Basin hydrological assessment method and system under climate change

Abstract

The present invention relates to the fields of hydrological assessment and environmental science and technology, and in particular to a method and system for watershed hydrological assessment under climate change. The method comprises: collecting multi-source heterogeneous data and generating fused feature data and causal path data through multimodal fusion; optimizing the watershed hydrological network structure based on reinforcement learning, using an enhanced hydrological model to simulate runoff and generate optimized runoff prediction data; extracting drought feature data by combining runoff prediction data with run theory, and generating scenario data describing extreme scenarios through a path-constrained generative adversarial network; constructing a dynamic joint distribution model of drought features based on the scenario data, and conducting multi-scenario risk assessment. The present invention achieves accurate modeling and risk quantification of watershed drought features under complex climate scenarios, significantly improving the accuracy and applicability of hydrological assessments and providing a scientific basis for watershed water resources management.

Description

Technical Field

[0001] The present invention relates to the field of hydrological assessment and environmental science and technology, and in particular to a method and system for basin hydrological assessment under climate change. Background Art

[0002] Climate change has had a profound impact on the hydrological processes of river basins, including significant changes in runoff patterns, drought frequency, and intensity, which has placed higher demands on the scientific management and sustainable utilization of river basin water resources. Existing technologies (Chinese invention patent, CN111797129B, publication number: A method for assessing hydrological drought under climate change scenarios) generally rely on hydrological models (such as the SWAT model) combined with statistical methods (such as the Copula function) to assess hydrological drought. Although these methods can provide relatively accurate assessments under historical climate conditions, they have obvious shortcomings under climate change scenarios: (1) they do not fully utilize multi-source heterogeneous data and ignore the dynamic causal relationship between meteorological, hydrological, and ecological variables; (2) they cannot dynamically optimize the river basin network structure and are difficult to adapt to network disturbances caused by climate change; (3) they lack high-reliability extreme scenario data generation methods, resulting in limited ability to quantify drought characteristics. Therefore, a new method combining multimodal data fusion, dynamic model optimization, and extreme scenario analysis is urgently needed to improve the accuracy and applicability of assessments. Summary of the Invention

[0003] To address the numerous issues with the aforementioned existing technologies, the present invention provides a method and system for assessing watershed hydrological conditions under climate change. This system uses multimodal data fusion to extract causal relationships between meteorological, hydrological, and ecological variables, optimizes the watershed hydrological network structure, and utilizes a path-constrained generative adversarial network to generate extreme scenario data. This system, combined with a dynamic joint distribution model, assesses multi-scenario risks. This system accurately captures the comprehensive response of a watershed's hydrological state to climate change, providing scientific support for drought quantification and risk management.

[0004] A watershed hydrological assessment method under climate change includes the following steps:

[0005] Collect meteorological data, geographic data, groundwater dynamics data, vegetation cover dynamics data, and social water use behavior data, clean and standardize the collected data, and generate fusion feature data and causal path data through a multimodal fusion model. The causal path data reflects the causal relationship and dynamic association between meteorological, hydrological, and ecological variables;

[0006] Based on the fusion of feature data and causal path data, a directed graph structure of the watershed hydrological network is constructed. Reinforcement learning is used to optimize the node and edge weights in the directed graph to generate optimized watershed hydrological network data. This data is then input into an enhanced hydrological model for runoff simulation. The runoff simulation results are corrected by dynamically adjusting the model parameters to generate optimized runoff prediction data.

[0007] Based on the optimized runoff prediction data, the drought severity and duration are quantified using the standardized runoff index. Drought event characteristics are extracted using run theory to generate drought characteristic data. Furthermore, the drought characteristic data and the optimized runoff prediction data are combined to generate scenario data describing extreme scenarios using a path-constrained generative adversarial network.

[0008] Based on the scenario data, a dynamic joint distribution model of drought characteristics is constructed. The scenario data is combined with the dynamic joint distribution model to conduct multi-scenario risk assessment. Based on the assessment results, basin hydrological assessment data is generated to evaluate the comprehensive response of drought characteristics and hydrological conditions in the basin.

[0009] Preferably, the collected meteorological data include hourly recorded precipitation data, temperature data and humidity data; geographic data include digital elevation models and land use type maps; groundwater dynamic data include monitoring well water level data; vegetation cover dynamic data include vegetation cover change data calculated based on the normalized vegetation index; social water use behavior data include regional statistics of agricultural irrigation water volume and industrial water volume.

[0010] Preferably, the multimodal fusion model generates fusion feature data through the following process:

[0011] Extract spatial features through convolutional neural networks;

[0012] Extract time series features using long short-term memory networks;

[0013] The extracted spatial features and time series features are input into the multi-layer perceptron network for fusion to generate fused feature data.

[0014] Preferably, the causal path data is generated by a causal network model, which is based on Granger causality analysis and self-supervised learning algorithm to extract the causal relationship between meteorological data, hydrological data and ecological variables, and represent the causal path through a directed graph. The causal path data includes causal relationship weights and lag time parameters.

[0015] Preferably, the steps of constructing a directed graph structure of a watershed hydrological network include:

[0016] The meteorological variables, hydrological variables and ecological variables in the causal path data are defined as network nodes;

[0017] Using reinforcement learning agents to adjust the weights of edges in the network to optimize the connection strength between nodes;

[0018] Output optimized watershed hydrological network data for subsequent runoff simulation.

[0019] Preferably, the steps of performing runoff simulation using the enhanced hydrological model include:

[0020] Input the optimized watershed hydrological network data into the enhanced hydrological model;

[0021] Dynamically adjust soil infiltration rate, evapotranspiration rate and groundwater recharge rate;

[0022] The optimization algorithm is combined to minimize the error between the runoff simulation results and the historical measured runoff data, and the optimized runoff prediction data is generated.

[0023] Preferably, the step of extracting drought event characteristics includes:

[0024] The standardized runoff index is calculated based on the optimized runoff prediction data. The calculation formula of the standardized runoff index is:

[0025]

[0026] Where SRI is the standardized runoff index, Q is the time series runoff value, μ is the mean of the time series runoff value, and σ is the standard deviation of the time series runoff value;

[0027] The cumulative intensity S and duration D of drought events are calculated in combination with the run theory. The calculation formula for the cumulative intensity is:

[0028]

[0029] Where S is the cumulative intensity of drought events, SRI i is the normalized runoff index during the i-th period of the drought period, and D is the duration of the drought event.

[0030] Preferably, the step of generating scenario data by using a path-constrained generative adversarial network includes:

[0031] The generator combines the node and edge weights of the watershed hydrological network to generate extreme scenario data that meets the constraints;

[0032] The discriminator verifies the credibility of the generated scenario data through optimal transmission theory to ensure the physical consistency of the scenario data.

[0033] Preferably, the step of constructing a dynamic joint distribution model of drought characteristics based on scenario data includes:

[0034] The probability distribution of drought characteristics is calculated using the dynamic kernel density estimation method. The calculation formula of the dynamic kernel density estimation is:

[0035]

[0036] Among them, f(x) is the probability density value at the point x to be estimated, n is the number of samples, h is the smoothing parameter, and x i is the value of the i-th sample point, K is the kernel function;

[0037] The dynamic joint distribution model performs multi-scenario risk assessment on scenario data and calculates the occurrence probability and distribution characteristics of drought characteristics under different scenarios.

[0038] A system for implementing the watershed hydrological assessment method under climate change, comprising:

[0039] The data acquisition module is used to collect meteorological data, geographic data, groundwater dynamic data, vegetation cover dynamic data and social water use behavior data, and clean and standardize the collected data;

[0040] a data fusion module, connected to the data acquisition module, for generating fused feature data and causal path data through a multimodal fusion model, wherein the causal path data reflects the causal relationship and dynamic association between meteorological variables, hydrological variables, and ecological variables;

[0041] A watershed network optimization module, connected to the data fusion module, is used to construct a directed graph structure of the watershed hydrological network based on the fused feature data and causal path data, and optimize the node and edge weights in the directed graph by combining reinforcement learning to generate optimized watershed hydrological network data;

[0042] A runoff simulation module is connected to the watershed network optimization module and is used to input the optimized watershed hydrological network data into the enhanced hydrological model to perform runoff simulation, and to correct the runoff simulation results by dynamically adjusting the model parameters to generate optimized runoff prediction data;

[0043] A drought feature extraction module is connected to the runoff simulation module and is used to calculate the standardized runoff index based on the optimized runoff prediction data, extract drought event characteristics in combination with the run theory, and generate drought feature data;

[0044] a scenario generation module, connected to the drought feature extraction module, for combining drought feature data and optimized runoff prediction data to generate scenario data describing extreme scenarios through a path-constrained generative adversarial network;

[0045] An assessment module is connected to the scenario generation module and is used to construct a dynamic joint distribution model of drought characteristics based on scenario data, perform multi-scenario risk assessment on the scenario data in combination with the dynamic joint distribution model, and generate watershed hydrological assessment data based on the assessment results. The watershed hydrological assessment data includes distribution information of drought characteristics, dynamic changes in runoff, and comprehensive response analysis of multiple scenario superpositions.

[0046] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0047] The present invention uses multimodal fusion technology to achieve deep fusion of multi-source heterogeneous data such as meteorological, geographical, and groundwater dynamics, generating fused feature data and causal path data that can reflect causal relationships, effectively solving the problem of insufficient causal analysis in existing technologies.

[0048] This paper optimizes the watershed hydrological network structure through reinforcement learning, realizes the dynamic adjustment of node and edge weights, and generates optimized watershed hydrological network data that conforms to the impact of climate change, overcoming the lack of adaptability caused by the fixed network structure in traditional methods.

[0049] This paper generates extreme scenario data through a path-constrained generative adversarial network, achieving high-confidence modeling of drought scenarios. It also verifies the physical consistency of the data through optimal transmission theory, addressing the limitation of the ability to describe extreme scenarios.

[0050] The present invention combines the dynamic joint distribution model with kernel density estimation technology to achieve accurate assessment of the probability of occurrence of drought characteristics under multiple scenarios and enhance the adaptability to complex scenario risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the process of the present invention;

[0052] Figure 2 Schematic diagram of the multimodal data fusion process in the present invention;

[0053] Figure 3 Schematic diagram of the scenario generation process of the path-constrained generative adversarial network in the present invention;

[0054] Figure 4 Schematic diagram of multi-scenario risk assessment of the dynamic joint distribution model in the present invention;

[0055] Figure 5 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0056] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0057] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise," "include," etc. used herein indicate the presence of the features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0058] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0059] like Figure 1 As shown in FIG, a watershed hydrological assessment method under climate change includes the following steps:

[0060] Collect meteorological data, geographic data, groundwater dynamics data, vegetation cover dynamics data, and social water use behavior data, clean and standardize the collected data, and generate fusion feature data and causal path data through a multimodal fusion model. The causal path data reflects the causal relationship and dynamic association between meteorological, hydrological, and ecological variables;

[0061] Preferably, the collected meteorological data include hourly recorded precipitation data, temperature data and humidity data; geographic data include digital elevation models and land use type maps; groundwater dynamic data include monitoring well water level data; vegetation cover dynamic data include vegetation cover change data calculated based on the normalized vegetation index; social water use behavior data include regional statistics of agricultural irrigation water volume and industrial water volume.

[0062] In principle, the data collection aims to encompass the multi-dimensional key factors influencing watershed hydrology. Meteorological data (such as hourly precipitation, temperature, and humidity) directly influence evapotranspiration, soil moisture, and runoff processes. Geographic data (such as digital elevation models and land-use type maps) provide information on the basin's topography and human activity. Groundwater dynamics data (such as monitoring well water levels) reflect groundwater level changes and recharge dynamics. Vegetation cover dynamics data (such as vegetation cover change data calculated based on the Normalized Difference Vegetation Index) describes the relationship between vegetation and evapotranspiration. Social water use behavior data (such as agricultural irrigation water and industrial water consumption) describe the direct impact of human activities on water resource consumption.

[0063] Preprocessing includes data cleaning, standardization, and missing value filling to ensure data quality and usability. Cleaning improves data accuracy by removing outliers and redundant information, while standardization uses mean normalization:

[0064]

[0065] Where X′ represents the normalized data, X represents the original data, mean(X) represents the mean, and std(X) represents the standard deviation. This method aligns the scales of different physical quantities, making it suitable for subsequent multimodal fusion. Missing values ​​are filled using spatiotemporal interpolation methods, such as kriging-based interpolation to estimate the data distribution in unobserved areas.

[0066] In effect, this step ensures high-quality input of multi-dimensional data and lays a solid foundation for subsequent modeling.

[0067] The core of the multimodal fusion model lies in fusing different data types (time series, raster data, and vector data) in a unified computing framework. First, a convolutional neural network (CNN) is used to extract spatial features from raster data, such as terrain elevation and the distribution pattern of land use types. In specific implementations, CNN captures local features of the data through convolution kernels and extracts high-level semantic information through multi-layer pooling operations. Subsequently, a long short-term memory network (LSTM) is used to analyze time series data (such as precipitation and groundwater level changes). The network effectively models long-term dependencies in time series by introducing memory units.

[0068] The fusion model feeds the extracted spatial and time series features into a multi-layer perceptron (MLP) to fuse heterogeneous data. This involves feature concatenation, weight normalization, and multi-layer nonlinear mapping to generate fused feature data. This fused feature data is a high-dimensional vector containing multidimensional information, reflecting both the patterns of individual data and the interactions between them.

[0069] The fused feature data provides a high-dimensional representation that fully captures the interactive characteristics of space, time and variables, providing a rich information basis for causal analysis and hydrological network modeling.

[0070] In this example, hourly precipitation, temperature, and humidity data were collected for a specific watershed to generate a daily-averaged meteorological time series. High-resolution remote sensing imagery was used to generate a digital elevation model (10-meter resolution) and a land use type map. Monitoring well data recorded water level fluctuations at different observation points within the watershed. The Normalized Difference Vegetation Index (NDVI) was extracted from MODIS data to provide dynamic trends in vegetation cover. This preprocessing resulted in standardized and temporally consistent data input.

[0071] The multimodal fusion model integrated these data to generate fused feature data. Causal path analysis revealed that precipitation significantly impacted runoff dynamics, with a three-day lag effect on groundwater levels. Ultimately, the generated causal path data was used in subsequent watershed hydrological network optimization, validating the model's reliability.

[0072] Preferably, Figure 2 As shown, the multimodal fusion model generates fusion feature data through the following process:

[0073] Extract spatial features through convolutional neural networks;

[0074] Extract time series features using long short-term memory networks;

[0075] The extracted spatial features and time series features are input into the multi-layer perceptron network for fusion to generate fused feature data.

[0076] Spatial Feature Extraction (CNN): Convolutional neural networks excel at extracting local features from rasterized spatial data. In this solution, the input data includes a digital elevation model (DEM) and a land use type map. By setting convolution kernels of different sizes (for example, 3×3 or 5×5), CNN is able to extract local spatial features from raster data, such as slope, aspect, and land use distribution patterns. Multi-layer convolution operations combined with pooling layers further capture high-level semantic features of spatial data, such as terrain complexity and regional land use intensity.

[0077] Time Series Feature Extraction (LSTM): Long Short-Term Memory (LSTM) networks are used to extract dynamic features from time series data, such as hourly precipitation, groundwater level changes, and temperature fluctuations. By introducing memory units and gating mechanisms, LSTM effectively learns long-term dependencies in time series. In this solution, LSTM can identify the lagged effects of meteorological changes on runoff dynamics, such as the lagged response time of precipitation to runoff. Input time series data is segmented into fixed-length windows (e.g., data from the past seven days), encoded through an embedding layer, and fed into the LSTM network for training.

[0078] Feature fusion (MLP): The extracted spatial features and time series features are each formed into high-dimensional vectors, which are then concatenated into a joint feature vector. A multi-layer perceptron is used to perform nonlinear mapping and feature fusion on this joint feature vector. The fusion process includes feature normalization, weight adjustment, and multi-layer activation operations. The final output fused feature data can comprehensively reflect the interactive relationship between spatial and temporal variables. For example, the combined effect of slope changes (spatial features) and continuous precipitation events (time series features) in a certain area can trigger extreme runoff events. These interactive relationships are modeled and explicitly represented using MLP.

[0079] The model unifies data from different modalities (spatial and temporal) into high-dimensional fused feature data. This data can reflect the complex interactions between meteorology, hydrology, and ecosystems, providing accurate and reliable input for causal path analysis. By extracting time series features using LSTM, the model accurately captures the lagged responses of meteorological variables (such as precipitation and temperature) to hydrological variables (such as groundwater levels and runoff), supporting in-depth analysis of watershed hydrological dynamics. By fusing spatial and temporal features using MLP, the model can identify interactions between spatial features (such as land use change) and temporal dynamics (such as long-term drought trends).

[0080] Embodiment, in actual application, the multimodal data of a certain watershed includes a high-resolution digital elevation model (10-meter grid resolution), a land use type map, hourly precipitation, and groundwater dynamic data. First, the digital elevation model and the land use type map are input into CNN, and local and global spatial features, such as slope and land use intensity distribution, are extracted through a three-layer convolutional network. Secondly, the hourly precipitation data and groundwater level data are divided into 7-day windows and input into LSTM for time series modeling to capture lag effects and periodic changes. Finally, the spatial feature vector output by CNN is spliced ​​with the temporal feature vector output by LSTM, and input into MLP for fusion to generate fused feature data containing spatial and temporal interaction information.

[0081] For example, during a sustained rainfall event, the LSTM identified a delayed response of a significant increase in groundwater levels three days after the peak rainfall, while land use features extracted by the CNN revealed that the slope and vegetation cover in the area had a significant impact on runoff paths. Through the fusion of the MLP, the model accurately characterized the dynamic relationship between precipitation, groundwater, and land use, providing precise input for causal pathway analysis and watershed hydrological network modeling.

[0082] Preferably, the causal path data is generated by a causal network model, which is based on Granger causality analysis and self-supervised learning algorithm to extract the causal relationship between meteorological data, hydrological data and ecological variables, and represent the causal path through a directed graph. The causal path data includes causal relationship weights and lag time parameters.

[0083] Granger causality analysis is a classic method for detecting causality in time series. In this approach, we assume two variables, X and Y. If using past X values ​​to predict Y is significantly better than using only past Y values, then X is considered to Granger cause Y. The formula is as follows:

[0084]

[0085] Among them, Y t is the target variable value at time ttt, X t-j is the value of the dependent variable at time lag jjj, β i and γ j is the regression coefficient, ∈ t is the error term, p and q are the number of lag terms respectively.

[0086] In this scheme, Granger causality analysis is used to preliminarily screen the direct causal relationship between meteorological variables (such as precipitation and temperature) and hydrological variables (such as groundwater level changes and runoff), and to estimate the lag time parameters. For example, precipitation may have a lag response of several hours or days to groundwater level changes.

[0087] To address implicit causal relationships in multimodal data (such as the indirect impact of vegetation cover changes on runoff), this solution introduces a self-supervised learning algorithm to construct a causal network model. Self-supervised learning uses unlabeled data to generate pseudo-labels for specific tasks, such as predicting whether there are synergistic effects between the change patterns of predictive variables. Specific implementations include:

[0088] Define input data characteristics, such as precipitation time series, normalized difference vegetation index, and groundwater dynamics;

[0089] Construct positive samples (pairs of variables that change together) and negative samples (pairs of variables that change independently) through contrastive learning methods;

[0090] The neural network model is trained to output causal weights, which reflect the direct or indirect causal strength between different variables.

[0091] Causal path data is stored as a directed graph, where nodes represent meteorological variables (hydrological variables and ecological variables such as precipitation, temperature, runoff, and vegetation cover), and edges represent causal relationships between variables. Edge weights indicate the strength of the causal relationship, and lag parameters characterize the time delay of the causal relationship. For example, the causal weight of precipitation on runoff is 0.8, and the lag time is 6 hours, indicating that changes in precipitation have a high correlation with runoff dynamics and a certain lag.

[0092] The causal network model quantifies the direct effects of meteorological variables on hydrological variables and the indirect effects of ecological variables using causal pathway data. For example, precipitation has a weight of 0.7 on groundwater levels with a lag of 48 hours, while changes in vegetation cover indirectly affect runoff dynamics by regulating evapotranspiration.

[0093] The combination of Granger causality analysis and self-supervised learning enables the model to capture complex interactions between nonlinear and cross-modal variables, such as the combined effects of temperature fluctuations and vegetation cover changes on runoff.

[0094] Causal path data provide accurate input for basin hydrological network optimization and the construction of dynamic joint distribution models, supporting further runoff prediction and risk assessment.

[0095] In this example, in a watershed with significant climate change, the input data included hourly recorded precipitation and temperature time series, daily groundwater dynamics data, vegetation cover change data (NDVI calculated based on MODIS imagery), and runoff observation data. First, Granger causality analysis was used to test the causality of precipitation and groundwater level changes. The results showed that precipitation had a significant lagged effect on groundwater level, with a lag time of 48 hours. Then, a self-supervised learning model was used to model the implicit relationship between vegetation cover and runoff. The results showed that changes in vegetation cover indirectly affected runoff dynamics by regulating evapotranspiration.

[0096] The directed graph generated from the causal path data includes nodes such as precipitation, temperature, groundwater level, runoff, and vegetation cover. Edge weights and lag parameters represent the causal strength and time delay between variables, respectively. For example, the weight of precipitation to runoff is 0.85, with a lag of 6 hours; the weight of vegetation cover to runoff is 0.6, with no significant lag. This causal path data provides a reliable basis for subsequent optimization of the watershed hydrological network and generation of extreme scenarios.

[0097] Based on the fusion of feature data and causal path data, a directed graph structure of the watershed hydrological network is constructed. Reinforcement learning is used to optimize the node and edge weights in the directed graph to generate optimized watershed hydrological network data. This data is then input into an enhanced hydrological model for runoff simulation. The runoff simulation results are corrected by dynamically adjusting the model parameters to generate optimized runoff prediction data.

[0098] The fusion of characteristic data and causal path data provides high-quality input data for constructing the directed graph structure of the watershed hydrological network. The fusion of characteristic data captures the complex interactions between space, time, and variables, while the causal path data clarifies the causal relationship and its lagged effects of meteorological variables (hydrological variables and ecological variables). In the directed graph:

[0099] Nodes: Represent key hydrological variables (such as precipitation, groundwater level, runoff, and evapotranspiration).

[0100] Edges represent causal relationships between nodes, edge weights quantify the strength of causal relationships, and lag parameters quantify the time delay of causal effects. By integrating this information, the directed graph structure not only provides an intuitive representation of the associations between variables but also dynamically captures the transmission effects of the hydrological network.

[0101] This solution uses an intelligent optimization method based on reinforcement learning, which uses intelligent agents to dynamically adjust the weights of nodes and edges in the graph to optimize the performance of the basin hydrological network. The core of reinforcement learning is that the intelligent agents learn the optimal strategy for decision-making through interaction:

[0102] State: the nodes, edge weights, and structure of the current directed graph.

[0103] Action: Adjust node properties or edge weights, such as increasing edge weights to enhance the effect of precipitation on runoff.

[0104] Reward function: Inversely optimize the error between runoff simulation results and measured data, minimizing the error to improve the rationality of the network structure. During iterative learning, the agent gradually optimizes the directed graph by comparing measured and simulated results, generating optimized watershed hydrological network data.

[0105] The optimized watershed hydrological network data is loaded as input into the enhanced hydrological model for runoff simulation. The enhanced hydrological model achieves higher accuracy through dynamic adjustment of the following parameters:

[0106] Soil permeability: reflects the impact of different soil types on runoff.

[0107] Evapotranspiration ratio: describes the regulation of evapotranspiration by vegetation and meteorological conditions.

[0108] Groundwater recharge rate: Quantifies the contribution of groundwater level dynamics to runoff. During the runoff simulation process, the model adjusts the above parameters through an optimization algorithm to minimize the error between the simulation results and historical measured data, ultimately generating optimized runoff forecast data.

[0109] In an example application for a specific watershed, the input data includes meteorological characteristics (hourly precipitation and temperature), hydrological characteristics (groundwater level dynamics and runoff records), and ecological characteristics (vegetation cover changes). First, a directed graph is constructed based on the fused feature data and causal path data, where the nodes include precipitation, temperature, groundwater level, runoff, and evapotranspiration. The edge weights and lag parameters are optimized using a reinforcement learning agent. For example, the edge weight from precipitation to runoff is optimized from 0.7 to 0.85, with a lag of 6 hours.

[0110] The optimized directed graph is fed into the enhanced hydrological model to dynamically adjust soil infiltration rate, evapotranspiration ratio, and groundwater recharge rate, for example:

[0111] The soil permeability was adjusted to 0.25 (representing sandy soil) and the evapotranspiration ratio was adjusted to 0.6.

[0112] The groundwater recharge rate is set at 0.1 mm / h according to the seasonal fluctuation of groundwater level.

[0113] After multiple rounds of iterations, the mean square error between the runoff simulation results output by the model and the measured runoff was reduced to 2%, verifying the predictive ability of the optimized basin hydrological network.

[0114] Preferably, the steps of constructing a directed graph structure of a watershed hydrological network include:

[0115] The meteorological variables, hydrological variables and ecological variables in the causal path data are defined as network nodes;

[0116] Using reinforcement learning agents to adjust the weights of edges in the network to optimize the connection strength between nodes;

[0117] Output optimized watershed hydrological network data for subsequent runoff simulation.

[0118] In a watershed hydrological network, nodes represent key variables that have direct or indirect effects on runoff dynamics, including:

[0119] Meteorological variables: such as precipitation, temperature, and humidity.

[0120] Hydrological variables: such as groundwater level, runoff, and evapotranspiration.

[0121] Ecological variables: such as vegetation coverage and soil moisture. These nodes are derived from causal path data, which clearly defines the dynamic associations and causal relationships between variables.

[0122] Edges in a network represent relationships between nodes. Their weights reflect the strength of causal relationships, and the lag parameter characterizes the time delay of the causal effect. For example, the edge weight from precipitation to runoff might be 0.8, with a lag of 6 hours, indicating that precipitation has a significant impact on runoff, but with a certain time lag.

[0123] To improve the structural rationality of the basin hydrological network, this solution introduces a reinforcement learning agent to dynamically optimize the weights of the edges in the network. The reinforcement learning process includes the following elements:

[0124] Status: the nodes, edge weights and structure of the current watershed hydrological network.

[0125] Action: Adjust the weights of the edges between nodes. For example, increase the weight of an edge to reflect a stronger causal relationship.

[0126] Reward function: By comparing the error between simulated runoff and measured runoff (such as mean square error), the agent is guided to optimize the edge weights so that the runoff simulation results are closer to actual observations.

[0127] The reinforcement learning agent continuously optimizes the network structure through interactive learning in multiple iterations to ensure that the node and edge weights can truly reflect the interactions between variables in the watershed.

[0128] In actual application, the specific execution steps include:

[0129] (1) Input the causal path data into the system, including the initial weights and lag time parameters of each node variable. For example, the initial edge weights of precipitation and temperature in the meteorological variables and runoff in the hydrological variable are 0.6 and 0.4, respectively, and the lag time parameters are 3 hours and 1 hour.

[0130] (2) The agent randomly initializes the edge weights to generate the initial watershed hydrological network.

[0131] This network was input into the enhanced hydrological model to simulate runoff output.

[0132] Calculate the error between the simulated runoff and the measured runoff, for example using the mean square error formula:

[0133]

[0134] Among them, MSE is the mean square error, Q sim,i is the simulated runoff value at the i-th moment, Q obs,i is the measured runoff value at the i-th moment, and n is the total number of time steps.

[0135] The edge weights are adjusted inversely according to the error and the network structure is updated.

[0136] Iterate the training until the reward function converges (i.e. the error reaches the minimum value), and output the optimized watershed hydrological network data.

[0137] (3) The optimized watershed hydrological network data includes the final edge weights for each node variable and its connections. For example, the edge weight from precipitation to runoff is optimized to 0.85, with a lag time of 6 hours; the edge weight from temperature to evapotranspiration is 0.7, with a lag time of 12 hours. These optimized parameters will be used in the subsequent runoff simulation process.

[0138] The basin hydrological network optimized through reinforcement learning more accurately reflects the causal relationship between meteorological, hydrological and ecological variables, making the network structure more consistent with the actual characteristics of the basin.

[0139] Because the reinforcement learning process uses runoff simulation error as the optimization target, the optimized network significantly improved prediction accuracy in subsequent runoff simulations. For example, in actual tests, the mean squared error of runoff predictions after optimization was reduced by 20% compared to the initial network.

[0140] Reinforcement learning enables the network to adapt to changes in watershed characteristics and input data. For example, when precipitation intensity increases significantly or vegetation cover changes drastically, the network can respond quickly by adjusting edge weights, providing more robust support for subsequent model predictions.

[0141] In this example, in a watershed study, the input causal path data included node variables such as precipitation, temperature, runoff, and groundwater level, along with their initial edge weights. During the optimization process, the reinforcement learning agent recognized that the impact of groundwater level changes on runoff was more significant than initially assumed and optimized its edge weight from 0.5 to 0.75. Furthermore, the agent detected that the effect of vegetation cover on evapotranspiration had a lag of 12 hours and adjusted its edge weight to 0.6.

[0142] The optimized watershed hydrological network data was fed into an enhanced hydrological model for runoff simulation, reducing the mean square error between the simulated results and measured runoff to 2%. The study demonstrated that the optimized network not only improved simulation accuracy but also accurately captured the combined impacts of meteorological and ecological variables on watershed hydrology, providing important support for subsequent scenario analysis.

[0143] Preferably, the steps of performing runoff simulation using the enhanced hydrological model include:

[0144] Input the optimized watershed hydrological network data into the enhanced hydrological model;

[0145] Dynamically adjust soil infiltration rate, evapotranspiration rate and groundwater recharge rate;

[0146] The optimization algorithm is combined to minimize the error between the runoff simulation results and the historical measured runoff data, and the optimized runoff prediction data is generated.

[0147] The optimized watershed hydrological network data provides reinforcement learning-optimized causal relationships between variables within the watershed, including node (such as precipitation, temperature, and runoff), edge weights (such as the intensity of precipitation's direct impact on runoff), and lag time parameters. This data is fed into the enhanced hydrological model as initial conditions and driving data, providing accurate starting states and dynamic correlation information for runoff simulation.

[0148] The enhanced hydrological model introduces dynamically adjusted key parameters that directly affect the simulation accuracy of the runoff process:

[0149] Soil permeability (P s ) is a parameter that represents the soil's ability to infiltrate precipitation, and its magnitude is affected by soil type, pore structure, and other factors. Adjusting this parameter allows for accurate simulation of the precipitation infiltration process and its effect on reducing surface runoff.

[0150] Evaporation ratio (E r ) describes the ratio of precipitation lost through evaporation and plant transpiration. Dynamically adjusting the evapotranspiration ratio can reflect the combined effects of vegetation cover, temperature, and other factors on the evapotranspiration process.

[0151] Groundwater recharge rate (G r ): Quantifies the contribution of groundwater to surface runoff recharge. Adjusting this parameter helps simulate groundwater-surface water interactions.

[0152] The runoff simulation results are compared with historical measured runoff data through an optimization algorithm. The error (e.g., mean square error) is calculated and the model parameters are adjusted inversely to minimize the error. The optimization algorithm can use gradient descent or genetic algorithm. Taking gradient descent as an example, its update formula is:

[0153]

[0154] Among them, θ represents the parameter to be optimized (such as P s 、E r , G r ), J(θ) represents the objective function (such as runoff simulation error), and α represents the learning rate. Through continuous iterative updates, the model parameters gradually converge to the optimal value.

[0155] In practice, the optimized watershed hydrological network data includes node variables such as precipitation and runoff, along with their edge weights (e.g., the weight of precipitation on runoff is 0.85, with a lag time of 6 hours). This data is input as initial conditions into the enhanced hydrological model to guide the simulation of the runoff process.

[0156] Initialize soil permeability P s , evaporation ratio E rand groundwater recharge rate G r For example, P s =0.3 (sand soil characteristics), E r =0.5 (medium vegetation cover), G r =0.1mm / h.

[0157] During the simulation process, the parameter values ​​are dynamically adjusted according to the input data and model output. For example, when the precipitation increases significantly and the runoff response lag time exceeds the expected value, the P s Improve penetration capacity and shorten lag time.

[0158] The simulation results are compared with the historical measured runoff data, and the error (such as mean square error) is calculated: The parameter values ​​are adjusted in reverse through the optimization algorithm, the model is rerun and the above process is repeated until the error meets the convergence condition.

[0159] The final optimized runoff prediction data includes time series data such as surface runoff and groundwater recharge, which can accurately reflect the comprehensive impact of meteorological and ecological changes on runoff dynamics.

[0160] Dynamic adjustment of soil permeability, evapotranspiration ratio and groundwater recharge rate makes the model highly adaptable to watershed characteristics and input conditions, significantly reducing the error between runoff simulation results and measured data.

[0161] The enhanced hydrological model combined with the optimization algorithm can accurately capture the runoff dynamics under different climatic conditions and watershed characteristics, for example, it performs particularly well under high-intensity precipitation or long-term drought scenarios.

[0162] The optimized runoff prediction data can be directly used as basic data for basin hydrological assessment and risk analysis, providing a scientific basis for water resources allocation, drought warning, etc.

[0163] In this embodiment, in a basin with severe precipitation fluctuations, the optimized basin hydrological network data is input, and the initial parameters are set to P s =0.3, E r =0.6, G r =0.15mm / h. During the runoff simulation, the model dynamically adjusts parameter values ​​based on the input data, for example:

[0164] When the precipitation intensity increases, the P s To 0.2, reducing the precipitation infiltration ratio.

[0165] When vegetation cover decreases, increase E rAfter 50 optimization iterations, the mean square error between the model output runoff prediction data and the measured runoff data increased from 10 mm to 0.7, reflecting a higher evapotranspiration ratio. 2 / h reduced to 2mm 2 / h, which verifies the effectiveness of the optimization process and the reliability of the prediction results.

[0166] Based on the optimized runoff prediction data, the drought severity and duration are quantified using the standardized runoff index. Drought event characteristics are extracted using run theory to generate drought characteristic data. Furthermore, the drought characteristic data and the optimized runoff prediction data are combined to generate scenario data describing extreme scenarios using a path-constrained generative adversarial network.

[0167] The Standardized Runoff Index (SRI) is a standardized indicator used to measure abnormal runoff conditions. By normalizing optimized runoff forecast data, the runoff data is converted into a dimensionless, standardized indicator that describes whether runoff conditions at different time steps deviate from normal values ​​(e.g., above or below the long-term average). Drought severity is defined as the cumulative value of runoff anomalies, while drought duration indicates the length of time that runoff remains below a certain threshold.

[0168] Run theory is a statistical method used to analyze the occurrence characteristics of continuous events. In this scheme, based on run theory, when the standardized runoff index falls below a set drought threshold, it is marked as the beginning of a drought event, and the index remains above the threshold until it rises back above the threshold. By analyzing the intensity (cumulative runoff anomaly) and duration (duration of the drought state) of each drought event, data describing drought characteristics is generated.

[0169] Generative Adversarial Networks (GANs) generate high-quality extreme scenario data through adversarial training of generators and discriminators. Path-constrained GANs introduce physical constraints to ensure that the generated scenario data conforms to the actual hydrological dynamics and physical laws of the basin:

[0170] Generator: Generates potential extreme scenario data based on the input drought characteristic data and optimized runoff prediction data.

[0171] Discriminator: Determines whether the generated data is consistent with the real scene data and improves the performance of the generator through adversarial training.

[0172] Path constraints: Ensure that the temporal and spatial characteristics of the generated data are consistent with the actual hydrological processes, for example, the changing trend of runoff should be consistent with the dynamics of precipitation and evapotranspiration.

[0173] In actual application, the specific implementation steps include:

[0174] (1) The optimized runoff prediction data is standardized to generate a standardized runoff index time series. The index value at each time step indicates the degree of deviation from the current runoff state. For example, a standardized runoff index below -1 indicates a drought state.

[0175] (2) Setting a drought threshold (e.g., normalized runoff index below -1).

[0176] Based on the run theory, the time periods continuously below the threshold are marked as drought events. The intensity (cumulative runoff anomaly) and duration (the length of time continuously below the threshold) of each drought event are calculated.

[0177] (3) Drought characteristic data (including intensity and duration) and optimized runoff prediction data are combined as input to the generator. Potential extreme scenario data are generated, such as generating anomaly sequences of runoff in a specific watershed under a historical drought event. The generated scenario data are compared with the actual historical data, and the generator parameters are adjusted through feedback. Runoff dynamics and watershed physical constraints (such as water conservation conditions) are introduced to ensure that the generated scenario data are physically consistent.

[0178] (4) The final output scenario data includes extreme runoff dynamics in time series format and descriptions of regional hydrological conditions, such as runoff anomaly magnitude and drought impact range. These data can be directly used to assess the risk and response capacity of the basin under extreme scenarios.

[0179] Through the standardized runoff index and run theory, the intensity and duration of drought events can be quantified into high-dimensional feature data, significantly improving the detailed description of the drought process.

[0180] The scenario data generated by the path-constrained GAN is not only consistent with the actual historical data in statistical distribution, but also can generate high-confidence scenarios that conform to hydrological dynamics under the physical constraints of the basin.

[0181] The generated scenario data provides an important basis for subsequent basin hydrological assessment and extreme scenario response analysis, and supports practical applications such as drought early warning and resource scheduling.

[0182] In this example, in a river basin with frequent droughts, the optimized runoff prediction data is used to generate a standardized runoff index, the drought threshold is set to -1, and the characteristics of drought events in the past 10 years are extracted. For example:

[0183] In 2012, a drought event had an intensity of -15 and lasted 30 days;

[0184] A drought event in 2015 had an intensity of -25 and lasted 50 days.

[0185] These drought characteristic data, along with runoff forecast data, were fed into a path-constrained GAN. Through multiple rounds of adversarial training, extreme scenario data was generated to simulate possible future drought scenarios. For example, the GAN generated an extreme drought scenario that could occur over the next 50 years, with runoff anomalies reaching -40° and affecting 70% of the basin. This data validated the generative power of the path-constrained GAN and provided a reliable scenario-based basis for developing future basin drought response strategies.

[0186] Preferably, the step of extracting drought event characteristics includes:

[0187] The standardized runoff index is calculated based on the optimized runoff prediction data. The calculation formula of the standardized runoff index is:

[0188]

[0189] Where SRI is the standardized runoff index, Q is the time series runoff value, μ is the mean of the time series runoff value, and σ is the standard deviation of the time series runoff value;

[0190] The Standardized Runoff Index (SRI) is an indicator used to measure the degree to which runoff data deviates from the historical average. By standardizing the optimized runoff forecast data, runoff data at different time scales are converted into a dimensionless standardized index. The standardization formula is as follows: This formula makes runoff data statistically consistent by removing the mean (μ) and normalizing the standard deviation (σ). A negative SRI indicates drought conditions, with lower values ​​indicating more severe droughts; a positive value indicates wet conditions.

[0191] The cumulative intensity S and duration D of drought events are calculated in combination with the run theory. The calculation formula for the cumulative intensity is:

[0192]

[0193] Where S is the cumulative intensity of drought events, SRI i is the normalized runoff index during the i-th period of the drought period, and D is the duration of the drought event.

[0194] Run theory is a statistical tool used to analyze the characteristics of continuous events in a time series. In this scenario, run theory is used to identify periods of time in the standardized runoff index series where the index continuously falls below a set drought threshold (e.g., SRI < 1). The cumulative severity, S, is the sum of the absolute values ​​of the standardized runoff index during the drought period; the duration, D, is the total length of time from the start to the end of the drought event, i.e., the length of the run.

[0195] In actual application, the specific implementation steps include:

[0196] (1) The optimized runoff prediction data is a high-precision runoff time series generated by the enhanced hydrological model after correction, with a consistent time step (such as daily or hourly) and basin range.

[0197] (2) Calculate the mean μ and standard deviation σ of the runoff data; apply the normalization formula to the runoff value Q at each time step to generate the corresponding SRI sequence. For example, for the runoff data recorded daily during a certain period:

[0198] The average daily runoff μ = 50 mm / day; the standard deviation σ = 10 mm / day; if the runoff value Q on a certain day is 30 mm / day, then SRI = (30-50) / 10 = -2.0, indicating that the runoff on that day is significantly lower than the historical average.

[0199] (3) setting a drought threshold (e.g., SRI < -1) to mark the time step of entering the drought state;

[0200] The time period that is continuously below the threshold is identified as a drought event; for each drought event, the cumulative intensity S and duration D are calculated. For example, if the SRI sequence of a drought event is -1.5, -2.0, -1.8, -1.2, -1.1, then:

[0201] Cumulative intensity S = |-1.5|+|-2.0|+|-1.8|+|-1.2|+|-1.1|=7.6;

[0202] Duration D = 5 days.

[0203] (4) The characteristics of each drought event (cumulative intensity S and duration D) are stored as drought characteristic data to form a structured output. For example:

[0204] Drought event 1: cumulative intensity 7.6, duration 5 days;

[0205] Drought event 2: cumulative intensity 12.3, duration 10 days.

[0206] By combining the standardized runoff index and run theory, drought events can be effectively distinguished from normal fluctuations, and the start and end time and intensity of each drought event can be identified.

[0207] The cumulative intensity and duration of drought events provide a quantitative description of the drought situation in the basin. These indicators can reflect the severity of a single drought event and, through statistical analysis, reveal long-term drought trends.

[0208] Drought characteristic data provides key input for subsequent scenario generation, risk assessment, and decision support. For example, by analyzing historical drought events, we can infer the likely frequency and impact of future droughts.

[0209] In this example, in a typical drought basin, daily runoff forecast data was used to generate a standardized runoff index sequence, and drought event characteristics over the past five years were extracted:

[0210] Event 1: duration 7 days, cumulative intensity 9.1;

[0211] Event 2: duration 14 days, cumulative intensity 15.7;

[0212] Event 3: Duration: 5 days, cumulative intensity: 5.4.

[0213] These data are further used for extreme scenario analysis to predict more severe drought events that may occur in the future, providing a scientific basis for basin managers to formulate drought response strategies.

[0214] Preferably, Figure 3 As shown in Figure 2, the steps of generating scenario data through a path-constrained generative adversarial network include:

[0215] The generator combines the node and edge weights of the watershed hydrological network to generate extreme scenario data that meets the constraints;

[0216] The discriminator verifies the credibility of the generated scenario data through optimal transmission theory to ensure the physical consistency of the scenario data.

[0217] Generative adversarial networks consist of two parts: a generator and a discriminator, and generate high-quality data through adversarial training:

[0218] The Generator generates data for potential extreme scenarios based on input watershed hydrological network data (including node and edge weights). The Discriminator compares the distribution of generated data with real data, using feedback to guide the Generator to produce more credible data. In this solution, GAN not only generates scenario data but also improves the rationality and consistency of the data by introducing path constraints and physical rules.

[0219] Path constraints are the innovation of this solution, ensuring that the generated scenario data conforms to the physical constraints and dynamic characteristics of the basin hydrological network. The key path constraints include:

[0220] Node constraints: The dynamic changes of node variables (such as precipitation and runoff) must conform to historical observations and statistical laws. Edge weight constraints: Edge weights (such as the intensity of precipitation's impact on runoff) must be consistent with the optimized watershed hydrological network data to ensure that the causal relationship between variables remains stable.

[0221] The discriminator verifies the credibility of the generated scenario data using optimal transmission theory. Optimal transmission theory calculates the transmission distance between the generated data distribution and the real data distribution. The closer the generated data is to the real data, the smaller the transmission distance and the higher the credibility. This process ensures that the generated data is not only statistically consistent with the real data but also physically conforms to the hydrological laws of the watershed.

[0222] In actual application, the specific implementation steps include:

[0223] (1) The optimized watershed hydrological network data, including node variables (such as precipitation, temperature, runoff, etc.) and edge weights (such as the intensity of the direct impact of precipitation on runoff), serve as the initial input of the generator.

[0224] (2) The node variables and edge weights are used as input feature vectors and fed into the generator model;

[0225] Based on the input features, the generator generates data for potential extreme scenarios. For example, the precipitation intensity in the node variable is simulated to be 1.2 times the historical extreme value; the runoff change is simulated as a downward trend for 5 consecutive days based on historical observations.

[0226] The generated scenario data include time series runoff changes, watershed responses under extreme weather conditions, etc.

[0227] (3) The discriminator calculates the distance (transmission distance) between the distribution of the generated scenario data and the distribution of the real data based on the optimal transmission theory. The discriminator provides feedback on the generated data. If the credibility is low, the generator adjusts the parameters and regenerates the data until the transmission distance reaches the set threshold, ensuring the credibility of the generated data.

[0228] (4) During the generator generation process, a path constraint function is added to impose physical rules on the generated data. For example, the conversion of precipitation to runoff must comply with edge weight constraints to reflect the true contribution of precipitation to runoff. Runoff dynamics must satisfy the water conservation relationship to avoid abnormal data that does not conform to physical laws.

[0229] (5) The final output scenario data includes the basin hydrological dynamics under extreme meteorological conditions, such as the runoff change sequence under continuous drought conditions and the flood response caused by extreme precipitation. These data can be directly used for basin hydrological risk assessment and decision support.

[0230] The generated scenario data can truly reflect the basin's hydrological dynamics under extreme meteorological conditions, with high credibility and physical consistency. The introduction of path constraints ensures that the generated data conforms to the physical laws of the basin's hydrological network, avoiding invalid data generated by instabilities in GAN training. High-quality scenario data can be used to analyze the potential impacts of climate change on the basin, providing a scientific basis for drought response, flood warning, and water resource scheduling.

[0231] In an embodiment, in a drought basin, the optimized basin hydrological network data is input, and the generator generates extreme scenario data, for example:

[0232] For 30 consecutive days, precipitation was 0.8 times the historical minimum, runoff decreased by 50%, and groundwater level decreased by 20%; under the extreme precipitation scenario, single-day precipitation reached 1.5 times the historical record, and runoff peak increased by 30%.

[0233] The discriminator verified the credibility of the generated data using optimal transmission theory. After multiple parameter adjustments, the generator reduced the distribution error between the final output scenario data and the actual historical data to 5%. The generated scenario data was used to analyze the impact of extreme weather events that may occur over the next 50 years on river basin hydrology, providing a decision-making basis for regional water resources planning.

[0234] like Figure 4 As shown in the figure, based on the scenario data, a dynamic joint distribution model of drought characteristics is constructed, and the scenario data are evaluated for multi-scenario risk in combination with the dynamic joint distribution model. Based on the evaluation results, basin hydrological assessment data are generated to evaluate the comprehensive response of drought characteristics and hydrological conditions in the basin.

[0235] Preferably, the step of constructing a dynamic joint distribution model of drought characteristics based on scenario data includes:

[0236] The probability distribution of drought characteristics is calculated using the dynamic kernel density estimation method. The calculation formula of the dynamic kernel density estimation is:

[0237]

[0238] Among them, f(x) is the probability density value at the point x to be estimated, n is the number of samples, h is the smoothing parameter, and x i is the value of the i-th sample point, K is the kernel function;

[0239] The dynamic joint distribution model performs multi-scenario risk assessment on scenario data and calculates the occurrence probability and distribution characteristics of drought characteristics under different scenarios.

[0240] The dynamic joint distribution model is used to quantify the joint probability distribution of drought characteristics (such as severity and duration) under different scenarios. By considering the dynamic nature of time series data, the model not only describes the marginal probability distribution of drought characteristics, but also describes the dependencies between them (such as the positive correlation between severity and duration).

[0241] Dynamic Kernel Density Estimation is a non-parametric method for calculating probability distribution. Its core is to use the kernel function KKK to smooth sample data and calculate the probability density of the point to be estimated. Its expression is:

[0242] Dynamic kernel density estimation generates a dynamic joint distribution model by performing probability density estimation on historical scenario data, which is used to describe the distribution characteristics of drought characteristics under different scenarios.

[0243] Based on the dynamic joint distribution model, a multi-scenario risk assessment is conducted on scenario data to calculate the probability of occurrence and distribution characteristics of drought characteristics under different scenarios. For example:

[0244] In scenario A, the distribution of drought severity is concentrated in the lower value range and the duration is shorter;

[0245] In scenario B, the distribution of drought severity is concentrated in the higher value range and the duration is significantly prolonged. Multi-scenario risk assessment helps quantify the risk level and potential impact by calculating the probability distribution of drought characteristics under each scenario.

[0246] The specific implementation includes the following steps:

[0247] (1) The input data include optimized runoff prediction data and drought characteristic data, which contain key variables (such as precipitation intensity and runoff volume) under different scenarios. The data are normalized to eliminate dimensionality effects and ensure the comparability of model calculation results.

[0248] (2) Probability density estimates of drought intensity and duration are obtained using dynamic kernel density estimation. Using the Gaussian kernel function as a smoothing tool, the joint distribution of drought characteristics under each scenario is calculated. The output dynamic joint distribution model includes a probability distribution matrix of drought characteristics, which is used to describe the distribution characteristics and interdependencies of drought characteristics under different scenarios.

[0249] (3) Using the dynamic joint distribution model, calculate the joint probability of drought intensity and duration in each scenario. For example, in scenario C:

[0250] The probability of an extreme drought event with an intensity greater than 20 and lasting more than 10 days is 15%, and the probability of an extreme drought event with an intensity greater than 30 is 5%. The risk levels under different scenarios are quantified and ranked to provide a basis for assessing the occurrence trend and impact range of drought events.

[0251] (4) Based on the multi-scenario risk assessment results, watershed hydrological assessment data are generated, including the distribution of drought characteristics, analysis of differences between scenarios, and the comprehensive response of the watershed hydrological conditions. The output data can be presented as probability distribution maps, cumulative probability curves, or spatial distribution maps to facilitate visual analysis.

[0252] The dynamic joint distribution model provides a clear mathematical description of the probability distribution of drought characteristics, and through multi-scenario risk assessment, it can quantify the intensity and scope of drought risks under different scenarios.

[0253] Basin hydrological assessment data provides a scientific basis for basin managers, such as predicting possible extreme drought events in the future and formulating response measures for water resource allocation and risk prevention and control.

[0254] Since the dynamic kernel density estimation method is highly adaptable to time series changes, the model can capture the nonlinear characteristics in scenario data and ensure the accuracy of the evaluation results.

[0255] In this embodiment, in a typical drought basin, scenario A and scenario B are generated based on historical data and optimized runoff prediction data:

[0256] Scenario A: 20% reduction in precipitation and 15% decrease in runoff;

[0257] Scenario B: 50% reduction in precipitation and 45% decrease in runoff.

[0258] The joint distribution model is constructed by dynamic kernel density estimation, and the calculation results show:

[0259] In scenario A, the probability of drought severity reaching 15 is 30%, and the probability of duration being 10 days is 25%;

[0260] In scenario B, the probability of drought severity exceeding 30 is 10%, and the probability of duration exceeding 20 days is 5%.

[0261] The results were converted into watershed hydrological assessment data, which showed that the drought risk under Scenario B was significantly higher than that under Scenario A, potentially leading to greater ecological and economic losses. Based on this, managers developed a zoning water restriction plan and an emergency water resource scheduling plan.

[0262] like Figure 5 As shown, a system for implementing the watershed hydrological assessment method under climate change comprises:

[0263] The data acquisition module is used to collect meteorological data, geographic data, groundwater dynamics data, vegetation cover dynamics data and social water use behavior data, and clean and standardize the collected data; collect multi-source heterogeneous data such as meteorological, geographical, groundwater dynamics data, vegetation cover data and social water use behavior data, and eliminate noise and dimensional differences through data cleaning and standardization, providing high-quality input for subsequent analysis.

[0264] The data fusion module is connected to the data acquisition module and is used to generate fused feature data and causal path data through a multimodal fusion model. The causal path data reflects the causal relationship and dynamic association between meteorological variables, hydrological variables and ecological variables; uses the multimodal fusion model to convert multi-source data into unified fused feature data, combines causal analysis to extract the dynamic causal relationship between meteorological variables, hydrological variables and ecological variables, and generates causal path data for describing the complex interactions between variables in the basin.

[0265] The watershed network optimization module, connected to the data fusion module, constructs a directed graph structure of the watershed hydrological network based on the fused feature data and causal path data. It optimizes the node and edge weights in the directed graph using reinforcement learning to generate optimized watershed hydrological network data. Based on the fused feature data and causal path data, it constructs a directed graph structure of the watershed hydrological network. Reinforcement learning is used to optimize node and edge weights, dynamically adjusting the network structure to more accurately reflect the impact of climate change on watershed hydrology.

[0266] The runoff simulation module is connected to the watershed network optimization module and is used to input the optimized watershed hydrological network data into the enhanced hydrological model for runoff simulation, and to correct the runoff simulation results by dynamically adjusting the model parameters to generate optimized runoff prediction data; the optimized watershed network data is input into the enhanced hydrological model, and the runoff simulation results are corrected by dynamically adjusting the model parameters (such as soil permeability, evapotranspiration ratio, etc.) to generate optimized runoff prediction data to ensure the reliability of the simulation results.

[0267] The drought characteristic extraction module is connected to the runoff simulation module and is used to calculate the standardized runoff index based on the optimized runoff prediction data, extract the characteristics of the drought event in combination with the run theory, and generate drought characteristic data; use the optimized runoff prediction data, combined with the standardized runoff index and run theory, to quantify the intensity and duration of the drought event, generate drought characteristic data, and provide key input for subsequent extreme scenario analysis.

[0268] A scenario generation module is connected to the drought feature extraction module and is used to combine drought feature data and optimized runoff prediction data to generate scenario data describing extreme scenarios through a path-constrained generative adversarial network; generate extreme scenario data that conforms to physical consistency through the path-constrained generative adversarial network; and perform multi-scenario risk assessment on the scenario data based on a dynamic joint distribution model to quantify the probability of occurrence and distribution characteristics of drought features under different scenarios.

[0269] An assessment module, connected to the scenario generation module, is configured to construct a dynamic joint distribution model of drought characteristics based on the scenario data, perform a multi-scenario risk assessment on the scenario data using the dynamic joint distribution model, and generate basin hydrological assessment data based on the assessment results. The basin hydrological assessment data includes information on the distribution of drought characteristics, dynamic changes in runoff, and a comprehensive response analysis of multiple scenarios. Combining the outputs of all modules, comprehensive basin hydrological assessment data is generated, including drought characteristic distribution, dynamic changes in runoff, and a response analysis of multiple scenarios, providing a scientific basis for climate adaptation management and decision-making.

[0270] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware.

[0271] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for watershed hydrological assessment under climate change, characterized in that: The following steps are involved: Collect meteorological data, geographic data, groundwater dynamics data, vegetation cover dynamics data, and social water use behavior data, clean and standardize the collected data, and generate fusion feature data and causal path data through a multimodal fusion model. The causal path data reflects the causal relationship and dynamic association between meteorological, hydrological, and ecological variables; Based on the fusion of feature data and causal path data, a directed graph structure of the watershed hydrological network is constructed. Reinforcement learning is used to optimize the node and edge weights in the directed graph to generate optimized watershed hydrological network data. This data is then input into an enhanced hydrological model for runoff simulation. The runoff simulation results are corrected by dynamically adjusting the model parameters to generate optimized runoff prediction data. The steps for constructing a directed graph structure for a watershed hydrological network include: defining meteorological, hydrological, and ecological variables in the causal path data as network nodes; using a reinforcement learning agent to adjust the weights of edges in the network to optimize the connection strength between nodes; and outputting the optimized watershed hydrological network data for subsequent runoff simulation. Based on the optimized runoff prediction data, the drought severity and duration are quantified using the standardized runoff index. Drought event characteristics are extracted using run theory to generate drought characteristic data. Furthermore, the drought characteristic data and the optimized runoff prediction data are combined to generate scenario data describing extreme scenarios using a path-constrained generative adversarial network. The steps for generating scenario data using a path-constrained generative adversarial network include: the generator combines the node and edge weights of the watershed hydrological network to generate extreme scenario data that meets the constraints; the discriminator verifies the credibility of the generated scenario data using optimal transmission theory to ensure the physical consistency of the scenario data; Based on the scenario data, a dynamic joint distribution model of drought characteristics is constructed. The scenario data is combined with the dynamic joint distribution model to conduct multi-scenario risk assessment. Based on the assessment results, basin hydrological assessment data is generated to evaluate the comprehensive response of drought characteristics and hydrological conditions in the basin.

2. The basin hydrological assessment method under climate change according to claim 1, characterized in that: The collected meteorological data include hourly recorded precipitation data, temperature data and humidity data; geographic data include digital elevation models and land use type maps; groundwater dynamic data include monitoring well water level data; vegetation cover dynamic data include vegetation cover change data calculated based on the normalized vegetation index; social water use behavior data include regional statistics of agricultural irrigation water volume and industrial water volume.

3. The basin hydrological assessment method under climate change according to claim 1, characterized in that: The multimodal fusion model generates fusion feature data through the following process: Extract spatial features through convolutional neural networks; Extract time series features using long short-term memory networks; The extracted spatial features and time series features are input into the multi-layer perceptron network for fusion to generate fused feature data.

4. The basin hydrological assessment method under climate change according to claim 1, characterized in that: The causal path data is generated through a causal network model. The causal network model is based on Granger causality analysis and a self-supervised learning algorithm to extract the causal relationship between meteorological data, hydrological data and ecological variables, and represents the causal path through a directed graph. The causal path data includes causal relationship weights and lag time parameters.

5. The basin hydrological assessment method under climate change according to claim 1, characterized in that: The steps for runoff simulation using the enhanced hydrological model include: Input the optimized watershed hydrological network data into the enhanced hydrological model; Dynamically adjust soil infiltration rate, evapotranspiration rate and groundwater recharge rate; The optimization algorithm is combined to minimize the error between the runoff simulation results and the historical measured runoff data, and the optimized runoff prediction data is generated.

6. The basin hydrological assessment method under climate change according to claim 1, characterized in that: The steps to extract drought event characteristics include: The standardized runoff index is calculated based on the optimized runoff prediction data. The calculation formula of the standardized runoff index is: Where SRI is the standardized runoff index, Q is the time series runoff value, μ is the mean of the time series runoff value, and σ is the standard deviation of the time series runoff value; The cumulative intensity S and duration D of drought events are calculated in combination with the run theory. The calculation formula for the cumulative intensity is: Where S is the cumulative intensity of drought events, SRI i is the normalized runoff index during the i-th period of the drought period, and D is the duration of the drought event.

7. The basin hydrological assessment method under climate change according to claim 1, characterized in that: The steps for constructing a dynamic joint distribution model of drought characteristics based on scenario data include: The probability distribution of drought characteristics is calculated using the dynamic kernel density estimation method. The calculation formula of the dynamic kernel density estimation is: Among them, f(x) is the probability density value at the point x to be estimated, n is the number of samples, h is the smoothing parameter, and x i is the value of the i-th sample point, K is the kernel function; The dynamic joint distribution model performs multi-scenario risk assessment on scenario data and calculates the occurrence probability and distribution characteristics of drought characteristics under different scenarios.

8. A system for implementing the watershed hydrological assessment method under climate change according to any one of claims 1 to 7, characterized in that: include: The data acquisition module is used to collect meteorological data, geographic data, groundwater dynamic data, vegetation cover dynamic data and social water use behavior data, and clean and standardize the collected data; a data fusion module, connected to the data acquisition module, for generating fused feature data and causal path data through a multimodal fusion model, wherein the causal path data reflects the causal relationship and dynamic association between meteorological variables, hydrological variables, and ecological variables; A watershed network optimization module, connected to the data fusion module, is used to construct a directed graph structure of the watershed hydrological network based on the fused feature data and causal path data, and optimize the node and edge weights in the directed graph by combining reinforcement learning to generate optimized watershed hydrological network data; A runoff simulation module is connected to the watershed network optimization module and is used to input the optimized watershed hydrological network data into the enhanced hydrological model to perform runoff simulation, and to correct the runoff simulation results by dynamically adjusting the model parameters to generate optimized runoff prediction data; A drought feature extraction module is connected to the runoff simulation module and is used to calculate the standardized runoff index based on the optimized runoff prediction data, extract drought event characteristics in combination with the run theory, and generate drought feature data; a scenario generation module, connected to the drought feature extraction module, for combining drought feature data and optimized runoff prediction data to generate scenario data describing extreme scenarios through a path-constrained generative adversarial network; An assessment module is connected to the scenario generation module and is used to construct a dynamic joint distribution model of drought characteristics based on scenario data, perform multi-scenario risk assessment on the scenario data in combination with the dynamic joint distribution model, and generate watershed hydrological assessment data based on the assessment results. The watershed hydrological assessment data includes distribution information of drought characteristics, dynamic changes in runoff, and comprehensive response analysis of multiple scenario superpositions.

Images