A Hydrological Trend Prediction Method Based on Big Data Analysis

Abstract

This invention proposes a hydrological trend prediction method based on big data analysis. It constructs an initial graph structure with monitoring stations as nodes and geographical distance and water system connectivity as edge weights. A message passing mechanism is used to iteratively aggregate node features, capturing the spatial interaction of hydrological variables between stations. Based on rainfall changes, the water system connectivity parameters and edge weights are dynamically adjusted to generate an adaptive graph structure. Combining graph neural networks, dynamic Bayesian networks, and spatiotemporal co-Kriging interpolation algorithms, hourly-level hydrological dynamic transmission features are extracted to form a spatiotemporally coupled prediction model. A prediction deviation threshold triggers model parameter optimization and message passing mechanism adjustments, achieving an adaptive prediction process of "dynamic modeling - feature fusion - closed-loop optimization." This method can accurately characterize the spatiotemporal correlation of hydrological elements, improve the analytical accuracy of complex hydrological processes, and enhance adaptability to sudden hydrological events and changes in the watershed environment.

Description

Technical Field

[0001] This invention relates to the field of hydrological data analysis, and in particular to a method for predicting hydrological trends based on big data analysis. Background Technology

[0002] Time series analysis of hydrological data is a key area for water resources management and flood control and disaster reduction. By dynamically modeling variables such as rainfall, flow, and soil moisture, it is possible to reveal the spatiotemporal patterns of watershed hydrological processes, providing important support for scientific decision-making. In recent years, neural networks and deep learning technologies have been widely used in hydrological time series forecasting due to their powerful nonlinear modeling capabilities. However, existing methods mostly focus on forecasting at a single station or the entire watershed, neglecting the spatial correlation between monitoring stations and the dynamic transmission process of hydrological variables. This leads to insufficient characterization of the spatial heterogeneity of hydrological processes within the watershed, especially in the dynamic adjustment of connectivity caused by changes in rainfall intensity, where model accuracy and adaptability significantly decrease.

[0003] First, the transmission of hydrological variables between stations is constrained by both geographical distance and water system connectivity, making it difficult for traditional models to integrate these complex spatial relationships. Second, the dynamic adjustment of edge weights caused by changes in rainfall intensity increases the complexity of time series modeling, and existing methods lack a real-time response mechanism to this dynamic. Furthermore, the multivariate interactions of nodal features such as rainfall, flow, and soil moisture further exacerbate the modeling difficulty, and the unresolved problem of dynamic spatial correlation modeling limits the accuracy of watershed hydrological predictions. Summary of the Invention

[0004] The embodiments of the present invention provide a hydrological trend prediction method based on big data analysis, which aims to solve the problems that existing technologies are unable to capture the spatial correlation and dynamic evolution process between stations within a watershed, and the hydrological trend analysis model has low accuracy.

[0005] To achieve the above objectives, the present invention provides a hydrological trend prediction method based on big data analysis, the method comprising:

[0006] The geographical coordinates, rainfall, flow rate and soil moisture data of monitoring stations within the watershed are obtained, and an initial graph structure is constructed. The initial graph structure uses the stations as nodes, and the feature vectors of the nodes include the rainfall, flow rate and soil moisture. The edge weights are calculated by the geographical distance between stations and the connectivity of the water system.

[0007] The initial graph structure is aggregated using a message passing mechanism. The feature vectors of the nodes are updated through iterative calculation to capture the mutual influence of hydrological variables between stations. The hydrological variables include rainfall, flow rate, and soil moisture.

[0008] If the change in rainfall in the node features exceeds a preset threshold, the water system connectivity parameters are adjusted and the edge weights are recalculated to generate a dynamically adjusted graph structure.

[0009] The dynamically adjusted graph structure is input into the graph neural network prediction layer, and the predicted values ​​of rainfall, flow rate and soil moisture for each station are calculated through mapping by the fully connected layer.

[0010] Time series analysis is performed on the predicted values ​​to extract hourly hydrological dynamic transmission characteristics, and a coupled model based on dynamic Bayesian network and spatiotemporal co-Kriging interpolation algorithm is constructed.

[0011] Based on the spatial distribution characteristics of the output of the coupled model, the parameters of the graph neural network are updated and the message passing mechanism is optimized;

[0012] If the deviation of the predicted value output by the optimized model is lower than the preset threshold, the final hydrological variable prediction result is output; otherwise, the dynamic adjustment and parameter optimization process is repeated.

[0013] Furthermore, the edge weight is calculated using the following formula:

[0014] W ij =D ij (1+α(1-C ij )),

[0015] In the formula, W ij D represents the edge weight between site i and site j; ij α is the geographical distance between station i and station j; α is the adjustment parameter; C ij C is the connectivity factor; if there is a waterway connection between station i and station j, then C ij =1, otherwise C ij =0.

[0016] Furthermore, the initial graph structure is aggregated using a message passing mechanism, and the node feature vectors are updated through iterative calculation to capture the mutual influence of hydrological variables between stations. These hydrological variables include rainfall, flow rate, and soil moisture, including:

[0017] By using a message passing mechanism, the neighbor node information of each node is obtained from the initial graph structure, and the influence of the neighbor node's feature vector on the current node is calculated.

[0018] Based on the message vector of each node, an aggregation function is used to fuse the message vectors;

[0019] From the aggregated node feature vectors, determine whether the convergence condition is met. If it is met, stop the iteration and determine the node feature vectors; if not, continue to the next round of iteration to obtain the updated node feature vectors.

[0020] Calculate the feature change trend of each node based on the updated node feature vector;

[0021] Based on the trend of feature changes, the K-means algorithm is used to group the nodes;

[0022] From the grouping results of the nodes, obtain the central feature vector of each group, calculate the feature similarity between groups, and obtain the similarity matrix between groups;

[0023] Based on the similarity matrix between groups, the intensity of interaction between hydrological variables is determined, and the grouping relationship reflecting the dynamic interaction of hydrological variables is obtained.

[0024] Furthermore, if the change in rainfall intensity in the node features exceeds a preset threshold, the water system connectivity parameters are adjusted and the edge weights are recalculated to generate a dynamically adjusted graph structure, including:

[0025] If the change in rainfall intensity in the node features exceeds a preset threshold, the water system connectivity parameters are adjusted and the edge weights are updated based on the change in rainfall intensity.

[0026] Based on the dynamically adjusted graph structure, obtain the set of neighboring nodes for each node, calculate the influence of the edge weights of the neighboring nodes on the current node, and obtain the edge weight influence value of each node.

[0027] If the influence value of the edge weight exceeds the preset influence threshold, the edge weight influence value is normalized by standardization to obtain the normalized edge weight influence value.

[0028] Based on the normalized edge weight influence values, a hierarchical clustering algorithm is used to group the nodes to obtain the node grouping results;

[0029] From the grouping results of the nodes, obtain the influence value of the center edge weight of each group, calculate the edge weight similarity between groups, and obtain the similarity matrix between groups;

[0030] If an element in the similarity matrix between groups exceeds a preset similarity threshold, it is determined that the groups have strong connectivity, the water system connectivity parameter is updated, and the grouping relationship reflecting the dynamic connectivity of the water system is obtained.

[0031] Based on the grouping relationships that reflect the dynamic connectivity of the water system, a weighted average method is used to calculate the comprehensive connectivity strength of each group, thus obtaining the comprehensive connectivity strength of each group.

[0032] Further, the step of inputting the dynamically adjusted graph structure into the graph neural network prediction layer, and calculating the predicted values ​​of rainfall, flow rate, and soil moisture for each station through a fully connected layer mapping, includes:

[0033] The dynamic graph structure is processed by a graph neural network, and the message passing mechanism is used to update the features of each site to obtain the updated site features.

[0034] Based on the updated site characteristics, a fully connected layer mapping is used to calculate the initial predicted value for each site;

[0035] If the initial predicted value deviates from the preset threshold by more than the limit, the weight matrix and bias parameters are optimized by the gradient descent algorithm to obtain the optimized weight matrix and bias parameters.

[0036] Based on the optimized weight matrix and bias parameters, the predicted values ​​for each site are updated by remapping through the fully connected layer;

[0037] If the updated forecast deviates significantly from historical data, the forecast will be adjusted through time series analysis.

[0038] Based on the adjusted forecast values, a weighted average method is used to integrate the forecast values ​​of rainfall, flow, and soil moisture to obtain the comprehensive forecast value for each station.

[0039] Cluster analysis was used to group the comprehensive predicted values ​​to obtain the grouping results of the site predicted values.

[0040] Furthermore, the step of performing time series analysis on the predicted values, extracting hourly hydrological dynamic transmission characteristics, and constructing a coupled model based on a dynamic Bayesian network and a spatiotemporal co-Kriging interpolation algorithm includes:

[0041] Hourly dynamic features were extracted from hydrological data through time series analysis, and the hourly dynamic features were obtained by using the sliding window method.

[0042] Based on hourly dynamic characteristics, a dynamic Bayesian network is constructed, node state transition probabilities are defined, and the node state probability distribution is obtained.

[0043] The spatiotemporal co-kriging interpolation algorithm is adopted to calculate spatial correlation based on the node state probability distribution and obtain spatial interpolation results;

[0044] If the spatial interpolation result deviates from the preset threshold by more than a certain range, the weight coefficients are optimized by gradient descent.

[0045] Based on the optimized interpolation weights, the spatial interpolation results are recalculated;

[0046] By using data fusion methods, combining the node state probability distribution and the updated spatial interpolation results, a comprehensive predicted value is obtained;

[0047] If there is a significant deviation between the overall predicted value and historical data, the model parameters will be adjusted through a dynamic update mechanism.

[0048] Furthermore, updating the parameters of the graph neural network and optimizing the message passing mechanism based on the spatial distribution characteristics of the coupled model output includes:

[0049] The interaction strength between nodes can be obtained by analyzing their spatial distribution characteristics.

[0050] The set of neighboring nodes is determined based on the strength of interactions between nodes;

[0051] Obtain the node feature vector by using the set of neighboring nodes;

[0052] Node messages are calculated using a message passing mechanism based on node feature vectors.

[0053] Calculate the loss function based on node messages;

[0054] If the loss function value is greater than the preset threshold, gradient descent is used to optimize the weight matrix and bias.

[0055] Update the graph neural network parameters based on the optimized weight matrix and bias.

[0056] Furthermore, if the deviation of the predicted value output by the optimized model is lower than a preset threshold, the final hydrological variable prediction result is output; otherwise, the dynamic adjustment and parameter optimization process is repeated, including:

[0057] Hydrological variable characteristics are obtained by inputting data, and feature vectors are calculated using feature extraction functions.

[0058] Initial prediction results are generated using a machine learning model based on the feature vectors of hydrological variables.

[0059] The deviation of the predicted value is calculated by comparing the initial prediction results with the actual value data;

[0060] If the predicted value deviation is lower than the preset threshold, the prediction result will be output.

[0061] If the deviation exceeds the preset threshold, dynamic adjustment will be performed;

[0062] Based on the deviation of the predicted values, the model structure is dynamically adjusted, and the feature weights are redistributed using a structure update function;

[0063] The parameters are optimized using the gradient descent algorithm based on the adjusted model structure.

[0064] Based on the optimized parameters, the prediction results are regenerated;

[0065] Based on the new prediction results, the deviation calculation and threshold judgment are repeated, and the process is repeated until the deviation is lower than the preset threshold, and the final hydrological variable prediction results are output.

[0066] The above technical solution has the following technical effects:

[0067] (1) The initial graph structure is constructed by geographical distance and water system connectivity, and the edge weights are dynamically adjusted according to the rainfall change threshold, so that the model can respond in real time to the sudden changes in water system connectivity caused by extreme weather, accurately capture the dynamic correlation of elements such as rainfall, flow, and soil moisture in the basin, and significantly improve the dynamic modeling capability for sudden hydrological events.

[0068] (2) By iteratively aggregating the features of neighboring nodes through the message passing mechanism, the physical transmission process of hydrological variables between stations is explicitly modeled. At the same time, by combining graph neural networks and dynamic Bayesian networks with spatiotemporal kriging interpolation, the dual features of spatial topology and time series are fused to realize the analysis of hourly dynamic transmission features of hydrological variables.

[0069] (3) By setting a prediction deviation threshold, a closed-loop process is triggered to dynamically adjust the edge weights of the graph structure, update the parameters of the graph neural network, and optimize the message passing mechanism, forming an adaptive mechanism of "prediction-feedback-optimization". When the hydrological conditions in the watershed change, the model can automatically calibrate the parameters through iterative optimization to continuously reduce the prediction deviation and ensure that it can maintain high prediction accuracy and stability under different environments. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating a hydrological trend prediction method based on big data analysis in one embodiment of the present invention.

[0071] Figure 2 This is a flowchart illustrating the process of dynamically adjusting the model structure based on the model prediction results in one embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram illustrating the process of updating graph neural network parameters based on the spatial distribution characteristics output by the coupled model in one embodiment of the present invention. Detailed Implementation

[0073] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0074] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0075] like Figure 1-3 This embodiment of a method for hydrological data mining and trend prediction based on big data may specifically include:

[0076] Step S101: Obtain the geographical coordinates, rainfall, flow rate and soil moisture data of monitoring stations within the watershed, and construct an initial graph structure, wherein the stations are used as nodes, the node feature vectors include the rainfall, flow rate and soil moisture, and the edge weights are calculated based on the geographical distance between the stations and the connectivity of the water system.

[0077] Geographic coordinates, rainfall, flow rate, and soil moisture data are obtained from monitoring stations within the watershed. A raw monitoring dataset is obtained through database queries and sensor data acquisition. Based on the raw monitoring dataset, the geographic coordinates of each station are extracted, and the geographic distance between stations is calculated using the Euclidean distance formula, resulting in a geographic distance matrix. The distances between stations are obtained from the geographic distance matrix. Combined with a pre-established water system connectivity topology map, if there is a water system connection between stations i and j, the connectivity factor C is determined. ij =1, otherwise C ij =0, using formula W ij =D ij (1+α(1-C ij The edge weights are calculated, where Wij is the edge weight and α is an adjustment parameter, resulting in an edge weight matrix. Based on the original monitoring dataset, rainfall, flow rate, and soil moisture are extracted for each station, normalized, and used to form feature vectors, resulting in a node feature matrix. Nodes are generated based on the stations, and an initial graph structure is constructed by combining the edge weight matrix and the node feature matrix. Nodes are associated with feature vectors, and edges are associated with edge weights. An adjacency matrix is ​​used to represent the graph structure, resulting in initial graph structure data. For the initial graph structure data, a Graph Convolutional Network (GCN) algorithm is used for feature extraction, resulting in an enhanced node feature matrix. Features are extracted from the enhanced node feature matrix, combined with the edge weight matrix, and the K-means algorithm is used to cluster the stations, resulting in the station clustering results.

[0078] For example, data on the geographical coordinates, rainfall, flow rate, and soil moisture of monitoring stations within a watershed can be acquired in real time through database queries combined with sensor data.

[0079] For example, a river basin has five monitoring stations located in the upper, middle, and lower reaches of the river. The sensors record data hourly, and the database stores the station's latitude and longitude, such as station A (116.5, 39.8) and station B (116.7, 39.9). Rainfall, flow rate, and soil moisture are expressed in millimeters, cubic meters per second, and percentages, respectively. This method ensures comprehensive and real-time data, providing a reliable foundation for subsequent analysis.

[0080] In one possible implementation, the Euclidean distance formula is used when calculating the geographical distance between stations.

[0081] For example, substituting the latitude and longitude of stations A and B into the formula yields a distance of approximately 22 kilometers. After calculating the distance between all stations pairwise, a geographic distance matrix is ​​formed. This matrix visually reflects the spatial relationships between stations, providing a basis for subsequent weight calculations.

[0082] It should be noted that Euclidean distance ignores terrain factors, but it is simple to calculate and suitable for quickly constructing distance matrices.

[0083] Specifically, a water system connectivity topology map can be generated from a watershed hydrological map.

[0084] For example, stations A and B are located in the same branch, C ij =1; Station C is located on another branch, C ij =0. Edge weight formula W ij In this context, α is set to 0.5. If there is no water system connection, the weight increases to reflect the impact of hydrological isolation.

[0085] For example, W from A to B ij The value is 22, W from A to C. ij The weight is 33. This weighting design balances geographical distance and water system relationships, which helps to accurately model watershed characteristics.

[0086] Preferably, the node feature matrix is ​​generated by normalizing rainfall, flow rate, and soil moisture.

[0087] For example, station A has a rainfall of 50 mm, a flow rate of 100 cubic meters per second, and a soil moisture of 60%, with maximum values ​​of 100 mm, 200 cubic meters per second, and 80%, respectively. The normalized feature vector is [0.5, 0.5, 0.75]. This normalization eliminates dimensional differences, enhances feature comparability, and provides a unified scale for graph structure modeling.

[0088] In one embodiment, the initial graph structure uses stations as nodes, edge weight matrices as edges, and node feature matrices as attributes.

[0089] For example, 5 sites generate a 5-node graph, and the adjacency matrix records W. ij The nodes are associated with feature vectors. This graph structure integrates spatial, hydrological, and environmental information, providing input for graph convolutional networks. The advantage of the graph structure is that it can simultaneously capture topological and feature relationships.

[0090] Understandably, graph convolutional networks enhance node features through multi-layer feature aggregation.

[0091] For example, the feature vector of station A, after being processed by GCN, incorporates rainfall and flow information from neighboring stations to generate a more global feature. This enhanced feature reflects the overall hydrological dynamics of the watershed and helps in discovering hidden patterns.

[0092] For example, K-means clustering, based on an enhanced feature matrix, divides sites into three classes.

[0093] For example, upstream stations are grouped into one category based on high rainfall, while midstream stations are grouped into another based on their flow characteristics. The clustering results can be used for zoned management, such as optimizing flood control measures for high-rainfall areas. This clustering improves the targeting and efficiency of watershed management.

[0094] In one embodiment, the overall solution forms a complete chain from data acquisition to clustering, with the core being the construction of the graph structure and GCN feature extraction. Extended solutions can incorporate time-series features to enrich dynamic analysis. Multiple aspects, such as distance, connectivity, and feature integration, collectively support the efficiency of watershed zoning, significantly improving water resource management and disaster prevention capabilities.

[0095] Step S102: The initial graph structure is aggregated with node features using a message passing mechanism. The node feature vector is updated by iterative calculation to capture the mutual influence of hydrological variables between stations. The hydrological variables include rainfall, flow rate and soil moisture.

[0096] Using a message passing mechanism, the neighbor node information of each node is obtained from the initial graph structure, and the influence of the neighbor node's feature vector on the current node is calculated. The formula is as follows:

[0097]

[0098] Among them The message vector of node i in the (t+1)th iteration, where N is the set of neighboring nodes of node i. Let W be the feature vector of neighbor node j in the t-th iteration. ij Given the edge weights between nodes i and j, we obtain the message vector for each node. Based on the message vectors of each node, we fuse the message vectors using an aggregation function, as shown in the formula:

[0099]

[0100] in Let i be the aggregated feature vector of node i in the (t+1)th iteration; Let k be the feature vector of node i in the t-th iteration, and k be the activation function, to obtain the aggregated node feature vector.

[0101] From the aggregated node feature vectors, determine whether the convergence condition is met. If it is, stop the iteration and determine the node feature vectors as follows. If the conditions are not met, the next iteration continues to obtain the updated node feature vectors. Based on the updated node feature vectors, the feature change trend of each node is calculated. Using these trends, the nodes are grouped using the K-means algorithm. From the grouping results, the central feature vector of each group is obtained, and the feature similarity between groups is calculated (using cosine similarity) to obtain the similarity matrix between groups. Based on the similarity matrix, the interaction strength of hydrological variables between groups is determined, thus obtaining the grouping relationship reflecting the dynamic interaction of hydrological variables.

[0102] For example, the message passing mechanism updates feature vectors through feature interactions between nodes. In a watershed monitoring scenario, assume there are four monitoring stations within the watershed, recording rainfall, flow rate, and soil moisture, respectively. In the initial graph structure, nodes represent stations, and edge weights are based on geographical distance and water system connectivity. The core of message passing is that each node collects information from its neighboring nodes.

[0103] For example, site A's neighbors are B and C. Site B has a rainfall of 40 mm and a flow rate of 80 cubic meters per second, while site C has a rainfall of 60 mm and a flow rate of 100 cubic meters per second. The message vector is generated by weighted summarization of neighbor features, with weight W... AB and W AC It reflects distance and water system relationships. This mechanism ensures that node characteristics incorporate information from the surrounding environment.

[0104] In one possible implementation, the aggregation function combines the message vector with the node's own features. The initial feature vector of site A is [0.5, 0.4, 0.6]. After summing the features of B and C, an activation function such as ReLU is applied to generate the aggregated feature vector. The aggregation process emphasizes the influence of neighbors on the current node.

[0105] For example, the rainfall feature of node A is enhanced by the high rainfall of nodes B and C. This aggregation method makes the node features more global.

[0106] It should be noted that the convergence condition is determined based on the change in the feature vector. Assuming the threshold ε is set to 0.01, if the difference between the aggregated feature vectors of station A and the difference between two iterations is less than 0.01, the iteration stops and the final feature vector is determined.

[0107] For example, if the feature vector of A is updated from [0.5, 0.4, 0.6] to [0.52, 0.42, 0.61], and the difference satisfies the condition, the iteration terminates. This method ensures feature stability.

[0108] Specifically, the feature change trend is calculated by comparing the initial and final feature vectors. The initial feature of station A is [0.5, 0.4, 0.6], the final feature is [0.52, 0.42, 0.61], and the change trend is [0.02, 0.02, 0.01]. The change trend reflects the dynamic evolution of the node in the hydrological environment.

[0109] For example, the slightly increased rainfall and flow characteristics of area A indicate that it was affected by the high rainfall in its neighboring area.

[0110] Preferably, the K-means algorithm groups stations based on their characteristic change trends. Assuming the four stations are divided into two groups, upstream stations A and B are grouped together due to similar rainfall change trends, while downstream stations C and D are grouped together. This grouping reflects the similarity of the stations in their hydrological characteristics.

[0111] For example, the high rainfall trends in A and B indicate active hydrology upstream.

[0112] In one embodiment, the similarity of the group center feature vectors is calculated using cosine similarity. Assuming the center features of group 1 are [0.03, 0.02, 0.01] and those of group 2 are [0.01, 0.01, 0.02], the similarity S... 12 A higher similarity indicates that the two groups of hydrological characteristics are similar. This similarity matrix quantifies the relationship between the groups.

[0113] For example, the strength of interaction between hydrological variables is determined based on a similarity threshold θ. If θ is set to 0.8, S 12 If the threshold is exceeded, it indicates that there is a strong interaction between group 1 and group 2.

[0114] For example, high rainfall in upstream group 1 may affect the flow in downstream group 2. Grouping relationships reveal the correlation of hydrological dynamics within the watershed. This approach, from message passing to interactive analysis, forms a complete scheme; extended schemes can introduce a time dimension to enrich dynamic interactive analysis.

[0115] Step S103: If the change in rainfall intensity in the node features exceeds a preset threshold, the water system connectivity parameters are adjusted and the edge weights are recalculated to generate a dynamically adjusted graph structure.

[0116] If the change in rainfall intensity in a node's features exceeds a preset threshold, the water system connectivity parameters are adjusted based on the rainfall intensity change, and the edge weights are updated to obtain a dynamically adjusted graph structure. Based on the dynamically adjusted graph structure, the set of neighboring nodes for each node is obtained, and the influence of the edge weights of the neighboring nodes on the current node is calculated to obtain the edge weight influence value for each node. If the edge weight influence value exceeds a preset influence threshold, standardization is used to normalize the edge weight influence value. Based on the normalized edge weight influence value, a hierarchical clustering algorithm is used to group the nodes. From the node grouping results, the influence value of the central edge weight of each group is obtained, and the edge weight similarity between groups is calculated (using cosine similarity) to obtain a similarity matrix between groups. If an element S in the similarity matrix between groups... kl If the similarity exceeds a preset threshold, group k and group l are determined to have strong connectivity. The water system connectivity parameters are then updated to obtain the grouping relationships that reflect the dynamic connectivity of the water system. Based on these grouping relationships, a weighted average method is used to calculate the overall connectivity strength of each group.

[0117] Specifically, in the context of watershed hydrological monitoring, dynamically adjusting the graph structure is key to optimizing water system connectivity analysis.

[0118] For example, changes in rainfall intensity can trigger edge weight updates. Taking a watershed as an example, suppose there are four monitoring stations A, B, C, and D. The initial edge weights are based on geographical distance, and the edge weight W between stations A and B is... AB It is 0.6. Rainfall intensity change ΔR A The value is 0.2, the adjustment coefficient α is 0.5, and the update formula W is... AB '=W AB (1+αΔR A After calculation, W AB The value is changed to 0.66. This adjustment reflects the dynamic impact of rainfall on water system connectivity and enhances the environmental adaptability of the edge weights.

[0119] Specifically, the calculation of the edge weight impact value focuses on the contribution of neighboring nodes. Site A's neighbors are B and C, and the updated edge weight W... AB ' is 0.66, W AC The value is 0.72. Calculate the edge weight influence value E of A. A The result is 0.66 + 0.72 = 1.38. This method quantifies the comprehensive impact of neighbors on the nodal hydrological environment, providing a data foundation for subsequent analysis.

[0120] It should be noted that if the edge weight has too high an impact value, standardization processing ensures data comparability, facilitates cross-site comparison, and highlights connectivity differences.

[0121] In one possible implementation, hierarchical clustering groups sites based on their normalized edge weight influence values. For example, sites A and B, with edge weight influence values ​​of 1 and 0.8 respectively, are grouped together due to their proximity; sites C and D, with edge weight influence values ​​of 0.4 and 0 respectively, are grouped into another group. This grouping reflects the similarity of the sites in hydrological connectivity, facilitating the identification of characteristics in different areas within the watershed.

[0122] Preferably, the edge weight similarity between groups is calculated using the influence value of the center edge weight. The center C1 of group 1 is 0.9, and the center C2 of group 2 is 0.2. The relationship between the two groups is quantified using the cosine similarity formula. If S 12 The value was 0.85, exceeding the threshold of 0.8, indicating strong connectivity between group 1 and group 2, necessitating an update of the river system parameters. This analysis reveals the intensity of hydrological interactions within the watershed.

[0123] Understandably, this approach progresses step-by-step from dynamic adjustments to connectivity analysis, ensuring a comprehensive capture of the watershed's hydrological characteristics. An extended scheme could incorporate time series data to analyze long-term trends in rainfall changes, further optimizing water system management.

[0124] Step S104: Input the dynamically adjusted graph structure into the graph neural network prediction layer, and calculate the predicted values ​​of rainfall, flow rate and soil moisture for each station through the fully connected layer mapping.

[0125] A graph neural network is used to process the dynamic graph structure, employing a message-passing mechanism to update the features of each station. Based on the updated station features, a fully connected layer is used for mapping to calculate the initial predicted value for each station. If the initial predicted value deviates from a preset threshold by more than a certain range, the weight matrix and bias parameters are optimized using a gradient descent algorithm. Based on the optimized weight matrix and bias parameters, the predicted values ​​for each station are updated again through a fully connected layer. If the updated predicted values ​​deviate significantly from historical data, time series analysis is used to adjust the predicted values. Based on the adjusted predicted values, a weighted average method is used to fuse the predicted values ​​of rainfall, flow, and soil moisture to obtain the comprehensive predicted value for each station. Cluster analysis is then used to group the comprehensive predicted values.

[0126] Specifically, in the context of watershed hydrological monitoring, graph neural networks update station features through a message passing mechanism, which can effectively capture the dynamic interaction relationships between stations.

[0127] For example, the core of the message passing mechanism lies in updating the characteristics of the target station using the characteristics of neighboring stations. Assume there are stations A, B, and C in the watershed, where A's neighbors are B and C. Station B has a rainfall characteristic of 10 mm, and station C has 15 mm. The message passing weight matrix is ​​set based on hydrological correlations. Through message passing, A's characteristics integrate the rainfall information from B and C, generating a more comprehensive feature representation. This method enhances the adaptability of station features to the watershed environment, providing a reliable foundation for subsequent predictions.

[0128] Specifically, the fully connected layer mapping transforms the updated features into initial predictions.

[0129] In one embodiment, the updated features of station A include rainfall and flow information. A predicted rainfall value for station A, such as 12 mm, is generated through mapping using a fully connected layer. This predicted value reflects the transformation of station features into specific hydrological variables, facilitating subsequent analysis.

[0130] It should be noted that if the predicted value deviates significantly from the actual observed value, for example, if the predicted value is 12 mm but the actual value is 15 mm, then the weight matrix and bias parameters need to be optimized.

[0131] Preferably, the gradient descent algorithm reduces prediction bias by iteratively adjusting the weights.

[0132] For example, the adjusted weights make the predicted values ​​closer to the actual observed values, thus improving the model accuracy.

[0133] In one possible implementation, the optimized weights and biases are used for remapping to update the predicted values.

[0134] For example, the rainfall forecast for station A was adjusted from 12 mm to 14 mm, which is closer to the actual value. This update ensures the dynamic adaptability of the forecast results.

[0135] Understandably, if the predicted values ​​deviate significantly from historical data, time series analysis can be used for further correction.

[0136] For example, if the historical average rainfall for region A is 13 mm and the adjustment factor is 0.5, the predicted value can be generated by weighted fusion of historical data to produce a more stable result. This method effectively reduces the interference of short-term fluctuations.

[0137] For example, the generation of composite forecasts is achieved by merging multivariate forecasts using a weighted average.

[0138] In one embodiment, the predicted rainfall for station A is 14 mm, the predicted flow rate is 200 m³ / s, and the predicted soil moisture is 1.2 m, with weights of 0.4, 0.3, and 0.3, respectively. The combined predicted value reflects multidimensional hydrological characteristics. This fusion improves the comprehensiveness of the prediction and facilitates watershed management.

[0139] Preferably, the cluster analysis groups the data based on the overall predicted values.

[0140] For example, stations A and B have similar composite forecast values ​​and are grouped together, while stations C and D are grouped together in another group. This grouping reveals the hydrological similarities among stations within the watershed, which is helpful for targeted management.

[0141] In one possible implementation, an extended scheme for time series analysis can further optimize forecasting.

[0142] For example, by analyzing the rainfall trend at site A over the past week and combining it with current forecasts, a more accurate long-term prediction can be generated. This method captures the dynamic changes in hydrological variables, improving the robustness of the prediction.

[0143] Understandably, the extended application of clustering and grouping can be combined with geographical location to refine watershed zoning management.

[0144] For example, A and B are located upstream, while C and D are located downstream; the grouping results guide water resource allocation in different regions. This progressive analytical logic, from feature updates to comprehensive predictions, fully supports watershed hydrological management.

[0145] Step S105: Perform time series analysis on the predicted values, extract hourly hydrological dynamic transmission features, and construct a coupled model based on dynamic Bayesian network and spatiotemporal co-Kriging interpolation algorithm.

[0146] Time series analysis was used to extract hourly dynamic features from hydrological data using a sliding window method. Based on these hourly dynamic features, a dynamic Bayesian network was constructed, and node state transition probabilities were defined to obtain node state probability distributions. A spatiotemporal co-kriging interpolation algorithm was employed to calculate spatial correlations based on the node state probability distributions, yielding spatial interpolation results. If the spatial interpolation results deviated from a preset threshold by more than a certain range, the weight coefficients were optimized using gradient descent to obtain optimized interpolation weights. Based on the optimized interpolation weights, the spatial interpolation results were recalculated to obtain updated spatial interpolation results. A data fusion method was used to combine the node state probability distributions and the updated spatial interpolation results, employing a weighted average to obtain a comprehensive predicted value. If the comprehensive predicted value deviated significantly from historical data, the model parameters were adjusted through a dynamic update mechanism to obtain optimized model parameters.

[0147] Specifically, in the context of watershed hydrological monitoring, time series analysis uses the sliding window method to extract hourly dynamic features, which can capture the short-term fluctuation patterns of variables such as rainfall and flow.

[0148] For example, the sliding window method takes rainfall data from the past 6 hours as input and combines it with a weight matrix to generate a feature representation for the current hour.

[0149] For example, if the rainfall at a certain station in the previous 6 hours was 8 mm, 10 mm, 12 mm, 15 mm, 14 mm, and 13 mm, a weighted fusion method can be used to generate the dynamic characteristics for the current hour, reflecting the rainfall trend. This method is convenient for capturing rapidly changing hydrological patterns and provides a foundation for subsequent analysis.

[0150] Specifically, a dynamic Bayesian network can be constructed based on hourly dynamic features to describe the dependencies between site states.

[0151] In one embodiment, assume there are stations A and B within the watershed, and the rainfall status of station A is affected by the flow status of station B. A Bayesian network calculates the probability of station A being in a high rainfall state using dynamic features.

[0152] For example, if the rainfall characteristic of region A is 12 mm and the flow characteristic of region B is 180 cubic meters per second, the network infers that the probability of high rainfall in region A is 0.7. This probability distribution provides a state basis for spatial analysis.

[0153] Preferably, the spatiotemporal co-Kriging interpolation algorithm uses the node state probability to calculate the hydrological values ​​of unmonitored points within the watershed.

[0154] For example, rainfall observations at stations A and B are 14 mm and 16 mm, respectively, while there is no observation data for target station C. By calculating the spatial correlation between A, B, and C, the interpolation estimate for rainfall at C is 15 mm. This method effectively fills in data gaps and improves watershed coverage.

[0155] It should be noted that if the interpolation result deviates significantly from the actual value, for example, if the actual rainfall in C is 17 mm, then the weight coefficients are optimized using gradient descent.

[0156] In one possible implementation, the initial weights are based on distance settings, which, after optimization, more accurately reflect the terrain influence between stations. After re-interpolation, the estimated value of C is adjusted to 16.5 mm, with a significant reduction in bias.

[0157] In one embodiment, the data fusion method integrates Bayesian network predictions and interpolation results to generate a comprehensive prediction value.

[0158] For example, A's Bayesian prediction of rainfall is 13 mm, the interpolation result is 14 mm, and the weighting coefficients are 0.6 and 0.4 respectively, resulting in a combined prediction of 13.4 mm. This fusion balances the advantages of probabilistic inference and spatial estimation, enhancing the stability of the prediction.

[0159] Understandably, if the overall predicted value deviates significantly from historical data, for example, if the historical average is 15 millimeters, the dynamic update mechanism will adjust the model parameters.

[0160] For example, by weighting the predictions with historical data, the predicted value was adjusted to 14.2 mm, which is closer to the long-term trend. This dynamic correction reduces the impact of short-term anomalies and improves the robustness of the model.

[0161] For example, in the extended scheme, time series analysis is combined with long-term trend optimization forecasting.

[0162] For example, by analyzing rainfall fluctuations at station A over the past 24 hours and combining them with current comprehensive forecasts, a more stable hourly forecast can be generated. This method captures the periodic changes in hydrological variables, providing a reliable basis for watershed management.

[0163] Step S106: Update the parameters of the graph neural network and optimize the message passing mechanism based on the spatial distribution characteristics of the output of the coupled model.

[0164] The interaction strength between nodes is obtained by leveraging spatial distribution characteristics. Based on this interaction strength, a set of neighboring nodes is determined. Node feature vectors are then obtained from the neighboring node set. Node messages are calculated using a message passing mechanism based on these feature vectors. A loss function is calculated using these node messages. If the loss function value exceeds a preset threshold, gradient descent is used to optimize the weight matrix and biases. The graph neural network parameters are updated based on the optimized weight matrix and biases, resulting in the optimized graph neural network parameters.

[0165] Specifically, in the context of watershed hydrological monitoring, the analysis of spatial distribution characteristics and the calculation of the interaction intensity between nodes are the core of constructing hydrological prediction models.

[0166] For example, spatial distribution characteristics can be understood as the distribution patterns of geographical locations, topographic elevations, or hydrological variables such as rainfall and flow rates at various monitoring stations within the watershed. By analyzing these characteristics, the mutual influences between stations can be quantified.

[0167] For example, station A is located on a hillside, where rainfall easily forms runoff, while station B is located downstream, and its flow is affected by A. The interaction intensity calculation reflects the degree of correlation between stations through spatial distance and topographical differences.

[0168] In one possible implementation, assuming that A and B are 5 kilometers apart and have a terrain elevation difference of 200 meters, the interaction intensity is weighted based on distance and elevation difference, generating a numerical representation of A's hydrological impact on B.

[0169] Specifically, the set of neighboring nodes is determined based on the interaction strength.

[0170] Preferably, a threshold is set to filter weakly associated nodes.

[0171] For example, if there are stations A, B, C, and D within a watershed, and the interaction strength between A and B / C is higher than a threshold, while that between A and D is lower than a threshold, then B and C are neighbors of A. This filtering ensures that only strongly associated nodes are considered, reducing computational redundancy.

[0172] In one embodiment, A's neighbor set includes B and C, which provide data on rainfall of 12 mm and flow rate of 150 cubic meters per second, respectively, for subsequent feature extraction. Obtaining the node feature vector involves integrating spatial distribution data and neighbor information.

[0173] Understandably, the feature vector is not only hydrological data from a single station, but also incorporates the dynamics of neighboring stations.

[0174] For example, if rainfall at location A is 10 mm, the flow rate at neighboring location B is 160 cubic meters per second, and rainfall at location C is 13 mm, a feature vector for location A is generated through weighted fusion, reflecting its hydrological status within the watershed network. The message passing mechanism further utilizes this feature vector to transmit dynamic information between stations.

[0175] In one possible implementation, the traffic characteristics of station B are converted into messages through a weight matrix and passed to A, reflecting the impact of B's ​​traffic changes on A's rainfall forecast.

[0176] For example, an increase in flow rate B might indicate increased rainfall upstream, and A adjusts its forecast accordingly upon receiving this information. The loss function is calculated to assess the deviation between the model's predictions and the actual values.

[0177] For example, A's predicted rainfall was 11 mm, but the actual rainfall was 12 mm. The loss function quantifies the error and guides parameter optimization.

[0178] In one embodiment, when the loss value exceeds a threshold, gradient descent is triggered to optimize the weight matrix and bias.

[0179] For example, the initial weights are based on distance settings, and after optimization, they more accurately reflect the relationship between topography and hydrology. After optimizing the weight matrix and bias, the graph neural network parameters are updated to improve prediction accuracy.

[0180] Specifically, the updated parameters better capture spatial dependencies between sites.

[0181] For example, the predicted rainfall for model A was adjusted from 11 mm to 11.8 mm, making it closer to the actual value. This dynamic adjustment enhances the model's adaptability to complex watershed environments.

[0182] For example, the extended scheme incorporates long-term spatial distribution trends to optimize forecasts. It analyzes the spatial fluctuations of rainfall over the past 24 hours in area A, integrates current information transmission results, and generates more stable hourly forecasts. This method effectively balances short-term fluctuations with long-term trends, providing a reliable basis for watershed hydrological management.

[0183] Step S107: If the deviation of the predicted value output by the optimized model is lower than the preset threshold, the final hydrological variable prediction result is output; otherwise, the dynamic adjustment and parameter optimization process is repeated.

[0184] Hydrological variable features are obtained from input data, and feature vectors are calculated using a feature extraction function. Based on these feature vectors, a machine learning model generates initial predictions. The prediction deviation is calculated using the initial predictions and actual values. If the deviation is below a preset threshold, the final hydrological variable prediction is output; otherwise, dynamic adjustment is performed. The model structure is dynamically adjusted based on the deviation, with feature weights redistributed using a structure update function. The parameters are optimized using a gradient descent algorithm based on the adjusted model structure. Prediction results are regenerated using the optimized parameters and calculated using the updated model. The deviation calculation and threshold judgment are repeated using the new predictions, iteratively executing until the deviation falls below the preset threshold, at which point the final hydrological variable prediction is output.

[0185] Specifically, in watershed hydrological prediction scenarios, the extraction of hydrological variable features is the foundation for building prediction models.

[0186] For example, hydrological variables include rainfall, flow rate, and evaporation. The role of the feature extraction function is to transform these raw data into feature vectors that reflect the hydrological state of the station.

[0187] Specifically, a monitoring station recorded 15 mm of rainfall and a flow rate of 200 cubic meters per second. The feature extraction function weighted and integrated these data to generate a multi-dimensional vector containing the station's hydrological characteristics. This method effectively captures the station's dynamic features within the watershed, providing reliable input for subsequent predictions. Using a machine learning model to generate initial predictions is one of the core steps.

[0188] In one possible implementation, the model uses feature vectors to calculate predicted values ​​using weight matrices and biases.

[0189] For example, a site's feature vector contains rainfall and traffic information. The model, based on a weight matrix trained using historical data, predicts that the rainfall for the next hour will be 12 millimeters.

[0190] It should be noted that the initial values ​​of the weight matrix are set based on the correlation between watershed topography and hydrology, while the bias is used to correct model bias. This approach ensures that the prediction results are closely related to the actual hydrological environment. The calculation of prediction bias is used to evaluate model performance.

[0191] Preferably, the deviation calculation function quantifies the error by comparing the predicted value with the actual value.

[0192] For example, if the predicted rainfall is 12 mm and the actual rainfall is 13 mm, the deviation calculation function generates an error value for subsequent optimization. This deviation assessment can intuitively reflect the accuracy of the model's predictions and provide a basis for dynamic adjustment. When the deviation of the predicted value exceeds a preset threshold, dynamically adjusting the model structure becomes a necessary step.

[0193] In one embodiment, the structure update function reallocates feature weights based on the deviation.

[0194] For example, if the flow characteristics of a particular station have a significant impact on the prediction, the update function will increase its weight. This adjustment can enhance the model's sensitivity to key hydrological variables and improve the adaptability of the prediction. The gradient descent algorithm is used to optimize the model parameters.

[0195] Understandably, the optimization process involves gradually adjusting the parameters by calculating the gradient of the bias with respect to the weights and biases.

[0196] For example, the initial weight matrix is ​​set based on distance and topographic differences, and after optimization, it more accurately reflects the hydrological relationships between stations. This method ensures that the model parameters can dynamically adapt to changes in the watershed environment. Regenerating the prediction results is a crucial step after optimization.

[0197] For example, the optimized model adjusts the predicted rainfall at a certain station from 12 mm to 12.8 mm, which is closer to the actual value. This iterative prediction can continuously approximate the true hydrological conditions. The process iteratively calculates the deviation and determines the threshold until the deviation falls below the threshold, at which point the final prediction result is output.

[0198] For example, the final forecast shows that the rainfall at a certain station will be 12.5 mm in the next hour, providing a reliable basis for hydrological management. This cyclical optimization mechanism can effectively improve the model's prediction accuracy. In the extended scheme, the forecast is further optimized by incorporating long-term hydrological trends.

[0199] In one embodiment, rainfall fluctuations at monitoring stations over the past 48 hours are analyzed and fused with current feature vectors to generate more stable prediction results. This method balances short-term dynamics with long-term patterns, providing more comprehensive support for watershed water resource management.

[0200] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A hydrological trend prediction method based on big data analysis, characterized in that, The method includes: The geographical coordinates, rainfall, flow rate and soil moisture data of monitoring stations within the watershed are obtained, and an initial graph structure is constructed. The initial graph structure uses the stations as nodes, and the feature vectors of the nodes include the rainfall, flow rate and soil moisture. The edge weights are calculated by the geographical distance between stations and the connectivity of the water system. The initial graph structure is aggregated using a message passing mechanism. The feature vectors of the nodes are updated through iterative calculation to capture the mutual influence of hydrological variables between stations. The hydrological variables include rainfall, flow rate, and soil moisture. If the change in rainfall in the node features exceeds a preset threshold, the water system connectivity is adjusted and the edge weights are recalculated to generate a dynamically adjusted graph structure. The dynamically adjusted graph structure is input into the graph neural network prediction layer, and the predicted values ​​of rainfall, flow rate and soil moisture for each station are calculated through mapping by the fully connected layer. Time series analysis is performed on the predicted values ​​to extract hourly hydrological dynamic transmission characteristics. A coupled model based on dynamic Bayesian networks and spatiotemporal co-Kriging interpolation algorithm is constructed, including: Hourly dynamic features were extracted from hydrological data through time series analysis, and the hourly dynamic features were obtained by using the sliding window method. Based on hourly dynamic characteristics, a dynamic Bayesian network is constructed, node state transition probabilities are defined, and the node state probability distribution is obtained. The spatiotemporal co-kriging interpolation algorithm is adopted to calculate spatial correlation based on the node state probability distribution and obtain spatial interpolation results; If the spatial interpolation result deviates from the preset threshold by more than a certain range, the weight coefficients are optimized by gradient descent. Based on the optimized interpolation weights, the spatial interpolation results are recalculated; By using data fusion methods, combining the node state probability distribution and the updated spatial interpolation results, a comprehensive predicted value is obtained; If there is a significant deviation between the overall predicted value and historical data, the model parameters will be adjusted through a dynamic update mechanism. Based on the spatial distribution characteristics of the output of the coupled model, the parameters of the graph neural network are updated and the message passing mechanism is optimized; If the deviation of the predicted value output by the optimized model is lower than the preset threshold, the final hydrological variable prediction result is output; otherwise, the dynamic adjustment and parameter optimization process is repeated.

2. The hydrological trend prediction method based on big data analysis according to claim 1, characterized in that, The edge weight is calculated using the following formula: W ij =D ij (1+ α (1-C ij )), In the formula, W ij D represents the edge weight between site i and site j; ij The geographical distance between site i and site j; α To adjust the parameters; C ij C is the connectivity factor; if there is a waterway connection between station i and station j, then C ij =1, otherwise C ij =0.

3. The hydrological trend prediction method based on big data analysis according to claim 1, characterized in that, The initial graph structure is aggregated using a message passing mechanism, and the node feature vectors are updated iteratively to capture the mutual influence of hydrological variables between stations. These hydrological variables include rainfall, flow rate, and soil moisture. By using a message passing mechanism, the neighbor node information of each node is obtained from the initial graph structure, and the influence of the neighbor node's feature vector on the current node is calculated. Based on the message vector of each node, an aggregation function is used to fuse the message vectors; From the aggregated node feature vectors, determine whether the convergence condition is met. If it is met, stop the iteration and determine the node feature vectors; if not, continue to the next round of iteration to obtain the updated node feature vectors. Calculate the feature change trend of each node based on the updated node feature vector; Based on the trend of feature changes, the K-means algorithm is used to group the nodes; From the grouping results of the nodes, obtain the central feature vector of each group, calculate the feature similarity between groups, and obtain the similarity matrix between groups; Based on the similarity matrix between groups, the intensity of interaction between hydrological variables is determined, and the grouping relationship reflecting the dynamic interaction of hydrological variables is obtained.

4. The hydrological trend prediction method based on big data analysis according to claim 1, characterized in that, If the change in rainfall intensity in the node features exceeds a preset threshold, the water system connectivity parameters are adjusted and the edge weights are recalculated to generate a dynamically adjusted graph structure, including: If the change in rainfall intensity in the node features exceeds a preset threshold, the water system connectivity parameters are adjusted and the edge weights are updated based on the change in rainfall intensity. Based on the dynamically adjusted graph structure, obtain the set of neighboring nodes for each node, calculate the influence of the edge weights of the neighboring nodes on the current node, and obtain the edge weight influence value of each node. If the influence value of the edge weight exceeds the preset influence threshold, the edge weight influence value is normalized by standardization to obtain the normalized edge weight influence value. Based on the normalized edge weight influence values, a hierarchical clustering algorithm is used to group the nodes to obtain the node grouping results; From the grouping results of the nodes, obtain the influence value of the center edge weight of each group, calculate the edge weight similarity between groups, and obtain the similarity matrix between groups; If an element in the similarity matrix between groups exceeds a preset similarity threshold, it is determined that the groups have strong connectivity, the water system connectivity parameter is updated, and the grouping relationship reflecting the dynamic connectivity of the water system is obtained. Based on the grouping relationships that reflect the dynamic connectivity of the water system, a weighted average method is used to calculate the comprehensive connectivity strength of each group, thus obtaining the comprehensive connectivity strength of each group.

5. The hydrological trend prediction method based on big data analysis according to claim 1, characterized in that, The step of inputting the dynamically adjusted graph structure into the graph neural network prediction layer, and calculating the predicted values ​​of rainfall, flow rate, and soil moisture for each station through fully connected layers includes: The dynamic graph structure is processed by a graph neural network, and the message passing mechanism is used to update the features of each site to obtain the updated site features. Based on the updated site characteristics, a fully connected layer mapping is used to calculate the initial predicted value for each site; If the initial predicted value deviates from the preset threshold by more than the limit, the weight matrix and bias parameters are optimized by the gradient descent algorithm to obtain the optimized weight matrix and bias parameters. Based on the optimized weight matrix and bias parameters, the predicted values ​​for each site are updated by remapping through the fully connected layer; If the updated forecast deviates significantly from historical data, the forecast will be adjusted through time series analysis. Based on the adjusted forecast values, a weighted average method is used to integrate the forecast values ​​of rainfall, flow, and soil moisture to obtain the comprehensive forecast value for each station. Cluster analysis was used to group the comprehensive predicted values ​​to obtain the grouping results of the site predicted values.

6. The hydrological trend prediction method based on big data analysis according to claim 1, characterized in that, The step of updating the parameters of the graph neural network and optimizing the message passing mechanism based on the spatial distribution characteristics output by the coupled model includes: The interaction strength between nodes can be obtained by analyzing their spatial distribution characteristics. The set of neighboring nodes is determined based on the strength of interactions between nodes; Obtain the node feature vector by using the set of neighboring nodes; Node messages are calculated using a message passing mechanism based on node feature vectors. Calculate the loss function based on node messages; If the loss function value is greater than the preset threshold, gradient descent is used to optimize the weight matrix and bias. Update the graph neural network parameters based on the optimized weight matrix and bias.

7. The hydrological trend prediction method based on big data analysis according to claim 1, characterized in that, If the deviation of the predicted value output by the optimized model is lower than a preset threshold, the final hydrological variable prediction result is output; otherwise, the dynamic adjustment and parameter optimization process is repeated, including: Hydrological variable characteristics are obtained by inputting data, and feature vectors are calculated using feature extraction functions. Initial prediction results are generated using a machine learning model based on the feature vectors of hydrological variables. The deviation of the predicted value is calculated by comparing the initial prediction results with the actual value data; If the predicted value deviation is lower than the preset threshold, the prediction result will be output. If the deviation exceeds the preset threshold, dynamic adjustment will be performed; Based on the deviation of the predicted values, the model structure is dynamically adjusted, and the feature weights are redistributed using a structure update function; The parameters are optimized using the gradient descent algorithm based on the adjusted model structure. Based on the optimized parameters, the prediction results are regenerated; Based on the new prediction results, the deviation calculation and threshold judgment are repeated, and the process is repeated until the deviation is lower than the preset threshold, and the final hydrological variable prediction results are output.

Images