Drainage basin runoff simulation method and device with graph neural network fused with hydrological priori knowledge

Abstract

The invention relates to the technical field of hydrological simulation and flood forecasting, and particularly discloses a watershed runoff simulation method and device with a graph neural network fused with hydrological priori knowledge, and the method comprises the steps: carrying out the grid division of a target watershed, and building a watershed graph structure in combination with the hydrological priori knowledge; introducing a trainable weight into the graph neural network model, and performing time delay weighting based on a spatial distance to obtain a fusion network model; the meteorological driving data, the rainfall data and the drainage basin attributes serve as input, the convergence process is simulated through the fusion network model, and the predicted downstream runoff volume is obtained; and comparing the predicted downstream runoff volume with the actually measured downstream runoff volume, and optimizing all learnable parameters of the fusion network model. According to the method, the defects that a traditional physical model is large in calculation amount and long in consumed time are overcome, rapid simulation of the runoff process is achieved, and the accuracy and reliability of a data driving model in a runoff simulation task are improved by fusing hydrological priori knowledge and introducing physical constraints.

Description

Technical Field

[0001] This invention relates to the field of hydrological simulation and flood forecasting technology, specifically to a watershed runoff simulation method and apparatus that integrates graph neural networks with prior hydrological knowledge. Background Technology

[0002] Watershed runoff simulation is a core component of water resource management, flood control and disaster reduction, and ecological environmental protection. Current runoff simulation methods are mainly divided into two categories: physical process models and data-driven models. Physical process models, such as SWAT and VIC models, describe various processes of the hydrological cycle based on detailed physical laws (such as water balance and energy balance equations). The advantage of these models is their clear physical meaning, revealing the intrinsic mechanisms of hydrological processes. However, their disadvantages are also significant: firstly, the model structure is complex, requiring a large amount of multi-source input data (meteorological, soil, vegetation, topographic, etc.); secondly, the model has numerous parameters, making calibration complex and time-consuming, placing high demands on users; most importantly, model operation requires enormous computing resources, with a single simulation typically taking hours or even days, making it difficult to meet the needs of real-time forecasting and large-scale applications. With the development of artificial intelligence technology, machine learning models, represented by artificial neural networks (ANN) and long short-term memory networks (LSTM), have been applied to runoff simulation. These models can learn the complex nonlinear relationships between inputs (such as rainfall) and outputs (such as runoff) from historical data. Their advantages are relatively simple model structure and fast running speed. However, their disadvantages are that they usually treat the entire watershed as a lumped "black box," ignoring the spatial structure and physical processes of water flow within the watershed. For example, they cannot explicitly express how runoff from different locations converges downstream, nor can they incorporate prior knowledge about slope, flow direction, etc., already possessed by hydrologists. This results in insufficient generalization ability and physical interpretability of the models, and often low simulation accuracy under extreme weather events.

[0003] Therefore, there is an urgent need for a new runoff simulation method that can take into account both the mechanistic nature of physical process models and the efficiency of data-driven models, while also being able to explicitly simulate the spatial confluence process within a watershed.

[0004] Based on this technical background, this invention studies a method and device for simulating watershed runoff by integrating graph neural networks with prior hydrological knowledge. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a watershed runoff simulation method and apparatus that integrates graph neural networks with prior hydrological knowledge. This method overcomes the drawbacks of traditional physical models, such as large computational load and long processing time, and enables rapid simulation of runoff processes, greatly improving simulation efficiency. By integrating prior hydrological knowledge and introducing physical constraints, the accuracy and reliability of data-driven models in runoff simulation tasks are enhanced.

[0006] To achieve the above objectives, a first aspect of the present invention provides a watershed runoff simulation method that integrates graph neural networks with prior hydrological knowledge, comprising:

[0007] The target watershed is divided into grids, and a watershed map structure is established by combining prior hydrological knowledge.

[0008] Based on the aforementioned watershed graph structure, trainable weights are introduced into the graph neural network model, and time delay weighting based on spatial distance is performed to obtain a fusion network model.

[0009] Using meteorological driving data, precipitation data, and watershed attributes as inputs, the fusion network model is used to simulate the confluence process and obtain the predicted downstream runoff.

[0010] The predicted downstream runoff is compared with the measured downstream runoff, the loss function is calculated, and all learnable parameters of the fusion network model are optimized using the backpropagation algorithm.

[0011] A second aspect of the present invention provides a watershed runoff simulation device that integrates graph neural networks with prior hydrological knowledge, comprising:

[0012] The watershed map structure building module is used to divide the target watershed into grids and, in conjunction with prior hydrological knowledge, to build the watershed map structure.

[0013] The fusion network model construction module is used to introduce trainable weights into the graph neural network model based on the watershed graph structure, and perform time delay weighting based on spatial distance to obtain the fusion network model.

[0014] The simulation module is used to simulate the confluence process using meteorological driving data, precipitation data, and watershed attributes as inputs, and to obtain the predicted downstream runoff.

[0015] The parameter optimization module compares the predicted downstream runoff with the measured downstream runoff, calculates the loss function, and optimizes all learnable parameters of the fusion network model using the backpropagation algorithm.

[0016] The beneficial effects of this invention include:

[0017] (1) The watershed runoff simulation method proposed in this invention, which integrates graph neural network with hydrological prior knowledge, overcomes the shortcomings of traditional physical models, such as large computational load and long time consumption, and realizes rapid simulation of runoff process, greatly improving simulation efficiency: by integrating hydrological prior knowledge and introducing physical constraints, the accuracy and reliability of data-driven models in runoff simulation tasks are improved.

[0018] (2) The watershed runoff simulation method proposed in this invention, which integrates graph neural network with hydrological prior knowledge, changes the "black box" characteristics of traditional machine learning models. It explicitly integrates the watershed spatial structure and hydrological physical laws (such as upstream and downstream relationships and confluence time) into the model, enhances the interpretability and generalization ability of the model, and solves the defects of low computational efficiency of physical process models and lack of physical mechanism constraints in data-driven models in the prior art.

[0019] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0020] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating the watershed runoff simulation method based on graph neural networks that integrates prior hydrological knowledge, as proposed in this invention.

[0022] Figure 2 This is a schematic diagram of the model flow in a specific implementation of the watershed runoff simulation method that integrates graph neural network with hydrological prior knowledge proposed in this invention.

[0023] Figure 3 This diagram illustrates a comparison of efficiency between a traditional model and a specific implementation of the watershed runoff simulation method that integrates graph neural networks with prior hydrological knowledge proposed in this invention.

[0024] Figure 4 This is a schematic diagram comparing model accuracy in a specific implementation of the watershed runoff simulation method that integrates graph neural networks with hydrological prior knowledge proposed in this invention.

[0025] Figure 5 This is a schematic diagram of runoff simulation in the Yangtze River basin, representing a specific implementation of the watershed runoff simulation method that integrates graph neural networks with prior hydrological knowledge proposed in this invention. Detailed Implementation

[0026] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0027] This invention provides a watershed runoff simulation method that integrates graph neural networks with prior hydrological knowledge, such as... Figure 1 As shown, it includes:

[0028] The target watershed is divided into grids, and a watershed map structure is established by combining prior hydrological knowledge.

[0029] Based on the watershed graph structure, trainable weights are introduced into the graph neural network model, and time delay weighting based on spatial distance is performed to obtain a fusion network model.

[0030] Using meteorological driving data, precipitation data, and watershed attributes as inputs, a fusion network model is used to simulate the confluence process and obtain the predicted downstream runoff.

[0031] The predicted downstream runoff is compared with the measured downstream runoff, the loss function is calculated, and all learnable parameters of the fusion network model are optimized through the backpropagation algorithm.

[0032] This invention overcomes the shortcomings of traditional physical models, such as large computational load and long time consumption, and realizes rapid simulation of runoff processes, greatly improving simulation efficiency: by integrating hydrological prior knowledge and introducing physical constraints, the accuracy and reliability of data-driven models in runoff simulation tasks are improved.

[0033] According to the present invention, the target watershed is divided into grids, and a watershed map structure is established by combining prior hydrological knowledge, including:

[0034] The target watershed is divided into multiple gridded computational units, and each computational unit is regarded as a node in the watershed graph;

[0035] Based on prior hydrological knowledge, the connection relationships between each node are determined as edges;

[0036] The input features for each node are set to hydrological and meteorological data at a specific point in time.

[0037] According to the present invention, the method of dividing the target watershed into grids and establishing the watershed map structure by combining prior hydrological knowledge further includes:

[0038] A watershed-based digital elevation model is constructed using type equilaterality, and the direction of water flow in each unit is determined by analyzing terrain data.

[0039] When water flows directly from computing unit i to computing unit j, a directed edge is established from node i to node j, thereby constructing the entire watershed into a directed acyclic graph. The directed acyclic graph includes unidirectional causal connections where the upstream influences the downstream and the downstream does not influence the upstream.

[0040] One-way causal connections are encoded as adjacency matrices.

[0041] According to the present invention, hydrological and meteorological data include the runoff and rainfall of the corresponding calculation unit;

[0042] Static properties of the type include elevation and land.

[0043] According to the present invention, based on the watershed graph structure, trainable weights are introduced into the graph neural network, and time delay weighting based on spatial distance is performed to obtain a fusion network model including:

[0044] In a neural network model, the weight of each edge is set as a learnable parameter. During model training, the network learns and optimizes this parameter to obtain trainable weights, which reflect the relative importance of the upstream computing unit's contribution to the downstream computing unit.

[0045] According to the present invention, based on the watershed graph structure, trainable weights are introduced into the graph neural network, and time delay weighting based on spatial distance is performed to obtain a fusion network model, which further includes:

[0046] Calculate the spatial distance of each calculation unit along the water flow path to the total outlet of the watershed;

[0047] The reciprocal of the spatial distance is taken as the static weighting factor. The farther the calculation unit is from the total outlet, the smaller the reciprocal of its spatial distance. That is, its contribution to the total outlet runoff is delayed in time and has a more gradual impact.

[0048] The trainable weights are multiplied by the inverse of the spatial distance and applied to the weights of message passing to simulate the time difference of runoff converging from different locations to the outlet section, thus obtaining the fusion network model.

[0049] According to this invention, using meteorological driving data, precipitation data, and watershed attributes as inputs, a fusion network model is employed to simulate the confluence process, yielding downstream runoff including:

[0050] In each layer of the fusion network model, each node aggregates information passed from all its upstream neighbor nodes;

[0051] The attention coefficient of upstream nodes to downstream nodes is dynamically calculated, and this coefficient is combined with spatial distance-related weights to determine the intensity of information transmission.

[0052] Historical time-series meteorological driving data, precipitation data, and watershed attributes are input into the fusion network model. Through multi-step message passing, the predicted downstream runoff is output at the node representing the watershed outlet.

[0053] Preferably, the meteorological driving data includes rainfall, runoff, and watershed attribute data, including soil type, land use cover, normalized difference vegetation index, and slope within the watershed.

[0054] The watershed attributes are topography, flow direction, and distance;

[0055] The graph neural network model module can be a graph attention network (GAT), a graph convolutional network (GCN), or GraphSAGE.

[0056] According to the present invention, it further includes:

[0057] For artificial regulation facilities existing within the watershed, treat them as special nodes, or design an independent, rule-based or learnable sub-module to simulate their complex behavior of regulating runoff, and embed this sub-module into the framework of the graph neural network model.

[0058] Artificial regulation facilities include reservoirs and dams.

[0059] This invention changes the "black box" nature of traditional machine learning models by explicitly integrating the spatial structure of the watershed and hydrological physical laws (such as upstream-downstream relationships and confluence time) into the model, thereby enhancing the interpretability and generalization ability of the model. It also solves the defects of low computational efficiency of physical process models and lack of physical mechanism constraints in data-driven models in the prior art.

[0060] This invention also provides a watershed runoff simulation device that integrates graph neural networks with prior hydrological knowledge, comprising:

[0061] The watershed map structure building module is used to divide the target watershed into grids and, in conjunction with prior hydrological knowledge, to build the watershed map structure.

[0062] The fusion network model building module is used to introduce trainable weights into the graph neural network model based on the watershed graph structure, and perform time delay weighting based on spatial distance to obtain the fusion network model.

[0063] The simulation module is used to simulate the confluence process using meteorological driving data, precipitation data, and watershed attributes as inputs, and obtain the predicted downstream runoff.

[0064] The parameter optimization module compares the predicted downstream runoff with the measured downstream runoff, calculates the loss function, and optimizes all learnable parameters of the fusion network model through the backpropagation algorithm.

[0065] The present invention will be described in more detail below through embodiments.

[0066] Example 1:

[0067] This embodiment proposes a watershed runoff simulation method that integrates graph neural networks with prior hydrological knowledge, specifically including the following steps:

[0068] (I) Construction of the watershed map structure:

[0069] The target watershed is divided into several gridded computational units; each computational unit is regarded as a node in a graph; the connection between nodes is the edge, which is determined based on prior hydrological knowledge; the input features of each node are hydrological and meteorological data at a specific time point, mainly including the runoff and rainfall of that unit; in addition, other static attributes such as elevation and land use type are also included. The construction of edges is based on the digital elevation model of the watershed, and the direction of water flow in each unit is determined by analyzing topographic data; a directed edge from node i to node j is established only when the water flow is directly from unit i to unit j; in this way, the entire watershed is constructed as a directed acyclic graph, which naturally contains the unidirectional causal relationship of "upstream affects downstream, and downstream does not affect upstream"; this connection relationship is encoded as an adjacency matrix.

[0070] (II) Graph weighting mechanism that integrates prior knowledge:

[0071] To enable the model to learn a more physically consistent confluence process, this embodiment introduces a two-layer weighting mechanism on the edge weights of the Graph Neural Network (GNN): The first is a trainable weight; in the GNN model, the weight W of each edge is a learnable parameter. During model training, the network automatically learns and optimizes this weight to reflect the relative importance of the upstream unit's contribution to the downstream unit. The second is a time-delay weighting based on spatial distance; to simulate the time difference (i.e., confluence time) of runoff from different locations to the outlet section, this embodiment introduces a prior weight based on spatial distance. Specifically, the spatial distance d from each grid cell along the flow path to the total outlet of the basin is calculated; then, its reciprocal 1 / d is taken as a static weighting factor. The farther the cell is from the outlet, the smaller its 1 / d value, representing a more delayed and gradual impact on the outlet runoff. Finally, the weight applied to message passing is a combination of the trainable weight W and the distance weight 1 / d, for example, through multiplication.

[0072] (III) Simulation of the merging process based on graph neural networks:

[0073] This embodiment uses a graph neural network (GNN) as the core processor to simulate the runoff confluence process; the overall flow of the model is as follows: Figure 2As shown: The model's inputs include three main categories: meteorological driving data (such as rainfall and runoff in this scheme), precipitation data, and watershed attributes (such as topography, flow direction, and distance in this scheme). These inputs are fed into a Graph Neural Network (GNN) module for processing. The model focuses on simulating runoff. In each layer of the GNN, each node j aggregates information (such as runoff) from all its upstream neighbor nodes i. The information transfer process is modulated by the aforementioned weighting mechanism. Taking a Graph Attention Network (GAT) as an example, it dynamically calculates... The attention coefficient of downstream node i to downstream node j, combined with the distance weight, determines the strength of information transmission. Subsequently, the historical time series watershed grid input data (runoff, rainfall, etc.) are input into the model. Through multi-step message passing, the model finally outputs the predicted downstream runoff at the node representing the watershed outlet. The predicted runoff is compared with the measured downstream runoff, the loss function (such as mean squared error) is calculated, and the backpropagation algorithm is used to optimize all learnable parameters of the graph neural network (GNN) model (including the trainable weights of the edges).

[0074] The above method is also applicable to the simulation of large-scale complex watersheds; taking the Yangtze River Basin as an example, the spatial distribution of its runoff in 2023 was simulated.

[0075] (I) Study Area and Data Preparation:

[0076] The study area in this embodiment is the entire Yangtze River Basin, with a latitude and longitude range of approximately 90°-122°E and 24°-35°N. The input data uses a 10km x 10km resolution grid. The input data consists of daily gridded rainfall and runoff data for 2023 covering the entire basin. The topographic data is a DEM of the corresponding resolution.

[0077] (II) Implementation Steps:

[0078] 1. Construction of a super-large graph: The hundreds of thousands of grid cells in the entire Yangtze River Basin are constructed into a huge directed graph, with the number of nodes and edges far exceeding that of ordinary river basins;

[0079] 2. Model Application: The graph neural network (GNN) model structure of this invention is adopted; the model is fully trained using historical data (such as 1980-2021) to enable it to learn the complex runoff generation and confluence patterns of the Yangtze River Basin;

[0080] 3. Spatial distribution prediction: The input data for 2023 is fed into the trained model to make daily predictions; the model of this invention can not only output the runoff at the final outlet (Datong Station), but also output the runoff value of each grid cell in the watershed at each time step (each day);

[0081] (III) Results Presentation and Analysis:

[0082] Appendix Figure 5 This paper presents an example of the spatial distribution of runoff in the Yangtze River Basin on July 16, 2023, as output by the model (representing the simulation results for 2023; in the figure, runoff represents runoff, latitude represents latitude, and longitude represents longitude). The figure clearly shows that the runoff distribution is extremely uneven in space. High runoff areas (dark areas in the figure) clearly delineate the river network morphology of the Yangtze River's main stream and major tributaries (such as the Yalong River, Min River, Jialing River, and Han River). Simultaneously, the runoff value shows a cumulative effect along the river channel from upstream to downstream, consistent with hydrological laws. The runoff value in the vast non-channel areas is very low (light color), also consistent with physical reality. This refined spatial output demonstrates that the method of this invention can not only predict the total flow at the outlet section but also accurately capture and simulate the spatial distribution characteristics of runoff within the basin. This has significant application value for understanding flood formation processes, identifying key runoff-producing areas, and conducting regional flood forecasting.

[0083] Compared with the prior art, the technical solution of the present invention has the following significant advantages and positive effects:

[0084] (I) This invention uses a graph neural network (GNN) model based on efficient matrix operations, avoiding the need to solve complex physical differential equations; compared with traditional physical process models, the computation speed is improved by nearly 100 times while maintaining accuracy, such as... Figure 3 As shown, it can meet the application scenarios with high timeliness requirements, such as real-time flood forecasting;

[0085] (ii) The model has high accuracy in runoff simulation, such as Figure 4 The performance comparison results shown demonstrate that the method proposed in this invention achieves the best performance in terms of the accuracy of runoff simulation at stations in the Yangtze River Basin (corresponding to "GNN" in the figure, i.e., the model with added physical rule constraints) and the key hydrological evaluation index of Nash-Sutcliffe efficiency (NSE). Its median NSE reaches approximately 0.7, which is significantly higher than the LSTM model without added physical constraints of the confluence process and also higher than the Random Forest (RF) model. This proves that this invention effectively improves the simulation accuracy and reliability of the model by integrating prior hydrological knowledge.

[0086] (III) This invention is not a purely “black box” model; by explicitly constructing the watershed as a directed graph and introducing weights based on flow direction and distance, the design of the model itself contains the basic physical laws of hydrology (the causal law from upstream to downstream and the difference in confluence time); this makes the model’s decision-making process more in line with physical intuition and the results more interpretable.

[0087] Example 2:

[0088] This embodiment provides a watershed runoff simulation device that integrates graph neural networks with prior hydrological knowledge, including:

[0089] The watershed map structure building module is used to divide the target watershed into grids and, in conjunction with prior hydrological knowledge, to build the watershed map structure.

[0090] The fusion network model building module is used to introduce trainable weights into the graph neural network model based on the watershed graph structure, and perform time delay weighting based on spatial distance to obtain the fusion network model.

[0091] The simulation module is used to simulate the confluence process using meteorological driving data, precipitation data, and watershed attributes as inputs, and obtain the predicted downstream runoff.

[0092] The parameter optimization module compares the predicted downstream runoff with the measured downstream runoff, calculates the loss function, and optimizes all learnable parameters of the fusion network model through the backpropagation algorithm.

[0093] The artificial regulation module treats artificial regulation facilities existing in the watershed as special nodes, or designs them as independent, rule-based or learnable sub-modules to simulate their complex behavior of regulating runoff, and embeds the sub-module into the framework of the graph neural network model.

[0094] Artificial regulation facilities include reservoirs and dams;

[0095] In this embodiment, the target watershed is divided into grids, and the watershed map structure is established by combining prior hydrological knowledge, including:

[0096] The target watershed is divided into multiple gridded computational units, and each computational unit is regarded as a node in the watershed graph;

[0097] Based on prior hydrological knowledge, the connection relationships between each node are determined as edges;

[0098] The input features for each node are set to hydrological and meteorological data at a specific point in time;

[0099] In this embodiment, the process of dividing the target watershed into grids and establishing the watershed map structure by combining prior hydrological knowledge also includes:

[0100] A watershed-based digital elevation model is constructed using type equilaterality, and the direction of water flow in each unit is determined by analyzing terrain data.

[0101] When water flows directly from computing unit i to computing unit j, a directed edge is established from node i to node j, thereby constructing the entire watershed into a directed acyclic graph. The directed acyclic graph includes unidirectional causal connections where the upstream influences the downstream and the downstream does not influence the upstream.

[0102] Encode unidirectional causal connections into an adjacency matrix;

[0103] In this embodiment, the hydrological and meteorological data includes the flow rate and rainfall of the corresponding calculation unit;

[0104] The static attributes of the type include elevation and land area;

[0105] In this embodiment, based on the watershed graph structure, trainable weights are introduced into the graph neural network, and time delay weighting based on spatial distance is performed to obtain the fusion network model, which includes:

[0106] In a neural network model, the weight of each edge is set as a learnable parameter. During model training, the network learns and optimizes this parameter to obtain trainable weights, which reflect the relative importance of the contribution of upstream computing units to downstream computing units.

[0107] In this embodiment, based on the watershed graph structure, trainable weights are introduced into the graph neural network, and time delay weighting based on spatial distance is performed to obtain the fusion network model, which further includes:

[0108] Calculate the spatial distance of each calculation unit along the water flow path to the total outlet of the watershed;

[0109] The reciprocal of the spatial distance is taken as the static weighting factor. The farther the calculation unit is from the total outlet, the smaller the reciprocal of its spatial distance. That is, its contribution to the total outlet runoff is delayed in time and has a more gradual impact.

[0110] The trainable weights are multiplied by the inverse of the spatial distance and applied to the weights of message passing to simulate the time difference of runoff converging from different locations to the outlet section, and finally the fusion network model is obtained.

[0111] In this embodiment, meteorological driving data, precipitation data, and watershed attributes are used as inputs. A fusion network model is employed to simulate the confluence process, yielding downstream runoff including:

[0112] In each layer of the fusion network model, each node aggregates information passed from all its upstream neighbor nodes;

[0113] The attention coefficient of upstream nodes to downstream nodes is dynamically calculated, and this coefficient is combined with spatial distance-related weights to determine the intensity of information transmission.

[0114] Historical time-series meteorological driving data, precipitation data, and watershed attributes are input into the fusion network model. Through multi-step message passing, the predicted downstream runoff is output at the node representing the watershed outlet.

[0115] Meteorological driving data include rainfall, runoff, and watershed attribute data, including soil type, land use cover, normalized vegetation index, and slope within the watershed.

[0116] The watershed attributes are topography, flow direction, and distance;

[0117] The graph neural network model module can be a graph attention network (GAT), a graph convolutional network (GCN), or GraphSAGE.

[0118] The watershed runoff simulation method proposed in the embodiments of the present invention, which integrates graph neural network with hydrological prior knowledge, overcomes the shortcomings of traditional physical models, such as large computational load and long time consumption, and realizes rapid simulation of runoff processes, greatly improving simulation efficiency: by integrating hydrological prior knowledge and introducing physical constraints, the accuracy and reliability of data-driven models in runoff simulation tasks are improved.

[0119] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A watershed runoff simulation method based on graph neural networks and prior hydrological knowledge, characterized in that, include: The target watershed is divided into grids, and a watershed map structure is established by combining prior hydrological knowledge. Based on the aforementioned watershed graph structure, trainable weights are introduced into the graph neural network model, and time delay weighting based on spatial distance is performed to obtain a fusion network model. Using meteorological driving data, precipitation data, and watershed attributes as inputs, the fusion network model is used to simulate the confluence process and obtain the predicted downstream runoff. The predicted downstream runoff is compared with the measured downstream runoff, the loss function is calculated, and all learnable parameters of the fusion network model are optimized using the backpropagation algorithm.

2. The method according to claim 1, characterized in that, The target watershed is divided into grids, and a watershed map structure is established by combining prior hydrological knowledge, including: The target watershed is divided into multiple gridded computational units, and each computational unit is regarded as a node in the watershed graph; Based on prior hydrological knowledge, the connection relationships between each node are determined as edges; The input features for each node are set to hydrological and meteorological data at a specific point in time.

3. The method according to claim 2, characterized in that, The process of dividing the target watershed into grids and establishing the watershed map structure by incorporating prior hydrological knowledge also includes: A watershed-based digital elevation model is constructed using type equilaterality, and the direction of water flow in each unit is determined by analyzing terrain data. When the water flows directly from computing unit i to computing unit j, a directed edge is established from node i to node j, thereby constructing the entire watershed into a directed acyclic graph. The directed acyclic graph includes unidirectional causal connections where the upstream influences the downstream and the downstream does not influence the upstream. The unidirectional causal connection is encoded as an adjacency matrix.

4. The method according to claim 3, characterized in that, The hydrological and meteorological data include the runoff and rainfall of the corresponding calculation units; The static attributes of this type include elevation and land.

5. The method according to claim 2, characterized in that, Based on the aforementioned watershed graph structure, trainable weights are introduced into the graph neural network, and time-delay weighting based on spatial distance is performed to obtain a fusion network model including: In the neural network model, the weight of each edge is set as a learnable parameter. During the model training process, the network learns and optimizes this parameter to obtain trainable weights, so as to reflect the relative importance of the contribution of the upstream computing unit to the downstream computing unit.

6. The method according to claim 5, characterized in that, Based on the aforementioned watershed graph structure, trainable weights are introduced into the graph neural network, and time delay weighting based on spatial distance is performed to obtain the fusion network model, which further includes: Calculate the spatial distance of each calculation unit along the water flow path to the total outlet of the watershed; The reciprocal of the spatial distance is taken as a static weighting factor. The farther the calculation unit is from the total outlet, the smaller the reciprocal of its spatial distance. That is, its contribution to the total outlet runoff is delayed in time and has a more gradual impact. The trainable weights are multiplied by the reciprocal of the spatial distance and applied to the weights of message passing to simulate the time difference of runoff converging from different locations to the outlet section, thus obtaining the fusion network model.

7. The method according to claim 6, characterized in that, Using meteorological driving data, precipitation data, and watershed attributes as input, the fusion network model is used to simulate the confluence process, yielding downstream runoff including: In each layer of the fusion network model, each node aggregates information passed from all its upstream neighbor nodes; The attention coefficient of the upstream node to the downstream node is dynamically calculated, and this coefficient is combined with the spatial distance-related weight to determine the intensity of information transmission. Historical time-series meteorological driving data, precipitation data, and watershed attributes are input into the fusion network model. Through multi-step message passing, the predicted downstream runoff is output at the node representing the watershed outlet.

8. The method according to claim 1, characterized in that, The meteorological driving data includes rainfall, runoff, and watershed attribute data, including soil type, land use cover, normalized vegetation index, and slope within the watershed. The watershed attributes are topography, flow direction, and distance; The graph neural network model module is a graph attention network, a graph convolutional network, or GraphSAGE.

9. The method according to claim 2, characterized in that, Also includes: For artificial regulation facilities existing within the watershed, treat them as special nodes, or design an independent, rule-based or learnable sub-module to simulate their complex behavior of regulating runoff, and embed the sub-module into the framework of the graph neural network model. The artificial control facilities include reservoirs and dams.

10. A watershed runoff simulation device that integrates graph neural networks with prior hydrological knowledge, characterized in that, include: The watershed map structure building module is used to divide the target watershed into grids and, in conjunction with prior hydrological knowledge, to build the watershed map structure. The fusion network model construction module is used to introduce trainable weights into the graph neural network model based on the watershed graph structure, and perform time delay weighting based on spatial distance to obtain the fusion network model. The simulation module is used to simulate the confluence process using meteorological driving data, precipitation data, and watershed attributes as inputs, and to obtain the predicted downstream runoff. The parameter optimization module compares the predicted downstream runoff with the measured downstream runoff, calculates the loss function, and optimizes all learnable parameters of the fusion network model using the backpropagation algorithm.

Images