A method and equipment for runoff prediction that integrates hydrological spatiotemporal knowledge

Abstract

This invention provides a method and device for runoff prediction that integrates hydrological spatiotemporal knowledge, relating to the field of runoff prediction. The method includes: S1: acquiring hydrological station data and constructing a directed hydrological graph using the data; S2: integrating temporal information into the directed hydrological graph to obtain a spatiotemporal directed graph; S3: correcting the spatiotemporal directed graph for flow lag time to obtain a hydrological spatiotemporal directed graph; S4: constructing a time-delayed directed spatiotemporal graph convolutional network, inputting the hydrological station data and the hydrological spatiotemporal directed graph into the network for spatiotemporal feature learning to obtain runoff prediction results. This method utilizes prior knowledge of river flow in time and space to construct a unique hydrological spatiotemporal directed graph, and considers the correction of the hydrological spatiotemporal directed graph by flow lag time, ensuring that the hydrological spatiotemporal directed graph possesses both temporal and spatial information while avoiding the influence of flow lag time.

Description

Technical Field

[0001] This invention relates to the field of runoff prediction, and more particularly to a runoff prediction method and device that integrates hydrological spatiotemporal knowledge. Background Technology

[0002] River flow forecasting, especially long-term forecasting, is crucial for the optimal management of water resources. Many water resource utilization decisions, such as hydropower generation, flood control, optimized reservoir operation, rational water resource allocation, and environmental protection, rely on river flow forecasts. Because river flow is influenced by various factors, including natural elements and human activities, its variations are highly uncertain. Accurately forecasting river flow is both challenging and invaluable.

[0003] Early research relied on physical process-based models to predict river flow. These models used boundary-condition-constrained equations to simulate river hydrological processes, thereby predicting future river flows. These models required not only hydrological observation data but also extensive geographical data from the watershed, such as topography, soil texture, hydrology, and land use cover. In practice, these models faced several limitations. First, due to spatial heterogeneity across different regions, physical process-based prediction models required substantial geographical data to calibrate the model parameters. Collecting large amounts of accurate geographical data is typically very difficult. Furthermore, physical process-based models involve numerous parameters with wide ranges of values ​​and complex interactions between them, making it difficult to accurately model the relationships between parameters. This resulted in models that were not only less widely used but also exhibited significant prediction bias.

[0004] Benefiting from the superior performance of machine learning in simulating complex relationships, machine learning-based river flow prediction methods are constantly emerging. For example, Kratzert used LSTM to simulate daily flows in 241 watersheds and found that its results outperformed regional-level first-order hydrological models. Yan used past river flow data, weather data, and weather forecasts to train an LSTM model, showing that the LSTM model outperforms SVR in flood prediction. Zhang used four neural networks—GRU, LSTM, MLP, and WNN—to predict the daily flow of combined sewer overflows, verifying the strong capabilities of GRU and LSTM in time series prediction. To simultaneously capture river flow characteristics from different periods, Dong added a dynamic window to LSTM. Dai added an attention mechanism to LSTM to improve the accuracy of hydrological time series prediction. Man's proposed Enhanced LSTM dynamically changes the weights of peak and mean runoff through additional feature extractors and loss functions, effectively capturing abrupt changes in runoff. Although the models model the temporal relationships of river flow observation data, they fail to effectively utilize the spatial relationships between rivers geographically, and the prediction accuracy of the models still has some limitations.

[0005] In reality, different rivers may have geographical spatial relationships, which provide rich spatial information for river flow prediction. To address this, researchers have begun to integrate networks with temporal features into CNNs. For example, Ghimire proposed a hybrid model based on CNN and LSTM to predict the short-term future flow of Brisbane and Teewah Creek. Ghose applied a CNN-LSTM model to predict the runoff of the Barak River. Huang designed a robust loss function based on CNNs to capture the spatiotemporal features between multiple river sequences. However, the data processed by CNNs needs to be organized into images or grid structures, which differs significantly from the spatial layout of rivers and makes it difficult to describe the spatial relationships between them.

[0006] In recent years, researchers have proposed graph neural networks (GNNs), represented by graph convolutional networks (GCNNs), to address the limitations of CNNs. GCNNs are deep learning networks capable of performing convolution operations on graph-structured data. Most current GCNN models originate from spectral domain GCNNs. Similar to traditional convolutional neural networks, spectral domain GCNNs have multiple convolutional layers and input / output feature maps. However, in practical applications, the computation of such spectral domain-based GCNNs is too complex. To reduce computational complexity, Chebyshev polynomial estimation was applied to replace the spectral domain convolution kernel, resulting in ChebNet. To further simplify the computation, GCN retains only the first term of the Chebyshev polynomial compared to ChebNet. This retention not only improves the performance of GCN but also allows graph convolution to generalize to the aggregation of vertex information based on node adjacency, connecting the spectral and vertex domains. Furthermore, there are several structurally improved variants of GCN. For example, GraphSAGE adds neighbor node information for sampling and aggregation on top of GCN. GAT is a graph convolutional network defined on the vertex domain. This network uses an attention mechanism to distinguish and dynamically adjust the importance of each node. GCNII effectively solves the oversmoothing problem of GCN and GAT by adding initial residuals and identity mapping. Nowadays, graph neural networks are widely used in fields where data can be represented as graph structures, such as social analysis, traffic prediction, and node classification.

[0007] However, these data-driven methods neglect the hydrological spatiotemporal knowledge of river flow observation data. For example, river flow is directional and lag-dependent, meaning that the flow of upstream rivers takes time to affect the flow of downstream rivers. Existing river flow prediction models are limited in accuracy due to the lack of support from hydrological spatiotemporal knowledge. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a method and device for runoff prediction that integrates hydrological spatiotemporal knowledge, in order to solve the technical problem that the accuracy of existing river flow prediction models is limited due to the lack of support from hydrological spatiotemporal knowledge.

[0009] This invention provides a runoff prediction method that integrates hydrological spatiotemporal knowledge, comprising:

[0010] S1: Obtain hydrological station data and construct a directed hydrological graph using the hydrological station data;

[0011] S2: Integrate temporal information into the hydrological directed graph to obtain a spatiotemporal directed graph;

[0012] S3: Perform water flow lag time correction on the spatiotemporal directed graph to obtain the hydrological spatiotemporal directed graph;

[0013] S4: Construct a time-delayed directed spatiotemporal graph convolutional network. Input hydrological station data and hydrological spatiotemporal directed graphs into the time-delayed directed spatiotemporal graph convolutional network to learn spatiotemporal features and obtain runoff prediction results.

[0014] Preferably, step S1 specifically includes:

[0015] S11: The acquired hydrological station data includes: latitude and longitude of each hydrological station, river flow direction, and topological relationship;

[0016] S12: Construct an initial directed graph, and represent the location of each hydrological station on the initial directed graph by nodes based on the latitude and longitude of the hydrological stations;

[0017] S13: Construct a directed adjacency matrix based on river flow direction and topological relationships. Connect each node in the initial directed graph using directed edges through the directed adjacency matrix to obtain a hydrological directed graph. The expression for the hydrological directed graph G is G = (V, A).

[0018] Where V represents the set of nodes; A∈R N*N Let A represent a directed adjacency matrix, where N represents the number of nodes, and the rows of A represent downstream hydrological stations and the columns represent upstream hydrological stations.

[0019] Preferably, step S2 specifically includes:

[0020] S21: Extend the directed adjacency matrix A in the directed hydrological graph G=(V,A) with temporal information to obtain the spatiotemporal adjacency matrix.

[0021] Where V represents the set of nodes, T C Where N is the number of time steps, and N represents the number of nodes;

[0022] S22: Spatiotemporal adjacency matrix A * The tth k line t s The submatrix of the column is represented as Among them, t s Indicates the starting time, t k Indicates the target time;

[0023] S23: t s >t k All elements in all submatrices are set to 0, as expressed by:

[0024]

[0025] S24: t s ≤t k Add a self-loop to the nodes of all submatrices, expressed as:

[0026]

[0027] Obtain the spacetime directed graph G1.

[0028] Preferably, step S3 specifically includes:

[0029] S31: Extract a set of hydrological station pairs from the node set V of the spatiotemporal directed graph G1. The hydrological station pairs include: upstream hydrological station u and downstream hydrological station d.

[0030] S32: Obtain the distance L from u to d along the river channel. u→d Given the average flow velocity v of the river, the flow lag time t from u to d is calculated. u→d The calculation formula is:

[0031]

[0032] Among them, T C For time steps, Indicates to Round up to the nearest integer;

[0033] S33: If the spatiotemporal adjacency matrix A * If there exist u and d, and a directed edge connects u and d, then A * submatrix The corresponding value is set to 1, otherwise it is set to 0; where t r Indicates the time, where r is the time number;

[0034] S34: For the spatiotemporal adjacency matrix A * Time correction is performed using the following expression:

[0035]

[0036] Where A represents the directed adjacency matrix, A d,u Let A represent the topological relationship between u and d in the directed adjacency matrix A. d,u =1 indicates that the water flows from d to u. Representation submatrix The topological relationship between u and d;

[0037] S35: Repeat steps S31-S34 until all hydrological station pairs are traversed to obtain the hydrological spatiotemporal directed graph G2.

[0038] Preferred:

[0039] The delayed directed spatiotemporal graph convolutional network consists of an input feature transformation layer, a delayed directed spatiotemporal graph convolutional aggregation layer, and an output feature transformation layer connected in sequence.

[0040] The input feature transformation layer consists of a fully connected layer;

[0041] The time-delayed directed spatiotemporal graph convolutional aggregation layer consists of multiple time-delayed directed spatiotemporal graph convolutional layers;

[0042] The time-delayed directed spatiotemporal graph convolutional layer includes multiple time-delayed directed spatiotemporal graph convolutional modules;

[0043] The time-delayed directed spatiotemporal graph convolution module includes multiple graph convolutional layers and pooling layers connected in sequence;

[0044] The output feature transformation layer consists of a two-layer fully connected layer and an activation function.

[0045] Preferably, step S4 specifically includes:

[0046] S41: Up-dimensionalize the hydrological spatiotemporal directed graph by input feature transformation layer to obtain the up-dimensional hydrological spatiotemporal data;

[0047] S42: By using each delayed directed spatiotemporal graph convolution module in the delayed directed spatiotemporal graph convolution aggregation layer to learn features from the upgraded hydrological spatiotemporal data, the spatiotemporal dependencies in the hydrological data are captured, and feature vectors are obtained.

[0048] S43: The learned feature vectors are dimensionally adjusted by the output feature transformation layer to obtain runoff prediction results.

[0049] A storage medium storing instructions and data for implementing the runoff prediction method that integrates hydrological spatiotemporal knowledge.

[0050] A runoff prediction device integrating hydrological spatiotemporal knowledge includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the runoff prediction method integrating hydrological spatiotemporal knowledge.

[0051] The present invention has the following beneficial effects:

[0052] This study utilizes prior knowledge of river flow in time and space to construct a unique hydrological spatiotemporal directed graph. It also considers the impact of flow lag time on the hydrological spatiotemporal directed graph, ensuring that the graph contains both temporal and spatial information while avoiding the influence of flow lag time. Through the hydrological spatiotemporal directed graph and additional learnable parameters, it adaptively captures the spatial and temporal relationships of hydrological data. Furthermore, by stacking multiple graph convolutional modules in the temporal dimension using a time-delayed directed spatiotemporal graph convolutional network, it adapts to the spatiotemporal heterogeneity of hydrological data in different local time periods, thereby improving the accuracy of runoff prediction. Attached Figure Description

[0053] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0054] Figure 2This is a schematic diagram of the construction of a hydrological spatiotemporal directed graph;

[0055] Figure 3 This is a structural diagram of a time-delayed directed spatiotemporal graph convolutional network;

[0056] Figure 4 This is a structural diagram of the delayed directed spatiotemporal graph convolution module;

[0057] Figure 5 This is a schematic diagram of clipping and aggregation in the pooling layer;

[0058] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0059] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0060] Reference Figure 1 This invention provides a runoff prediction method that integrates hydrological spatiotemporal knowledge, comprising:

[0061] S1: Obtain hydrological station data and construct a directed hydrological graph using the hydrological station data;

[0062] S2: Integrate temporal information into the hydrological directed graph to obtain a spatiotemporal directed graph;

[0063] S3: Perform water flow lag time correction on the spatiotemporal directed graph to obtain the hydrological spatiotemporal directed graph;

[0064] S4: Construct a time-delayed directed spatiotemporal graph convolutional network. Input hydrological station data and hydrological spatiotemporal directed graphs into the time-delayed directed spatiotemporal graph convolutional network to learn spatiotemporal features and obtain runoff prediction results.

[0065] Specifically, the construction process of a hydrological spatiotemporal directed graph is as follows: Figure 2 As shown, to accurately predict river flow, a reasonable adjacency matrix needs to be constructed to effectively simulate the relationships between nodes. The adjacency matrix constructed in this paper can represent not only the spatial relationships between nodes but also their temporal relationships. Furthermore, the construction of this adjacency matrix incorporates spatiotemporal knowledge from the hydrological field, enabling the information aggregation process to effectively simulate the river's flow.

[0066] Furthermore, step S1 specifically includes:

[0067] S11: The acquired hydrological station data includes: latitude and longitude of each hydrological station, river flow direction, and topological relationship;

[0068] S12: Construct an initial directed graph, and represent the location of each hydrological station on the initial directed graph by nodes based on the latitude and longitude of the hydrological stations;

[0069] S13: Construct a directed adjacency matrix based on river flow direction and topological relationships. Connect each node in the initial directed graph using directed edges through the directed adjacency matrix to obtain a hydrological directed graph. The expression for the hydrological directed graph G is G = (V, A).

[0070] Where V represents the set of nodes; A∈R N*N Let A represent a directed adjacency matrix, where N represents the number of nodes, and the rows of A represent downstream hydrological stations and the columns represent upstream hydrological stations.

[0071] Specifically, in order to rationalize the information aggregation process of graph convolution by utilizing the actual river directions, this paper defines the hydrological station network as a directed hydrological graph G = (V, A), such as... Figure 2 The term "directed hydrological graph" is used. Here, V is a finite set of nodes representing all hydrological stations within the study area. The directed adjacency matrix is ​​constructed based on the connectivity of the geographical hydrological station network and the direction of rivers. Therefore, the directed adjacency matrix A expresses the spatial relationships and upstream-downstream relationships between all hydrological stations in the region. Specifically, when water can flow from station i to station j, A... ij Set to 1.

[0072] Furthermore, step S2 specifically involves:

[0073] S21: Extend the directed adjacency matrix A in the directed hydrological graph G=(V,A) with temporal information to obtain the spatiotemporal adjacency matrix.

[0074] Where V represents the set of nodes, T C Where N is the number of time steps, and N represents the number of nodes;

[0075] S22: Spatiotemporal adjacency matrix A * The tth k line t s The submatrix of the column is represented as Among them, t s Indicates the starting time, t k Indicates the target time;

[0076] S23: t s >t k All elements in all submatrices are set to 0, as expressed by:

[0077]

[0078] S24: t s ≤t k Add a self-loop to the nodes of all submatrices, expressed as:

[0079]

[0080] Obtain the spacetime directed graph G1.

[0081] Specifically, to express temporal information, we constructed a spatiotemporal adjacency matrix. Obtain the spacetime directed graph G1=(V,A) * ),like Figure 2 For ease of presentation, the self-loop representation of nodes is omitted from the "spacetime directed graph" diagram.

[0082] When t s =t k When, submatrix This represents the spatial relationships between nodes, i.e., the spatial relationships expressed by the directed hydrological graph G. And when t... s <t k When, submatrix Expression t s All nodes at time t k The spatiotemporal relationships of all nodes at time t. Considering that in the actual geographical environment, the observational information of future rivers has no significant impact on past river flows, therefore, when t... s >t k When, submatrix The spatiotemporal relationships represented are considered meaningless. To reduce the interference of meaningless relationships in the information aggregation process of the model, when t s >t k hour, All elements are set to 0, meaning that this submatrix cannot express any temporal, spatial, or spatiotemporal relationships between nodes.

[0083] Due to the spatiotemporal adjacency matrix A * It is derived from the directed adjacency matrix A, so A * It is also a directed adjacency matrix.

[0084] Furthermore, complex time dependencies exist between river flows at different times within the same site. To capture these time dependencies, when t s ≤t k At that time, the spacetime directed matrix A * The submatrices will have additional self-loops added.

[0085] Furthermore, step S3 specifically includes:

[0086] S31: Extract a set of hydrological station pairs from the node set V of the spatiotemporal directed graph G1. The hydrological station pairs include: upstream hydrological station u and downstream hydrological station d.

[0087] S32: Obtain the distance L from u to d along the river channel. u→d Given the average flow velocity v of the river, the flow lag time t from u to d is calculated.u→d The calculation formula is:

[0088]

[0089] Among them, T C For time steps, Indicates to Round up to the nearest integer;

[0090] S33: If the spatiotemporal adjacency matrix A * If there exist u and d, and a directed edge connects u and d, then A * submatrix The corresponding value is set to 1, otherwise it is set to 0; where t r Indicates the time, where r is the time number;

[0091] S34: For the spatiotemporal adjacency matrix A * Time correction is performed using the following expression:

[0092]

[0093] Where A represents the directed adjacency matrix, A d,u Let A represent the topological relationship between u and d in the directed adjacency matrix A. d,u =1 indicates that the water flows from d to u. Representation submatrix The topological relationship between u and d;

[0094] S35: Repeat steps S31-S34 until all hydrological station pairs are traversed to obtain the hydrological spatiotemporal directed graph G2.

[0095] Specifically, the impact of upstream water flow on downstream stations is not instantaneous; the river often needs a certain amount of time to flow along the channel before reaching the downstream hydrological station. This time is called the water delay time in hydrology. For simplicity, when calculating the water delay time between hydrological stations, only two factors are considered: the distance along the river channel between the stations and the average flow velocity of the river. Assume the distance along the river channel from upstream hydrological station u to downstream hydrological station d is L. u→d The average flow velocity of the river is v. Since the time steps that the spatiotemporal adjacency matrix can represent are finite, the calculated flow lag time may exceed T. C This can lead to problems in subsequent steps. Therefore, the water flow lag time from station u to station d will be chosen to be the smaller of the two.

[0096] The spatiotemporal adjacency matrix is ​​used to express the complete spatiotemporal relationship based on lag time. If there exists an upstream hydrological station u and a downstream hydrological station d in the directed adjacency matrix A, and there is a directed hydrological connection between the two stations, then the spatiotemporal adjacency matrix A... * submatrix The corresponding value will be set to 1, otherwise the corresponding value will be set to 0. Due to the existence of water flow lag time, when constructing the spatiotemporal adjacency matrix A... * It is necessary to correct the adjacency relationships between nodes based on the lag time to obtain the corrected delayed spatiotemporal directed adjacency matrix A. * Obtain the hydrological spatiotemporal directed graph G2=(V,A). * ),like Figure 2 For ease of presentation, the self-loop representation of nodes is omitted from the "Hydrological Spatiotemporal Directed Graph".

[0097] Further:

[0098] The delayed directed spatiotemporal graph convolutional network consists of an input feature transformation layer, a delayed directed spatiotemporal graph convolutional aggregation layer, and an output feature transformation layer connected in sequence.

[0099] The input feature transformation layer consists of a fully connected layer;

[0100] The time-delayed directed spatiotemporal graph convolutional aggregation layer consists of multiple time-delayed directed spatiotemporal graph convolutional layers;

[0101] The time-delayed directed spatiotemporal graph convolutional layer includes multiple time-delayed directed spatiotemporal graph convolutional modules;

[0102] The time-delayed directed spatiotemporal graph convolution module includes multiple graph convolutional layers and pooling layers connected in sequence;

[0103] The output feature transformation layer consists of a two-layer fully connected layer and an activation function.

[0104] Specifically, the structure of a time-delayed directed spatiotemporal graph convolutional network is as follows: Figure 3 As shown;

[0105] The time-delayed spatial-temporal directed graph convolutional layer (DSTDGCL) is obtained by horizontally stacking the time-delayed spatial-temporal directed graph convolutional module (DSTDGCM).

[0106] First, we introduce the delayed directed spatiotemporal graph convolution module. The structure of the delayed directed spatiotemporal graph convolution module is as follows: Figure 4 As shown.

[0107] 1. In order to more effectively capture the spatiotemporal characteristics between stations, the time-delayed spatiotemporal directed matrix A * An additional learnable parameter matrix W will be embedded. A* ∈(T C N)*(T C N) is used to adjust the spatiotemporal relationships between nodes. Specifically, the time-delayed spatiotemporal directed adjacency matrix A takes values ​​of only 0 or 1. * Will with W A The time-delayed spacetime directed parameter matrix is ​​obtained through the Hadamard multiplication.

[0108]

[0109] so, It can not only express whether there is a hydrological connection between different stations, but also dynamically adjust its own relation weights based on actual data during the training process, so as to more accurately reflect the spatiotemporal relationship of hydrological data.

[0110] 2. Use graph convolutional layers to capture the spatiotemporal dependencies of the input data. During graph convolution, because... It is a directed adjacency matrix; the spatiotemporal information in the input data will only traverse the directed adjacency matrix. The graph convolutional model aggregates the edges in the middle, thus simulating the actual process of water flowing from an upstream station to multiple downstream stations. Simultaneously, to transform the feature space of the output, a fully connected layer and a ReLU activation function are added to the graph convolution. Assuming the input data matrix... Where C represents the dimension of the feature. The operation of the graph convolutional layer in DSTDGCM can be represented as:

[0111]

[0112] Among them W f ∈R C*C' , denoted as the learnable weight parameters and bias parameters of the fully connected layer, respectively; h represents the output of the graph convolutional layer; and C' represents the transformed feature dimension.

[0113] 3. Stacking multiple graph convolutional layers to capture richer spatiotemporal dependencies, with the number of graph convolutional layers being K. To effectively aggregate the spatiotemporal dependencies captured by these graph convolutional layers, max pooling is used to fuse the outputs of different graph convolutional operations. If we represent the result of aggregating the outputs of all graph convolutional layers, then this process can be represented as:

[0114]

[0115] Where h (i)This represents the output of the i-th graph convolutional layer.

[0116] 4. Use pooling layers for dimensionality pruning. After the outputs of all graph convolutional layers are input into the pooling layer, the pooling layer first prunes the temporal dimension of each input. This is because the network stacks DSTDGCMs along the temporal dimension during construction, resulting in many time steps being modeled repeatedly. If the output of DSTDGCMs covers the entire time period, the structure formed by stacking DSTDGCMs will output results containing a lot of redundant information, thus affecting the prediction accuracy of the entire model. Therefore, only the intermediate time step of the graph convolutional output (or a smaller time step if none exists) is selected as the output of DSTDGCM. The pruning and aggregation process in the pooling layer is as follows: Figure 5 As shown.

[0117] The time-delayed spatiotemporal directed graph convolutional layer is obtained by horizontally stacking time-delayed spatiotemporal directed graph convolutional modules. Horizontal stacking refers to using multiple time-delayed spatiotemporal directed graph convolutional modules to model the spatiotemporal relationships of the entire input time period.

[0118] 1. The input sequence is divided along the time dimension, resulting in multiple local time-segment input sequences. Each DSTDGCM in the network model takes a specific local input sequence as input, allowing each DSTDGCM to independently model a specific local time-segment. This approach allows each DSTDGCM to avoid data interference from other time-segments, more effectively capturing local spatiotemporal dependencies. Furthermore, using multiple DSTDGCMs to capture the spatiotemporal relationships of different local time-segments is more effective at capturing the spatiotemporal heterogeneity of the input data compared to using a single shared DSTDGCM. This structure, composed of multiple DSTDGCMs capable of extracting the spatiotemporal relationships of the entire input time-segment, is called a Delayed Directed Spatiotemporal Convolutional Layer (DSTDGCL).

[0119] 2. A sliding window is used to divide local time periods, where the size of the sliding window is the spatiotemporal adjacency matrix. The number of time steps T that can be expressed C Assume X∈R T*N*C As input to DSTDGCL, the sliding window will divide X into TT along the time dimension. C +1 time length is T C Local input sequence At the same time, DSTDGCL will build TT C +1 DSTDGCM. Next, the dimension of each local input sequence will be changed by T. C *N*C transforms into (T) CThe N)*C values ​​are then fed into a specific DSTDGCM to capture the spatiotemporal dependencies of local time intervals. Finally, the outputs of all DSTDGCMs are concatenated along the time dimension to obtain the output of the DSTDGCL. The entire process can be represented as:

[0120]

[0121] X" p =transform(X' p p = 1, 2, ..., TT C +1

[0122]

[0123]

[0124] in, This represents the p-th local input sequence. These represent the input and output of the p-th DSTDGCM, respectively. This represents the output of DSTDGCL. C represents the feature dimension of the input, and C' represents the feature dimension of the output.

[0125] Furthermore, step S4 specifically involves:

[0126] S41: Up-dimensionalize the hydrological spatiotemporal directed graph by input feature transformation layer to obtain the up-dimensional hydrological spatiotemporal data;

[0127] S42: By using each delayed directed spatiotemporal graph convolution module in the delayed directed spatiotemporal graph convolution aggregation layer to learn features from the upgraded hydrological spatiotemporal data, the spatiotemporal dependencies in the hydrological data are captured, and feature vectors are obtained.

[0128] S43: The learned feature vectors are dimensionally adjusted by the output feature transformation layer to obtain runoff prediction results.

[0129] Specifically, the entire network model mainly consists of three components: an input feature transform layer, a time-delayed spatial-temporal-directed graph convolutional aggregation layer (DSTDGCAL), and an output feature transform layer. DSTDGCN first uses the input feature transform layer to transform the input sequence features to a higher dimension. Then, the transformed sequence is fed into the time-delayed spatial-temporal-directed graph convolutional aggregation layer to capture the spatiotemporal dependencies and heterogeneity in the hydrological data. Finally, the captured spatiotemporal dependency features are transformed by the output feature transform layer into the predicted river flow results for each station.

[0130] 1. Input Feature Transformation Layer: To enhance the feature information represented by the input data, an input feature transformation layer is needed before inputting the data into the model to transform the data to a higher dimension. The input feature transformation layer consists of a fully connected layer, and its transformation process can be represented as:

[0131] X'=XW+b

[0132] Where, X∈R N*T*1 This represents the input river flow data, X'∈R N*T*C Let W represent the data after dimensional transformation, and C represent the transformed feature dimension. Where W ∈ R 1*C ,b∈R N*T*C These are learnable parameters.

[0133] 2. Delayed-Directed Spatiotemporal Graph Convolutional Aggregation Layer: To enable the entire network model to learn deeper spatiotemporal information, a delayed-directed spatiotemporal graph convolutional aggregation layer, consisting of multiple delayed-directed spatiotemporal graph convolutional layers, is used to process the input data. During processing, the output of the previous convolutional layer is used as the input of the next convolutional layer. Assume H... (1) If H is the output of the first convolutional layer, then the output H of the l-th convolutional layer... (l) It can be represented as:

[0134]

[0135] Among them, H (l) H is the output of the l-th DSTDGCL in the time-delayed directed spatiotemporal graph convolutional aggregation layer. (0) It is the input to the time-delayed directed spatiotemporal graph convolutional aggregation layer.

[0136] 3. Output Feature Transformation Layer: For the output of the delayed directed spatiotemporal graph convolutional aggregation layer, the output feature transformation layer is constructed from two fully connected layers and a ReLU activation function. Assume the input to the output feature transformation layer is... The output feature transformation layer first transforms the dimension of the input from T*N*C' to N*(TC'), then... The input is fed into the output feature transformation layer to obtain the prediction results for all N nodes. This process can be represented as...

[0137]

[0138] in, This represents the predicted values ​​for all monitoring stations at the τ-th time step in the future. and These are learnable parameters. C h This represents the feature dimension of the output of the first fully connected layer. T' represents the number of time steps to be predicted. Then, the data for target site i over the next T' time steps will be concatenated into a vector, as shown below:

[0139]

[0140] in, This represents the final output of the entire model. This represents the predicted river flow at station i at the τ-th time step in the future.

[0141] The training objective of time-delayed directed spatiotemporal graph convolutional networks:

[0142] The traffic prediction results will be obtained at the output layer. To achieve accurate prediction of runoff, the mean squared error is used as the loss function during model training, and its definition is as follows:

[0143]

[0144] Where Q represents the actual value of the river flow. This represents the predicted river flow value from the model.

[0145] The constructed hydrological spatiotemporal directed graph and the flow monitoring data of each time series are input into the network architecture in chronological order for spatiotemporal feature capture and aggregation. Finally, the predicted runoff values ​​for a specific station from time t+1 to t+τ are output. That is, the time-delayed directed spatiotemporal graph convolutional network uses the river flow observations X = {X...} of all N nodes at T time steps. (t-T+1) ,X (t-T+2) …X (t)}∈R N*T*1Given the input, predict the river flow sequence for a specific node i in the hydrological map over the next τ time steps. Where X (t) ∈R N*1 This represents the river flow monitored by all monitoring stations at time t. This represents the river flow at station i at time t+τ.

[0146] A storage medium storing instructions and data for implementing the runoff prediction method that integrates hydrological spatiotemporal knowledge.

[0147] A runoff prediction device integrating hydrological spatiotemporal knowledge includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the runoff prediction method integrating hydrological spatiotemporal knowledge.

[0148] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0149] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0150] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A runoff prediction method integrating hydrological spatiotemporal knowledge, characterized in that, include: S1: Obtain hydrological station data and construct a directed hydrological graph using the hydrological station data; S2: Integrate temporal information into the hydrological directed graph to obtain a spatiotemporal directed graph, specifically: S21: Directed graph of hydrology Directed adjacency matrix in Extend the temporal information to obtain the spatiotemporal adjacency matrix. ; Where V represents the set of nodes, For time steps, Indicates the number of nodes; S22: Spatiotemporal Adjacency Matrix The Middle Line number The submatrix of the column is represented as ;in, Indicates the start time. Indicates the target time; S23: Will > All elements in all submatrices are set to 0, as expressed by: S24: Will ≤ Add a self-loop to the nodes of all submatrices, expressed as: Obtain the spacetime directed graph G1; S3: Perform water flow lag time correction on the spatiotemporal directed graph to obtain the hydrological spatiotemporal directed graph, specifically: S31: Extract a set of hydrological station pairs from the node set V of the spatiotemporal directed graph G1. The hydrological station pairs include: upstream hydrological stations. and downstream hydrological stations ; S32: Acquire arrive Distance along the river and the average flow velocity of the river Calculation obtained arrive Water flow lag time The calculation formula is: in, For time steps, Indicates to Round up to the nearest integer; S33: If the spatiotemporal adjacency matrix exist and ,and and If there is a directed edge connecting them, then... submatrix (r) The corresponding value is set to 1, otherwise it is set to 0; where Indicates the time, where r is the time number; S34: For the spatiotemporal adjacency matrix Time correction is performed using the following expression: in, Represents a directed adjacency matrix. In the directed adjacency matrix A, and Topological relationships, This indicates that the water flows from d to u. Representation submatrix middle and Topological relationships; S35: Repeat steps S31-S34 until all hydrological station pairs are traversed to obtain the hydrological spatiotemporal directed graph G2; S4: Construct a time-delayed directed spatiotemporal graph convolutional network. Input hydrological station data and hydrological spatiotemporal directed graphs into the time-delayed directed spatiotemporal graph convolutional network to learn spatiotemporal features and obtain runoff prediction results.

2. The runoff prediction method integrating hydrological spatiotemporal knowledge according to claim 1, characterized in that, Step S1 is as follows: S11: The acquired hydrological station data includes: latitude and longitude of each hydrological station, river flow direction, and topological relationship; S12: Construct an initial directed graph, and represent the location of each hydrological station on the initial directed graph by nodes based on the latitude and longitude of the hydrological stations; S13: Construct a directed adjacency matrix based on river flow direction and topological relationships. Connect all nodes in the initial directed graph using directed edges through this matrix to obtain a hydrological directed graph. The expression is ; Where V represents the set of nodes; Represents a directed adjacency matrix. Indicates the number of nodes. The rows represent downstream hydrological stations, and the columns represent upstream hydrological stations.

3. The runoff prediction method integrating hydrological spatiotemporal knowledge according to claim 1, characterized in that: The delayed directed spatiotemporal graph convolutional network consists of an input feature transformation layer, a delayed directed spatiotemporal graph convolutional aggregation layer, and an output feature transformation layer connected in sequence. The input feature transformation layer consists of a fully connected layer; The time-delayed directed spatiotemporal graph convolutional aggregation layer consists of multiple time-delayed directed spatiotemporal graph convolutional layers; The time-delayed directed spatiotemporal graph convolutional layer includes multiple time-delayed directed spatiotemporal graph convolutional modules; The time-delayed directed spatiotemporal graph convolution module includes multiple graph convolutional layers and pooling layers connected in sequence; The output feature transformation layer consists of a two-layer fully connected layer and an activation function.

4. The runoff prediction method integrating hydrological spatiotemporal knowledge according to claim 3, characterized in that, Step S4 is as follows: S41: Up-dimensionalize the hydrological spatiotemporal directed graph by input feature transformation layer to obtain the up-dimensional hydrological spatiotemporal data; S42: By using each delayed directed spatiotemporal graph convolution module in the delayed directed spatiotemporal graph convolution aggregation layer to learn features from the upgraded hydrological spatiotemporal data, the spatiotemporal dependencies in the hydrological data are captured, and feature vectors are obtained. S43: The learned feature vectors are dimensionally adjusted by the output feature transformation layer to obtain runoff prediction results.

5. A storage medium, characterized in that: The storage medium stores instructions and data to implement the runoff prediction method that integrates hydrological spatiotemporal knowledge as described in any one of claims 1 to 4.

6. A runoff prediction device integrating hydrological spatiotemporal knowledge, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the runoff prediction method integrating hydrological spatiotemporal knowledge as described in any one of claims 1 to 4.

Images