A watershed hydrology prediction method for multi-stage reservoir blocking

Abstract

The application belongs to the technical field of hydrological prediction, and specifically discloses a kind of basin hydrology prediction methods for multi-level reservoir blocking, comprising the following steps: the historical flow data of each node is normalized independently;The feature coding is carried out on the future rainfall forecast data and is fused with the normalized flow sequence and time coding feature;The fusion feature is input into the interactive diffusion graph convolution network tree structure, the input sequence is split according to the odd-even parallel processing and the diffusion graph convolution is generated dynamically to extract multi-scale spatio-temporal features;The initial result is obtained by decoding through the gate linear unit, and the future rainfall forecast data is decoded independently by the residual method to obtain the final result;The mean square error weighted sum optimization model of all nodes and core target nodes is adopted. The application can effectively improve the long-term runoff prediction accuracy of multi-level reservoir basin and inhibit the rapid decay of prediction accuracy with the increase of prediction length.

Description

Technical Field

[0001] This application belongs to the field of hydrological prediction technology, and more specifically, relates to a watershed hydrological prediction method for multi-level reservoir blockage. Background Technology

[0002] Hydrological forecasting is a core technology for water resource allocation and flood control and disaster reduction. Its main application is to predict future runoff changes at key sections or reservoirs based on historical hydrological and meteorological observation data from multiple stations within a watershed, thereby providing a scientific basis for reservoir storage and release decisions and flood warnings. With the development of deep learning, using graph neural networks to process the spatial topological relationships of multiple stations in a watershed has become an important technical direction for improving the accuracy of multi-node joint forecasting.

[0003] A prior art LSTM-based rainfall-runoff prediction scheme is disclosed. This scheme uses historical flow and rainfall sequences from a single hydrological station or reservoir, arranged by time step, as input vectors. The retention and updating of historical time step information are controlled through forget gates, input gates, and output gates within the LSTM unit. The hidden state of the last time step is mapped to runoff prediction values ​​for multiple future time steps through a fully connected layer. When facing multi-site prediction, this scheme typically adopts an independent modeling approach of "one model per station," or flattens the data from multiple stations and inputs them into the same LSTM network. However, LSTM is essentially a one-dimensional time series model and cannot explicitly model the spatial topology of the watershed. This leads to the neglect of the physical spatial dependence of rainfall and runoff from upstream tributaries propagating downstream along the river network, making it difficult to capture the "peak evolution" effect between upstream and downstream, thus limiting the accuracy of multi-node joint prediction.

[0004] The prior art also discloses a graph-based LSTM (Graph-LSTM) flood forecasting scheme. This scheme constructs a graph adjacency matrix based on the upstream and downstream connections of real river networks. For each time step, the hidden state of the current node is weighted and aggregated with the hidden states of its upstream neighbor nodes through graph convolution operations. The aggregated node features are then input into a traditional LSTM unit for time-dimensional state updates. However, this scheme has the following drawbacks: First, it strictly follows the serial paradigm of "first spatial graph aggregation, then time update," which disrupts the spatiotemporal coupling characteristic of "spatial rainfall distribution dynamically evolving over time" in hydrological processes, and fails to capture dynamic features that need to span both time and spatial scales simultaneously. Second, it relies on a predefined static river network topology matrix for information transmission, and cannot adaptively adjust the influence weights between nodes according to different hydrological conditions, resulting in distortion of spatial feature extraction under extreme flood conditions. Third, it lacks a mechanism for handling scale differences between nodes. For heterogeneous node networks that include large reservoirs (daily average flow can reach tens of thousands of cubic meters per second) and small and medium-sized hydrological stations (daily average flow is only tens of cubic meters per second), high-flow nodes are diluted by low-flow nodes in joint optimization, causing the accuracy of medium- and long-term predictions to rapidly decline as the prediction duration increases.

[0005] Therefore, how to improve the accuracy of medium- and long-term runoff forecasting in multi-level reservoir basins and suppress the rapid decline in forecast accuracy as the forecast period increases is an urgent problem to be solved. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a watershed hydrological prediction method for multi-stage reservoir blockage, which can effectively improve the accuracy of medium- and long-term runoff prediction in multi-stage reservoir watersheds and suppress the rapid decay of prediction accuracy with increasing prediction time.

[0007] To achieve the above objectives, in a first aspect, this application provides a watershed hydrological prediction method for multi-stage reservoir blockage, comprising the following steps: S10: Obtain historical flow data and future rainfall forecast data for multiple nodes within the basin. Perform independent normalization processing at the node level on the historical flow data of each node to obtain the normalized flow sequence of each node. S20, the future rainfall forecast data is encoded with rainfall features to obtain rainfall embedding features, and the normalized flow sequence, time coding features and rainfall embedding features are fused to obtain a fused feature tensor; S30, the fused feature tensor is input into a pre-constructed interactive diffusion graph convolutional network tree structure for spatiotemporal coupling modeling. The interactive diffusion graph convolutional network tree structure splits and processes the input sequence layer by layer according to the parity of time steps, and extracts multi-scale spatiotemporal features within each branch through dynamic graph generation and diffusion graph convolution to obtain a multi-scale spatiotemporal feature tensor. S40, the multi-scale spatiotemporal feature tensor is decoded through a gated linear unit to obtain a preliminary runoff prediction result. At the same time, the future rainfall forecast data is converted into a runoff correction amount through an independent decoder, and the runoff correction amount is injected into the preliminary runoff prediction result in a residual manner to obtain the final runoff prediction result. S50, during the model training phase, a weighted loss function is used to optimize the model parameters. The weighted loss function is the weighted sum of the mean squared errors of all nodes and the mean squared errors of the core target node.

[0008] As a further preferred embodiment, in step S10, the node-level independent normalization processing includes: for the first... i The node of the first j The value of each feature Using formula Normalization is performed, where For nodes i Features j In the time dimension T Independent means on, For nodes i Features j In the time dimension T The independent standard deviations of the time series were calculated; and the time series was divided into different stages according to the flood season and the dry season, and the mean and standard deviation of each node were calculated and normalized in each stage.

[0009] As a further preferred embodiment, in step S20, the rainfall feature encoding includes a spatial attention mechanism and a temporal attention mechanism, both of which employ a multi-head attention calculation formula, which is as follows: Where Q is the query matrix, K is the key matrix, and V is the value matrix. d k Let be the dimension of the key matrix; For the spatial attention mechanism, Q, K, and V are derived from the input features in the node dimension. The linear transformation is obtained; for the aforementioned temporal attention mechanism, Q, K, and V are derived from the input features in the time dimension. It is obtained by linear transformation.

[0010] As a further preferred embodiment, in step S20, the feature fusion specifically involves: concatenating the normalized flow sequence, the time-encoded feature, and the rainfall embedding feature in the channel dimension to obtain a concatenated feature; and then performing feature interaction and dimensionality reduction on the concatenated feature through two-dimensional convolution to obtain the fused feature tensor.

[0011] As a further preferred embodiment, in step S30, the dynamic graph generation specifically involves: extracting the dynamic relationships between nodes through adaptive learning to generate a dynamic adjacency matrix. The dynamic adjacency matrix automatically adjusts the influence weights between nodes based on the current hydrological state, and its calculation formula is as follows:

[0012]

[0013] in, X To fuse feature tensors; d The hidden dimension (i.e., number of channels) of the node features; W 1 represents the node projection weight matrix, used to map node features to the node association space; b 1 represents the first bias vector; The diffusion graph convolution specifically involves: based on the dynamic adjacency matrix... To conduct spatial information dissemination and obtain spatially aggregated features , ,in For a symmetric normalized Laplace matrix, For degree matrix, W 2 represents the spatial feature transformation weight matrix; b 2 is the second bias vector.

[0014] As a further preferred embodiment, in step S30, each interactive diffusion map convolutional unit further includes temporal convolution and residual connection: for the spatially aggregated features Causal convolution is applied to extract temporal dependencies, and residual connections are added to obtain... .

[0015] As a further preferred embodiment, in step S30, the layer-by-layer splitting and parallel processing according to the parity of time steps includes: the first layer splits the input sequence into even-time-step subsequences and odd-time-step subsequences, and inputs them into the first interactive diffusion graph convolutional unit to obtain the first output and the second output respectively; the second layer further splits the first output and the second output according to parity and inputs them into the second interactive diffusion graph convolutional unit; the third layer repeats the operation of the second layer; finally, the outputs of all branches of the tree structure are re-concatenated in the time dimension to obtain the multi-scale spatiotemporal feature tensor.

[0016] As a further preferred embodiment, in step S40, the future rainfall forecast data is converted into runoff correction data through an independent decoder and injected as residuals, specifically using the following formula: ,in For the final runoff prediction results, These are preliminary runoff prediction results. For future rainfall forecast data, α This is a learnable scaling factor.

[0017] As a further preferred embodiment, in step S50, the weighted loss function is specifically: Where λ is the weight coefficient, N is the total number of nodes, Q is the prediction step size, i is the node index, and q is the time step index. For the core target node (Danjiangkou) in time step Predicted runoff values For the core target node (Danjiangkou) in time step The actual runoff value For nodes At time step Predicted runoff values For nodes At time step The actual runoff value, The dynamic weighting function is used when the flow rate at the core target node (Danjiangkou) exceeds a set threshold. Or when traffic is on the rise, It increases exponentially, and its calculation formula is: ,in These are the weighting coefficients. This is the traffic threshold.

[0018] Secondly, this application provides a watershed hydrological prediction system for multi-stage reservoir blockage, used to implement the method described in any one of the above, comprising: The data acquisition and normalization module is used to acquire historical flow data and future rainfall forecast data of multiple nodes in the basin, and to perform independent node-level normalization processing on the historical flow data of each node to obtain the normalized flow sequence of each node. The rainfall encoding and fusion module is used to encode rainfall features into the future rainfall forecast data to obtain rainfall embedding features, and to fuse the normalized flow sequence, time encoding features and rainfall embedding features to obtain a fused feature tensor. The spatiotemporal coupling modeling module has a built-in interactive diffusion graph convolutional network tree structure, which is used to perform spatiotemporal coupling modeling on the fused feature tensor. It splits and processes the input sequence layer by layer according to the parity of time steps, and extracts multi-scale spatiotemporal features within each branch through dynamic graph generation and diffusion graph convolution to obtain a multi-scale spatiotemporal feature tensor. The bypass decoding module is used to decode the multi-scale spatiotemporal feature tensor through a gated linear unit to obtain a preliminary runoff prediction result. At the same time, the future rainfall forecast data is converted into a runoff correction amount through an independent decoder and injected into the preliminary runoff prediction result in a residual manner to obtain the final runoff prediction result. The training optimization module is used to optimize the model parameters using a weighted loss function, which is a weighted sum of the mean squared errors of all nodes and the mean squared errors of the core target node.

[0019] The beneficial effects of this application are as follows: (1) Node-independent and phased seasonal normalization design for heterogeneous node differences and temporal nonstationarity. Given the extremely large flow differences among nodes in large multi-level reservoir basins (e.g., the daily average flow of backbone reservoirs can reach tens of thousands of cubic meters per second, while some small and medium-sized hydrological stations only reach tens of cubic meters per second), and the significant "alternating flood and dry seasons" nonstationary characteristics of runoff influenced by monsoon climate, this application abandons the traditional global unified normalization method and innovatively adopts a "node-independent + time-phased seasonal" normalization strategy. This design not only avoids the problem of large flow nodes being "diluted" by small flow nodes due to global unified scaling, but also eliminates the statistical distribution conflict between low variance during the dry season and high variance during the flood season through phased processing, preserving the complete original numerical distribution characteristics for the model to accurately identify flood processes in the future. This is the cornerstone for solving the problem of inaccurate large flow prediction.

[0020] (2) IDGCN Tree-based Multi-scale Spatiotemporal Coupling Design Based on Odd-Even Splitting. To address the problem of information loss caused by the forced decoupling of spatiotemporal features in traditional spatiotemporal graph models (such as STGCN with graph convolution followed by temporal convolution), this application designs an Interactive Diffusion Graph Convolution (IDGCN) tree structure. This structure cleverly splits and processes the input time series layer by layer according to odd and even positions (similar to multi-resolution analysis), and embeds dynamic graph generation and diffusion graph convolution within each branch. This design enables the model to capture both rapid local changes (such as sudden changes in intraday flood peaks) and slow global trends (such as baseflow during the dry season across months) without significantly increasing the number of parameters, achieving true spatiotemporal coupling feature extraction.

[0021] (3) Design of Graph Adjacency Matrix Construction Integrating Spatial Location and Dynamic Hydrological Features. Addressing the shortcomings of existing graph models that rely on predefined static river network topology and cannot adapt to dynamic hydrological conditions, this application introduces a multi-source feature fusion strategy in the graph construction stage. This design overcomes the limitations of constructing adjacency matrices solely based on physical river network upstream-downstream connections or simple geographical distances. It deeply integrates the actual geographic spatial relationships of nodes with dynamic hydrological features such as the current flow status and water level difference of the nodes, generating a dynamic graph adjacency matrix through adaptive learning. This enables the model to adjust the information transmission weights between nodes in real time and adaptively according to different operating conditions, such as baseflow control during the dry season and specific rainstorm trends during the flood season, significantly improving the spatial feature expression capability under extreme conditions.

[0022] (4) A physically meaningful future rainfall bypass and a hydrological-meteorological coupling mechanism. Addressing the industry challenge of exponentially decreasing accuracy in medium- to long-term (7-15 days) forecasts due to purely extrapolating historical flow sequences, this application avoids forcibly incorporating future rainfall forecasts into the underlying spatiotemporal coding network (to prevent information leakage and training oscillations). Instead, it establishes a direct physical bypass at the final decoding output layer. This mechanism transforms future rainfall forecasts into runoff corrections via an independent decoder and directly superimposes them onto the preliminary forecast results as residuals. The introduction of this mechanism transforms the forecast model from traditional purely hydrological statistical extrapolation into a hydrological-meteorological coupling forecast model, resulting in a highly clear physical meaning (i.e., future rainfall directly corresponds to runoff increments) and effectively breaking through the information bottleneck of medium- to long-term forecasts, significantly suppressing accuracy decay.

[0023] (5) Design of differentiated loss functions for key objectives related to trend and flood peak tendency. In the joint optimization of multiple nodes in a watershed, the conventional mean square error (MSE) loss function tends to cause the model to fit nodes with large numbers but small variance (dry season), resulting in a severe "peak reduction" phenomenon during flood peak periods. This application designs a multi-dimensional differentiated weighting strategy. In addition to assigning high weights to the basic loss of core control reservoir nodes, a penalty mechanism specifically for flow change trends and flood peak tendency is introduced. During model training, an exponentially increasing loss weight is applied to the prediction error during flood peak periods (high flow intervals), forcing the model to pay more attention to the flood peak evolution process. This effectively solves the common problem of conventional models "fitting well at low flow rates but underestimating flood peaks," ensuring the reliability of the model's forecasts during the most critical period of flood control scheduling. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method provided in the embodiments of this application; Figure 2This is a comparison chart of the actual and predicted values ​​for 15 days provided in the embodiments of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] The key aspects of this application are: designing a node-level normalization method to preserve the dynamic range of heterogeneous sites; constructing an interactive diffusion graph convolutional network (IDGCN) tree structure to achieve multi-scale extraction of spatiotemporally coupled features; and introducing a future rainfall bypass mechanism and a weighted loss function for key targets. Through these technical measures, this application achieves the following results: under lightweight conditions with approximately 1.5 million model parameters, a 15-day average NSE of 0.72 and a first-day NSE of 0.83 are achieved, effectively meeting the medium- and long-term runoff forecasting needs of watershed flood control scheduling.

[0027] like Figure 1 As shown, the hydrological prediction method for the Hanjiang River basin proposed in this embodiment includes the following steps: (1) Data acquisition and node-level normalization: Historical flow data (P=32 days) of 35 nodes in the Hanjiang River Basin and multi-node rainfall forecast data (P+Q=47 days) were acquired, and node-level Z-score normalization was performed on the flow data.

[0028] To address the issue of flow rate differences of hundreds of times between large reservoirs and small hydrological stations within a river basin, this embodiment abandons global normalization and adopts node-independent normalization. For the first... The node of the first The normalized formula for each feature is:

[0029] in, For nodes feature In the time dimension Independent means on, It is its independent standard deviation. This algorithm ensures that high-flow nodes such as Danjiangkou maintain their complete dynamic range after normalization.

[0030] In this embodiment, the period from June to October of each year is specifically divided into the flood season and the period from November to May of the following year is divided into the dry season. The mean and standard deviation of each node in each stage are calculated and substituted into the above formula to eliminate the statistical distribution conflict between the low variance of the dry season and the high variance of the flood season.

[0031] (2) Rainfall feature encoding: The rainfall matrix of multiple nodes is encoded by a rainfall encoder. ( The process involves processing the 15-day forecast period. First, node-independent linear mapping is performed. get( The characteristics of rainfall are then extracted using spatial and temporal attention mechanisms, with the multi-head attention calculation formula being:

[0032] In spatial attention, Depend on At the node dimension The linear transformation yields the result; in time attention, Because of its time dimension The Q, K, and V matrices are obtained by mapping the input features through three independent linear layers along the node dimension. , , ,in Characteristics of rainfall This is a learnable projection weight matrix. The final output is the rainfall embedding. .

[0033] (3) Multi-source feature fusion: The normalized flow features, time-coded features and rainfall embedding features obtained in step (2) are concatenated in the channel dimension, and feature interaction and dimensionality reduction are performed through two-dimensional convolution (Conv2D) to obtain the fused feature tensor.

[0034] (4) Spatiotemporal modeling of IDGCN tree structure: The fused features are input into a multi-layer IDGCN tree structure to achieve spatiotemporally coupled multi-scale extraction. Assume the input features are... The execution process of each IDGCN unit is as follows: (4.1) Dynamic graph generation. Dynamic relationships between nodes are extracted through adaptive learning, calculated as follows:

[0035]

[0036] in It is a dynamic adjacency matrix that can automatically adjust the influence weights between nodes according to the current hydrological status.

[0037] (4.2) Diffusion map convolution. Based on To disseminate spatial information:

[0038] in For a symmetric normalized Laplace matrix, It is a degree matrix.

[0039] (4.3) Temporal Convolution and Residual Connections. Causal convolution is applied to the spatially aggregated features to extract temporal dependencies, and residual connections are added to prevent gradient vanishing.

[0040] (4.4) Tree-based multi-scale extraction. The individual IDGCN units described above are encapsulated into a tree structure, and the input sequence is split and processed in parallel according to the parity of time steps: First layer: Input Split into even-numbered time steps and odd time steps Input IDGCN-1 respectively to get and ; Second and third layers: Continue with... and Even-odd splitting is performed, processed using IDGCN-2 and IDGCN-3 respectively; Final output: The outputs of all branches of the tree structure are reassembled along the time dimension, i.e. .

[0041] (5) Future rainfall bypass decoding: The output of step (4) is output as a preliminary runoff prediction result through a gated linear unit (GLU) decoder. Simultaneously extract rainfall forecasts for the next Q steps. After passing through an independent decoder (two-layer perceptron), the residual is directly injected into the preliminary prediction result. The calculation formula for bypass correction is:

[0042] in This is a learnable scaling factor. The mechanism directly utilizes future weather forecast information to physically correct results derived solely from historical sequence extrapolation.

[0043] (6) Danjiangkou Weighted Loss Optimization: During the model training phase, in order to ensure the prediction accuracy of the core target node (Danjiangkou, node index 0) in multi-node joint optimization, the loss function calculation formula designed in this embodiment is as follows:

[0044] in, The dynamic weighting function is used when the flow rate at the core target node (Danjiangkou) exceeds a set threshold. Or when traffic is on the rise, It increases exponentially, and its calculation formula is: ,in These are the weighting coefficients. This represents the flow threshold. The first term is the mean square error of all nodes globally (N=35), and the second term is the mean square error of the Danjiangkou node. This is the weighting factor (default value is 5.0). During backpropagation, the algorithm forces the model to fit the flow variation pattern of the Danjiangkou Reservoir with a gradient strength of 5 times.

[0045] The beneficial effects of this application are as follows: (1) Node-independent and phased seasonal normalization design for heterogeneous node differences and temporal nonstationarity. Given the extremely large flow differences among nodes in large multi-level reservoir basins (e.g., the daily average flow of backbone reservoirs can reach tens of thousands of cubic meters per second, while some small and medium-sized hydrological stations only reach tens of cubic meters per second), and the significant "alternating flood and dry seasons" nonstationary characteristics of runoff influenced by monsoon climate, this application abandons the traditional global unified normalization method and innovatively adopts a "node-independent + time-phased seasonal" normalization strategy. This design not only avoids the problem of large flow nodes being "diluted" by small flow nodes due to global unified scaling, but also eliminates the statistical distribution conflict between low variance during the dry season and high variance during the flood season through phased processing, preserving the complete original numerical distribution characteristics for the model to accurately identify flood processes in the future. This is the cornerstone for solving the problem of inaccurate large flow prediction.

[0046] (2) IDGCN Tree-based Multi-scale Spatiotemporal Coupling Design Based on Odd-Even Splitting. To address the problem of information loss caused by the forced decoupling of spatiotemporal features in traditional spatiotemporal graph models (such as STGCN with graph convolution followed by temporal convolution), this application designs an Interactive Diffusion Graph Convolution (IDGCN) tree structure. This structure cleverly splits and processes the input time series layer by layer according to odd and even positions (similar to multi-resolution analysis), and embeds dynamic graph generation and diffusion graph convolution within each branch. This design enables the model to capture both rapid local changes (such as sudden changes in intraday flood peaks) and slow global trends (such as baseflow during the dry season across months) without significantly increasing the number of parameters, achieving true spatiotemporal coupling feature extraction.

[0047] (3) Design of Graph Adjacency Matrix Construction Integrating Spatial Location and Dynamic Hydrological Features. Addressing the shortcomings of existing graph models that rely on predefined static river network topology and cannot adapt to dynamic hydrological conditions, this application introduces a multi-source feature fusion strategy in the graph construction stage. This design overcomes the limitations of constructing adjacency matrices solely based on physical river network upstream-downstream connections or simple geographical distances. It deeply integrates the actual geographic spatial relationships of nodes with dynamic hydrological features such as the current flow status and water level difference of the nodes, generating a dynamic graph adjacency matrix through adaptive learning. This enables the model to adjust the information transmission weights between nodes in real time and adaptively according to different operating conditions, such as baseflow control during the dry season and specific rainstorm trends during the flood season, significantly improving the spatial feature expression capability under extreme conditions.

[0048] (4) A physically meaningful future rainfall bypass and a hydrological-meteorological coupling mechanism. Addressing the industry challenge of exponentially decreasing accuracy in medium- to long-term (7-15 days) forecasts due to purely extrapolating historical flow sequences, this application avoids forcibly incorporating future rainfall forecasts into the underlying spatiotemporal coding network (to prevent information leakage and training oscillations). Instead, it establishes a direct physical bypass at the final decoding output layer. This mechanism transforms future rainfall forecasts into runoff corrections via an independent decoder and directly superimposes them onto the preliminary forecast results as residuals. The introduction of this mechanism transforms the forecast model from traditional purely hydrological statistical extrapolation into a hydrological-meteorological coupling forecast model, resulting in a highly clear physical meaning (i.e., future rainfall directly corresponds to runoff increments) and effectively breaking through the information bottleneck of medium- to long-term forecasts, significantly suppressing accuracy decay.

[0049] (5) Design of differentiated loss functions for key objectives related to trend and flood peak tendency. In the joint optimization of multiple nodes in a watershed, the conventional mean square error (MSE) loss function tends to cause the model to fit nodes with large numbers but small variance (dry season), resulting in a severe "peak reduction" phenomenon during flood peak periods. This application designs a multi-dimensional differentiated weighting strategy. In addition to assigning high weights to the basic loss of core control reservoir nodes, a penalty mechanism specifically for flow change trends and flood peak tendency is introduced. During model training, an exponentially increasing loss weight is applied to the prediction error during flood peak periods (high flow intervals), forcing the model to pay more attention to the flood peak evolution process. This effectively solves the common problem of conventional models "fitting well at low flow rates but underestimating flood peaks," ensuring the reliability of the model's forecasts during the most critical period of flood control scheduling.

[0050] Compared with existing technologies, the technical solution proposed in this application has achieved significant benefits in terms of improved prediction accuracy, suppression of long-term degradation, targeted optimization of core nodes, and optimization of computing resources. The following detailed explanation is based on specific experimental data: (1) Experimental conditions and methods Dataset: Daily-scale data from 2020 to 2026 for 35 nodes (including 13 reservoirs and 22 hydrological stations) in the Hanjiang River Basin.

[0051] Input features: 32-day historical flow data (inflow and outflow), and 47-day (32-day historical data + 15-day future) multi-node rainfall forecast data.

[0052] Forecast target: Daily inflow over the next 15 days.

[0053] Model parameters: 64 hidden channels, Danjiangkou node loss weight set to 5.0, main branch learning rate 0.0005, rainfall branch learning rate 0.00005.

[0054] Evaluation metrics: Nash efficiency coefficient (NSE), mean absolute error (MAE).

[0055] (2) The prediction accuracy is significantly improved, effectively overcoming the node "dilution" effect. In multi-node joint prediction scenarios, existing global normalization methods often lead to a decrease in prediction accuracy for high-flow nodes. This application achieves excellent prediction results on the Danjiangkou Reservoir (the core node) through a "node-level normalization" and "key target weighting" mechanism. Figure 2 As shown, the predicted curve closely matches the measured curve.

[0056] Overall accuracy: The 15-day average NSE reached 0.7181 (achieving a "good" rating), and the mean absolute error (MAE) was only 650.60 m³ / s.

[0057] Extremely high short-term accuracy: The predicted NSE on day 1 is as high as 0.8285, which can capture the current traffic status with great accuracy and provide a high-confidence basis for daily scheduling.

[0058] (3) It has made a breakthrough in suppressing the decline in the accuracy of medium- and long-term forecasts. Existing deep learning methods often experience a sharp drop in NSE (Network Sequence Size) when the prediction period is extended from 1 day to 15 days (e.g., the NSE often drops below 0.5 on the 7th day). This application benefits from the "future rainfall bypass mechanism," which directly injects the physical increment of weather forecasts into the output layer, greatly mitigating the error accumulation caused by pure time series extrapolation. See the table below: Table 1. Gradual Forecast of NSE Indicators for Danjiangkou Node over the Next 15 Days

[0059] The medium-term forecast is robust: the NSE was 0.8005 on day 5 and remained at a high level of 0.7640 on day 7, with minimal fluctuations, indicating that the model has a very stable grasp of the runoff trend in the coming week.

[0060] Long-term forecasts are available: Even on the 15th day, when there is a lack of strong historical correlation, the NSE still reaches 0.5953, breaking through the bottleneck of traditional methods being "unreliable" in long-term forecasts and truly meeting the operational needs of medium- and long-term flood control scheduling.

[0061] (4) The model is highly interpretable and conforms to the laws of hydrophysics. Unlike existing purely data-driven "black box" models, the prediction results of this application have clear physical traceability. By using the attention weights of the rainfall encoder, the key upstream rainfall areas that caused a flood peak can be clearly identified; by using the residual output of the future rainfall bypass, the contribution of future weather forecasts to the correction of runoff results can be directly quantified. This design, which integrates the physical logic of "rainfall-runoff-confluence" into the network structure, greatly enhances the confidence of hydrological operators in the model's prediction results.

[0062] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A watershed hydrological prediction method for multi-stage reservoir blockage, characterized in that, Includes the following steps: S10: Obtain historical flow data and future rainfall forecast data for multiple nodes within the basin. Perform independent normalization processing at the node level on the historical flow data of each node to obtain the normalized flow sequence of each node. S20, the future rainfall forecast data is encoded with rainfall features to obtain rainfall embedding features, and the normalized flow sequence, time coding features and rainfall embedding features are fused to obtain a fused feature tensor; S30, the fused feature tensor is input into a pre-constructed interactive diffusion graph convolutional network tree structure for spatiotemporal coupling modeling. The interactive diffusion graph convolutional network tree structure splits and processes the input sequence layer by layer according to the parity of time steps, and extracts multi-scale spatiotemporal features within each branch through dynamic graph generation and diffusion graph convolution to obtain a multi-scale spatiotemporal feature tensor. S40, the multi-scale spatiotemporal feature tensor is decoded through a gated linear unit to obtain a preliminary runoff prediction result. At the same time, the future rainfall forecast data is converted into a runoff correction amount through an independent decoder, and the runoff correction amount is injected into the preliminary runoff prediction result in a residual manner to obtain the final runoff prediction result. S50, during the model training phase, a weighted loss function is used to optimize the model parameters. The weighted loss function is the weighted sum of the mean squared errors of all nodes and the mean squared errors of the core target node.

2. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 1, characterized in that, In step S10, the independent normalization process at the node level includes: for the first... i The node of the first j The value of each feature Using formula Normalization is performed, where For nodes i Features j In the time dimension T Independent means on, For node i Features j In the time dimension T The independent standard deviations of the time series were calculated; and the time series was divided into different stages according to the flood season and the dry season, and the mean and standard deviation of each node were calculated and normalized in each stage.

3. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 1, characterized in that, In step S20, the rainfall feature encoding includes a spatial attention mechanism and a temporal attention mechanism. Both the spatial attention mechanism and the temporal attention mechanism employ a multi-head attention calculation formula, which is as follows: Where Q is the query matrix, K is the key matrix, and V is the value matrix. d k Let be the dimension of the key matrix; For the spatial attention mechanism, Q, K, and V are derived from the input features in the node dimension. The linear transformation is obtained; for the aforementioned temporal attention mechanism, Q, K, and V are derived from the input features in the time dimension. It is obtained by linear transformation.

4. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 1, characterized in that, In step S20, the feature fusion specifically involves: concatenating the normalized flow sequence, the time-encoded feature, and the rainfall embedding feature in the channel dimension to obtain the concatenated feature; then performing feature interaction and dimensionality reduction on the concatenated feature through two-dimensional convolution to obtain the fused feature tensor.

5. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 1, characterized in that, In step S30, the dynamic graph generation specifically involves: extracting dynamic relationships between nodes through adaptive learning to generate a dynamic adjacency matrix. The dynamic adjacency matrix automatically adjusts the influence weights between nodes based on the current hydrological state, and its calculation formula is as follows: in, X To fuse feature tensors; d The hidden dimension (i.e., number of channels) of the node features; W 1 represents the node projection weight matrix, used to map node features to the node association space; b 1 represents the first bias vector; The diffusion graph convolution specifically involves: based on the dynamic adjacency matrix... To conduct spatial information dissemination and obtain spatially aggregated features , ,in For a symmetric normalized Laplace matrix, For degree matrix, W 2 represents the spatial feature transformation weight matrix; b 2 is the second bias vector.

6. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 5, characterized in that, In step S30, each interactive diffusion map convolutional unit further includes temporal convolution and residual connections: processing the spatially aggregated features... Causal convolution is applied to extract temporal dependencies, and residual connections are added to obtain... .

7. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 6, characterized in that, In step S30, the layer-by-layer splitting and parallel processing according to the parity of time steps includes: the first layer splits the input sequence into even-time-step subsequences and odd-time-step subsequences, and inputs them into the first interactive diffusion graph convolutional unit to obtain the first output and the second output respectively; the second layer further splits the first output and the second output according to parity and inputs them into the second interactive diffusion graph convolutional unit; the third layer repeats the operation of the second layer; finally, the outputs of all branches of the tree structure are re-concatenated in the time dimension to obtain the multi-scale spatiotemporal feature tensor.

8. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 1, characterized in that, In step S40, the future rainfall forecast data is converted into runoff correction data through an independent decoder and injected as residuals. The specific formula is as follows: ,in For the final runoff prediction results, These are preliminary runoff prediction results. For future rainfall forecast data, α This is a learnable scaling factor.

9. The watershed hydrological prediction method for multi-stage reservoir blockage as described in claim 1, characterized in that, In step S50, the weighted loss function is specifically as follows: In the formula, λ is the weight coefficient; N is the total number of nodes; Q is the prediction step size; i is the node index; and q is the time step index. For the core target node at time step The predicted runoff value; For the core target node at time step The actual runoff value; For nodes At time step The predicted runoff value; For nodes At time step The actual runoff value; It is a dynamic weighting function; when the traffic to the core target node exceeds a set threshold... Or when traffic is on the rise, It increases exponentially, and its calculation formula is: ,in These are the weighting coefficients. This is the traffic threshold.

10. A watershed hydrological prediction system for multi-stage reservoir blockage, characterized in that, To implement the method of any one of claims 1 to 9, comprising: The data acquisition and normalization module is used to acquire historical flow data and future rainfall forecast data of multiple nodes in the basin, and to perform independent node-level normalization processing on the historical flow data of each node to obtain the normalized flow sequence of each node. The rainfall encoding and fusion module is used to encode rainfall features into the future rainfall forecast data to obtain rainfall embedding features, and to fuse the normalized flow sequence, time encoding features and rainfall embedding features to obtain a fused feature tensor. The spatiotemporal coupling modeling module has a built-in interactive diffusion graph convolutional network tree structure, which is used to perform spatiotemporal coupling modeling on the fused feature tensor. It splits and processes the input sequence layer by layer according to the parity of time steps, and extracts multi-scale spatiotemporal features within each branch through dynamic graph generation and diffusion graph convolution to obtain a multi-scale spatiotemporal feature tensor. The bypass decoding module is used to decode the multi-scale spatiotemporal feature tensor through a gated linear unit to obtain a preliminary runoff prediction result. At the same time, the future rainfall forecast data is converted into a runoff correction amount through an independent decoder and injected into the preliminary runoff prediction result in a residual manner to obtain the final runoff prediction result. The training optimization module is used to optimize the model parameters using a weighted loss function, which is a weighted sum of the mean squared errors of all nodes and the mean squared errors of the core target node.

Images