River network greenhouse gas diffusion tracking method and system based on graph neural network

Abstract

The invention discloses a river network greenhouse gas diffusion tracking method and system based on a graph neural network. According to the method, historical greenhouse gas change data and hydrological monitoring data of a target river network are obtained, a greenhouse gas diffusion map neural network is constructed, and a map attention mechanism is utilized to aggregate spatio-temporal characteristics to generate a diffusion path. Hydrological data are analyzed in combination with the LSTM network, the flow velocity gradient is calculated, and the vortex generation probability is predicted. And when the probability exceeds a threshold value, establishing a gas retention effect evaluation matrix, and adjusting a diffusion weight coefficient to optimize a diffusion path. According to the method, the precision of greenhouse gas diffusion tracking is improved, and a scientific basis is provided for emission evaluation and environmental governance.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring and tracking technology, and in particular to a method and system for tracking greenhouse gas diffusion in a river network based on a graph neural network. Background Art

[0002] Greenhouse gas emissions (such as carbon dioxide and methane) have a significant impact on global climate change, and rivers and their tributaries are a key source of these gases. Due to the fluidity of water bodies and the complexity of environmental factors, the diffusion of greenhouse gases within river networks is influenced by multiple factors, including flow velocity, eddy currents, water temperature fluctuations, and river morphology. This makes accurately tracking their diffusion paths a significant challenge.

[0003] Currently, research on greenhouse gas diffusion in water bodies primarily relies on physical models and statistical analysis methods, such as numerical simulations based on fluid dynamics or empirical formula calculations. However, these methods suffer from high computational costs, difficulty meeting real-time requirements, and strong parameter dependence when dealing with complex flows in large-scale river networks, making it difficult to effectively track the diffusion of greenhouse gases.

[0004] In recent years, with the advancement of deep learning technology, spatiotemporal data modeling methods based on graph neural networks (GNNs) have demonstrated superior performance in areas such as traffic flow forecasting and water pollution source tracing. GNNs can leverage the topological structure of river networks to learn complex diffusion patterns through feature correlations between nodes, improving the accuracy of data-driven modeling. Furthermore, when combined with time series analysis methods such as long short-term memory (LSTM) networks, they can effectively capture the impact of dynamic changes in water flow on greenhouse gas diffusion, thereby enhancing the ability to predict diffusion paths.

[0005] To address the shortcomings of existing methods for tracking greenhouse gas diffusion in river networks, this paper proposes a method and system for tracking greenhouse gas diffusion based on graph neural networks. This method combines the spatial relationship modeling capabilities of graph neural networks, the feature aggregation capabilities of attention mechanisms, and the temporal analysis capabilities of LSTM networks to achieve accurate prediction and real-time correction of greenhouse gas diffusion paths, providing a scientific basis for greenhouse gas emission assessment and water environment management. Summary of the Invention

[0006] In order to solve at least one of the above technical problems, the present invention proposes a method and system for tracking greenhouse gas diffusion in river networks based on graph neural networks.

[0007] A first aspect of the present invention provides a method for tracking greenhouse gas diffusion in a river network based on a graph neural network, comprising:

[0008] Obtain historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network;

[0009] The spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network are aggregated through a graph attention mechanism to generate a real-time greenhouse gas diffusion path;

[0010] Based on the LSTM network, the real-time hydrological monitoring data of the target river network is analyzed in time series, the flow velocity gradient characteristics are extracted and input into the eddy generation probability calculation model to calculate the eddy generation probability of each river segment;

[0011] When the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix;

[0012] The real-time greenhouse gas diffusion path is corrected according to the greenhouse gas diffusion map neural network after adjusting the diffusion weight coefficient.

[0013] In this solution, the historical greenhouse gas change data and historical hydrological monitoring data of the target river network are obtained to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network, specifically:

[0014] By deploying distributed monitoring stations at river boundary points, the system periodically collects greenhouse gas concentration change data from each monitoring station in the target river network over a historical period, and simultaneously obtains hydrological monitoring data for the corresponding period, including water flow velocity, water temperature, river cross-sectional width, and water depth;

[0015] The topological structure data of the target river network is extracted based on the geographic information system. Each hydrological monitoring station is used as a graph node, and the river flow connection relationship is used as a directed edge to establish the river network graph structure.

[0016] The historical greenhouse gas concentration change data and hydrological monitoring data were temporally and spatially aligned to construct a graph node feature matrix containing dissolved gas concentration, temperature, flow velocity and cross-sectional morphological parameters;

[0017] Directed edge characteristics are defined based on the hydraulic parameters of the river segment, including the river channel length, average water depth, historical diffusion rate, and riverbed roughness coefficient between adjacent nodes;

[0018] Based on the graph convolutional network architecture, the graph node features are spatially associated with the directed edge features, and the diffusion weight parameters under different hydrological characteristics between nodes of the graph neural network are trained through supervised learning.

[0019] The sliding time window mechanism is used to perform multi-time step iterative optimization of the network, retaining the network parameters that meet the prediction error threshold, and forming a greenhouse gas diffusion graph neural network under different hydrological characteristics.

[0020] In this solution, the spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network are aggregated through the graph attention mechanism to generate a real-time greenhouse gas diffusion path, specifically:

[0021] Obtain current greenhouse gas concentration change data and real-time hydrological monitoring data from each monitoring station in the target river network, and generate real-time input vectors consistent with the characteristic dimensions of the graph neural network nodes;

[0022] Input the real-time input vector into the greenhouse gas diffusion graph neural network, and calculate the attention weight coefficients between adjacent nodes at the current moment through the multi-head attention layer in the graph attention mechanism. The attention weight coefficients are used to characterize the gas diffusion influence of different upstream nodes on downstream nodes.

[0023] Based on the attention weight coefficient, the spatiotemporal features of adjacent nodes are weightedly aggregated, where the spatial features include the current greenhouse gas concentration and cross-sectional morphological parameters. The temporal features are extracted by using a temporal convolutional network to extract the concentration change trend and fluctuation frequency of each node within a preset time window. The aggregated feature vector is input into a gated recurrent unit to generate a spatiotemporal fusion feature for each node.

[0024] A gas diffusion state transfer matrix is constructed based on the spatiotemporal fusion features, wherein each element in the matrix represents the probability of gas diffusion between adjacent nodes per unit time. A Gaussian mixture model is used to perform probability distribution modeling on the transfer matrix to generate a greenhouse gas diffusion characteristic probability distribution model including diffusion direction and diffusion rate.

[0025] The real-time greenhouse gas diffusion path in the target river network is determined based on the greenhouse gas diffusion characteristic probability distribution model.

[0026] In this solution, the real-time hydrological monitoring data of the target river network is analyzed in time series based on the LSTM network. The velocity gradient characteristics are extracted and input into the eddy current generation probability calculation model to calculate the eddy current generation probability of each river segment. Specifically,

[0027] The real-time hydrological monitoring data in the continuous time window is input into the LSTM network for time series modeling, and the flow velocity gradient change characteristics between adjacent monitoring points are extracted;

[0028] Acquiring historical vortex event data of the target river network, wherein the historical vortex event data includes water velocity variation data, vortex intensity, and vortex generation hydrological conditions;

[0029] Dividing the target river network into multiple finite element units, analyzing the historical vortex event data based on the Navier-Stokes equations, constructing a fluid mechanics equation, determining boundary conditions for each finite element unit, the boundary conditions including inlet flow velocity and outlet flow velocity, and solving the fluid mechanics equation based on the finite element method to determine the boundary condition values for each finite element unit;

[0030] Numerical simulation is performed on the boundary condition values of each finite element based on the Monte Carlo simulation method to determine the mapping relationship between the vortex generation intensity and vortex generation probability and the water velocity change, and a vortex generation probability calculation model is constructed based on the mapping relationship;

[0031] The flow velocity gradient change characteristics are introduced into the vortex generation probability calculation model, and the vortex intensity coefficient of each river segment within a preset time interval is calculated in the fluid mechanics model, and the vortex generation probability is determined based on the vortex intensity coefficient.

[0032] In this solution, when the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and the diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix. Specifically,

[0033] Obtain historical greenhouse gas diffusion data for river segments where the current vortex generation probability is greater than a preset value, extract the gas retention time parameters and diffusion rate attenuation coefficients for these segments during historical vortex events, calculate the gas retention intensity index based on the current vortex intensity coefficient, and construct the gas retention effect assessment matrix for the target river network using the gas retention intensity index.

[0034] When the gas retention intensity index exceeds a preset retention threshold, a three-dimensional vortex field model is constructed based on the cross-sectional morphological parameters and water flow velocity of the river segment to simulate the retention area distribution of the gas under the action of the vortex. The simulation results are spatially matched with the actual retention positions in the historical diffusion data to calculate the retention area matching degree;

[0035] If the retention area matching degree is greater than the matching threshold, the diffusion weight coefficient between the upstream and downstream nodes of the current river segment is extracted, and the weight attenuation factor is generated according to the product of the retention intensity index and the matching degree, and the diffusion weight coefficient between the upstream and downstream nodes is dynamically attenuated;

[0036] When the attenuated diffusion weight coefficient is lower than the minimum diffusion threshold, the diffusion connection path between the nodes is cut off and the attention weight coefficients of the adjacent nodes are recalculated;

[0037] If the retention area matching degree is not greater than the matching threshold, the original diffusion weight coefficient is kept unchanged, and the time series monitoring frequency of the node is increased to update the gas retention effect evaluation matrix.

[0038] In this solution, the greenhouse gas diffusion graph neural network adjusted according to the diffusion weight coefficient corrects the real-time greenhouse gas diffusion path, specifically:

[0039] The adjusted diffusion weight coefficient is obtained and, combined with the flow velocity gradient in the real-time hydrological monitoring data, the gas diffusion rate correction value between each node of the greenhouse gas diffusion graph neural network is calculated. When the deviation between the diffusion rate correction value and the original diffusion path prediction value exceeds a preset deviation threshold, the vortex generation probability and gas retention intensity index of the current river segment are extracted;

[0040] constructing a path correction factor matrix based on the vortex generation probability and the gas retention intensity index, and using a path optimization algorithm to replan the node connection paths in the greenhouse gas diffusion graph neural network to generate a set of candidate paths including the corrected diffusion direction and rate;

[0041] The candidate path set is input into the Monte Carlo simulator to simulate the gas diffusion trajectory of each candidate path within a preset time window. The matching degree between the diffusion coverage area of each path and the actual monitored concentration is obtained. If there is a path with a matching degree greater than a preset matching threshold, the path with the highest matching degree is selected as the corrected real-time diffusion path.

[0042] If the matching degree of all candidate paths does not reach the preset matching threshold, the dynamic backtracking mechanism is activated, and the diffusion weight coefficients and hydrological data of the N time steps before the current moment are re-extracted. The parameters of the graph neural network are recalibrated based on the time sliding window. After generating a new diffusion weight coefficient, the path correction process is executed again until a corrected path that meets the matching degree requirements is obtained. The greenhouse gases in the target river network are tracked according to the corrected real-time greenhouse gas diffusion path.

[0043] A second aspect of the present invention further provides a river network greenhouse gas diffusion tracking system based on a graph neural network. The system includes: a memory and a processor. The memory includes a river network greenhouse gas diffusion tracking method program based on a graph neural network. When the river network greenhouse gas diffusion tracking method program based on a graph neural network is executed by the processor, the following steps are implemented:

[0044] Obtain historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network;

[0045] The spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network are aggregated through a graph attention mechanism to generate a real-time greenhouse gas diffusion path;

[0046] Based on the LSTM network, the real-time hydrological monitoring data of the target river network is analyzed in time series, the flow velocity gradient characteristics are extracted and input into the eddy generation probability calculation model to calculate the eddy generation probability of each river segment;

[0047] When the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix;

[0048] The real-time greenhouse gas diffusion path is corrected according to the greenhouse gas diffusion map neural network after adjusting the diffusion weight coefficient.

[0049] This invention discloses a method and system for tracking greenhouse gas diffusion in river networks based on graph neural networks. This method constructs a greenhouse gas diffusion graph neural network by acquiring historical greenhouse gas change data and hydrological monitoring data from a target river network. It then uses a graph attention mechanism to aggregate spatiotemporal features and generate diffusion paths. This method then uses an LSTM network to analyze hydrological data, calculate flow velocity gradients, and predict the probability of eddy formation. When the probability exceeds a threshold, a gas retention effect evaluation matrix is established, and the diffusion weight coefficient is adjusted to optimize the diffusion path. This method improves the accuracy of greenhouse gas diffusion tracking, providing a scientific basis for emission assessment and environmental governance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of a method for tracking greenhouse gas diffusion in a river network based on a graph neural network is shown in the present invention;

[0051] Figure 2 A flow chart of adjusting the diffusion weight coefficient according to the present invention is shown;

[0052] Figure 3 A flow chart showing the correction of the real-time greenhouse gas diffusion path according to the present invention is shown;

[0053] Figure 4 A block diagram of a river network greenhouse gas diffusion tracking system based on a graph neural network of the present invention is shown. DETAILED DESCRIPTION

[0054] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0056] Figure 1 A flow chart of a method for tracking greenhouse gas diffusion in a river network based on a graph neural network is shown in the present invention.

[0057] like Figure 1 As shown, the first aspect of the present invention provides a method for tracking greenhouse gas diffusion in a river network based on a graph neural network, comprising:

[0058] S102, obtaining historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network;

[0059] S104, aggregating the spatiotemporal features of adjacent nodes of the greenhouse gas diffusion network through a graph attention mechanism to generate a real-time greenhouse gas diffusion path;

[0060] S106, performing time series analysis on the real-time hydrological monitoring data of the target river network based on the LSTM network, extracting flow velocity gradient characteristics and inputting them into the eddy generation probability calculation model to calculate the eddy generation probability of each river segment;

[0061] S108, when the vortex generation probability is greater than a preset value, establishing a gas retention effect evaluation matrix based on the vortex generation probability and historical diffusion data, and adjusting diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network based on the gas retention effect evaluation matrix;

[0062] S110, correcting the real-time greenhouse gas diffusion path according to the greenhouse gas diffusion map neural network after adjusting the diffusion weight coefficient.

[0063] It is important to note that by acquiring historical greenhouse gas (GHG) variation data and hydrological monitoring data from the target river network, a GHG diffusion graph neural network (GGDN) model is constructed under different hydrological characteristics. This model can adaptively adjust diffusion patterns to varying hydrodynamic conditions, improving its adaptability to complex river network environments. A graph attention mechanism is used to aggregate the spatiotemporal features of adjacent nodes in the GHG diffusion network, enabling real-time adjustments to diffusion paths, avoiding error accumulation due to changing hydrodynamic conditions and improving prediction accuracy. An LSTM network is used to perform time series analysis on real-time hydrological monitoring data, extracting velocity gradient features to identify velocity trends in different river sections. This is then used to calculate eddy formation probabilities and identify high-eddy regions, thereby capturing anomalous diffusion behavior. When the eddy formation probability exceeds a preset value, a gas retention effect assessment matrix is constructed using historical diffusion data to identify river sections prone to gas accumulation. The diffusion weight coefficients in the GGDN are then adjusted to accurately reflect gas retention and attenuation characteristics. Finally, the real-time GHG diffusion paths are corrected based on the adjusted diffusion weights, resulting in more accurate tracking results, avoiding path misjudgments caused by eddy effects, and improving the reliability of GHG flow monitoring.

[0064] According to an embodiment of the present invention, the acquisition of historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network is specifically as follows:

[0065] By deploying distributed monitoring stations at river boundary points, the system periodically collects greenhouse gas concentration change data from each monitoring station in the target river network over a historical period, and simultaneously obtains hydrological monitoring data for the corresponding period, including water flow velocity, water temperature, river cross-sectional width, and water depth;

[0066] The topological structure data of the target river network is extracted based on the geographic information system. Each hydrological monitoring station is used as a graph node, and the river flow connection relationship is used as a directed edge to establish the river network graph structure.

[0067] The historical greenhouse gas concentration change data and hydrological monitoring data were temporally and spatially aligned to construct a graph node feature matrix containing dissolved gas concentration, temperature, flow velocity and cross-sectional morphological parameters;

[0068] Directed edge characteristics are defined based on the hydraulic parameters of the river segment, including the river channel length, average water depth, historical diffusion rate, and riverbed roughness coefficient between adjacent nodes;

[0069] Based on the graph convolutional network architecture, the graph node features are spatially associated with the directed edge features, and the diffusion weight parameters under different hydrological characteristics between nodes of the graph neural network are trained through supervised learning.

[0070] The sliding time window mechanism is used to perform multi-time step iterative optimization of the network, retaining the network parameters that meet the prediction error threshold, and forming a greenhouse gas diffusion graph neural network under different hydrological characteristics.

[0071] It is important to note that by constructing a greenhouse gas diffusion graph neural network, accurate diffusion path tracking and dynamic prediction are achieved. First, river network topology data is extracted using a geographic information system. Combined with greenhouse gas concentration and hydrological data periodically collected from distributed monitoring stations, river flow relationships are modeled as a graph structure. This allows the diffusion model to truly reflect the connectivity and hydrodynamic characteristics of the river network, improving its adaptability to complex river environments. Subsequently, historical greenhouse gas concentration change data and hydrological monitoring data are spatiotemporally aligned to construct a graph node feature matrix containing parameters such as dissolved gas concentration, flow velocity, temperature, and cross-sectional morphology. Directed edge features are defined based on river hydraulic parameters, enabling the model to simultaneously capture both spatial patterns and temporal trends in gas diffusion, improving its ability to characterize diffusion behavior under dynamic hydrological conditions. Based on a graph convolutional network architecture, node and edge features are spatially associated, and diffusion weight parameters are trained through supervised learning. This allows the model to adapt to varying hydrodynamic conditions, reducing reliance on empirical parameters while avoiding the high computational complexity of traditional numerical simulations. A sliding time window mechanism is used for multi-time-step iterative optimization to ensure that model parameters can be continuously updated, maintain stable prediction performance under different hydrological conditions, and improve the accuracy and long-term reliability of greenhouse gas diffusion tracking.

[0072] According to an embodiment of the present invention, the graph attention mechanism is used to aggregate the spatiotemporal features of adjacent nodes of the greenhouse gas diffusion network to generate a real-time greenhouse gas diffusion path, specifically:

[0073] Obtain current greenhouse gas concentration change data and real-time hydrological monitoring data from each monitoring station in the target river network, and generate real-time input vectors consistent with the characteristic dimensions of the graph neural network nodes;

[0074] Input the real-time input vector into the greenhouse gas diffusion graph neural network, and calculate the attention weight coefficients between adjacent nodes at the current moment through the multi-head attention layer in the graph attention mechanism. The attention weight coefficients are used to characterize the gas diffusion influence of different upstream nodes on downstream nodes.

[0075] Based on the attention weight coefficient, the spatiotemporal features of adjacent nodes are weightedly aggregated, where the spatial features include the current greenhouse gas concentration and cross-sectional morphological parameters. The temporal features are extracted by using a temporal convolutional network to extract the concentration change trend and fluctuation frequency of each node within a preset time window. The aggregated feature vector is input into a gated recurrent unit to generate a spatiotemporal fusion feature for each node.

[0076] A gas diffusion state transfer matrix is constructed based on the spatiotemporal fusion features, wherein each element in the matrix represents the probability of gas diffusion between adjacent nodes per unit time. A Gaussian mixture model is used to perform probability distribution modeling on the transfer matrix to generate a greenhouse gas diffusion characteristic probability distribution model including diffusion direction and diffusion rate.

[0077] The real-time greenhouse gas diffusion path in the target river network is determined based on the greenhouse gas diffusion characteristic probability distribution model.

[0078] It is important to note that this input vector is fed into a greenhouse gas diffusion graph neural network (GHGDGNN). A multi-head attention layer is used to calculate attention weights between adjacent nodes. The multi-head attention layer is a mechanism that enhances information exchange and simultaneously focuses on multiple diffusion patterns, enabling the model to dynamically adjust the influence of different upstream nodes on downstream nodes, improving the accuracy of modeling diffusion relationships. Based on the calculated attention weights, the spatial and temporal features of adjacent nodes are weightedly aggregated. Spatial features include current greenhouse gas concentrations and river cross-sectional morphological parameters. Temporal features are extracted using a temporal convolutional network (TCN) to capture the concentration trends and fluctuation frequencies of each node within a specific time window, effectively capturing the temporal dynamics of the diffusion process. The aggregated feature vector is then fed into a gated recurrent unit (GRU), an improved recurrent neural network (RNN) architecture that mitigates long-term dependencies and enables the model to effectively utilize historical information, thereby generating more representative spatiotemporal fusion features. A gas diffusion state transition matrix is constructed based on these fused features. The matrix elements represent the probability of gas diffusion per unit time between adjacent nodes, clearly characterizing the diffusion direction and rate. A Gaussian mixture model (GMM) was used to model the probability distribution of the transfer matrix. This method can effectively fit complex diffusion patterns and improve the model's adaptability to different diffusion scenarios. Ultimately, a probability distribution model of greenhouse gas diffusion characteristics, including diffusion direction and diffusion rate, was generated. The attention weight coefficient calculation process involves nonlinearly mapping the river channel length, water depth difference, and real-time flow velocity gradient between nodes to generate a correlation score.

[0079] According to an embodiment of the present invention, the LSTM network is used to perform time series analysis on the real-time hydrological monitoring data of the target river network, extract the flow velocity gradient characteristics and input them into the vortex generation probability calculation model to calculate the vortex generation probability of each river segment, specifically:

[0080] The real-time hydrological monitoring data in the continuous time window is input into the LSTM network for time series modeling, and the flow velocity gradient change characteristics between adjacent monitoring points are extracted;

[0081] Acquiring historical vortex event data of the target river network, wherein the historical vortex event data includes water velocity variation data, vortex intensity, and vortex generation hydrological conditions;

[0082] Dividing the target river network into multiple finite element units, analyzing the historical vortex event data based on the Navier-Stokes equations, constructing a fluid mechanics equation, determining boundary conditions for each finite element unit, the boundary conditions including inlet flow velocity and outlet flow velocity, and solving the fluid mechanics equation based on the finite element method to determine the boundary condition values for each finite element unit;

[0083] Numerical simulation is performed on the boundary condition values of each finite element based on the Monte Carlo simulation method to determine the mapping relationship between the vortex generation intensity and vortex generation probability and the water velocity change, and a vortex generation probability calculation model is constructed based on the mapping relationship;

[0084] The flow velocity gradient change characteristics are introduced into the vortex generation probability calculation model, and the vortex intensity coefficient of each river segment within a preset time interval is calculated in the fluid mechanics model, and the vortex generation probability is determined based on the vortex intensity coefficient.

[0085] It is important to note that a time series analysis method based on a long short-term memory (LSTM) network, combined with a fluid dynamics model, calculates eddy formation probabilities, improving the accuracy of greenhouse gas diffusion path predictions within river networks. First, real-time hydrological monitoring data within a continuous time window is fed into the LSTM network. This network effectively handles long-term dependencies. By memorizing past hydrological data trends, it extracts the characteristics of flow velocity gradients between adjacent monitoring points, thereby identifying patterns of water velocity variation across different river sections. Subsequently, historical eddy event data for the target river network is obtained to establish statistical characteristics of eddy occurrence. To more accurately simulate the eddy formation process, the target river network is divided into multiple finite element cells. The finite element method is a numerical analysis method commonly used for computing continuous media problems, discretizing river networks and improving computational accuracy. A fluid dynamics model is established based on the Navier-Stokes equations, which describe the motion characteristics of fluids and are used to calculate velocity, pressure, and turbulence characteristics of the flow, and further determine the boundary conditions of each finite element. Based on this, the team used Monte Carlo simulation to perform random numerical simulations of boundary conditions. Through a large number of random sampling methods, they calculated the eddy intensity and established a mapping relationship between eddy generation probability and water velocity changes. Finally, based on this mapping relationship, they constructed a model for calculating the eddy generation probability. The velocity gradient features extracted by the LSTM were then input into the model. Within the framework of a fluid dynamics model, the eddy intensity coefficients for each river channel segment were calculated and used to determine the eddy generation probability. This method accurately identified river sections prone to eddy generation, improved the accuracy of greenhouse gas diffusion path predictions, and optimized the diffusion model's adaptability to complex hydrological conditions.

[0086] Figure 2 A flow chart of adjusting the diffusion weight coefficient according to the present invention is shown.

[0087] According to an embodiment of the present invention, when the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and the diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix. Specifically,

[0088] S202: Obtain historical greenhouse gas diffusion data for a river segment where the current vortex generation probability is greater than a preset value, extract the gas retention time parameter and diffusion rate attenuation coefficient of the segment during historical vortex events, calculate a gas retention intensity index based on the current vortex intensity coefficient, and construct the gas retention intensity index into a gas retention effect assessment matrix for the target river network;

[0089] S204: When the gas retention intensity index exceeds a preset retention threshold, a three-dimensional vortex field model is constructed based on the cross-sectional morphological parameters and water flow velocity of the river channel segment to simulate the retention area distribution of gas under the action of the vortex. The simulation results are spatially matched with the actual retention locations in the historical diffusion data to calculate the retention area matching degree;

[0090] S206, if the retention area matching degree is greater than the matching threshold, then extracting the diffusion weight coefficient between the upstream and downstream nodes of the current river segment, generating a weight attenuation factor based on the product of the retention intensity index and the matching degree, and dynamically attenuating the diffusion weight coefficient between the upstream and downstream nodes;

[0091] S208, when the attenuated diffusion weight coefficient is lower than the minimum diffusion threshold, the diffusion connection path between the nodes is cut off, and the attention weight coefficients of the adjacent nodes are recalculated;

[0092] S210: If the retention area matching degree is not greater than the matching threshold, the original diffusion weight coefficient is kept unchanged, and the time series monitoring frequency of the node is increased to update the gas retention effect evaluation matrix.

[0093] It should be noted that the diffusion of greenhouse gases in river networks is affected by multiple factors, including hydrodynamic conditions, topographical features, and the physical properties of gases. In particular, under the influence of eddies, gases may remain in local areas for long periods of time, resulting in a decrease in diffusion rate or even an accumulation effect, making it difficult for traditional uniform diffusion models to accurately characterize the diffusion process. Therefore, by obtaining river segments where the probability of eddy generation is greater than a preset value, and extracting the gas retention time parameter (i.e., the average time the gas stays in the area) and the diffusion rate attenuation coefficient (indicating the degree of decrease in the gas diffusion rate in the area) in historical eddy events for this segment, a gas retention intensity index is calculated based on the current eddy intensity, and a gas retention effect evaluation matrix is constructed based on this index to quantify the gas retention characteristics of different river sections. When the gas retention intensity index exceeds the preset retention threshold, a three-dimensional eddy field model is constructed based on the cross-sectional morphological parameters of the river section (such as river width and water depth) and water velocity to simulate the distribution of gas retention areas under the influence of eddies. The model predicts gas accumulation locations based on eddy flow field characteristics and spatially matches them with actual retention locations in historical diffusion data. The retention area matching degree (i.e., the degree of consistency between the simulated retention area and the historically monitored retention area) is calculated. If the retention area matching degree exceeds a preset matching threshold, the diffusion weight coefficients between upstream and downstream nodes in the current river segment are extracted. A weight decay factor is calculated based on the product of the retention intensity index and the matching degree, which is used to dynamically adjust the weights of the diffusion paths. After the diffusion weight coefficient decays, if it falls below the minimum diffusion threshold (i.e., the threshold at which gas cannot effectively diffuse), the diffusion connection paths between the nodes are disconnected, and the attention weight coefficients of adjacent nodes are recalculated, thereby adjusting the gas diffusion direction and avoiding mispredictions. If the retention area matching degree does not exceed the preset matching threshold, the original diffusion weight coefficients are retained, and the time series monitoring frequency of the node is increased to obtain more detailed gas concentration change data. The gas retention effect evaluation matrix is dynamically updated, improving the diffusion model's adaptability to future data.

[0094] Figure 3 A flow chart of the present invention for correcting the real-time greenhouse gas diffusion path is shown.

[0095] According to an embodiment of the present invention, the greenhouse gas diffusion graph neural network, after adjusting the diffusion weight coefficient, corrects the real-time greenhouse gas diffusion path, specifically:

[0096] S302: Obtaining the adjusted diffusion weight coefficient, combining it with the flow velocity gradient in the real-time hydrological monitoring data, calculating the gas diffusion rate correction value between each node of the greenhouse gas diffusion graph neural network, and extracting the eddy flow generation probability and gas retention intensity index of the current river segment when the deviation between the diffusion rate correction value and the original diffusion path prediction value exceeds a preset deviation threshold;

[0097] S304: constructing a path correction factor matrix based on the vortex generation probability and the gas retention intensity index, and replanning the node connection paths in the greenhouse gas diffusion graph neural network using a path optimization algorithm to generate a set of candidate paths including corrected diffusion directions and rates;

[0098] S306: Input the candidate path set into a Monte Carlo simulator to simulate the gas diffusion trajectory of each candidate path within a preset time window, and obtain the matching degree between the diffusion coverage area of each path and the actual monitored concentration. If there is a path with a matching degree greater than a preset matching threshold, the path with the highest matching degree is selected as the corrected real-time diffusion path;

[0099] S308: If the matching degree of all candidate paths does not reach the preset matching threshold, the dynamic backtracking mechanism is activated, and the diffusion weight coefficients and hydrological data of the N time steps before the current moment are re-extracted. The parameters of the graph neural network are recalibrated based on the time sliding window. After generating a new diffusion weight coefficient, the path correction process is executed again until a corrected path that meets the matching degree requirements is obtained. The greenhouse gases in the target river network are tracked according to the corrected real-time greenhouse gas diffusion path.

[0100] It should be noted that after obtaining the adjusted diffusion weight coefficients, the gas diffusion rate corrections between nodes in the greenhouse gas diffusion graph neural network are calculated using the flow velocity gradients in the real-time hydrological monitoring data. When the deviation between this correction value and the original diffusion path prediction exceeds a preset threshold, it indicates that the current path has a significant error and requires correction. Subsequently, a path correction factor matrix is constructed based on the vortex generation probability and the gas retention intensity index to guide path replanning. Using a path optimization algorithm, the node connection paths in the greenhouse gas diffusion graph neural network are replanned to generate a set of candidate paths with corrected diffusion directions and rates. A Monte Carlo simulator is then used to simulate the diffusion trajectories of each candidate path within a preset time window, and the degree of match between the diffusion coverage area of each path and the actual monitored concentration is calculated. When the matching degree of certain paths exceeds a preset matching threshold, the path with the highest matching degree is selected as the corrected real-time diffusion path. This process significantly improves the accuracy of greenhouse gas diffusion path corrections, ensuring the real-time and accuracy of the gas diffusion process. The path correction factor matrix calculates factors influencing path correction based on eddy current and retention effects, which are then used to modify the diffusion path. The path optimization algorithm uses optimization algorithms, such as A*, to search for the optimal diffusion path and generate a set of candidate paths that conform to actual diffusion patterns. The dynamic backtracking mechanism uses a mechanism that backtracks to the previous time step when the matching degree of a candidate path falls below a preset threshold, recalibrating network parameters to generate a more accurate path.

[0101] Figure 4A block diagram of a river network greenhouse gas diffusion tracking system based on a graph neural network of the present invention is shown.

[0102] A second aspect of the present invention further provides a river network greenhouse gas diffusion tracking system 4 based on a graph neural network. The system includes: a memory 41 and a processor 42. The memory includes a river network greenhouse gas diffusion tracking method program based on a graph neural network. When the graph neural network-based river network greenhouse gas diffusion tracking method program is executed by the processor, the following steps are implemented:

[0103] Obtain historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network;

[0104] The spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network are aggregated through a graph attention mechanism to generate a real-time greenhouse gas diffusion path;

[0105] Based on the LSTM network, the real-time hydrological monitoring data of the target river network is analyzed in time series, the flow velocity gradient characteristics are extracted and input into the eddy generation probability calculation model to calculate the eddy generation probability of each river segment;

[0106] When the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix;

[0107] The real-time greenhouse gas diffusion path is corrected according to the greenhouse gas diffusion map neural network after adjusting the diffusion weight coefficient.

[0108] This invention discloses a method and system for tracking greenhouse gas diffusion in river networks based on graph neural networks. This method constructs a greenhouse gas diffusion graph neural network by acquiring historical greenhouse gas change data and hydrological monitoring data from a target river network. It then uses a graph attention mechanism to aggregate spatiotemporal features and generate diffusion paths. This method then uses an LSTM network to analyze hydrological data, calculate flow velocity gradients, and predict the probability of eddy formation. When the probability exceeds a threshold, a gas retention effect evaluation matrix is established, and the diffusion weight coefficient is adjusted to optimize the diffusion path. This method improves the accuracy of greenhouse gas diffusion tracking, providing a scientific basis for emission assessment and environmental governance.

[0109] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0110] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0111] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0112] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0113] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for tracking greenhouse gas diffusion in river networks based on graph neural networks, characterized in that: The following steps are involved: Obtain historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network; The spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network are aggregated through a graph attention mechanism to generate a real-time greenhouse gas diffusion path; Based on the LSTM network, the real-time hydrological monitoring data of the target river network is analyzed in time series, the flow velocity gradient characteristics are extracted and input into the eddy generation probability calculation model to calculate the eddy generation probability of each river segment; When the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix; The real-time greenhouse gas diffusion path is corrected according to the greenhouse gas diffusion map neural network after adjusting the diffusion weight coefficient.

2. The method for tracking greenhouse gas diffusion in river networks based on graph neural networks according to claim 1, characterized in that: The method of obtaining historical greenhouse gas change data and historical hydrological monitoring data of the target river network and constructing a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network is specifically as follows: By deploying distributed monitoring stations at river boundary points, the system periodically collects greenhouse gas concentration change data from each monitoring station in the target river network over a historical period, and simultaneously obtains hydrological monitoring data for the corresponding period, including water flow velocity, water temperature, river cross-sectional width, and water depth; The topological structure data of the target river network is extracted based on the geographic information system. Each hydrological monitoring station is used as a graph node, and the river flow connection relationship is used as a directed edge to establish the river network graph structure. The historical greenhouse gas concentration change data and hydrological monitoring data were temporally and spatially aligned to construct a graph node feature matrix containing dissolved gas concentration, temperature, flow velocity and cross-sectional morphological parameters; Directed edge characteristics are defined based on the hydraulic parameters of the river segment, including the river channel length, average water depth, historical diffusion rate, and riverbed roughness coefficient between adjacent nodes; Based on the graph convolutional network architecture, the graph node features are spatially associated with the directed edge features, and the diffusion weight parameters under different hydrological characteristics between nodes of the graph neural network are trained through supervised learning. The sliding time window mechanism is used to perform multi-time step iterative optimization of the network, retaining the network parameters that meet the prediction error threshold, and forming a greenhouse gas diffusion graph neural network under different hydrological characteristics.

3. The method for tracking greenhouse gas diffusion in river networks based on graph neural networks according to claim 1, characterized in that: The graph attention mechanism is used to aggregate the spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network to generate a real-time greenhouse gas diffusion path, specifically: Obtain current greenhouse gas concentration change data and real-time hydrological monitoring data from each monitoring station in the target river network, and generate real-time input vectors consistent with the characteristic dimensions of the graph neural network nodes; Input the real-time input vector into the greenhouse gas diffusion graph neural network, and calculate the attention weight coefficients between adjacent nodes at the current moment through the multi-head attention layer in the graph attention mechanism. The attention weight coefficients are used to characterize the gas diffusion influence of different upstream nodes on downstream nodes. Based on the attention weight coefficient, the spatiotemporal features of adjacent nodes are weightedly aggregated, where the spatial features include the current greenhouse gas concentration and cross-sectional morphological parameters. The temporal features are extracted by using a temporal convolutional network to extract the concentration change trend and fluctuation frequency of each node within a preset time window. The aggregated feature vector is input into a gated recurrent unit to generate a spatiotemporal fusion feature for each node. A gas diffusion state transfer matrix is constructed based on the spatiotemporal fusion features, wherein each element in the matrix represents the probability of gas diffusion between adjacent nodes per unit time. A Gaussian mixture model is used to perform probability distribution modeling on the transfer matrix to generate a greenhouse gas diffusion characteristic probability distribution model including diffusion direction and diffusion rate. The real-time greenhouse gas diffusion path in the target river network is determined based on the greenhouse gas diffusion characteristic probability distribution model.

4. The method for tracking greenhouse gas diffusion in river networks based on graph neural networks according to claim 3 is characterized in that: The LSTM network is used to perform time series analysis on the real-time hydrological monitoring data of the target river network, extract the flow velocity gradient characteristics and input them into the eddy current generation probability calculation model to calculate the eddy current generation probability of each river segment. Specifically, The real-time hydrological monitoring data in the continuous time window is input into the LSTM network for time series modeling, and the flow velocity gradient change characteristics between adjacent monitoring points are extracted; Acquiring historical vortex event data of the target river network, wherein the historical vortex event data includes water velocity variation data, vortex intensity, and vortex generation hydrological conditions; Dividing the target river network into multiple finite element units, analyzing the historical vortex event data based on the Navier-Stokes equations, constructing a fluid mechanics equation, determining boundary conditions for each finite element unit, the boundary conditions including inlet flow velocity and outlet flow velocity, and solving the fluid mechanics equation based on the finite element method to determine the boundary condition values for each finite element unit; Numerical simulation is performed on the boundary condition values of each finite element based on the Monte Carlo simulation method to determine the mapping relationship between the vortex generation intensity and vortex generation probability and the water velocity change, and a vortex generation probability calculation model is constructed based on the mapping relationship; The flow velocity gradient change characteristics are introduced into the vortex generation probability calculation model, and the vortex intensity coefficient of each river segment within a preset time interval is calculated in the fluid mechanics model, and the vortex generation probability is determined based on the vortex intensity coefficient.

5. The method for tracking greenhouse gas diffusion in river networks based on graph neural networks according to claim 1, characterized in that: When the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and the diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix, specifically: Obtain historical greenhouse gas diffusion data for river segments where the current vortex generation probability is greater than a preset value, extract the gas retention time parameters and diffusion rate attenuation coefficients for these segments during historical vortex events, calculate the gas retention intensity index based on the current vortex intensity coefficient, and construct the gas retention effect assessment matrix for the target river network using the gas retention intensity index. When the gas retention intensity index exceeds a preset retention threshold, a three-dimensional vortex field model is constructed based on the cross-sectional morphological parameters and water flow velocity of the river segment to simulate the retention area distribution of the gas under the action of the vortex. The simulation results are spatially matched with the actual retention positions in the historical diffusion data to calculate the retention area matching degree; If the retention area matching degree is greater than the matching threshold, the diffusion weight coefficient between the upstream and downstream nodes of the current river segment is extracted, and the weight attenuation factor is generated according to the product of the retention intensity index and the matching degree, and the diffusion weight coefficient between the upstream and downstream nodes is dynamically attenuated; When the attenuated diffusion weight coefficient is lower than the minimum diffusion threshold, the diffusion connection path between the nodes is cut off and the attention weight coefficients of the adjacent nodes are recalculated; If the retention area matching degree is not greater than the matching threshold, the original diffusion weight coefficient is kept unchanged, and the time series monitoring frequency of the node is increased to update the gas retention effect evaluation matrix.

6. The method for tracking greenhouse gas diffusion in river networks based on graph neural networks according to claim 1, characterized in that: The greenhouse gas diffusion graph neural network, which is adjusted based on the diffusion weight coefficient, corrects the real-time greenhouse gas diffusion path, specifically: The adjusted diffusion weight coefficient is obtained and, combined with the flow velocity gradient in the real-time hydrological monitoring data, the gas diffusion rate correction value between each node of the greenhouse gas diffusion graph neural network is calculated. When the deviation between the diffusion rate correction value and the original diffusion path prediction value exceeds a preset deviation threshold, the vortex generation probability and gas retention intensity index of the current river segment are extracted; constructing a path correction factor matrix based on the vortex generation probability and the gas retention intensity index, and using a path optimization algorithm to replan the node connection paths in the greenhouse gas diffusion graph neural network to generate a set of candidate paths including the corrected diffusion direction and rate; The candidate path set is input into the Monte Carlo simulator to simulate the gas diffusion trajectory of each candidate path within a preset time window. The matching degree between the diffusion coverage area of each path and the actual monitored concentration is obtained. If there is a path with a matching degree greater than a preset matching threshold, the path with the highest matching degree is selected as the corrected real-time diffusion path. If the matching degree of all candidate paths does not reach the preset matching threshold, the dynamic backtracking mechanism is activated, and the diffusion weight coefficients and hydrological data of the N time steps before the current moment are re-extracted. The parameters of the graph neural network are recalibrated based on the time sliding window. After generating a new diffusion weight coefficient, the path correction process is executed again until a corrected path that meets the matching degree requirements is obtained. The greenhouse gases in the target river network are tracked according to the corrected real-time greenhouse gas diffusion path.

7. A river network greenhouse gas diffusion tracking system based on graph neural network, characterized by: The river network greenhouse gas diffusion tracking system based on graph neural network includes a storage and a processor. The storage includes a river network greenhouse gas diffusion tracking method program based on graph neural network. When the river network greenhouse gas diffusion tracking method program based on graph neural network is executed by the processor, the following steps are implemented: Obtain historical greenhouse gas change data and historical hydrological monitoring data of the target river network to construct a greenhouse gas diffusion graph neural network under different hydrological characteristics of the target river network; The spatiotemporal features of adjacent nodes in the greenhouse gas diffusion network are aggregated through a graph attention mechanism to generate a real-time greenhouse gas diffusion path; Based on the LSTM network, the real-time hydrological monitoring data of the target river network is analyzed in time series, the flow velocity gradient characteristics are extracted and input into the eddy generation probability calculation model to calculate the eddy generation probability of each river segment; When the vortex generation probability is greater than a preset value, a gas retention effect evaluation matrix is established based on the vortex generation probability and historical diffusion data, and diffusion weight coefficients between nodes in the greenhouse gas diffusion graph neural network are adjusted based on the gas retention effect evaluation matrix; The real-time greenhouse gas diffusion path is corrected according to the greenhouse gas diffusion map neural network after adjusting the diffusion weight coefficient.

Images