for unidirectional coupling and bidirectional coupling, and based on the composite hydrological process of lake inflow from natural and urban underlying surfaces, a conceptual framework for coupling large language models and hydrological models that is universal and innovative is constructed. Additionally, specific implementation procedures are provided, offering a reusable methodological reference for studying complex hydrological processes. [Conclusion] The coupling technology of large language models with hydrological models, through continuous exploration and research, has preliminarily demonstrated the application potential of integrating large language models with hydrological simulation and prediction. Large language models, leveraging their advantages in semantic parsing, dynamic parameter calibration, and multi-model synergistic scheduling, have expanded the functional boundaries of hydrological models in aspects such as real-time interactive response, cross-scale coupled simulation, and emergency decision-making support. This integration not only lowers the application threshold of professional hydrological models through intelligent interfaces, but also enhances the efficiency of simulating complex hydrological processes and the accuracy of emergency decision-making through dynamic parameter optimization and synergistic computing mechanisms. The coupling of these two models demonstrates rapid adaptability to novel environmental conditions and the characteristics of autonomous reasoning and optimization, paving an innovative path for establishing a “perception-simulation-decision” full-chain integrated digital twin watershed system.
大模型 / 水文模型 / 协同方式 / 单向耦合 / 双向耦合
large language models / hydrological models / synergistic approaches / unidirectional coupling / bidirectional coupling
| [1] |
罗贤, 李运刚, 季漩, 等. 中国国际河流水文地理研究进展[J]. 地理学报, 2023, 78(7):1703-1717.
中国发育了亚洲主要的国际河流，丰富的跨境水资源在区域“水—能源—粮食—生态”安全维持中发挥着重要作用。近几十年来，受全球气候变化特别是大规模水利水电工程建设驱动，国际河流区水文及生态过程变化与跨境影响等问题备受关注。国内对这些问题的研究，重点聚焦于水文及生态过程变化规律与变化归因、跨境影响与安全风险调控，探讨跨境流域“水—能源—粮食—生态”纽带关系，构建适应全球变化的跨境水资源协调机制等方面，取得了突出的研发成效。面对全球变化影响下日益突出的跨境水安全与生态安全风险问题，国际河流的水文地理研究，更需要借助空天地一体化精准监测技术、现代空间地理信息技术和智能技术等，通过提供可量化、可参与、可公开的研发成果，更好地为国家对国际河流的合理利用与地缘合作、健康维持与风险管控、跨境水外交与环境外交等提供科学依据和决策支持。
(
Most of Asian major international rivers originate from China. Their abundant transboundary water resources play an important role in regional "water-energy-food-ecology" security. In recent decades, influenced by global change, especially by the construction of large hydraulic and hydroelectric engineering, the changes in hydrological and ecological processes and their transboundary impacts in the international river regions have attracted more and more attention. The research on these issues in China has achieved prominent results in several aspects, including the changes in hydrological and ecological processes and their attributions, the transboundary impacts and risk regulation, the "water-energy-food-ecology" nexus in transboundary watersheds, and the construction of transboundary water resources coordination mechanism to adapt to global changes. In the face of the increasing risks of transboundary water security and ecological security under global change, the hydro-geographical research on international rivers needs to make use of space-air-ground integrated monitoring technology, modern spatial geographic information technology, intelligent technology, and so on. By providing quantifiable, participatory, and public results, these researches can better provide scientific basis and decision support for the rational utilization of international rivers and geopolitical cooperation, health maintenance and risk control, transboundary water diplomacy and environmental diplomacy. |
| [2] |
罗文兵, 王修贵, 乔伟, 等. 基于水文水动力耦合模型的平原湖区土地利用变化对排涝模数的影响[J]. raybet体育在线
院报, 2018, 35(1): 76-81.
为准确模拟平原湖区土地利用变化对排涝模数的影响,选取湖北省四湖流域螺山排区作为研究区,利用构建的SCS-MIKE11耦合模型计算不同时期土地利用类型下的排涝模数,分析土地利用变化对排涝模数的影响,并通过设置不同水旱比、水面率和地面硬化率的组合,对土地利用变化条件下的排涝措施进行模拟优化。结果表明:在10 a一遇的1 d暴雨3 d排除和3 d暴雨5 d排除的标准下, 2011年土地利用方式下求得的排涝模数比1994年求得的排涝模数大,分别增加了159.3%和33.6%;在保持水旱比和水面率不变的情况下,地面硬化率每增加1%,1 d和3 d暴雨下排涝模数分别增加0.005 m<sup>3</sup>/(s·km<sup>2</sup>)和0.003 m<sup>3</sup>/(s·km<sup>2</sup>);在保持水旱比和地面硬化率不变的情况下,水面率每增加1%,1 d和3 d暴雨下排涝模数分别减小0.016 m<sup>3</sup>/(s·km<sup>2</sup>)和0.012 m<sup>3</sup>/(s·km<sup>2</sup>);在保持水面率和地面硬化率不变的情况下,水旱比每增加0.1,1 d暴雨和3 d暴雨下排涝模数分别减小0.004 m<sup>3</sup>/(s·km<sup>2</sup>)和0.003 m<sup>3</sup>/(s·km<sup>2</sup>)。因此,除了增加排涝泵站的排涝流量外,减少地面硬化率(例如采用透水路面)、增加水面率和水旱比也是除涝减灾的有效措施。研究成果可为土地利用变化条件下平原湖区排涝模数的确定和排涝措施的制定提供参考。
(
An SCS-MIKE11 coupled model was built to calculate the drainage moduli of plain lake area with different land-use patterns, and the effect of land-use change on drainage modulus was analyzed. Drainage measures in the background of land-use change were simulated and optimized by setting different combinations among proportions of paddy field to dry land area, water surface ratio, and ground hardening rate. The Luoshan drainage area in the Four-lake Drainage Basin in Hubei Province was selected as study area. Results revealed that the drainage modulus in 2011 was larger than that in 1994 under the same land-use pattern. According to the water logging control standards (10-year rainstorm for one day and draining for three days, 10-year rainstorm for three days and draining for five days), the drainage modulus increased by 159.3% and 33.6% respectively from 1994 to 2011. When water surface ratio and proportion of paddy field to dry land area kept unchanged, the drainage modulus under one-day rainstorm and three-day rainstorm increased by 0.005 m<sup>3</sup>/(s·km<sup>2</sup>) and 0.003 m<sup>3</sup>/(s·km<sup>2</sup>) respectively as ground hardening rate increased by 1%; the drainage modulus under one-day rainstorm and three-day rainstorm decreased by 0.016 m<sup>3</sup>/(s·km<sup>2</sup>) and 0.012 m<sup>3</sup>/(s·km<sup>2</sup>) as water surface ratio increased by 1% when ground hardening rate and proportion of paddy field to dry land area kept unchanged; the drainage modulus under one-day rainstorm and three-day rainstorm decreased by 0.004 m<sup>3</sup>/(s·km<sup>2</sup>) and 0.003 m<sup>3</sup>/(s·km<sup>2</sup>) as proportion of paddy field to dry land area increased by 0.1 when ground hardening rate and water surface ratio remained unchanged. Therefore, reducing the ground hardening rate by permeable pavement for example, and increasing the water surface ratio and proportion of paddy field to dry land area in addition to increasing the drainage discharge of pumping stations are effective measures of reducing waterlogging disaster loss. The research results could be used as a reference for reasonable determination of the drainage modulus and formulation of waterlogging control measures under the condition of land-use change in plain lake areas.
|
| [3] |
苟娇娇, 缪驰远, 徐宗学, 等. 大尺度水文模型参数不确定性分析的挑战与综合研究框架[J]. 水科学进展, 2022, 33(2): 327-335.
(
|
| [4] |
|
| [5] |
张赛亚, 张珂, 晁丽君, 等. 考虑地形与理化性质的土壤关键水力特性多种模型构建与比较[J]. 河海大学学报(自然科学版), 2024, 52(3): 42-50.
(
|
| [6] |
郭怡, 吴鑫淼, 郄志红, 等. 基于BP神经网络的MIKE SHE模型参数率定[J]. raybet体育在线
院报, 2019, 36(3): 26-30.
为了更精细地对水文全过程进行描述和解析,更准确地构建分布式水文模型,以丹麦Karup流域为例,对MIKE SHE模型的饱和导水率、饱和带水平水力传导系数、河床透水系数进行了参数率定,模拟流域的日径流过程。结果表明:基于BP神经网络反分析的参数率定方法比MIKE SHE模型参数自动率定计算得到的均方根误差RMSE小,模型效率系数Ens更接近1;采用BP神经网络反演率定参数后,3组测试样本的日径流模拟过程的RMSE分别为0.04,0.03,0.08 m<sup>3</sup>/s,Ens均为0.99,且模拟结果能较好地反映径流的实际变化趋势。因此,这种基于BP神经网络反分析的参数率定方法对构建分布式水文模型具有一定的价值。
(
In order to describe and interpret hydrological processes in more detail, and at the same time to construct a more accurate distributed hydrological model, we took the Karup watershed in Denmark as an example and calibrated three parameters of MIKE SHE model, namely, saturated hydraulic conductivity, saturated horizontal hydraulic conductivity, and leakage coefficient of river bank, and simulated the daily runoff process in the watershed. Results demonstrate that the root mean square error (RMSE) obtained by the method of parameter calibration based on BP neural network is smaller than that by automatic parameter calibration in MIKE SHE model, with the model efficiency coefficient Ens closer to 1. Having been treated by parameter calibration by BP neural network, the values of RMSE of daily runoff of three test samples are 0.04 m<sup>3</sup>/s, 0.03 m<sup>3</sup>/s, and 0.08 m<sup>3</sup>/s, respectively, and the value of Ens is 0.99. As the simulated runoff displays a trend in agreement with the real runoff, the back analysis method of parameter calibration based on BP neural network is of certain value in runoff simulation.
|
| [7] |
李琳, 文雄飞, 谭德宝, 等. 盐湖流域空-天-地立体监测系统结合VIC模型径流模拟初探[J]. raybet体育在线
院报, 2022, 39(8):126-132.
盐湖流域属于无资料地区且基础资料稀缺,使用传统方法开展水文模拟存在困难。为了研究盐湖流域生态环境、水文气象等变化,综合利用卫星遥感、无人船等技术,建立盐湖流域空-天-地立体监测系统,构建盐湖水位-面积-容积关系曲线,并通过月尺度遥感数据解译了盐湖月尺度湖泊面积,推算了盐湖蓄变量和盐湖入湖月尺度径流量。通过收集并整理中国区域地面气象要素驱动数据集CMFD和MODIS LAI叶面积指数等数据,构建了基于多源数据的VIC模型,开展盐湖流域湖泊径流模拟研究。结果表明,盐湖流域径流模拟纳什系数NSE和相对误差分别为 0.87和11.81%,表现出实际径流的丰枯变化情况。盐湖流域空-天-地立体监测系统结合VIC模型能够模拟盐湖流域的月尺度径流变化过程,对无资料区域河湖水文模拟具有一定参考意义。
(
Hydrological simulation for the Salt Lake basin encounters many difficulties by using traditional methods due to the lack of basic data.An air-space-ground stereoscopic monitoring system for the Salt Lake was established using satellite remote sensing technology,unmanned ships and other technologies to study the changes of ecological environment and hydrometeorology.The water level-area-volume relationship curve was obtained,the monthly-scale lake area was interpreted through monthly remote sensing data,and the storage variable as well as the monthly-scale runoff into the lake were deduced.Furthermore,a VIC(Variable Infiltration Capacity)model based on multi-source data was constructed to simulate the lake runoff by collecting and sorting out the Chinese regional ground meteorological element-driven data set CMFD and MODIS LAI(leaf area index).Results reveal that the Nash coefficient NSE and relative error of the simulated runoff of the Salt Lake basin are 0.87 and 11.81%,respectively,which reflects the change of the actual runoff.The air-space-ground stereo monitoring system combined with the VIC model could simulate the monthly-scale runoff change process of the Salt Lake basin,and hence offering reference for the hydrological simulation of rivers and lakes in areas with no data.
|
| [8] |
李晓英, 张金辉, 赵洪杰. CASC2D水文模型参数全局敏感性分析[J]. raybet体育在线
院报, 2025, 42(6):111-117.
目前CASC2D水文模型的调整参数工作主要为人工试错法,缺乏模型参数的全局敏感性分析。为解决此问题,研究采用Sobol指数法,以贺兰山东麓流域苏峪口水文站以上区域为研究区,基于洪峰峰现时间、洪峰流量以及确定性系数3个指标,对CASC2D水文模型的8个主要参数进行全局敏感性分析。结果表明:对峰现时间指标全局影响最大的3个参数分别为饱和水力传导度K<sub>s</sub>、河道糙率系数n<sub>c</sub>以及土壤缺水量M<sub>d</sub>,且3个参数与峰现时间指标呈正相关关系;对洪峰流量指标和确定性系数指标全局影响最大的3个参数为饱和水力传导度K<sub>s</sub>、毛细管压力水头H<sub>c</sub>以及河道糙率系数n<sub>c</sub>,且参数与洪峰流量指标呈正相关关系;而饱和水力传导度K<sub>s</sub>与确定性系数呈负相关关系,毛细管压力水头H<sub>c</sub>以及河道糙率系数n<sub>c</sub>与确定性系数呈正相关关系。由于存在参数间的互相影响,在参数率定过程中需要根据一阶影响指数、总效应指数结果得到的各参数独立敏感度和互相影响程度,对参数变化方向和变化范围进行合理选择。研究成果可为模型参数率定工作提供参考。
(
|
| [9] |
汪涛, 徐杨, 曹辉, 等. 基于LSTM的三峡-葛洲坝梯级电站超短期水位预测[J]. raybet体育在线
院报, 2025, 42(4):80-86.
三峡-葛洲坝梯级电站的水位预测关系到电站安全稳定运行和综合效益发挥,然而在动静库容计算体系转换关系复杂、电站下游非恒定流等多种因素的综合影响下,传统方法在短期水位预测过程时难以跟踪,在电站承担调峰、调频任务及复杂工况下有突破调度规程及开闸的风险,从而引发工程安全风险和经济损失。采用长短时记忆网络(LSTM)深度学习方法,建立了三峡-葛洲坝梯级电站超短期水位预测模型,利用水位、入库流量、出力数据预测电站超短期的水位过程,并通过大调峰工况数据对模型预测精度进行应用分析。研究结果表明该模型总体精度较高、稳定性和适应性较好,在不同调峰工况下预测精度稳定,但在水位极值处预测结果往往会出现均化现象。三峡、葛洲坝上游水位24 h预测平均误差均<0.05 m。研究成果可为梯级电站精细化调度提供技术支撑。
(
|
| [10] |
由宇军, 白云岗, 卢震林, 等. 缺资料地区可解释性混合机器学习模型中长期径流预报[J]. raybet体育在线
院报, 2025, 42(7):52-59.
在无资料地区气象、水文等观测资料缺乏,影响径流预报的准确性,直接影响水文预报和防汛抗旱工作的开展。分析无资料地区现有的降水、气温和径流数据在中长期预报中的适用性,进而实现径流预报非常重要。分别采用卷积神经网络算法(CNN)、双向门控循环神经网络(BiGRU)、注意力机制(Attention)和优化粒子群算法(IPSO)构建CNN-BiGRU-Attention、IPSO-CNN-BiGRU-Attention组合模型,再与门控循环单元模型(GRU)和ABCD水量平衡模型进行对比分析,并在玉龙喀什河进行综合评估,并结合SHAP可解释性机器学习方法探究最优模型中输入特征对径流影响的贡献程度。结果表明:加入降水和气温的组合模型IPSO-CNN-BiGRU-Attention预测精度整体优于CNN-BiGRU-Attention、GRU模型,与实际值能够较好地吻合;随着预见期的增加,提出的组合模型在验证期内均方根误差(RMSE)、平均绝对误差(MAE)、平均绝对百分误差(MAPE)、纳什效率系数(NSE)分别为2.11、1.32 m<sup>3</sup>/s、73.76%和0.94,并且在前3个月预报精度最高。该方法在缺资料地区月径流预报中具有较好的效果。
(
|
| [11] |
兰小机, 贺永兰, 武帅文. 基于RF-BiLSTM模型的河流水质预测[J]. raybet体育在线
院报, 2024, 41(7):57-63,71.
水环境中过量的氮、磷和高锰酸盐会对流域造成严重污染,准确预测这三类指标的含量对流域污染治理具有重要意义。然而,现有的模型预测精度低,输入因子的选择缺乏数理依据。基于此,以邕江为研究区域,提出一种RF-BiLSTM的混合网络模型。该模型具有利用RF算法提取水质指标最优特征和利用BiLSTM模型提取输入数据的时间特征的优势,采用先降维后预测的方式对TN、TP和 COD<sub>Mn</sub>进行预测,并将深度学习中的CNN、LSTM、BiLSTM和RF-LSTM作为基准模型与本研究所提模型作对比研究。研究结果表明,本研究模型预测TN、TP和COD<sub>Mn</sub>的平均绝对百分比误差(MAPE)分别达到了4.330%、6.781%和7.384%,均低于其他基准模型,预测结果具有较高的准确性和实用性,可为水环境的污染治理提供有效的技术支持。
(
|
| [12] |
|
| [13] |
|
| [14] |
We introduce Graph of Thoughts (GoT): a framework that\nadvances prompting capabilities in large language models\n(LLMs) beyond those offered by paradigms such as \nChain-of-Thought or Tree of Thoughts (ToT). The key idea and \nprimary advantage of GoT is the ability to model the information \ngenerated by an LLM as an arbitrary graph, where units of \ninformation (\"LLM thoughts\") are vertices, and edges correspond\nto dependencies between these vertices. This approach enables \ncombining arbitrary LLM thoughts into synergistic outcomes, \ndistilling the essence of whole networks of thoughts,\nor enhancing thoughts using feedback loops. We illustrate\nthat GoT offers advantages over state of the art on different\ntasks, for example increasing the quality of sorting by 62%\nover ToT, while simultaneously reducing costs by >31%.\nWe ensure that GoT is extensible with new thought \ntransformations and thus can be used to spearhead new prompting\nschemes. This work brings the LLM reasoning closer to human \nthinking or brain mechanisms such as recurrence, both\nof which form complex networks
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
Recently, tool learning with large language models (LLMs) has emerged as a promising paradigm for augmenting the capabilities of LLMs to tackle highly complex problems. Despite growing attention and rapid advancements in this field, the existing literature remains fragmented and lacks systematic organization, posing barriers to entry for newcomers. This gap motivates us to conduct a comprehensive survey of existing works on tool learning with LLMs. In this survey, we focus on reviewing existing literature from the two primary aspects (1) why tool learning is beneficial and (2) how tool learning is implemented, enabling a comprehensive understanding of tool learning with LLMs. We first explore the “why” by reviewing both the benefits of tool integration and the inherent benefits of the tool learning paradigm from six specific aspects. In terms of “how”, we systematically review the literature according to a taxonomy of four key stages in the tool learning workflow: task planning, tool selection, tool calling, and response generation. Additionally, we provide a detailed summary of existing benchmarks and evaluation methods, categorizing them according to their relevance to different stages. Finally, we discuss current challenges and outline potential future directions, aiming to inspire both researchers and industrial developers to further explore this emerging and promising area. |
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
| [37] |
|
| [38] |
|
| [39] |
|
| [40] |
|
| [41] |
|
| [42] |
|
| [43] |
|
| [44] |
|
| [45] |
|
/
| 〈 |
|
〉 |
