A method for predicting water level in coastal estuary under multi-element coupling

Abstract

The application discloses a kind of multi-element coupling under coastal area estuary water level prediction method, comprising: the time series representation of short-term dynamics and long-term rhythm is established, and spatial hydrological dependence information transmission mechanism is established by time series representation, to output estuary response prediction;The application models the spatial topology and physical coupling relationship between monitoring stations through "heterogeneous correlation body network structure", realizes the multi-level hydrological dependence expression from precipitation driving to water level response.This mechanism significantly improves the explainability and propagation accuracy of spatial characteristics, enabling the model to truly reproduce the transmission and diffusion behavior of hydrological quantities in natural basins.The proposed "periodic spectrum mapping mechanism" projects precipitation and water level time series from time domain to frequency spectrum space, and through principal frequency energy decomposition and feature reconstruction, realizes the extraction of periodic hydrological law.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of estuary water level prediction methods under the coupling of multiple elements in coastal areas. BACKGROUND

[0002] The existing hydrological prediction method generally relies on linear statistics or empirical model, and the feature extraction method is too simplified, which cannot identify the multi-level periodic behavior hidden in natural hydrological system. Especially under the interaction of tide and precipitation, the periodic signal is often covered by noise, making it difficult for the model to capture seasonal fluctuations and long-period change rules, resulting in insufficient prediction accuracy and stability. SUMMARY

[0003] The present application aims to provide a kind of estuary water level prediction methods under the coupling of multiple elements in coastal areas, which realizes the accurate identification and structured reconstruction of the periodic pattern in time series through the synergistic effect of multi-dimensional feature mapping mechanism and periodic spectrum mapping technology. The method projects the original time series into a multi-dimensional feature space, reconstructs the time marker relationship and energy distribution pattern in it, and realizes the adaptive integration of periodic characteristics with multi-domain perception fusion array, thereby forming time sequence representation with short-term dynamics and long-term rhythm.

[0004] To achieve the above purpose, the scheme of the present application is:

[0005] A kind of estuary water level prediction method under the coupling of multiple elements in coastal areas, comprising: establishing time sequence representation with short-term dynamics and long-term rhythm, then establishing spatial hydrological dependent information transmission mechanism from time sequence representation, to output estuary response prediction;Wherein:

[0006] The establishment of time sequence representation with short-term dynamics and long-term rhythm includes:

[0007] Step 1, time marker decomposition:

[0008] Input time series data of multiple precipitation stations and water level stations to form time sequence characteristics, the time series data includes time stamp and corresponding observation value;

[0009] Step 2, periodic pattern refining:

[0010] Map the time sequence characteristics to the frequency spectrum space, analyze the energy distribution of different periodic components, and form a two-dimensional feature matrix;

[0011] Step 3, periodic feature reconstruction:

[0012] Convert the two-dimensional feature matrix to one-dimensional time sequence structure to form the dynamic pattern of energy in short cycle, cross cycle and long cycle level;

[0013] The establishment of heterogeneous hydrological dependent information transmission mechanism from time sequence representation includes:

[0014] Step 1, Heterogeneous correlation network construction:

[0015] Obtain the geographic coordinates, hydrological properties and mutual spatial adjacency information of each monitoring station to form a heterogeneous hydrological dependency graph G = {N, E, R} based on the geographic coordinates, hydrological properties and mutual spatial adjacency information of each monitoring station;

[0016] Step 2, Dynamic update of node characteristics:

[0017] Form the state of each monitoring station node after spatial information enhancement according to the heterogeneous hydrological dependency graph G and the initial characteristic representation of each monitoring station node;

[0018] Step 3: Relationship-specific information fusion and global update:

[0019] Form the fused global spatial representation by locally aggregating the multi-relation node features formed by the multi-relation nodes;

[0020] Step 4, Estuary response prediction output:

[0021] Perform water level prediction and tidal response analysis in the estuary area according to the fused global spatial representation, output the estuary tidal water level prediction value and potential flood risk index under different time scales, and realize the end-to-end closed loop from spatial propagation to final physical quantity prediction.

[0022] The scheme further comprises:

[0023] I. Time stamp decomposition:

[0024] Multi-level analysis is performed on the time stamp information, which is decomposed from the "year / month / day / hour" format into interpretable time features, including for day resolution data, three attributes are extracted, the three attributes are:

[0025] (1) Week cycle identifier, represented by week number;

[0026] (2) Intra-month sequence, represented by date number;

[0027] (3) Annual position, represented by annual day sequence number;

[0028] For time resolution data, further extract the hour sequence number as a short cycle identifier;

[0029] The decomposed time stamp is represented as:

[0030] ,

[0031] Wherein: T is the time step, D is the attribute dimension; R represents the total set of real numbers, for daily resolution data, D=3, for time resolution data, D=4;

[0032] II. Observation value embedding:

[0033] Observation sequence representation:

[0034] Wherein: T is the time step, and T in them is the same;

[0035] First, the timestamp And are mapped to a 16-dimensional space to obtain And The time sequence position of the observation value is encoded to obtain The three features are fused to obtain the time sequence feature, that is, the time sequence embedding vector :

[0036]

[0037] Wherein: f is a mapping function, Xt is an observation sequence, is the timestamp, and PositionalEncoding represents the position identifier operator.

[0038] The scheme further is: the energy distribution of the different period components is analyzed to form a two-dimensional feature matrix:

[0039] I. Select the first K frequency components with the most significant energy in the interval [1, T / 2], as shown in the following formula:

[0040]

[0041] Wherein: is the frequency component set, represents the amplitude of the i-th frequency, X is the time sequence embedding vector obtained above, and TopK represents the operation of selecting the first K values after arranging by amplitude;

[0042] II. Then, the periods corresponding to these frequency components are calculated:

[0043]

[0044] III. Finally, the frequency components and their corresponding periods are integrated to reshape the two-dimensional feature matrix:

[0045] ,

[0046] Wherein: Pd i is the i-th period, f iis the ith frequency component, R is the real set;

[0047] The row vector represents the trend of cross-period variation, and the column vector captures the fluctuation within a single period.

[0048] The scheme further is: the expression of the formation of the dynamic pattern of energy at the short, cross-period and long period levels is:

[0049]

[0050]

[0051] wherein: represents the weight of the ith frequency component, represents the amplitude of the ith frequency component, represents the characteristic response operator of the multi-domain perception fusion array to the ith periodic matrix.

[0052] The scheme further is: the hydrological properties include three types of hydrological relationships:

[0053] I. Precipitation-precipitation relationship: used to describe the rainfall characteristics of different stations under atmospheric circulation or weather system;

[0054] II. Water level-water level relationship: reflects the water flow transmission and tidal feedback between upstream and downstream stations;

[0055] III. Precipitation-water level relationship: used to express the causal coupling and time delay characteristics in the rainfall runoff process.

[0056] The scheme further is: the state of each monitoring station node after the spatial information is enhanced is obtained by performing an information transmission operation in the graph structure, so that the state update of the node depends not only on the observation sequence itself, but also on the state of its associated nodes and the relationship type. The information transmission between nodes and the update rule of the node are represented as:

[0057]

[0058] wherein: represents the neighbor node set of node v and its neighbor node u under the relationship type r; is a message passing operator under a specific relationship r, used to simulate the transmission mechanism of different hydrological processes; is a nonlinear activation operator; represents the number of layers, represents the feature of node u on the ith layer network, the original feature of the node is its , and represents the feature of node v in the same layer; is the total set of relations between all nodes (a total of three types: rainfall-rainfall, water level-water level, rainfall-water level), represents the representation of the layer node v.

[0059] The scheme is further: in the formation of the fused global spatial representation, the fusion is a multi-relation hierarchical fusion architecture, which is divided into two levels:

[0060] I. Intra-relation fusion: feature aggregation within the same type of relation, maintaining its independent physical meaning;

[0061] II. Inter-relation fusion: integrate features of each relation category across layers to form a unified node global representation;

[0062] The final update of the node is represented as:

[0063]

[0064] wherein, and are learnable linear transformation matrices, the subscript r represents binding with the relation type, and MEAN is a neighbor aggregator that takes the mean; represents the representation of the layer node v.

[0065] After updating, each node feature not only contains its own observation information, but also fuses the spatio-temporal hydrological characteristics of the neighborhood, realizing dynamic information coupling at the regional scale.

[0066] The scheme is further: the output of the estuary tidal water level prediction value and the potential flood risk index under different time scales is obtained by fusing the global spatial representation, obtaining the fused features of all nodes, and predicting the water level of the nodes of the estuary at different prediction periods.

[0067] The beneficial effects of the present application are: through the "heterogeneous associated network structure", the spatial topology and physical coupling relationship between monitoring stations are modeled, realizing multi-level hydrological dependence expression from rainfall driving to water level response. This mechanism significantly improves the explainability and propagation accuracy of spatial features, enabling the model to truly reproduce the transmission and diffusion behavior of hydrological quantities in natural basins.

[0068] The proposed "periodic spectrum mapping mechanism" projects the rainfall and water level time series from the time domain to the frequency spectrum space, realizes the extraction of periodic hydrological laws through frequency energy decomposition and feature reconstruction.

[0069] The following further detailed explanation is made in combination with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a schematic diagram of the technical solution of the present application;

[0071] Figure 2 is a list of prediction ability evaluation indicators of the present application and other deep learning algorithms;

[0072] Figure 3 is a schematic diagram of the comparison between the prediction and the actual observation results under different prediction periods of an example. DETAILED DESCRIPTION

[0073] A multi-element coupling effect under coastal area estuary water level prediction method, through the synergistic effect of multi-dimensional feature mapping mechanism and periodic spectrum mapping technology, realizes the accurate identification and structured reconstruction of the periodic pattern implied in the time series. The core idea of this method is: project the original time series into a multi-dimensional feature space, reconstruct the time marker relationship and energy distribution pattern in it, and realize the adaptive integration of periodic characteristics with a multi-domain perception fusion array; Through the "heterogeneous correlation network structure", the spatial topology and physical coupling relationship between monitoring stations are modeled, realizing the multi-level hydrological dependence expression from precipitation driving to water level response. This mechanism significantly improves the explainability and propagation accuracy of spatial features, enabling the model to truly reproduce the transmission and diffusion behavior of hydrological quantities in natural basins; the prediction method comprises: establishing a time series representation of short-term dynamics and long-term rhythm, and then establishing a spatial hydrological dependence information transmission mechanism from the time series representation, thereby outputting the estuary response prediction; wherein:

[0074] The time series representation of short-term dynamics and long-term rhythm comprises:

[0075] Step 1, time marker decomposition:

[0076] Input time series data of multiple precipitation stations and water level stations to form time series features, the time series data including time stamps and corresponding observation values;

[0077] Wherein: the formation process of the time series features comprises:

[0078] I. Time stamp disassembly:

[0079] The time stamp information is analyzed in multiple levels, which is decomposed from the "year / month / day / time" format into interpretable time features, including for daily resolution data, three attributes are extracted, the three attributes are:

[0080] (1) Weekly period identifier, represented by the number of weeks;

[0081] (2) Intra-month order, represented by the number of days;

[0082] (3) Annual position, represented by the annual day sequence number;

[0083] For time resolution data, the hour sequence number is further extracted as a short period identifier;

[0084] The deconstructed timestamp is represented as:

[0085] ,

[0086] Where: T is the time step, D is the attribute dimension, R represents the real number set, for daily resolution data, D = 3, for time resolution data, D = 4;

[0087] II. Observation value embedding:

[0088] The observation sequence is represented as:

[0089] Where: T is the time step, and T in the above formula is the same as T in the formula;

[0090] First, the timestamp and are mapped to a 16-dimensional space to obtain and Then, the time sequence position of the observation value is encoded to obtain The three features are fused to obtain the time sequence feature, i.e., the time sequence embedding vector :

[0091]

[0092] Where: f is a mapping function, Xt is an observation sequence, is a timestamp, and PositionalEncoding represents a position identifier operator. The resulting fused feature is used as the input for subsequent period extraction.

[0093] Step 2, Periodic pattern extraction:

[0094] Map the time sequence feature to the frequency spectrum space to analyze the energy distribution of different period components and form a two-dimensional feature matrix.

[0095] The analysis of the energy distribution of different period components to form a two-dimensional feature matrix is as follows:

[0096] I. Select the first K frequency components with the most significant energy in the interval [1, T / 2], as shown in the following formula:

[0097]

[0098] Where: is the set of frequency components; represents the amplitude of the i-th frequency, X is the time sequence embedding vector obtained above, and TopK represents the operation of selecting the first K values after arranging by amplitude.

[0099] II. Then, the periods corresponding to these frequency components are calculated:

[0100]

[0101] III. Finally, these frequency components and their corresponding periods are integrated to reshape a two-dimensional feature matrix: ,

[0102] where: Pd i is the i-th period, f i is the i-th frequency component, and R is the real set.

[0103] The row vectors represent the trend of cross-period changes, and the column vectors capture the fluctuations within a single period.

[0104] This structure realizes the reconstruction of space across the time domain to the spectral domain, enabling the model to reveal the periodic characteristics driven by tidal and precipitation alternately from the perspective of energy distribution.

[0105] Step 3: Reconstruction of periodic characteristics:

[0106] The two-dimensional feature matrix is converted to a one-dimensional time series structure, forming the dynamic patterns of energy at short, cross, and long periods.

[0107] To achieve efficient fusion of multi-scale periodic characteristics, this embodiment designs a new structure "multi-domain perception fusion array". The array consists of six parallel perception channels, each with independent perception scales and weight parameters to simultaneously detect the dynamic patterns of the period matrix at short, cross, and long periods. The feature responses of each channel output by the array are fused through a "spectral energy guided weighting mechanism". This mechanism automatically assigns weighting coefficients based on the energy amplitude of spectral components, allowing dominant energy higher period components (such as principal tides and seasonal waves) to have higher expression weights in the final fusion result, thereby achieving adaptive enhancement based on physical energy;

[0108] The weighting relationship can be formally expressed as follows: the expression of the dynamic patterns of energy at short, cross, and long periods is:

[0109]

[0110]

[0111] where: represents the weight of the i-th frequency component, represents the amplitude of the i-th frequency component, represents the characteristic response operator of the multi-domain perception fusion array to the i-th periodic matrix. The fused high-dimensional periodic representation is regressed into a one-dimensional time series structure via a "residual remapping unit". This unit ensures that the enhanced representation maintains the energy conservation and dynamic consistency of the original signal while integrating multi-scale features through a residual information compensation mechanism. This design significantly reduces gradient oscillation during the training phase, improving model stability and convergence speed, while ensuring that the final embedding vector is both expressive and faithful to the underlying hydrological dynamic evolution law.

[0112] The spatial hydrological dependency information transmission mechanism established by the time series representation includes:

[0113] Step 1, Heterogeneous correlation network construction:

[0114] Obtain the geographic coordinates, hydrological properties and mutual spatial adjacency information of each monitoring station to form a heterogeneous hydrological dependency graph G = {N, E, R} based on the geographic coordinates, hydrological properties and mutual spatial adjacency information of each monitoring station;

[0115] The hydrological properties include three types of hydrological relationships:

[0116] I. Precipitation-precipitation relationship: used to describe the rainfall characteristics of different stations under atmospheric circulation or weather systems;

[0117] II. Water level-water level relationship: reflects the water flow transmission and tidal feedback between upstream and downstream stations;

[0118] III. Precipitation-water level relationship: used to express the causal coupling and time delay characteristics in the rainfall runoff process.

[0119] This network structure is constructed through a digital terrain model and a hydrological zoning boundary recognition algorithm, ensuring that each edge has physical meaning and avoiding misinformation caused by invalid connections.

[0120] Step 2, Dynamic update of node features:

[0121] Form the state of each monitoring station node after spatial information enhancement based on the heterogeneous hydrological dependency graph G and the initial feature representation of each monitoring station node;

[0122] The state of each monitoring station node after spatial information enhancement is formed by performing information transmission operations in the graph structure, so that the state update of the node not only depends on its own observation sequence, but also is related to the state of its associated nodes and the relationship type. The information transmission and node update rules are represented as:

[0123]

[0124] Where: represents the neighbor node set of node v and its neighbor node u under relationship type r. For a specific relation r, a message passing operator is used to simulate the transmission mechanism of different hydrological processes; It is a nonlinear activation operator; Represents the number of floors. Representing the The characteristics of node u on a layer network, the original characteristics of the node are its Similarly, This represents the characteristics of node v at the same level; It is the total set of relationships between all nodes (in three categories: rainfall-rainfall, water level-water level, and rainfall-water level). Indicates the first The representation of layer node v. This design realizes hierarchical modeling of multiple types of hydrological relationships, making information propagation directional and consistent in physical semantics.

[0125] Step 3: Relationship-specific information fusion and global update:

[0126] The multi-relation node features, which are formed by the local aggregation of multi-relation nodes, are used to form a fused global spatial representation.

[0127] In the resulting fused global spatial representation, the fusion is a multi-relational hierarchical fusion architecture, which consists of two levels:

[0128] I. Intra-relational fusion: Feature aggregation within the same type of relation, while maintaining its independent physical meaning;

[0129] II. Relationship Integration: Integrate the features of each relationship category across layers to form a unified global representation of nodes;

[0130] The final update of a node can be represented as:

[0131]

[0132] in, and It is a learnable linear transformation matrix, where the subscript r indicates binding to the relation type. MEAN is a neighbor aggregator that takes the mean. Indicates the first Representation of layer node v;

[0133] After the update is completed, the features of each node not only include its own observation information, but also integrate the spatiotemporal hydrological characteristics of the surrounding area, realizing dynamic information coupling at the regional scale.

[0134] Step 4, Estuary Response Prediction Output:

[0135] According to the fused global spatial representation, water level prediction and tidal response analysis of the estuary area are performed, and estuary tidal water level prediction values and potential flood risk indicators at different time scales are output, realizing an end-to-end closed loop from spatial propagation to final physical quantity prediction.

[0136] The output of the estuary tidal water level prediction values and potential flood risk indicators at different time scales is obtained by fusing the global spatial representation, obtaining the fusion features of all nodes, and predicting the water level of the nodes of the estuary at different prediction periods. After sufficient fusion of spatial dependent information, the model finally performs water level result inference through the self-designed spatio-temporal prediction output unit. Based on the fused global node features, the unit performs water level prediction and tidal response analysis of the estuary area. Through the internal time sequence mapping layer and the nonlinear combination module, the estuary tidal water level prediction values and potential flood risk indicators at different time scales can be output, realizing an end-to-end closed loop from spatial propagation to final physical quantity prediction.

[0137] A case analysis, Ho Chi Minh City in Vietnam is the largest city and economic center in Vietnam, located in the lower reaches of the Saigon River and the Dong Nai River. When heavy rainfall and storm surges overlap, the risk of flood events is particularly prominent. With the continuous rise of sea level, the threat and severity of these disasters are increasing. The case uses two sets of observation data. The first set of data comes from the daily rainfall records of seven rain stations in Ho Chi Minh City and its surrounding areas, and the second set of data contains daily water level observation data of the upper reaches of the Dong Nai River and the Cam Ranh estuary (the river mouth). The recording period is from January 1, 2008 to December 31, 2010 for three consecutive years, and these data are used for model training and verification. We use four commonly used indicators for analyzing and predicting the effect (MSE mean square error, MAE mean absolute error, R2 determination coefficient, PE peak prediction error) to evaluate the model and other baseline models in time series prediction (RNN recurrent neural network, LSTM long short-term memory model, TCN time convolution network, Informer attention mechanism-based time series prediction model). The method proposed in this embodiment is the best in both short-term prediction and long-term prediction; the results are shown in Figure 2 .

[0138] The table shows the performance of different models (RNN, LSTM, TCN, Informer, and the model of the present patent) on the same test set across different prediction lengths. The evaluation metrics include mean squared error (MSE), mean absolute error (MAE), coefficient of determination R², and peak error percentage (PE) (%). The best performance is marked in red, and the suboptimal results are underlined. In short-term prediction, the model of the present patent performs best in all indicators, with the lowest MSE, MAE, and PE, and an average R² value of 0.983. The TCN model ranks second, with stable performance in each length interval, and occasionally surpasses the model of the present patent in MSE and MAE. The Informer model ranks third, showing strong peak capture ability. In long-term prediction, PhaST-Net continues to dominate all competitors, with the lowest MSE of 0.966 and PE of 8.53%, ranking first. Although the Informer ranks second, there is still a significant gap. Overall, the model of the present patent performs well in both short-term and long-term prediction, with prediction accuracy (MSE, MAE) and reliability (PE, R²) at the industry-leading level.

[0139] In addition, the model of the present method performs well on non-training sets. We selected a storm surge disaster that occurred on April 1, 2012 in Vietnam, triggered by storm Pahkar, with multiple factors. It was found that the model could capture different prediction periods for this event and performed well, as shown in Figure 3 .

[0140] The above-mentioned embodiment method combines time marker decomposition and multi-dimensional feature embedding to form a high-dimensional time representation that can express both short-term fluctuations and long-term rhythms. This technology introduces spectral information in the physical energy sense into hydrological prediction modeling, which is different from traditional statistical regression or black-box neural networks. The core is to replace simple "time convolution" with "period energy mapping."

[0141] The proposed multi-domain perception fusion array analyzes the period feature matrix in parallel on different perception scales, and proposes a spectral energy guided weighting mechanism to automatically adjust the weights of each period component according to the spectral energy intensity. This is a structural layer innovation on top of the time series feature reconstruction mechanism, and is an original design in the model architecture field. Unlike existing multi-scale convolution structures, the array uses "spectral energy weight" as the dynamic fusion basis to form a physically oriented adaptive feature fusion method, which is innovative and significant.

[0142] The proposed heterogeneous hydrological dependency graph construction explicitly describes three types of hydrological relationships: precipitation-precipitation, water level-water level, and precipitation-water level in the same graph structure. This graph structure is automatically generated based on the digital terrain model and the basin topology, reflecting the true spatial structure and hydrodynamic direction of the natural water system. It is a spatial modeling scheme that combines physical topology and graph modeling, distinguishing it from existing GNN methods. Through physical consistency constraints (such as upstream-downstream flow direction and basin boundary), this technology ensures that each edge of the graph has a clear geographical and hydrodynamic semantic, which is an engineering structural innovation in spatial relationship modeling.

[0143] In the hydrological relationship propagation engine and multi-relation hierarchical fusion mechanism, the hydrological relationship propagation engine is proposed to realize directional information transmission based on relationship types in the heterogeneous graph structure. Through relationship-aware aggregation and multi-relation hierarchical fusion mechanism, the model can distinguish different physical processes (such as rainfall synchronization vs. water flow propagation) and dynamically adjust the transmission strength of information flow. It is an algorithm-level innovation in spatial information propagation. Unlike traditional graph neural propagation "uniform convolution", this scheme realizes differentiated propagation paths for hydrological physical processes through "relationship-aware propagation + hierarchical weighted fusion", with obvious innovation and physical rationality.

[0144] The proposed spatio-temporal collaborative prediction closed loop and estuary response output mechanism seamlessly connects the time domain module (periodic spectrum mapping and fusion array) and the spatial domain module (heterogeneous dependency graph and propagation engine) to form a unified spatio-temporal collaborative prediction framework. Through the spatio-temporal prediction output unit, the tidal water level prediction and flood risk quantification results are directly generated, realizing an end-to-end inference closed loop from rainfall input to tidal output. Unlike traditional step-by-step tidal prediction methods, this scheme realizes the full-link automatic process of multi-source observation → spectrum analysis → spatial propagation → tidal response prediction.

Claims

1. A method for predicting water level in a coastal estuary under the action of multiple factors, comprising: The application relates to a method for predicting the response of a river mouth, and belongs to the technical field of hydrology. The method comprises the following steps: Step 1, time mark decomposition: Input time series data of multiple precipitation stations and water level stations to form time series features, wherein the time series data comprises time stamps and corresponding observation values; Step 2, periodic pattern extraction: Map the time series features to a frequency spectrum space, analyze the energy distribution of different periodic components, and form a two-dimensional feature matrix; Step 3, periodic feature reconstruction: Convert the two-dimensional feature matrix into a one-dimensional time series structure to form dynamic patterns of energy in short, cross and long periods; The method comprises the following steps: Step 1, heterogeneous correlation network construction: Obtain geographical coordinates, hydrological properties and mutual spatial adjacency information of each monitoring station to form a heterogeneous hydrological dependency graph G={N, E, R} according to the geographical coordinates, hydrological properties and mutual spatial adjacency information of each monitoring station, The hydrological properties comprise three types of hydrological relationships: (1) precipitation-precipitation relationship: used for describing the rainfall characteristics of different stations under atmospheric circulation or weather systems; (2) water level-water level relationship: reflecting the water flow transmission and tidal feedback between upstream and downstream stations; (3) precipitation-water level relationship: used for expressing the causal coupling and time delay characteristics in the rainfall runoff process; Step 2, dynamic updating of node feature: Form the node state of each monitoring station after spatial information enhancement according to the heterogeneous hydrological dependency graph G and the initial feature representation of each monitoring station node; Step 3, relationship-specific information fusion and global updating: Form the global spatial representation after fusion by locally aggregating the multi-relation node features of the multi-relation nodes; Step 4, river mouth response prediction output: Perform water level prediction and tidal response analysis in the river mouth area according to the global spatial representation after fusion to output river mouth tidal water level prediction values and potential flood risk indicators at different time scales, and realize an end-to-end closed loop from spatial propagation to final physical quantity prediction.

2. The prediction method of claim 1, wherein, The formation process of the time series features comprises: (1) time stamp decomposition: Decompose the time stamp information into interpretable time features by decomposing the time stamp information from the "year / month / day / hour" format, including extracting three attributes for daily resolution data, namely: (1) week period identifier, represented by the number of weeks; (2) intra-month order, represented by the number of days; (3) annual position, represented by the annual day sequence number; For time resolution data, further extract the hour sequence number as a short period identifier; The decomposed time stamp is represented as: ， Wherein: T is the time step, D is the attribute dimension; R represents the real number set, for daily resolution data, D=3, for time resolution data, D=4; (2) observation value embedding: The observation sequence is represented as: wherein: T is a time step, and T in (1) is the same as in (2); First, the timestamps and are mapped to a 16-dimensional space, obtaining and , then the time series position of the observation value is encoded, obtaining , the three features are fused to obtain the time series feature, that is, the time series embedding vector : where: f is a mapping function, Xt is the observation sequence, is the timestamp, and PositionalEncoding represents a position identification operator.

3. The prediction method of claim 2, wherein, The energy distribution of different periodic components is analyzed to form a two-dimensional feature matrix, which comprises: (1) select the first K frequency components with the most significant energy in the interval [1, T / 2], and the formula is as follows: wherein: is a set of frequency components, denotes the amplitude of the i-th frequency, X is the aforementioned time-series embedding vector, and TopK denotes the operation of selecting the top K values after sorting by amplitude. (2) then calculate the periods corresponding to these frequency components: Thirdly, the frequency components and their corresponding periods are integrated to reshape the two-dimensional feature matrix: ， where: Pd i is the ith cycle, f i is the ith frequency component, R is the set of real numbers; The row vectors represent the cross-period trend, and the column vectors capture the fluctuations within a single period.

4. The prediction method of claim 3, wherein, The expression of the dynamic pattern of the formation energy at the short, cross, and long periods is: wherein: represents the weight of the i-th frequency component, represents the amplitude of the i-th frequency component, represents the characteristic response operator of the multi-domain perception fusion array to the i-th periodic matrix.

5. The prediction method of claim 4, wherein, The enhanced node state of each monitoring station forms spatial information, which is executed in the graph structure to perform information transmission operations. The state update of the node depends not only on its own observation sequence, but also on the state of its associated nodes and the relationship type. The information transmission between nodes and the update rule of the node are represented as: in: This represents the set of neighbor nodes of node v and its neighbor node u under relation type r; For a specific relation r, a message passing operator is used to simulate the transmission mechanism of different hydrological processes; It is a nonlinear activation operator; Represents the number of floors. Representing the The characteristics of node u on a layer network, the original characteristics of the node are its Similarly, This represents the characteristics of node v at the same level; It is the total set of relationships between all nodes, categorized into three types: rainfall-rainfall, water level-water level, and rainfall-water level. Indicates the first Representation of layer node v.

6. The prediction method of claim 5, wherein, The fusion of the global spatial representation is a multi-relation hierarchical fusion architecture, which is divided into two levels:

1. Intra-relation fusion: feature aggregation within the same type of relationship to maintain its independent physical meaning; 2. Inter-relation fusion: integrate features of each relationship category across layers to form a unified global representation of nodes; The final update of the node is represented as: wherein, and is a learnable linear variation matrix whose subscript r indicates binding to a relation type, MEAN is a neighbor aggregator taking the mean; denotes the representation of the layer node v. After the update, each node's features not only contain its own observation information, but also integrate the spatiotemporal hydrological characteristics of the neighborhood, achieving dynamic information coupling at the regional scale.

7. The prediction method of claim 1, wherein, The output of the estuary tidal water level prediction value and potential flood risk index at different time scales is obtained by fusing the global spatial representation to get the fused features of all nodes, and predicting the water level of the estuary nodes at different forecast periods.

Images