Monitoring scarce point water quality prediction and early warning method based on liquid diagram neural network

Abstract

The invention provides a scarce point monitoring water quality prediction and early warning method based on a liquid diagram neural network, and relates to the technical field of water environment big data. The liquid graph neural network model is constructed based on the following method: according to river channel data, historical hydrological data and hydrological communication data of nodes, determining an initial connection weight, and constructing an adjacent matrix, a degree matrix and a feature matrix; adopting a liquid neuron continuous time evolution mechanism and a preset liquid differential equation to realize dynamic coding of a node hidden state; a two-channel network and a causal attention mechanism are introduced to update the initial adjacency matrix in real time, and a dynamic adjacency matrix is generated; and inputting a node hiding state and the updated adjacent matrix into the model, performing multi-objective optimization based on a graph Laplacian regular term, and optimizing the precision of the water quality prediction model of the data scarce point location. The method shows higher prediction precision and generalization ability in water quality prediction of data sparse point locations.

Description

Technical Field

[0001] The present invention relates to the technical field of water environment big data, and in particular, to a monitoring sparse point water quality prediction and early warning method based on a liquid graph neural network. Background Art

[0002] The prevention and control of basin water pollution is an important guarantee to ensure the health of the regional water environment. With the continuous enhancement of human activities, cross-regional pollution diffusion and sudden pollution incidents occur frequently. However, in the sparse areas of the monitoring network, there are monitoring blind spots and data continuity problems, and it is difficult for traditional methods to capture the deterioration of basin water quality in a timely manner. Therefore, it is urgent to construct a water quality prediction model suitable for sparse data points to break through the technical bottleneck of the coupling of hydrological topology and dynamic environment.

[0003] With the deep penetration of artificial intelligence into the water environment field, data-driven modeling provides a new reliable tool for basin water quality early warning. Compared with traditional mechanism models that rely on prior mechanisms and have complex parameters, machine learning models can mine the correlation features in data, are more flexible in modeling and have stronger adaptability, and have been widely used in scenarios such as basin water quality prediction and mutation early warning in recent years. However, existing machine learning algorithms have application limitations in the environment of sparse monitoring networks. Under the conditions of discontinuous or insufficiently covered monitoring data, related algorithms rely on historical data under static topological constraints for information reasoning, which easily leads to the accumulation of errors between nodes, restricting the robustness and generalization performance of the model, and it is difficult to meet the accurate water quality early warning requirements under scarce monitoring. Summary of the Invention

[0004] The problem to be solved by the present invention is that water quality prediction mostly depends on high-quality time series data of dense monitoring points, and the characterization ability of spatial heterogeneity features such as hydrological topological structure and upstream-downstream correlation response is insufficient. There is insufficient spatio-temporal coupling in the data-scarce scenario, and the prediction accuracy is poor.

[0005] To solve the above problems, the present invention provides a monitoring sparse point water quality prediction and early warning method based on a liquid graph neural network, and the liquid graph neural network model is constructed based on the following method: According to the river channel data, historical hydrological data and hydrological connectivity data of the nodes, determine the initial connection weights between each pair of adjacent nodes, and construct the initial adjacency matrix, degree matrix and feature matrix of each node, wherein each node includes the target site and the surrounding monitoring sites, and each pair of adjacent nodes are two nodes with a downstream relationship; Determine the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node, and generate a comprehensive input of each node for a preset liquid differential equation according to the feature matrix of each node, the initial adjacency matrix, and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node; Construct a topological channel update network according to the hidden state of each node at the previous moment and the degree matrix, construct a hydrological channel update network according to the real-time hydrological data, update the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introduce a mask matrix through a joint attention mechanism to obtain a dynamic adjacency matrix of each node, where the mask matrix is used to mask the reverse flow relationship and future moment data in the initial adjacency matrix, and the degree matrix is updated based on the update of the initial adjacency matrix; Input the hidden state of each node at the previous moment and the dynamic adjacency matrix into the initial neural network model, output the predicted values of the water quality parameters of each node, and optimize and verify the parameters of the initial neural network model to obtain a liquid graph neural network model.

[0006] A method for predicting and warning the water quality of sparse points based on a liquid graph neural network provided by the present invention, and the liquid graph neural network model is constructed based on the following method: According to the river channel data, historical hydrological data, and hydrological connectivity data of nodes, determine the initial connection weights between each pair of adjacent nodes, and construct the initial adjacency matrix, degree matrix, and feature matrix of each node, providing structural and feature information for the liquid neural network; By dynamically adjusting the connection weights, combined with the static topological structure (adjacency matrix, degree matrix), simulate the real-time update process of the node state through differential equations under the liquid neural network, generate the spatio-temporal feature matrix after dynamic update, that is, the hidden state of each node, and realize the full expression of features caused by data loss; Ensure the physical causality constraint of pollutant migration through a dual-channel network, avoid future information interference, and fuse the updated degree matrix, thereby obtaining the updated dynamic adjacency matrix, which can simulate the non-steady-state transmission process of pollutants changing with hydrological conditions. Finally, use the generated hidden state and the updated dynamic adjacency matrix as inputs for multi-objective dynamic prediction. Since the hidden state provides the node's own and its historical feature information, reflecting the heterogeneity of the spatial distribution within the basin at the network structure level, and the dynamic adjacency matrix is used to adjust the weights of information propagation between nodes, making the aggregated features (hidden state) more in line with the current hydrological environment. The combination of these two ensures that the constructed liquid graph neural network model considers both the local states of each node and the global spatio-temporal dependence relationship during prediction. Generally speaking, by introducing continuous-time differential equations and liquid neurons under the liquid graph neural network, the dynamic evolution of node states and adaptive weight updates are realized, and the flexible adjustment of the network structure can be driven by real-time hydrological parameters, overcoming the problem of insufficient spatio-temporal coupling under data scarcity conditions in related technologies, showing good robustness. At the same time, the liquid graph neural network embeds an attention mechanism with physical constraints to more accurately capture the lag effect and path dependence of pollutants, thereby improving the physical interpretability of the model and ensuring that the prediction results are more in line with the actual migration law of pollutants. Based on the above characteristics, the model constructed in the embodiments of the present invention shows higher prediction accuracy and generalization ability at sparse data points when used for predicting and warning the water quality of sparse points, providing a new technical path for basin pollution prediction and having significant application prospects.

[0007] The present invention also provides a construction system for a liquid graph neural network model, including: A construction module, configured to determine the initial connection weights between each pair of adjacent nodes according to the river channel data, historical hydrological data, and hydrological connectivity data of the nodes, and construct the initial adjacency matrix, degree matrix, and feature matrix of each node, where each of the nodes includes a target site and the surrounding monitoring sites, and each pair of adjacent nodes are two nodes with a downstream relationship; A feature encoding module, configured to determine the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node, generate a comprehensive input of each node for a preset liquid differential equation according to the feature matrix of each node, the initial adjacency matrix, and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node; A perception update module, configured to construct a topological channel update network according to the hidden state of each node at the previous moment and the degree matrix, construct a hydrological channel update network according to the real-time hydrological data, update the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introduce a mask matrix through a joint attention mechanism to obtain a dynamic adjacency matrix of each node, where the mask matrix is used to mask the countercurrent relationship and future moment data in the initial adjacency matrix, and the degree matrix is updated based on the update of the initial adjacency matrix; An optimization module, configured to input the hidden state of each node at the previous moment and the dynamic adjacency matrix into an initial neural network model, output the predicted values of the water quality parameters of each node, and optimize and verify the parameters of the initial neural network model to obtain a liquid graph neural network model.

[0008] For the construction system of the liquid graph neural network model provided by the present invention, its beneficial effects can refer to the beneficial effects of the construction of the above-mentioned liquid graph neural network model, and will not be described herein again. Description of the Drawings

[0009] Figure 1 Shows the basin space topological map in the embodiment of the present invention; Figure 2 Shows the schematic flow chart of the construction method of the liquid graph neural network model in the embodiment of the present invention; Figure 3 Shows the schematic diagram of the spatio-temporal feature dynamic encoding mechanism driven by liquid neurons in the embodiment of the present invention; Figure 4 Shows the schematic diagram of the liquid causal perception mechanism in the embodiment of the present invention; Figure 5 Shows the NSE graph of the total nitrogen concentration prediction model based on the liquid graph neural network in the embodiment of the present invention; Figure 6 Shows the comparison graph of the predicted NSE effects of different models in the embodiment of the present invention; Figure 7 Shows the comparison graph of the predicted and measured values of total nitrogen in the water quality parameters in the embodiment of the present invention; Figure 8The structural schematic diagram of the construction system of the liquid graph neural network model in the embodiment of the present invention is shown; Figure 9 The flowchart of the water quality prediction and early warning method in the embodiment of the present invention is shown; Figure 10 The node influence Sankey diagram in the embodiment of the present invention is shown. Detailed implementation manners

[0010] To make the above objects, features, and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0011] It should be noted that relational terms such as "first" and "second" in the present invention are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article, or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device including the element.

[0012] In the description of this specification, the descriptions referring to terms such as "embodiment", "one embodiment", and "one implementation manner" mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or implementation manner are included in at least one embodiment or implementation manner of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or implementation manner. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or implementation manners in a suitable manner.

[0013] Refer to Figure 1 And Figure 2 As shown, an embodiment of the present invention proposes a monitoring scarce point water quality prediction and early warning method based on a liquid graph neural network, and the liquid graph neural network model is constructed based on the following method: S100: Multi-source heterogeneous spatio-temporal data fusion and construction of a watershed spatial topology map. According to the river channel data, historical hydrological data, and hydrological connectivity data of the nodes, determine the initial connection weights between each pair of adjacent nodes, and construct the initial adjacency matrix, degree matrix, and feature matrix of each node, where each node includes a target site and the surrounding monitoring sites, and each pair of adjacent nodes are two nodes with a downstream relationship.

[0014] Specifically, the target sites are those with less or missing water quality data. According to the spatial location of monitoring points, the upstream and downstream flow directions, and the river channel topological relationship, etc., a heterogeneous graph of the watershed space is constructed with surrounding monitoring sites as known nodes, data-sparse points as target sites (nodes to be inverted), and hydrological associations as edges, as shown in Figure 1 shown below. Figure 1 In the figure, 1 - 10 are surrounding monitoring sites, and points A, B, and C are target sites. Obtain the spatial relationships between the sites, such as the geographical locations of the nodes, the distances between them, and the water flow directions, etc. These information helps to understand the spatial correlation between the sites. In addition, it is also necessary to collect watershed characteristic data, including but not limited to the topography and landforms of the watershed, land use types, vegetation coverage, etc. These characteristics have an important impact on water quality. By integrating these data, comprehensive basic information is provided for subsequent model construction and analysis, that is, a feature matrix of the nodes is constructed. The initial connection weight is used to characterize the relationship between the water flow velocity and the distance attenuation coefficient, which directly reflects the weight of information transfer between nodes. A higher initial connection weight means that node j contributes more to node i. The initial adjacency matrix represents the adjacent relationship between nodes in the heterogeneous Figure 1 graph of the watershed space, reflecting the connection relationships between water flow directions, water flow paths, and diffusion paths. The degree matrix is a diagonal matrix, and its diagonal elements are the sum of the connection numbers of the corresponding nodes in the adjacency matrix. All the above data provide structural and feature information for the liquid neural network and can be used to construct a liquid graph neural network model.

[0015] S200: Spatiotemporal feature dynamic encoding. Determine the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node. Generate the comprehensive input of each node for a preset liquid differential equation according to the feature matrix of each node, the initial adjacency matrix, and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node.

[0016] Specifically, real-time hydrological data generally includes data such as rainfall and flow velocity between nodes. These data will affect the connection weights. Therefore, by combining real-time hydrological data, the initial connection weights can be updated to obtain the dynamic connection weights between nodes. By combining the dynamic connection weights, the feature matrix, and the initial adjacency matrix, through the continuous-time evolution mechanism of liquid neurons, the spatial heterogeneity constraint is dynamically fused (that is, the connection weights between nodes are dynamically adjusted in combination with hydrological parameters, and the water quality impact signals of upstream and downstream nodes are aggregated to generate a fused spatial heterogeneity constraint. The adjacency matrix is a static matrix that is the basis of spatial heterogeneity. When the model calculates the hidden state of a node, it will aggregate the water quality impact signals from adjacent nodes, especially upstream nodes, according to the adjacency matrix and the dynamic connection weights, so as to reflect the spatial heterogeneity within the basin at the network structure level. Spatial heterogeneity refers to the difference in the impact of nodes at different spatial positions within the basin on pollutant transport reflected in the process of generating the hidden state by dynamically adjusting the connection weights in combination with the static topological structure adjacency matrix and degree matrix). And time series features, that is, using the liquid differential equation to simulate the real-time update process of the node state, dynamically adjusting the connection weights between nodes in combination with hydrological parameters, aggregating the water quality impact signals of upstream and downstream nodes, fusing the spatial heterogeneity constraint and the time dynamic response, generating a dynamically updated spatio-temporal feature matrix, that is, the hidden state of each node, to achieve the full expression of features caused by data loss.

[0017] S300: Spatiotemporal propagation perception under causal constraints. According to the hidden state of each node at the previous moment and the degree matrix, construct a topological channel update network. According to the real-time hydrological data, construct a hydrological channel update network. According to the topological channel update network and the hydrological channel update network, update the initial adjacency matrix, and introduce a mask matrix through a joint attention mechanism to obtain the dynamic adjacency matrix of each node, where the mask matrix is used to mask the reverse flow relationship and future moment data in the initial adjacency matrix, and the degree matrix is updated based on the update of the initial adjacency matrix.

[0018] Specifically, based on the real-time hydrological data driving the dynamic update of the connection weights between nodes (essentially the update of the degree matrix), introduce an attention mask that only depends on historical moments and upstream nodes to ensure the physical causal constraint of pollutant migration and avoid interference from future information. After fusing the two, the non-steady-state transport process of pollutants with changes in hydrological conditions can be simulated, and thus the updated dynamic adjacency matrix can be obtained.

[0019] S400: Dynamic prediction with multi-objective optimization. Input the hidden states of each node at the previous moment and the dynamic adjacency matrix into the initial neural network model, output the predicted values of the water quality parameters of each node, and optimize and validate the parameters of the initial neural network model to obtain a liquid graph neural network model, which is used for water quality prediction and early warning.

[0020] Specifically, use the generated hidden states and the updated dynamic adjacency matrix as inputs for multi-objective dynamic prediction. The hidden states provide information about the nodes themselves and their historical features, while the dynamic adjacency matrix is used to adjust the weights of information propagation between nodes, making the feature aggregation more in line with the current hydrological environment. The combination of these two ensures that the model after parameter optimization and validation takes into account both the local states of each node and the global spatio-temporal dependence relationships during prediction, thus effectively improving the prediction accuracy, and making the early warning based on multi-objective dynamic prediction more accurate.

[0021] When this embodiment is applied in practice, by determining the initial connection weights between each pair of adjacent nodes based on the channel data, historical hydrological data, and hydrological connectivity data of the nodes, constructing the initial adjacency matrix, degree matrix, and feature matrix of each node, it provides structural and feature information for the liquid neural network; by dynamically adjusting the connection weights and combining the static topological structure (adjacency matrix, degree matrix), the real-time update process of the node state is simulated through differential equations under the liquid neural network to generate the spatio-temporal feature matrix after dynamic update, that is, the hidden state of each node, realizing the full expression of features caused by data loss; by ensuring the physical causality constraint of pollutant migration through a dual-channel network, avoiding future information interference, and fusing the updated degree matrix, the updated dynamic adjacency matrix can be obtained, which can simulate the non-steady-state transport process of pollutants changing with hydrological conditions. Finally, taking the generated hidden state and the updated dynamic adjacency matrix as inputs for multi-objective dynamic prediction. Since the hidden state provides the node's own and its historical feature information, reflecting the heterogeneity of the spatial distribution within the basin at the network structure level, while the dynamic adjacency matrix is used to adjust the weights of information propagation between nodes, making the aggregated features (hidden state) more in line with the current hydrological environment. The combination of these two ensures that the constructed liquid graph neural network model takes into account both the local states of each node and the global spatio-temporal dependence relationship during prediction. In summary, by introducing continuous-time differential equations and liquid neurons under the liquid graph neural network, the dynamic evolution of node states and adaptive weight updates are realized, enabling the flexible adjustment of the network structure driven by real-time hydrological parameters, overcoming the problem of insufficient spatio-temporal coupling under data-scarce conditions in related technologies. At the same time, the attention mechanism with physical constraints embedded in the liquid graph neural network can more accurately capture the lag effect and path dependence of pollutants, thereby improving the physical interpretability of the model and ensuring that the prediction results are more in line with the actual migration law of pollutants at data-scarce points. Based on the above characteristics, the model constructed in the embodiment of the present invention shows higher prediction accuracy and generalization ability at data-scarce points when used for monitoring the prediction and early warning of rare-point water quality, providing a new technical path for basin pollution prediction and having significant application prospects.

[0022] As an optional embodiment of the present invention, the determining the initial connection weights between each pair of adjacent nodes based on the channel data, historical hydrological data, and hydrological connectivity data of the nodes, and constructing the initial adjacency matrix, degree matrix, and feature matrix of each node includes: Determining the initial connection weights between each pair of adjacent nodes according to the channel distance, river flow velocity, and riverbed permeability coefficient between each pair of adjacent nodes; Specifically, the initial connection weight adopts an improved attenuation coefficient formula, ; In the formula: represents the initial connection weight; represents the river channel distance from node i to j, where i is any node and j is any node other than i, and it is cyclic. For example, when i = 1, j is 2 - 10; when i = 2, j is 1 and 3 - 10; represents the river flow velocity, represents the permeability coefficient of the riverbed.

[0023] Construct the initial adjacency matrix of each node according to the river channel distance between each pair of adjacent nodes, the average node distance of each node, and the distance attenuation coefficient. Construct the degree matrix of each node according to the initial adjacency matrix of each node; Specifically, the initial adjacency matrix is as follows: ; where, represents the initial adjacency matrix, represents the river channel distance from node i to j, represents the average distance between each pair of adjacent nodes in the basin, α represents the distance attenuation exponent, taking 1.5 - 2.0, which reflects the diffusion intensity of pollutants with distance. For the determination of, for example, under the premise of having a connection relationship, such as are 1km, 5km, and 10km respectively, then =(1 + 5 + 10) / 3 = 5.33.

[0024] Construct the feature matrix of each node according to the historical hydrological data and the hydrological connectivity data of each node; among them, the historical hydrological data includes historical meteorological time series data, historical water quality data, and vegetation coverage index, the river channel data includes the river channel distance, the river flow velocity, the permeability coefficient of the riverbed, the average node distance, and the distance attenuation coefficient, the hydrological connectivity data includes the node spatial topology relationship, and the historical water quality data of the target site is generated by an adversarial network.

[0025] Specifically, the feature matrix X contains the attributes of each node, such as water quality parameters, meteorological factors, etc. The attributes of each node are 22 - dimensional, where static attributes, dynamic time series, spatial topology, etc. are integrated into multi - dimensional features. Table 1 below gives the detailed content.

[0026] Table 1:

[0027] Construct the degree matrix according to the adjacency matrix. The degree matrix is a diagonal matrix, and its diagonal elements are the sum of the connection numbers of the corresponding nodes in the adjacency matrix.

[0028] The exemplary adjacency matrix is shown in Table 2 below.

[0029] Table 2: 。

[0030] Exemplarily, the degree matrix obtained according to Table 2 is shown in Table 3 below.

[0031] Table 3: 。

[0032] When this embodiment is applied in practice, historical water quality data, vegetation coverage index, meteorological time series data, etc. of monitoring points in the basin are collected, and the above data are preprocessed, including: finding and deleting the error values of the various data; setting reasonable upper and lower threshold values and performing capping processing on the data exceeding the threshold; filling in the missing values with a random forest regression model; performing normalization processing to eliminate the dimension difference; ensuring that the data time step lengths of each site are consistent, averaging and downscaling the hourly data to the daily scale data, and generating a spatially and temporally aligned model input data set; collecting the vegetation coverage index and meteorological time series data, etc. of data-sparse points (target sites), performing the above data preprocessing, using an adversarial network to generate historical water quality auxiliary pseudo-node data, forming a spatio-temporal data set covering the whole region, thereby obtaining a feature matrix, and the generation of the adjacency matrix and the degree matrix can take into account the dynamic connection weights (including water flow velocity, distance attenuation coefficient).

[0033] As Figure 3 shown, as an optional embodiment of the present invention, the determining of the dynamic connection weight between each pair of adjacent nodes according to the initial connection weight between each pair of adjacent nodes and the real-time hydrology Number data of each node includes: Determining a dynamic correction factor between each pair of adjacent nodes according to the real-time hydrology data of each node and the learning parameters, where the dynamic correction factor is used to characterize the driving relationship of the linear combination of the flow velocity gradient and the rainfall change on the initial connection weight.

[0034] Specifically, the dynamic adjustment of the correction factor is expressed as follows: 。

[0035] In the formula: : The node and the node between the time of the dynamic correction factor; : Hyperbolic tangent function; : Cross-sectional flow velocity, including ; : rainfall intensity; , : learnable parameter; : channel length from node i to j.

[0036] Determine the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights and the dynamic correction factors.

[0037] Specifically, it is expressed as the following formula: .

[0038] represents the initial connection weight, represents the dynamic connection weight. The dynamic connection weight is a quantitative representation of the information transfer strength between nodes in the model. The connection strength is the weight of information transfer between nodes. A higher means that the contribution degree of node j to node i is greater, and the reference value of the model is higher.

[0039] In practical application of this embodiment, the dynamic adjustment factor drives the change of the dynamic connection weight through the real-time hydrological response equation ( formula), and realizes the mechanism-data hybrid drive of hydrodynamics principle and machine learning through the linear combination of flow velocity gradient and rainfall change. When > , it indicates the existence of accelerating flow, that is, the enhancement of pollutant transport (there is a case of less than), and the correction factor increases positively; through to capture the driving force of rainfall events on surface runoff. A positive value indicates that rainfall is increasing, which will increase the input intensity of non-point source pollution, so as to ensure the dynamic update of the dynamic connection weights between each pair of adjacent nodes.

[0040] As an alternative embodiment of the present invention, generating the comprehensive input of each node for the preset liquid differential equation according to the feature matrix of each node, the initial adjacency matrix and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node includes: Determine the eigenvectors of the feature matrix of each node, and generate the comprehensive input of each node for the preset liquid differential equation according to the eigenvectors and the dynamic connection weights, so as to output the hidden state of each node; Among them, the preset liquid differential equation is:

[0041] ; represents the node at time in the hidden state, used to map node features to the hidden space; represents a non-linear activation function, namely the ReLU function; represents a learnable bias term represents a learnable parameter matrix, which is automatically learned through model training without manual setting, and is used for mapping the feature vector of the node to the hidden space; represents the node at time of the said feature vector, represents the dynamic connection weight, represents the feature concatenation operation, represents a learnable bias term, represents the time delay, represents the said comprehensive input, j represents the node adjacent to the node represents the node adjacent to the node, and the set of the said nodes.

[0042] Specifically, the above uses an improved liquid time constant network for liquid neuron design, uses the liquid differential equation to simulate the real-time update process of the node state, dynamically adjusts the connection weight between nodes in combination with hydrological parameters, aggregates the water quality impact signals of upstream and downstream nodes, and generates a spatio-temporal feature matrix of data sparse points that integrates spatial heterogeneity constraints (achieved by dynamically adjusting the dynamic connection weight, the adjacency matrix is a static matrix which is the basis of spatial heterogeneity, when the model calculates the node hidden state, it will aggregate the water quality impact signals from adjacent nodes (especially upstream nodes) according to the adjacency matrix and the dynamic connection weight, so as to reflect the spatial heterogeneity within the basin at the network structure level) and time dynamic response (i.e., the hidden state matrix of the node, the set of the hidden states of all nodes at the current moment (or a time series)), and solves the problem of insufficient feature expression caused by data missing.

[0043] When this embodiment is applied in practice, through the continuous time evolution mechanism of the liquid neuron (reflected in the preset liquid differential equation), it can dynamically fuse the spatial topological dependence relationship and the time series characteristics, use the liquid differential equation to simulate the real-time update process of the node state, dynamically encode the original data such as node attributes, adjacency matrix and degree matrix through the liquid time constant network, generate the hidden state reflecting the spatio-temporal characteristics of the node, and finally output a dynamically updated spatio-temporal feature matrix, that is, the hidden state of each node.

[0044] As an alternative embodiment of the present invention, constructing a topological channel update network according to the hidden states of the nodes at the previous moment and the degree matrix, constructing a hydrological channel update network according to the real-time hydrological data, updating the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introducing a mask matrix through a joint attention mechanism to obtain a dynamic adjacency matrix includes: Determine attention scores based on the hidden states of the nodes at the previous moment, construct a spatial mask matrix and a temporal mask matrix, and fuse and process the spatial mask matrix and the temporal mask matrix and apply them to the attention scores to obtain joint attention.

[0045] Specifically, select the hidden state at the previous moment. The attention score reflects the importance of different elements in the input sequence. In the attention mechanism, the query matrix, key matrix, and value matrix are all obtained by linear transformation of the hidden state. Then, the dot product of the query and the key is scaled and processed by Softmax to obtain the attention score, and combined with the mask matrix to obtain a new hidden state.

[0046] Capture the hidden states of the nodes at the previous moment through a graph convolutional network, and combine with the degree matrix to construct the topological channel update network. Capture the spatio-temporal convolution of the real-time flow velocity field matrix and the turbulence feature matrix through the graph convolutional network to construct the hydrological channel update network, where the real-time flow velocity field matrix and the turbulence feature matrix are determined according to the real-time hydrological data.

[0047] Specifically, based on real-time hydrological data-driven dynamic update of the connection weights between nodes, simulate the unsteady transmission process of pollutants with changes in hydrological conditions. Capture local water flow mutations caused by events such as rainstorms and gate control through spatio-temporal convolution of the flow velocity field and turbulence features to construct a hydrological channel network; use graph convolution to analyze the propagation mode of node states and identify potential pollution paths to construct a topological channel network.

[0048] Merge and activate the topological channel update network and the hydrological channel update network, and based on the initial adjacency matrix, fuse the joint attention to obtain the dynamic adjacency matrix of each node.

[0049] Specifically, construct a dual-channel update network including a hydrological channel network and a topological channel network, and combine joint attention to generate a dynamic adjacency matrix reflecting the real-time system state on the basis of the initial adjacency matrix, providing a more accurate graph structure representation for prediction and control tasks in complex environments, that is, based on real-time hydrological data-driven dynamic update of the connection weights between nodes, simulate the unsteady transmission process of pollutants with changes in hydrological conditions.

[0050] Among them, the mask matrix includes a spatial mask matrix and a temporal mask matrix. The spatial mask matrix is used to mask the countercurrent relationship in the initial adjacency matrix, and the temporal mask matrix is used to mask the data at future times in the initial adjacency matrix.

[0051] Specifically, the causal attention mask is designed as follows: Through the upper triangular mask matrix , enforce compliance with the unidirectionality of the water flow, avoid the spread of countercurrent pollution, and construct a spatial constraint mask: ; In the formula: : The spatial mask matrix, which only allows information to flow from upstream to downstream.

[0052] As Figure 4 shown, adopt a strictly causal convolutional kernel to ensure causality and avoid interference from future data, and define a temporal constraint mask: ; In the formula: : The temporal mask matrix, which only uses previous data to update subsequent data.

[0053] Through the mask mechanism, ensure that the model decision complies with the laws of water flow dynamics, and can constrain the search space, shortening the model training time.

[0054] As an optional embodiment of the present invention, the dynamic adjacency matrix is represented by the formula: ; Among them, represents the dynamic adjacency matrix, represents the initial adjacency matrix, represents an activation function, which is used to map the input variable to between [0, 1], normalize and adjust the feature weights, helps to stabilize the training and keep the updated weights within a reasonable range; represents the hydrological channel update network, and ; represents the topological channel update network, and ; represents a graph convolutional network, represents the real-time flow velocity field matrix, indicating the cross-sectional flow velocity of each node at time t, represents the turbulent feature matrix (which can be represented by the three-dimensional eddy viscosity coefficient and reflects the intensity of river channel turbulent mixing), represents the hidden state at the previous moment; represents the degree matrix, which contains the real-time in-degree / out-degree of nodes; ; where respectively represent the query matrix, the key matrix, and the value matrix, and the three are respectively obtained by linear transformation of the hidden state at the previous moment, represents the vector dimension of the key matrix, represents the spatial mask matrix, represents the temporal mask matrix, represents the normalized exponential function, which is used to convert each element in the input vector into a probability distribution.

[0055] Specifically, For the query matrix, the hidden state of the node is input as the query vector; The key matrix represents the historical state of each node or the features of its adjacent nodes it depends on; represents the value matrix, which contains the node features (hidden states) that can be passed to the target site; represents the dimension of the key vector, which is used to scale the dot product result; represents , which converts the attention scores into a probability distribution, Q (query matrix), indicating which information the current node i needs to pay attention to, is obtained by projecting and changing the hidden state of node i through to obtain the query vector; K (key matrix), indicating which information other nodes j can provide, comes from the hidden state of node j through projecting and changing to obtain the query vector; V (value matrix), indicating the specific feature information actually passed to the target site, comes from the hidden state of node j through projecting and changing to obtain the query vector. Q (query matrix), K (key matrix), and V (value matrix) all indirectly come from the hidden state generated in step three, and only achieve feature projection through different linear transformations, , and are the weight matrices of Q, K, and V respectively, and the previous all belong to the built-in tuning parameters.

[0056] In summary, by adding the above-mentioned spatial mask matrix and temporal mask matrix and processing the attention scores, the requirements of temporal causality and spatial unidirectionality are satisfied (that is, the causal attention mask is applied simultaneously to block reverse flow information and future moment information), ensuring that the model conforms to hydrological unidirectionality and temporal causality, and the priority of information propagation between nodes can be determined according to spatio-temporal characteristics, improving the model training efficiency and prediction accuracy.

[0057] As an alternative embodiment of the present invention, inputting the hidden states of the nodes at the previous moment and the dynamic adjacency matrix into the initial neural network model, outputting the predicted values of the water quality parameters of the nodes, and optimizing and validating the parameters of the initial neural network model to obtain the liquid graph neural network model includes: Based on the predicted values of the water quality parameters of the nodes, a loss function is constructed, and based on the loss function and the K-fold cross-validation mechanism, the optimal hyperparameters of the initial neural network model are determined to obtain the liquid graph neural network model.

[0058] Specifically, through loss function optimization and spatial generalization verification in the training stage: based on the K-fold cross-validation framework (K = 10), perform "leave-one-out" testing in the region - each iteration masks 1 known surrounding monitoring site (sites numbered 1-10 in total), and trains the model with the topological relationships and historical data of the remaining K - 1 sites. When the accuracy of cross-validation meets the preset accuracy, predict the pollutant concentration time series of the masked points, calculate the Nash-Sutcliffe Efficiency coefficient (Nash-Sutcliffe Efficiency, NSE). This process traverses all surrounding monitoring points, and finally obtains the NSE of each node, see Figure 5 the circled part in

[0059] Subsequently, perform prediction verification on the target sites (i.e., scarce sites): after obtaining the optimal model parameters, for the three data-scarce sites A, B, and C that did not participate in the training (each site has measured data for ≤ 5% of the time steps), input the initial adjacency matrix, degree matrix, and feature matrix, output the predicted values of the water quality parameters for each day in the next 7 days, and then compare the predicted values with the sparse historical measured data to obtain the NSE of the three data-scarce sites A, B, and C, see Figure 5 the triangular part in

[0060] Figure 6 represents the comparison chart of the predicted NSE effects of different models. By comparing random forest, long short-term memory network, graph neural network, and the liquid neural network in the present invention, it can be found from Figure 6 that the technical solution under the liquid neural network in the present invention has the highest prediction efficiency.

[0061] Figure 7 The comparison of the prediction results of total nitrogen in water quality parameters and the measured values is given. As can be seen from Figure 7 the prediction and effect comparison of the technical solution of the present invention based on a liquid neural network, most of the predicted values follow the measured values, indicating the accuracy of the prediction of the present invention.

[0062] As an optional embodiment of the present invention, the loss function is: ; where represents the total loss function, which is an index for measuring the prediction error and the regularization effect, N represents the node the set of the nodes adjacent to the current moment represents the predicted value of the water quality parameter, represents the measured value of the water quality parameter, represents the optimal hyperparameter, L represents the Laplacian matrix, which is used to quantify the structural information between nodes in the graph and to constrain the spatial smoothness of the predicted values of adjacent nodes, represents the matrix composed of the predicted values of the water quality parameters of each of the nodes, represents the sum of all elements on the diagonal of the matrix, represents the Laplacian regularization term.

[0063] Combined with the foregoing K-fold cross-validation framework, each time one fold is used as the validation set, and the other K - 1 folds are used as the training set. Through repeated training and validation, K loss evaluations are obtained. Finally, the hyperparameters and model parameters are selected when the average value of L is the smallest, so as to determine the optimal model. For example, the node numbers are from 1 to 10. For the first time, nodes 1 - 9 are used as the training set, and node 10 is used as the validation set; for the second time, nodes 2 - 10 are used as the training set, and node 1 is used as the validation set. Repeat this 10 times in total. Each node will be used as the validation set once in turn. After 10 repeated validations, 10 validation values and NSE will be obtained. The average of these 10 results is used as the overall performance index of the current model and hyperparameter configuration. By comparing the average validation losses under different hyperparameters, the hyperparameter configuration that makes the validation value the smallest is selected, so as to determine the optimal model.

[0064] Specifically, the multi-objective loss function can be used to optimize the prediction accuracy of the model and the spatial gradient consistency of the constrained hydrological topology. The mean squared error can quantify the deviation between the model prediction value and the measured value, ensuring the fitting ability of the model to historical water quality data (such as pH, dissolved oxygen, etc.). Based on the static topological adjacency matrix of the basin, a graph Laplacian matrix is constructed. By constraining the spatial gradient consistency of the prediction value matrix, the prediction results of adjacent nodes (i.e., monitoring stations with direct hydrological connection) are forced to satisfy the continuity law of pollutant diffusion. Cross-validation is performed on the hyperparameters to balance the weights of data fitting and physical constraints, enhancing the generalization ability of the model in complex hydrological scenarios. This loss function significantly improves the prediction reliability of data-scarce points by integrating data-driven and mechanism modeling, and provides a theoretical guarantee for the generation of spatio-temporal confidence interval warning signals.

[0065] As Figure 9 shown, the method for predicting and warning the water quality of data-scarce points based on a liquid graph neural network further includes: inputting the initial adjacency matrix, degree matrix, and feature matrix of the target site into the liquid graph neural network model constructed by the method for predicting and warning the water quality of data-scarce points based on a liquid graph neural network as described in the foregoing embodiment, and outputting the predicted value of the water quality parameter of the target site; specifically, inputting the initial adjacency matrix, degree matrix, and feature matrix of the target site into the constructed liquid graph neural network model. Since the target site is a data-scarce site and its water quality data is scarce, historical water quality auxiliary pseudo-node data can be generated through an adversarial network, and then the corresponding feature matrix can be generated. Through the trained liquid graph neural network model, the predicted value of the water quality parameter can be determined.

[0066] Analyze the liquid graph neural network model, the initial adjacency matrix, the degree matrix, the feature matrix of each node, and the predicted value of the water quality parameter of the target site to determine the contribution degree of the surrounding monitoring stations and real-time hydrological data to the predicted value of the water quality parameter of the target site, and issue a warning based on the predicted value of the water quality parameter of the target site; quantitatively explain the pollution key nodes and driving factors based on the graph structure interpretation tool; specifically, input the relevant model and the code of the predicted value of the water quality parameter into the liquid graph neural network model constructed by the liquid neural network, and then interpret through the interpretation tool. For example, for the graph structure interpretation tool GNNExplainer, calculate the contribution degree of each node and edge to the prediction, and inversely locate the key upstream nodes of pollution source tracing. Set the target site to be explained as three data-scarce sites A, B, and C, respectively obtain the importance scores of the edges and nodes, analyze which monitoring stations (1-10) have the highest contribution degree to the water quality prediction of the data-sparse points, and give the contribution degree of the important edges and nodes, such as Figure 10As shown, by combining environmental variables such as land use types (urban, forest, industrial, etc.), upstream input fluxes, river channel lengths, and total nitrogen amounts, it is possible to further explain why certain features have a high contribution to the predicted values of A, B, and C.

[0067] Figure 10 The explanation of the surrounding monitoring stations for the target station is given in [reference]. On the left side are the surrounding monitoring stations from 1 to 10, and on the right side are the target stations A, B, and C. The thicker the connection line, the greater the contribution of the monitoring station to the water quality prediction of the target station; the thinner the connection line, the relatively smaller the contribution. It can be seen that when the hydrological distance is closer and the terrain position is directly or indirectly upstream, the connection line is often thicker, indicating that the nearby and upstream stations in the basin have a higher prediction contribution to the data-sparse points. Furthermore, when the predicted value of the water quality parameter at the target station exceeds the warning value, a warning can be issued for the monitoring stations that exceed the preset contribution value (corresponding to the preset line thickness) to improve the water quality of the corresponding target station downstream; the warning value is set according to which water quality category (from class I to class V) the target station belongs to in the Surface Water Environment Quality Standard, and warning values are set for different water quality indicators. For example, if this target station is a class III water body (mainly applicable to the secondary protection area of the centralized drinking water surface water source, the overwintering ground of fish and shrimp, the migration channel, the aquaculture area and other fishery waters, and the swimming area), then the permanganate index of the target station is less than or equal to 6 mg / L, and the permanganate index warning value of the current target station is set to 6 mg / L.

[0068] As Figure 8 shown, the present invention also provides a construction system 200 for a liquid graph neural network model, including: A construction module 210, configured to determine the initial connection weights between each pair of adjacent nodes according to the river channel data, historical hydrological data, and hydrological connectivity data of the nodes, and construct the initial adjacency matrix, degree matrix, and feature matrix of each node, where each node includes a target station and the surrounding monitoring stations, and each pair of adjacent nodes are two nodes with a downstream relationship.

[0069] A feature encoding module 220, configured to determine the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node, and generate the comprehensive input of each node to a preset liquid differential equation according to the feature matrix, the initial adjacency matrix of each node, and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node.

[0070] A perception update module 230, configured to construct a topological channel update network according to the hidden states of the nodes at the previous moment and the degree matrix, construct a hydrological channel update network according to the real-time hydrological data, update the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introduce a mask matrix through a joint attention mechanism to obtain a dynamic adjacency matrix of each node, where the mask matrix is used to mask the countercurrent relationship and future moment data in the initial adjacency matrix, and the degree matrix is updated based on the update of the initial adjacency matrix.

[0071] An optimization module 240, configured to input the hidden states of the nodes at the previous moment and the dynamic adjacency matrix into an initial neural network model, output predicted values of water quality parameters of each node, and optimize and verify parameters of the initial neural network model to obtain a liquid graph neural network model.

[0072] Specifically, the specific implementation manner of this embodiment may refer to the corresponding construction method of the liquid graph neural network model described above, and will not be elaborated here.

[0073] The above are only specific implementation manners of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A monitoring and early warning method for predicting rare point water quality based on a liquid graph neural network, characterized in that, The liquid graph neural network model is constructed based on the following method: According to the river channel data, historical hydrological data, and hydrological connectivity data of the nodes, determine the initial connection weights between each pair of adjacent nodes, and construct the initial adjacency matrix, degree matrix, and feature matrix of each node. Among them, each node includes a target site and surrounding monitoring sites, and each pair of adjacent nodes is two nodes with a downstream relationship; According to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node, determine the dynamic connection weights between each pair of adjacent nodes. According to the feature matrix, the initial adjacency matrix of each node, and the dynamic connection weights between each pair of adjacent nodes, generate the comprehensive input of each node to the preset liquid differential equation to output the hidden state of each node; Construct a topological channel update network according to the hidden state of each node at the previous moment and the degree matrix, construct a hydrological channel update network according to the real-time hydrological data, update the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introduce a mask matrix through a joint attention mechanism to obtain the dynamic adjacency matrix of each node. Among them, the mask matrix is used to mask the reverse flow relationship and future time data in the initial adjacency matrix, and the degree matrix is updated based on the update of the initial adjacency matrix; Input the hidden state of each node at the previous moment and the dynamic adjacency matrix into the initial neural network model, output the predicted water quality parameter values of each node, and optimize and verify the parameters of the initial neural network model to obtain the liquid graph neural network model.

2. The monitoring and prediction warning method for rare point water quality based on the liquid graph neural network according to claim 1, characterized in that The determining the initial connection weights between each pair of adjacent nodes according to the river channel data, historical hydrological data, and hydrological connectivity data of the nodes, and constructing the initial adjacency matrix, degree matrix, and feature matrix of each node includes: Determine the initial connection weights between each pair of adjacent nodes according to the river channel distance, river flow velocity, and riverbed permeability coefficient between each pair of adjacent nodes; Construct the initial adjacency matrix of each node according to the river channel distance between each pair of adjacent nodes, the average node distance of each node, and the distance attenuation coefficient, and construct the degree matrix of each node according to the initial adjacency matrix of each node; Construct the feature matrix of each node according to the historical hydrological data and the hydrological connectivity data of each node; Among them, the historical hydrological data includes historical meteorological time series data, historical water quality data, and vegetation cover index. The river channel data includes the river channel distance, the river flow velocity, the riverbed permeability coefficient, the average node distance, and the distance attenuation coefficient. The hydrological connectivity data includes the node spatial topology relationship, and the historical water quality data of the target site is generated by an adversarial network.

3. A method for predicting and warning the water quality of scarce points based on a liquid graph neural network according to claim 1, characterized in that, Determining the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node includes: Determining a dynamic correction factor between each pair of adjacent nodes according to the real-time hydrological data of each node and a learning parameter, where the dynamic correction factor is used to characterize the driving relationship of the linear combination of the flow velocity gradient and the rainfall change on the initial connection weights; Determining the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights and the dynamic correction factor.

4. The monitoring and prediction warning method for rare point water quality based on a liquid graph neural network according to claim 3, characterized in that, Generating a comprehensive input of each node for a preset liquid differential equation according to the feature matrix of each node, the initial adjacency matrix, and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node includes: Determining the eigenvectors of the feature matrix of each node, and generating a comprehensive input of each node for a preset liquid differential equation according to the eigenvectors and the dynamic connection weights, so as to output the hidden state of each node; Among them, the preset liquid differential equation is: ； Represents a node at time the hidden state of represents a non-linear activation function represents a learnable bias term represents a learnable parameter matrix represents a node at time the feature vector of represents a dynamic connection weight represents a feature concatenation operation represents a learnable bias term represents a time delay represents the comprehensive input, j represents the node adjacent to node the said adjacent node represents the node adjacent to node the set of the said adjacent nodes 5. A method for predicting and warning of rare - point water quality based on a liquid graph neural network according to claim 3 or 4, characterized in that, Constructing a topological channel update network according to the hidden state of each node at the previous moment and the degree matrix, constructing a hydrological channel update network according to the real-time hydrological data, updating the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introducing a mask matrix through a joint attention mechanism to obtain a dynamic adjacency matrix includes: Based on the hidden state of each node at the previous moment, determining an attention score, constructing a spatial mask matrix and a temporal mask matrix, and fusing the spatial mask matrix and the temporal mask matrix and applying them to the attention score to obtain a joint attention; Capturing the hidden state of each node at the previous moment through a graph convolutional network, and combining with the degree matrix to construct the topological channel update network, and capturing the spatio-temporal convolution of the real-time flow velocity field matrix and the turbulence feature matrix through the graph convolutional network to construct the hydrological channel update network, where the real-time flow velocity field matrix and the turbulence feature matrix are determined according to the real-time hydrological data; Merging and activating the topological channel update network and the hydrological channel update network, and fusing the joint attention based on the initial adjacency matrix to obtain the dynamic adjacency matrix of each node; Among them, the mask matrix includes the spatial mask matrix and the temporal mask matrix, the spatial mask matrix is used to mask the reverse flow relationship in the initial adjacency matrix, and the temporal mask matrix is used to mask the data at the future moment in the initial adjacency matrix.

6. The monitoring and prediction and early warning method for scarce point water quality based on a liquid graph neural network according to claim 5, characterized in that The dynamic adjacency matrix is represented by the formula: ； Among them, represents the dynamic adjacency matrix, represents the initial adjacency matrix, represents the activation function; represents the hydrological channel update network, and ; represents the topological channel update network, and ; represents the said graph convolutional network, represents the said real-time flow velocity field matrix, represents the said turbulence feature matrix, represents the said hidden state at the previous moment, represents the said degree matrix; ； respectively represent a query matrix, a key matrix, and a value matrix, and the three are respectively obtained by linear transformation of the hidden state at the previous moment, represents the vector dimension of the key matrix, represents the spatial mask matrix, represents the temporal mask matrix, represents the normalized exponential function.

7. A method for predicting and warning of rare-point water quality based on a liquid graph neural network according to claim 5, characterized in that Inputting the hidden state of each node at the previous moment and the dynamic adjacency matrix into an initial neural network model, outputting the predicted values of the water quality parameters of each node, and optimizing and verifying the parameters of the initial neural network model to obtain a liquid graph neural network model includes: Based on the predicted water quality parameter values of each of the nodes, a loss function is constructed, and based on the loss function and the K-fold cross-validation mechanism, the optimal hyperparameters of the initial neural network model are determined to obtain the liquid graph neural network model.

8. The monitoring and prediction warning method for scarce point water quality based on a liquid graph neural network according to claim 7, characterized in that The loss function is: ； Among them, represents the total loss function, and N represents the node the set of the nodes adjacent to the current moment, represents the predicted value of the water quality parameter, represents the measured value of the water quality parameter, represents the optimal hyperparameter, and L represents the graph Laplacian matrix, represents the matrix composed of the predicted values of the water quality parameters of each of the nodes, represents the sum of all elements on the diagonal of the matrix, represents the graph Laplacian regularization term.

9. A method for predicting and warning the water quality of rare and scarce points based on a liquid graph neural network according to any one of claims 1-4, characterized in that It further includes: Input the initial adjacency matrix, degree matrix, and feature matrix of the target site into the liquid graph neural network model, and output the predicted water quality parameter values of the target site; Analyze the liquid graph neural network model, the initial adjacency matrix of each node, the degree matrix, the feature matrix, and the predicted water quality parameter values of the target site, determine the contribution degrees of the surrounding monitoring sites and real-time hydrological data to the predicted water quality parameter values of the target site, and issue an early warning based on the predicted water quality parameter values of the target site.

10. A construction system for a liquid graph neural network model, characterized in that, Applying a monitoring and prediction warning method for sparse-point water quality based on a liquid graph neural network as described in any one of claims 1-9, includes: A construction module, configured to determine the initial connection weights between each pair of adjacent nodes according to the river channel data, historical hydrological data, and hydrological connectivity data of the nodes, and construct the initial adjacency matrix, degree matrix, and feature matrix of each node, where each node includes a target site and surrounding monitoring sites, and each pair of adjacent nodes are two nodes with a downstream relationship; A feature encoding module, configured to determine the dynamic connection weights between each pair of adjacent nodes according to the initial connection weights between each pair of adjacent nodes and the real-time hydrological data of each node, and generate the comprehensive input of each node to a preset liquid differential equation according to the feature matrix, the initial adjacency matrix of each node, and the dynamic connection weights between each pair of adjacent nodes, so as to output the hidden state of each node; A perception update module, configured to construct a topological channel update network according to the hidden state of each node at the previous moment and the degree matrix, construct a hydrological channel update network according to the real-time hydrological data, update the initial adjacency matrix according to the topological channel update network and the hydrological channel update network, and introduce a mask matrix through a joint attention mechanism to obtain the dynamic adjacency matrix of each node, where the mask matrix is used to mask the upstream relationship and future moment data in the initial adjacency matrix, and the degree matrix is updated based on the update of the initial adjacency matrix; An optimization module, configured to input the hidden state of each node at the previous moment and the dynamic adjacency matrix into the initial neural network model, output the predicted water quality parameter values of each node, and perform parameter optimization and verification on the initial neural network model to obtain the liquid graph neural network model.

Images