A residual water level calculation method, system, device and medium for a virtual tide station

Abstract

This invention discloses a method, system, equipment, and medium for estimating the residual water level of a virtual tide gauge station. The method first acquires real-time measured total water level data, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data from each reference tide gauge station, as well as the spatial location information of each reference tide gauge station and the target virtual tide gauge station. Then, the data is preprocessed to obtain a spatiotemporal data cube and residual water level data for each reference tide gauge station. Finally, based on the spatial location information, the spatiotemporal data cube, and the residual water level data, a hybrid neural network model including convolutional neural network branches, long short-term memory neural network branches, position encoding branches, feature fusion layers, and decoding estimation layers is used to estimate the residual water level, obtaining the estimated residual water level sequence for the target virtual tide gauge station. This invention effectively integrates spatiotemporal features, improves the generalization ability for unknown areas, and thus enhances the accuracy and real-time performance of residual water level estimation.

Description

Technical Field

[0001] This invention relates to the field of marine surveying and mapping technology, and in particular to a method, system, equipment and medium for calculating the residual water level of a virtual tide gauge station. Background Technology

[0002] Accurate and real-time measurement of ocean water levels is crucial for resource development and risk control in nearshore areas. Traditional water level monitoring mainly relies on fixed physical tide gauge stations. However, physical tide gauge stations have inherent drawbacks such as high construction and maintenance costs and sparse and uneven spatial distribution, failing to meet the urgent need for high spatiotemporal resolution water level monitoring in vast nearshore areas. Therefore, virtual tide gauge technology has emerged. This technology aims to use readily available driving data to extrapolate water level information in areas without existing stations through mathematical models. In existing technologies, a typical approach is the virtual tide gauge residual water level extrapolation method based on Long Short-Term Memory (LSTM) networks. This method typically involves: collecting historical water level data from physical tide gauge stations and performing tidal separation to obtain residual water level sequences; constructing an LSTM neural network model with residual water level time-series data as input and predicted water level as output; training the model using historical data to learn the nonlinear mapping relationship from input to output; and finally, combining tidal models and spatial interpolation techniques to achieve residual water level extrapolation for the target virtual station.

[0003] However, while LSTM models excel at time-dependent modeling, they are insufficient at capturing spatial correlations between features. Nearshore water level changes are driven by multiple spatially significant factors, including wind fields, pressure fields, wave propagation, and tidal wave propagation. Relying solely on single or multiple independent time series as input leads to a one-sided understanding of complex hydrodynamic processes, limiting further improvements in estimation accuracy. Furthermore, when the target virtual tide gauge is located in an area with sparse physical tide gauge distribution or far from the model's training data coverage, the generalization performance of LSTM models relying purely on time series features significantly decreases, easily resulting in large estimation errors in areas without gauge stations or in the open sea. For sudden, nonlinear, and drastic water level changes caused by typhoons, storm surges, etc., LSTM models primarily rely on information within a limited historical time window for extrapolation prediction, making it difficult to quickly capture and respond to abrupt changes triggered by rapidly evolving spatial weather systems. This often leads to time lags or underestimations of peak water levels. Moreover, the performance of this method is highly dependent on the continuity and quality of the input station's time series data. If key reference tide gauge data is missing or abnormal, it will directly affect the stability of the entire model's output. In addition, the model has failed to effectively integrate static geographic information (such as seabed topography) with dynamic spatial field data, and its end-to-end automation and real-time processing capabilities need to be strengthened.

[0004] In summary, existing methods for estimating residual water levels have technical shortcomings, including failure to effectively integrate spatiotemporal characteristics, weak generalization ability in unknown areas, and low accuracy and real-time performance in estimating residual water levels. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, equipment, and medium for estimating the residual water level of a virtual tide gauge station, which can effectively integrate spatiotemporal characteristics, improve the generalization ability of unknown areas, and thus improve the accuracy and real-time performance of residual water level estimation.

[0006] To achieve the above objectives, this invention provides a method for calculating the residual water level of a virtual tide gauge station, comprising: Acquire real-time multi-source data from multiple reference tide gauge stations, as well as spatial location information of each reference tide gauge station and the target virtual tide gauge station; wherein, the real-time multi-source data includes real-time measured total water level data, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data from each reference tide gauge station; The multi-source data is preprocessed to obtain a spatiotemporal data cube and residual water level data of each reference tide gauge station; Based on the spatial location information, the spatiotemporal data cube, and the remaining water level data, a hybrid neural network model is used to calculate the remaining water level, thereby obtaining the calculated remaining water level sequence of the target virtual tide gauge station; wherein, the hybrid neural network model includes a convolutional neural network branch, a long short-term memory neural network branch and a position encoding branch, a feature fusion layer and a decoding calculation layer.

[0007] Optionally, the step of using a hybrid neural network model to calculate the remaining water level based on the spatial location information, the spatiotemporal data cube, and the remaining water level data to obtain the calculated remaining water level sequence of the target virtual tide gauge station includes: Based on the spatial location information, the spatiotemporal data cube, and the residual water level data, a hybrid neural network model is used to extract features, thereby obtaining the temporal features, spatial features, and location features of each reference tide gauge station, as well as the spatial features and location features of the target virtual tide gauge station. Attention weights are calculated based on the temporal, spatial, and locational characteristics of each reference tide gauge station and the spatial and locational characteristics of the target virtual tide gauge station. The attention weights are used to fuse the temporal, spatial, and locational features of each reference tide gauge station to obtain fused features. The fusion features are decoded to output the estimated residual water level sequence of the target virtual tide gauge station.

[0008] Optionally, based on the spatial location information, the spatiotemporal data cube, and the residual water level data, a hybrid neural network model is used for feature extraction to obtain the temporal features, spatial features, and location features of each reference tide gauge station, and the spatial features and location features of the target virtual tide gauge station, including: The residual water level data is processed by a branch of a long short-term memory neural network to extract the time characteristics of each reference tide gauge station. The spatiotemporal data cube is processed by a convolutional neural network branch to extract the spatial features of each reference tide gauge station and the target virtual tide gauge station; The spatial location information is processed by the location encoding branch to extract the location features of each reference tide gauge station and the target virtual tide gauge station.

[0009] Optionally, the step of calculating attention weights based on the temporal, spatial, and locational characteristics of each reference tide gauge station and the spatial and locational characteristics of the target virtual tide gauge station includes: The temporal, spatial, and locational characteristics of each reference tide gauge station are pieced together to form the comprehensive characteristics of each reference tide gauge station. The spatial and locational features of the target virtual tide gauge station are spliced ​​together to form the comprehensive features of the target virtual tide gauge station; The similarity between the comprehensive characteristics of the target virtual tide gauge station and the comprehensive characteristics of each reference tide gauge station is calculated to obtain the attention weight.

[0010] Optionally, the hybrid neural network model is trained through the following steps: Acquire historical multi-source data from multiple reference tide gauge stations, as well as spatial location information of each reference tide gauge station and the target virtual tide gauge station, to construct training samples; The training samples are input into a network structure containing the convolutional neural network branch, the long short-term memory neural network branch, and the position encoding branch, and the remaining water level is predicted by forward propagation. Calculate the loss function value between the predicted residual water level and the actual residual water level label; Based on the loss function value, the parameters of the network structure are optimized using the backpropagation algorithm to obtain a trained hybrid neural network model.

[0011] Optionally, the preprocessing of the multi-source data to obtain the spatiotemporal data cube and the residual water level data of each reference tide gauge station includes: Based on the real-time measured total water level data and the real-time astronomical tide forecast data, the initial residual water level data of each reference tide gauge station is calculated. The initial residual water level data, the real-time climate reanalysis data, and the pre-stored seabed topography data are spatiotemporally aligned to obtain the residual water level data of each reference tide gauge station, as well as the spatiotemporally aligned climate reanalysis data and seabed topography data. The spatiotemporal data cube is obtained by overlaying the spatiotemporal aligned climate reanalysis data and seabed topography data along the channel dimension.

[0012] Optionally, the step of spatiotemporally aligning the initial residual water level data, the real-time climate reanalysis data, and the pre-stored seabed topography data includes: The initial residual water level data of each reference tide gauge station will be uniformly resampled to a preset time interval; The real-time climate reanalysis data is resampled to the preset time interval and interpolated to a preset spatial grid. The seabed topography data is cropped and sampled into the preset spatial grid.

[0013] To achieve the above objectives, the present invention also provides a virtual tide gauge residual water level calculation system, comprising: The data acquisition module is used to acquire real-time multi-source data from multiple reference tide gauge stations, as well as the spatial location information of each reference tide gauge station and the target virtual tide gauge station; wherein, the real-time multi-source data includes real-time measured total water level data of each reference tide gauge station, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data; The data preprocessing module is used to preprocess the multi-source data to obtain a spatiotemporal data cube and residual water level data of each reference tide gauge station. The residual water level estimation module is used to estimate the residual water level based on the spatial location information, the spatiotemporal data cube, and the residual water level data, using a hybrid neural network model to obtain the estimated residual water level sequence of the target virtual tide gauge station; wherein, the hybrid neural network model includes a convolutional neural network branch, a long short-term memory neural network branch and a position encoding branch, a feature fusion layer and a decoding estimation layer.

[0014] To achieve the above objectives, the present invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the virtual tide gauge residual water level estimation method as described above.

[0015] To achieve the above objectives, the present invention also provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the virtual tide gauge residual water level estimation method as described above.

[0016] Compared with existing technologies, the present invention provides a method, system, equipment, and medium for estimating residual water levels at virtual tide gauge stations. It dedicates a convolutional neural network branch to extracting spatial structural features driving water level changes from the grid data, thus overcoming the limitations of pure long short-term memory (LSTM) network models in spatial modeling. This allows the model to more comprehensively understand the spatial synergistic effects of factors such as wind, pressure, waves, and topography. The LSTM branch continues to leverage its advantage in capturing the temporal evolution of water levels, while the location encoding branch explicitly embeds geospatial relationships into the model. By fusing the parallel extracted spatiotemporal features with location features and utilizing a spatiotemporal data cube constructed from multi-source data, the model achieves deep learning of the joint spatial-temporal mapping relationship, improving generalization ability and estimation accuracy in unknown areas with sparse or no data at physical tide gauge stations. Simultaneously, by real-time fusion of static topography and dynamic meteorological fields, it reduces over-reliance on the quality of data from a single station, enhancing system robustness. Furthermore, by instantly capturing spatial abrupt change signals, it improves the real-time performance of estimating extreme water level events caused by typhoons and storm surges. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for calculating the residual water level of a virtual tide gauge station provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a spatiotemporal data cube provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the temporal replication process of seabed topographic data provided in this embodiment of the invention. Figure 4 This is another schematic diagram of a spatiotemporal data cube provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a hybrid neural network model architecture provided in an embodiment of the present invention; Figure 6 This is a structural block diagram of a virtual tide gauge residual water level estimation system provided in an embodiment of the present invention; Figure 7 This is a structural block diagram of a terminal device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 , Figure 1 This is a flowchart of a method for calculating the residual water level of a virtual tide gauge station according to an embodiment of the present invention. The method for calculating the residual water level of a virtual tide gauge station includes steps S1 to S3: Step S1: Obtain real-time multi-source data from multiple reference tide gauge stations, as well as the spatial location information of each reference tide gauge station and the target virtual tide gauge station; wherein, the real-time multi-source data includes real-time measured total water level data, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data from each reference tide gauge station.

[0021] The real-time measured total water level data is obtained in real-time from reference tide gauge stations distributed in the nearshore area via data link, with 24-hour and 5-minute intervals for a specified date; the real-time astronomical tide forecast data is dynamically calculated based on the harmonic constant of the ETO20 global tidal model; the real-time climate reanalysis data includes sea surface wind speed, sea level pressure, and sea surface temperature data from the global climate reanalysis dataset and effective wave height data from the ENVISAT Advanced Synthetic Aperture Radar (ASAR) real-time wave dataset, and the real-time climate reanalysis data is transmitted in NetCDF format; the seabed topography data is static data from the 15″ resolution global depth topography model (SRTM15); and the spatial location information is the latitude and longitude coordinates of each reference tide gauge station and the target virtual tide gauge station.

[0022] Step S2: Preprocess the multi-source data to obtain a spatiotemporal data cube and residual water level data of each reference tide gauge station; In an optional embodiment, step S2 includes steps S201 to S203: Step S201: Calculate the initial residual water level data of each reference tide gauge station based on the real-time measured total water level data and the real-time astronomical tide forecast data. Understandably, to eliminate the periodic interference of astronomical tides, tidal separation can be performed on the measured water levels at each reference tide gauge station, i.e., measured water level = astronomical tide + residual water level. The astronomical tide level is calculated using the harmonic constant extracted from the ETO20 global tidal model, and the core formula is as follows: ; in, yes Astronomical tide level at any given time; It is a tide index; and The amplitude and lag angle were extracted from ETO20; It is the tidal angular velocity (a known constant); It is the astronomical phase angle of the tidal equinox, which can be calculated using astronomical parameters based on time and location; It is the intersection factor (which changes slowly and can be approximated as 1 in short-term forecasts).

[0023] Then, combining the measured water levels at each reference tide gauge station, the residual water level after removing the astronomical tide level is calculated: ; in, For reference, the actual water level measured at the tide gauge station.

[0024] Step S202: Spatiotemporally align the initial residual water level data, the real-time climate reanalysis data, and the seabed topography data to obtain the residual water level data of each reference tide gauge station, as well as the spatiotemporally aligned climate reanalysis data and seabed topography data. In one optional embodiment, the spatiotemporal alignment of the initial residual water level data, the real-time climate reanalysis data, and the pre-stored seabed topography data includes steps a to c: Step a. Resample the initial residual water level data of each reference tide gauge station to a preset time interval; For example, the initial residual water levels calculated by each reference tide gauge station are standardized to a 5-minute interval, and the original sequence is set as follows: ,in, It is the timestamp of the original data. Raw data Remaining water level at any given time; target 5-minute sequence: ,in, It is the timestamp of the target every 5 minutes. Target 5-minute interval The remaining water level at any given time.

[0025] To unify the residual water level data and real-time climate reanalysis data to a 5-minute time interval, corresponding data interpolation / averaging methods were adopted to address the differences between the original sampling interval and the target 5-minute interval, as follows: When the original sampling interval At the minute mark, cubic spline interpolation is used to complete the data, ensuring the continuity and smoothness of the time series data. Cubic spline interpolation function: ; Wherein, the coefficients of the interpolation function , , , The following conditions must be met to find a solution: Functional continuity: ; The first derivative is continuous: ; The second derivative is continuous: ; Boundary conditions (natural boundaries): ; By solving using the above interpolation function, the target 5-minute sequence is obtained as follows: .

[0026] When the original sampling interval At the minute mark, the arithmetic mean method is used to downsample the data to avoid data redundancy. First, define a 5-minute window function to determine the range of raw data corresponding to each target time point: ; Then, the arithmetic mean of the raw data in each window is calculated and used as the data value for the target time point corresponding to that window. The calculation formula is as follows: ; in, It is a window The number of data points in the data.

[0027] In addition, for the spatiotemporally aligned residual water level data, the 3σ principle can be used to detect and remove outliers within the time window.

[0028] For windows Data points in the data, if they satisfy: ; Then mark the point as an outlier and remove it, where, , for window The average of the remaining water level; , for window The standard deviation of the residual water level.

[0029] Step b. Resample the real-time climate reanalysis data to the preset time interval and interpolate it to the preset spatial grid; Step c. Cropping and sampling the seabed topography data into the preset spatial grid.

[0030] It should be noted that, in order to achieve the fusion and networking of dynamic real-time climate reanalysis data and seabed topography data, the real-time climate reanalysis data are first interpolated with the same spatial grid (1'×1') and temporal resolution (5 minutes), and then combined with the processed seabed topography data to construct a spatiotemporal data cube.

[0031] It should be noted that since the real-time climate reanalysis data in the embodiments of the present invention is mainly meteorological and oceanographic data, the following description will use meteorological and oceanographic data as an example.

[0032] Specifically, firstly, let the original meteorological and oceanographic data be... That is, the first Meteorological variables (such as customs, air pressure) in time Spatial location The value; the original time resolution is The original spatial grid is Let the target spatiotemporal data cube be... The target time resolution is minutes; target space grid is .

[0033] Then, the meteorological and oceanographic data are resampled to the target temporal resolution and interpolated to the target spatial grid.

[0034] For example, for each target space grid point and each meteorological variable First, extract the complete time series of the grid point from the original data to obtain: ; The original time series is interpolated to the target time resolution, i.e., a target time point of 5 minutes, using linear interpolation. The interpolation formula is: ; in, .

[0035] After interpolation, the final time-aligned data is as follows: .

[0036] For example, for each target time point and each variable Find the grid points containing the target space. The original mesh cell to which it belongs, that is, finding the one that satisfies Original mesh vertices .

[0037] Calculate the interpolation weighting coefficients based on the relative positions of the target space grid points within the original grid points: This weight reflects the distance between the target space grid point and the four vertices of the original grid; the closer the distance, the greater the weight of the corresponding vertex.

[0038] Furthermore, the values ​​of the four original grid vertices are weighted and summed to obtain the spatial interpolation result of the target space grid points: ; The interpolated data from all time steps and spatial grid points are integrated into a three-dimensional tensor, forming a spatiotemporal data cube corresponding to each meteorological variable k: ; in, This represents the number of time steps, corresponding to the length of the time series with 5-minute intervals. The spatial grid has dimensions of 1'×1'; the generated spatiotemporal data cube is as follows: Figure 2 As shown, Figure 2 In, each time step (e.g.) ~ ) corresponds to one The spatial layers are 4, which represent the number of meteorological variable channels, corresponding to meteorological variables such as wind speed, air pressure, significant wave height, and sea surface temperature. After being stacked in multiple layers, they form a cubic structure.

[0039] Then, the seafloor topography data was cropped and sampled into a spatial grid of the exact same size as the meteorological oceanographic field, i.e., 1'×1'. Grid. Because the seabed topography is a static background field that does not change over time, the processed water depth data is treated as a constant value channel, with the following structure: See also Figure 3 , Figure 3 This is a schematic diagram illustrating the temporal replication process of seabed topographic data provided in an embodiment of the present invention. For example... Figure 3 As shown in the figure, multiple The grid, with its edge-trimmed depth channel, is replicated at every time step to ensure that terrain information matches meteorological data at all points in time.

[0040] Step S203: Overlay the spatiotemporally aligned climate reanalysis data and seabed topography data by channel dimension to obtain the spatiotemporal data cube.

[0041] Specifically, the dynamic meteorological and oceanographic data for each time step was originally... The structure contains 4 channels, and then a static water depth grid is formed. Overlaying the meteorological and oceanographic grids at each time step along the channel dimension yields the grid structure for each time step. (4 meteorological channels + 1 topographic channel), such as Figure 4 As shown, all time steps When the grids are stacked, a complete spatiotemporal data cube is formed.

[0042] Step S3: Based on the spatial location information, the spatiotemporal data cube, and the remaining water level data, a hybrid neural network model is used to calculate the remaining water level and obtain the calculated remaining water level sequence of the target virtual tide gauge station; wherein, the hybrid neural network model includes a convolutional neural network branch, a long short-term memory neural network branch, a location encoding branch, a feature fusion layer, and a decoding and calculation layer.

[0043] In an optional embodiment, step S3 includes steps S301 to S304: Step S301: Based on the spatial location information, the spatiotemporal data cube, and the residual water level data, a hybrid neural network model is used to extract features to obtain the temporal features, spatial features, and location features of each reference tide gauge station, as well as the spatial features and location features of the target virtual tide gauge station. Step S302: Calculate the attention weights based on the temporal, spatial, and locational characteristics of each reference tide gauge station and the spatial and locational characteristics of the target virtual tide gauge station; Step S303: Use the attention weights to fuse the temporal, spatial, and locational features of each reference tide gauge station to obtain fused features; Step S304: Decode the fused features and output the estimated residual water level sequence of the target virtual tide gauge station.

[0044] In one optional embodiment, step S301 includes: The residual water level data is processed by a branch of a long short-term memory neural network to extract the time characteristics of each reference tide gauge station. The spatiotemporal data cube is processed by a convolutional neural network branch to extract the spatial features of each reference tide gauge station and the target virtual tide gauge station; The spatial location information is processed by the location encoding branch to extract the location features of each reference tide gauge station and the target virtual tide gauge station.

[0045] In one optional embodiment, step S302 includes: The temporal, spatial, and locational characteristics of each reference tide gauge station are pieced together to form the comprehensive characteristics of each reference tide gauge station. The spatial and locational features of the target virtual tide gauge station are spliced ​​together to form the comprehensive features of the target virtual tide gauge station; The similarity between the comprehensive characteristics of the target virtual tide gauge station and the comprehensive characteristics of each reference tide gauge station is calculated to obtain the attention weight.

[0046] In one alternative embodiment, the hybrid neural network model is trained through the following steps: Step 1: Obtain historical multi-source data from multiple reference tide gauge stations, as well as the spatial location information of each reference tide gauge station and the target virtual tide gauge station, and construct training samples; wherein, the historical multi-source data includes historical measured total water level data, historical astronomical tide forecast data, historical climate reanalysis data, and pre-stored seabed topography data of each reference tide gauge station; It should be noted that the core logic of data preprocessing in real-time residual water level estimation and model training is completely consistent. Both require tidal separation of water level data to obtain residual water level data, spatiotemporal alignment and outlier detection processing of residual water level data, meteorological and oceanographic reanalysis data and seabed topography data, normalization of multi-source data after alignment and purification to eliminate differences in dimensions and numerical scales, and superposition of spatiotemporally aligned meteorological and oceanographic reanalysis data and adapted seabed topography data in the channel dimension to construct a spatiotemporal data cube with unified structure and dimensions. The pruning and resampling processing standards of static seabed topography data are also completely consistent in both stages. Only the normalization parameters statistically saved in the model training stage are directly reused in the real-time estimation stage to achieve efficient real-time data processing.

[0047] Specifically, the training samples are constructed as follows: Using a sliding time window as the core approach, each training sample is extracted from historical data. For each training sample (time step t, unit is 5 minutes), the input includes: The residual water level data of each reference tide gauge station A, B, C... over a period of time ([1, time length, number of reference tide gauge stations]), for example, [1, 288, N] represents the water level data of N actual tide gauge stations every 5 minutes over the past 24 hours; The location information of reference tide gauge stations A, B, C... is ([1, number of reference tide gauge stations, 2]), where 2 represents the latitude and longitude of the reference tide gauge station; Location information of the target virtual tide gauge station X ([1, 2]), where 2 represents the latitude and longitude of the target virtual tide gauge station; Reference tide gauge station and target virtual tide gauge station area meteorological and topographic data ([time length, grid height, grid width, data channel]), for example, [288, P, Q, 5] represents the sequence of the past 24 hours at 5-minute intervals, covering P×Q grid points, and containing 5 channels (wind speed, air pressure, wave height, water temperature, water depth); The output of each sample is the water level data of the target virtual tide gauge station X over a period of time ([1, time step, 1]). For example, [1, 288, 1] represents the water level data of the target virtual tide gauge station X every 5 minutes over the past 24 hours.

[0048] See Figure 5 , Figure 5 This is a schematic diagram of a hybrid neural network model architecture provided in an embodiment of the present invention. Figure 5 As shown, the standard training sample inputs during the training phase of the hybrid neural network model include: Reference tide gauge area meteorological topography: [B, 288, N, P, Q, 5] (time × station × grid height × grid width × number of channels); where B is Batch_size.

[0049] Reference tide gauge data: [B, 288, N] (24-hour historical data, 5-minute intervals, N reference tide gauges); Reference tide gauge location: [B, N, 2] (latitude and longitude coordinates); Location of the target virtual tide gauge station: [B, 2] (latitude and longitude coordinates); Meteorological topography of the target virtual tide gauge area: [B, 288, P, Q, 5] (time × grid height × grid width × number of channels).

[0050] Step 2: Input the training samples into a network structure containing the convolutional neural network branch, the long short-term memory neural network branch, and the position encoding branch, and calculate the predicted water level using forward propagation; Among them, such as Figure 5 As shown, the convolutional neural network branch consists of two convolutional blocks and one global average pooling layer connected in series. The combined structure of the convolutional blocks is a three-dimensional convolutional layer (Conv3d) and a ReLU activation function.

[0051] The specific design of each module is as follows: Convolution block 1: Input: Meteorological topography of the reference tide gauge area [B, 288, N, P, Q, 5] and meteorological topography of the target virtual tide gauge area [B, 288, P, Q, 5]; Operation: A three-dimensional convolutional layer is used, configured with 5 input channels, 32 output channels, a kernel size of 3×3×3, and a padding value of 1 to ensure that the spatiotemporal dimension remains unchanged after convolution; a ReLU activation function is then applied to introduce nonlinearity.

[0052] Output: Spatial characteristics of the reference tide gauge station (B, 288, N, P, Q, 32) and spatial characteristics of the target virtual tide gauge station (B, 288, P, Q, 32).

[0053] Function: To preliminarily extract local correlation features of meteorological and topographic data in the "time-space-channel" dimension.

[0054] Convolutional block 2: Input: The output of convolution block 1.

[0055] Operation: Continue using a 3D convolutional layer with 32 input channels, 64 output channels, a kernel size of 3×3×3, and padding of 1, followed by a ReLU activation function.

[0056] Output: Spatial characteristics of the reference tide gauge station (B, 288, N, P, Q, 64) and the target virtual tide gauge station (B, 288, P, Q, 64).

[0057] Function: To further extract deeper spatial features and enhance feature expression capabilities.

[0058] Global average pooling (averaging P×Q values ​​from each station): Input: The output of convolution block 2.

[0059] Operation: Perform global average pooling on the channel dimension for the spatial grid (P×Q) of each station to compress the spatial features of each station into a fixed-length feature vector.

[0060] Output: Spatial characteristics of the reference tide gauge station (B, 288, N, 64), spatial characteristics of the target virtual tide gauge station (B, 288, 64).

[0061] Function: To convert spatial grid features into site-level features, which facilitates subsequent fusion with time series features.

[0062] The long short-term memory neural network branch includes an input projection layer and a 3-layer LSTM temporal encoder, specifically designed as follows: Input projection layer: Input: Reference tide gauge residual water level data (B, 288, N, 1).

[0063] Operation: A fully connected layer (Linear) is used to project the water level value of each station onto a 32-dimensional feature space, followed by a ReLU activation function.

[0064] Output: Projected features (B, 288, N, 32).

[0065] Function: Maps the original water level sequence to a high-dimensional feature space, enhancing the feature representation capability.

[0066] 3-layer LSTM timing encoder: Input: Projected features (B, 288, N, 32).

[0067] Operation: An LSTM network configured with an input dimension of 32, a hidden state dimension of 64, 3 layers, and a dropout rate of 0.2 is used for stepwise encoding according to the time step.

[0068] Output: The final hidden state (3, B, N, 64) is taken and copied and expanded along the time dimension to obtain the temporal feature (B, 288, N, 64).

[0069] Function: To capture the long-term dependence of the residual water level at a reference tide gauge station over time.

[0070] Among them, such as Figure 5 As shown, the location encoding branch includes a shared fully connected unit and a replication extension unit, which respectively handle the location of the reference tide gauge station / target virtual tide gauge station. The specific design is as follows: Shared fully connected unit: Input: Reference tide gauge location (B, N, 2) and target virtual tide gauge location (B, 2).

[0071] Operation: A fully connected layer is used to map 2D coordinates into 32D feature vectors, followed by a ReLU activation function.

[0072] Output: Location features of the reference tide gauge station (B, N, 32), location features of the target virtual tide gauge station (B, 32).

[0073] Function: To encode geographic location information into learnable feature representations.

[0074] Copy extension unit: Input: Location features of the reference tide gauge station (B, N, 32), location features of the target virtual tide gauge station (B, 32).

[0075] Operation: Copy the location features of the reference tide gauge station 288 times along the time dimension to obtain (B, 288, N, 32); copy the location features of the target virtual tide gauge station in the same way to obtain (B, 288, 32).

[0076] Function: To align location features with other features in the time dimension.

[0077] Among them, such as Figure 5As shown, the feature fusion layer aligns the spatial features output by the Convolutional Neural Network (CNN) with the temporal features output by the Long Short-Term Memory (LSTM) neural network in the time dimension, incorporating position-encoded features to enhance spatial generalization ability. Furthermore, it introduces a spatiotemporal similarity attention mechanism to weightedly fuse features from multiple reference tide gauge stations into a single feature set. The feature fusion layer mainly includes feature concatenation and a spatiotemporal similarity attention mechanism, specifically designed as follows: ①Feature splicing: Reference tide gauge station feature concatenation: The spatial features (B, 288, N, 64) output by the CNN branch, the temporal features (B, 288, N, 64) output by the LSTM branch, and the positional features (B, 288, N, 32) output by the positional encoding branch are concatenated in the last dimension to obtain the comprehensive feature (B, 288, N, 160).

[0078] Feature concatenation of the target virtual tide gauge station: The spatial features (B,288, 64) and the location features (B, 288, 32) of the target virtual tide gauge station output by the CNN branch are concatenated to obtain the comprehensive features (B, 288, 96).

[0079] Function: To integrate features from multiple sources and provide a foundation for subsequent fusion.

[0080] ② Spatiotemporal similarity attention mechanism: Feature mapping: The comprehensive features (B, 288, N, 160) of the reference tide gauge station are mapped to the same feature dimension as the target virtual tide gauge station through a fully connected layer, resulting in (B, 288, N, 96).

[0081] Similarity calculation: Calculate the dot product similarity between the features of the target virtual tide gauge station and the features of each reference tide gauge station to obtain the similarity matrix (B, 288, 1, N).

[0082] Characteristics of the target virtual tide gauge station: (B, 288, 96) → (B,288,1,96).

[0083] Reference tide gauge characteristics: (B,288,N,96) → transpose the last two dimensions → (B,288,96,N).

[0084] Similarity = [B,288,(1,96)·(96,N)].

[0085] Output: (B,288,1,N).

[0086] Normalization to "attention weights": (B,288,1,N) → Softmax normalization → (B,288,1,N1) (the sum of vector N is 1).

[0087] Weight normalization: The similarity matrix is ​​normalized using Softmax in the dimension of the reference tide gauge to obtain the attention weights (B,288,1,N), with the weights summing to 1.

[0088] Weighted fusion: The comprehensive features of the reference tide gauge station are weighted and summed using attention weights to obtain the fused features of the reference tide gauge station (B,288,160).

[0089] Input: Reference tide gauge station comprehensive features (B,288,N,160), spatiotemporal similarity attention weights (B,288,1,N1).

[0090] Weighted fusion: Weighted feature_i = Σ(weight_i × reference tide gauge feature_i).

[0091] Output: Weighted fusion feature of reference tide gauge station (B,288,160).

[0092] Function: Based on the similarity between the target virtual tide gauge station and each reference tide gauge station in the spatiotemporal environment, the contribution of the reference tide gauge station is dynamically adjusted to enhance the physical interpretability and generalization ability of the model.

[0093] Final feature stitching: Input: Comprehensive features of the target virtual tide gauge station (B,288,96) and weighted fusion features of the reference tide gauge station (B,288,160).

[0094] Operation: Concatenate the last dimension to obtain the final fused feature (B,288,256).

[0095] Function: It combines the characteristics of the target virtual tide gauge station with the weighted characteristics of the reference tide gauge station to provide comprehensive input for the decoder.

[0096] Among them, such as Figure 5 As shown, the decoding and estimation layer includes an LSTM decoder and an output projection layer. It primarily uses the LSTM decoder to decode the fused features in the time dimension, ultimately outputting the residual water level sequence of the target virtual tide gauge station. The specific design is as follows: LSTM decoder: Input: Final fused features (B,288,256).

[0097] Operation: An LSTM network configured with 256 input dimensions, 256 hidden state dimensions, 2 layers, and a Dropout rate of 0.3 is decoded step by step according to the time step.

[0098] Output: Decoded feature sequence (B,288,256).

[0099] Function: To capture the evolution of fusion features over time and provide a temporal context for residual water level estimation.

[0100] Output projection layer: Input: Decoder output (B,288,256).

[0101] Operation: A fully connected layer is used to project the 256-dimensional features into 1 dimension, corresponding to the residual water level value of the target virtual tide gauge station every 5 minutes.

[0102] Output: Target virtual tide gauge residual water level estimation sequence (B,288,1).

[0103] Function: To map high-dimensional features to specific water level values ​​and complete the task of estimating the remaining water level.

[0104] In summary, step 2 specifically involves inputting a small batch of training data (e.g., containing 32 of the aforementioned samples) into the hybrid neural network model. The data sequentially passes through the input layer, convolutional neural network branch, long short-term memory neural network branch, position encoding branch, feature fusion layer, LSTM decoder, and output layer. For each sample in this batch of data, the model will output a calculated 288-step residual water level sequence Y_pred (shape: (B,288,1)).

[0105] Step 3: Calculate the loss function value between the predicted water level and the actual water level label; For example, a loss function is first used to quantify the error between the model-predicted value Y_pred and the true value Y_true. For regression tasks such as the prediction of the remaining water level, the mean squared error (MSE) is selected as the loss function. Its core characteristic is to apply a higher penalty weight to larger errors. By amplifying the impact of extreme deviations, the model is forced to prioritize correcting samples with significant prediction errors, thereby improving the overall prediction accuracy.

[0106] Step 4: Based on the loss function value, optimize the parameters of the network structure using the backpropagation algorithm to obtain the trained hybrid neural network model.

[0107] It's important to note that the total loss calculated by the loss function is a scalar value. Based on the chain rule, we iterate backwards from the output of the loss function throughout the entire network, calculating the gradient of the loss value relative to each trainable parameter in the model (including weights W and biases b). The core of this process is to identify "which parameters caused the error" and quantify the contribution of each parameter to the total loss, providing a precise basis for subsequent parameter adjustments.

[0108] Based on the gradient information obtained from backpropagation, all trainable parameters of the model are adjusted by an optimizer. In the residual water level estimation model, the Adam optimizer is preferred. This optimizer will adaptively update the weights W and biases b by combining the gradient magnitude and the preset learning rate. The learning rate controls the step size of parameter updates, and the gradient direction determines the direction of parameter adjustment. Ultimately, iterative optimization of model parameters is achieved, gradually reducing the loss value and improving the estimation accuracy.

[0109] To enable those skilled in the art to better understand the specific implementation process of the present invention, the following describes the specific process of the virtual tide gauge station residual water level estimation method.

[0110] First, data acquisition and aggregation were performed, pulling the latest data from various data sources. Tide level data was obtained in real-time from distributed physical reference tide gauge stations via data links, providing 24-hour, 5-minute interval water level data for a specified date. Meteorological and oceanographic data were acquired synchronously with the water level data, including climate reanalysis data, seafloor topography data, and the location information of all tide gauge stations covering the study area. Next, data preprocessing and spatiotemporal alignment were performed. The real-time acquired raw data, which may have different spatiotemporal resolutions, were uniformly interpolated to a standard 5-minute time interval and a 1'×1' spatial grid. Dynamic meteorological and oceanographic forecast data and static water depth and topography data were overlaid in the channel dimension to form a standard spatiotemporal data cube (288, P, Q, 5) for the next 24 hours. The mean and standard deviation saved during model training were used to normalize all input data. Then, the model input tensor was constructed, assembling complete data samples for each target virtual station to be extrapolated, including the reference tide gauge water level sequence (1, 288, N), the reference tide gauge location (1, N, 2), and the target virtual station location (1, 2). The system consists of a target virtual tide gauge station (1,288,P,Q,5) and a reference tide gauge station (1,288,N,P,Q,5). The system then performs forward propagation, loading the trained model with fixed parameters into memory or a GPU. The constructed input tensor is fed into the model for inference calculations. The model directly outputs a normalized target virtual tide gauge station water level sequence (1,288,1) every 5 minutes for a specified date over 24 hours. In the post-processing stage, the normalized water level values ​​output by the model are denormalized using the water level mean and standard deviation saved during training, converting them into real water level values ​​in meters. The denormalized 288 water level data points are then combined with their corresponding timestamps and encapsulated into a standard data format. Finally, the system publishes and visualizes the results, publishing them in real-time to downstream users or systems via API interfaces, message queues, or database writes. Simultaneously, the location of the virtual tide gauge station and the real-time changing water level curves are visualized on electronic nautical charts, providing users with intuitive water level information.

[0111] In summary, the virtual tide gauge residual water level estimation method provided by this invention introduces a hybrid neural network model comprising convolutional neural network branches, long short-term memory network branches, and location encoding branches. This model collaboratively processes residual water level sequences from multiple reference tide gauges, meteorological, oceanographic, and seabed topographic grid data of the covered area, and spatial location information, effectively overcoming the shortcomings of existing technologies and producing significant beneficial effects. Dedicated to extracting spatial structural features driving water level changes from the grid data using convolutional neural network branches can compensate for the insufficient spatial modeling capabilities of pure long short-term memory network models, enabling the model to more comprehensively understand the spatial synergistic effects of factors such as wind, pressure, waves, and topography. The long short-term memory network branch continues to leverage its advantage in capturing the temporal evolution of water levels. The location encoding branch explicitly embeds geographic spatial relationships into the model. By fusing the parallel extracted spatiotemporal features with location features and utilizing a spatiotemporal data cube constructed from multi-source data, the model achieves deep learning of the "space-time" joint mapping relationship, improving generalization ability and estimation accuracy in unknown areas with sparse or no data from physical tide gauges. Meanwhile, by integrating static terrain and dynamic meteorological fields in real time, the over-reliance on the data quality of a single site can be reduced, enhancing the system's robustness. By capturing spatially abrupt change signals in real time, the real-time performance of the prediction of extreme water level events caused by typhoons, storm surges, and other events can be improved.

[0112] Based on the above method items, the present invention provides corresponding system items embodiments.

[0113] See Figure 6 , Figure 6 This is a structural block diagram of a virtual tide gauge station residual water level estimation system provided in an embodiment of the present invention. The virtual tide gauge station residual water level estimation system includes: The data acquisition module 21 is used to acquire real-time multi-source data from multiple reference tide gauge stations and spatial location information of each reference tide gauge station and the target virtual tide gauge station; wherein, the real-time multi-source data includes real-time measured total water level data of each reference tide gauge station, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data; Data preprocessing module 22 is used to preprocess the multi-source data to obtain a spatiotemporal data cube and residual water level data of each reference tide gauge station; The residual water level estimation module 23 is used to estimate the residual water level based on the spatial location information, the spatiotemporal data cube, and the residual water level data using a hybrid neural network model, to obtain the estimated residual water level sequence of the target virtual tide gauge station; wherein, the hybrid neural network model includes a convolutional neural network branch, a long short-term memory neural network branch and a position encoding branch, a feature fusion layer and a decoding estimation layer.

[0114] In one optional embodiment, the data preprocessing module 22 includes: The tidal separation unit is used to calculate the initial residual water level data of each reference tide gauge station based on the real-time measured total water level data and the real-time astronomical tide forecast data. The spatiotemporal alignment unit is used to spatiotemporally align the initial residual water level data, the real-time climate reanalysis data, and the pre-stored seabed topography data to obtain the residual water level data of each reference tide gauge station, as well as the spatiotemporally aligned climate reanalysis data and seabed topography data. The spatiotemporal data cube construction unit is used to overlay spatiotemporally aligned climate reanalysis data and seabed topography data by channel dimension to obtain the spatiotemporal data cube.

[0115] In one alternative embodiment, the spatiotemporal alignment unit is configured to: The initial residual water level data of each reference tide gauge station will be uniformly resampled to a preset time interval; The real-time climate reanalysis data is resampled to the preset time interval and interpolated to a preset spatial grid. The seabed topography data is cropped and sampled into the preset spatial grid.

[0116] In one optional embodiment, the residual water level calculation module 23 includes: The feature extraction unit is used to extract features based on the spatial location information, the spatiotemporal data cube, and the residual water level data using a hybrid neural network model, to obtain the temporal features, spatial features, and location features of each reference tide gauge station, as well as the spatial features and location features of the target virtual tide gauge station. The weight calculation unit is used to calculate the attention weight based on the time characteristics, spatial characteristics, and location characteristics of each reference tide gauge station and the spatial characteristics and location characteristics of the target virtual tide gauge station. The feature fusion unit is used to fuse the temporal features, spatial features, and location features of each reference tide gauge station using the attention weights to obtain fused features. The residual water level estimation unit is used to decode the fused features and output the estimated residual water level sequence of the target virtual tide gauge station.

[0117] In one alternative embodiment, the feature extraction unit is configured to: The residual water level data is processed by a branch of a long short-term memory neural network to extract the time characteristics of each reference tide gauge station. The spatiotemporal data cube is processed by a convolutional neural network branch to extract the spatial features of each reference tide gauge station and the target virtual tide gauge station; The spatial location information is processed by the location encoding branch to extract the location features of each reference tide gauge station and the target virtual tide gauge station.

[0118] In one alternative embodiment, the weight calculation unit is configured to: The temporal, spatial, and locational characteristics of each reference tide gauge station are pieced together to form the comprehensive characteristics of each reference tide gauge station. The spatial and locational features of the target virtual tide gauge station are spliced ​​together to form the comprehensive features of the target virtual tide gauge station; The similarity between the comprehensive characteristics of the target virtual tide gauge station and the comprehensive characteristics of each reference tide gauge station is calculated to obtain the attention weight.

[0119] It should be noted that the virtual tide gauge station residual water level estimation system provided in this embodiment of the invention is used to execute all the process steps of the virtual tide gauge station residual water level estimation method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0120] This invention also provides a terminal device, such as... Figure 7 The diagram shown is a structural block diagram of a preferred embodiment of a terminal device provided by the present invention. The terminal device includes a processor 31, a memory 32, and a computer program stored in the memory 32 and configured to be executed by the processor 31. When the processor 31 executes the computer program, it implements the virtual tide gauge residual water level estimation method as described in any of the above embodiments.

[0121] In addition, embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the virtual tide gauge residual water level estimation method as described in any of the above embodiments.

[0122] When the processor 31 executes the computer program, it implements the steps in the above-described embodiment of the virtual tide gauge residual water level estimation method, for example... Figure 1 The method for calculating the residual water level of a virtual tide gauge station, as shown, includes all the steps. Alternatively, when the processor 31 executes the computer program, it implements the functions of each module in the above-described embodiment of the virtual tide gauge station residual water level calculation system, for example... Figure 6 The functions of each module in the virtual tide gauge station residual water level estimation system are shown.

[0123] Preferably, the computer program can be divided into one or more modules / units, which are stored in the memory 32 and executed by the processor 31 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0124] The processor 31 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 31 can be any conventional processor. The processor 31 is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0125] The memory 32 mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., and the data storage area can store related data, etc. In addition, the memory 32 can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, and a flash card, etc., or the memory 32 can also be other volatile solid-state storage devices.

[0126] It should be noted that the aforementioned terminal devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 7 The structural block diagram shown is merely a structural example of the terminal device described above and does not constitute a limitation on the structure of the terminal device. The terminal device may include more or fewer components than shown, or combine certain components, or use different components.

[0127] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for estimating the residual water level at a virtual tide gauge station, characterized in that, include: Acquire real-time multi-source data from multiple reference tide gauge stations, as well as spatial location information of each reference tide gauge station and the target virtual tide gauge station; wherein, the real-time multi-source data includes real-time measured total water level data, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data from each reference tide gauge station; The multi-source data is preprocessed to obtain a spatiotemporal data cube and residual water level data of each reference tide gauge station; Based on the spatial location information, the spatiotemporal data cube, and the remaining water level data, a hybrid neural network model is used to calculate the remaining water level, thereby obtaining the calculated remaining water level sequence of the target virtual tide gauge station; wherein, the hybrid neural network model includes a convolutional neural network branch, a long short-term memory neural network branch and a position encoding branch, a feature fusion layer and a decoding calculation layer.

2. The method for calculating the residual water level of a virtual tide gauge station as described in claim 1, characterized in that, The method of calculating the residual water level based on the spatial location information, the spatiotemporal data cube, and the residual water level data using a hybrid neural network model to obtain the calculated residual water level sequence of the target virtual tide gauge station includes: Based on the spatial location information, the spatiotemporal data cube, and the residual water level data, a hybrid neural network model is used to extract features, thereby obtaining the temporal features, spatial features, and location features of each reference tide gauge station, as well as the spatial features and location features of the target virtual tide gauge station. Attention weights are calculated based on the temporal, spatial, and locational characteristics of each reference tide gauge station and the spatial and locational characteristics of the target virtual tide gauge station. The attention weights are used to fuse the temporal, spatial, and locational features of each reference tide gauge station to obtain fused features. The fusion features are decoded to output the estimated residual water level sequence of the target virtual tide gauge station.

3. The method for calculating the residual water level of a virtual tide gauge station as described in claim 2, characterized in that, Based on the spatial location information, the spatiotemporal data cube, and the residual water level data, a hybrid neural network model is used for feature extraction to obtain the temporal, spatial, and location features of each reference tide gauge station and the spatial and location features of the target virtual tide gauge station, including: The residual water level data is processed by a branch of a long short-term memory neural network to extract the time characteristics of each reference tide gauge station. The spatiotemporal data cube is processed by a convolutional neural network branch to extract the spatial features of each reference tide gauge station and the target virtual tide gauge station; The spatial location information is processed by the location encoding branch to extract the location features of each reference tide gauge station and the target virtual tide gauge station.

4. The method for calculating the residual water level of a virtual tide gauge station as described in claim 2, characterized in that, The calculation of attention weights based on the temporal, spatial, and locational characteristics of each reference tide gauge station and the spatial and locational characteristics of the target virtual tide gauge station includes: The temporal, spatial, and locational characteristics of each reference tide gauge station are pieced together to form the comprehensive characteristics of each reference tide gauge station. The spatial and locational features of the target virtual tide gauge station are spliced ​​together to form the comprehensive features of the target virtual tide gauge station; The similarity between the comprehensive characteristics of the target virtual tide gauge station and the comprehensive characteristics of each reference tide gauge station is calculated to obtain the attention weight.

5. The method for calculating the residual water level of a virtual tide gauge station as described in claim 1, characterized in that, The hybrid neural network model is trained through the following steps: Acquire historical multi-source data from multiple reference tide gauge stations, as well as spatial location information of each reference tide gauge station and the target virtual tide gauge station, to construct training samples; The training samples are input into a network structure containing the convolutional neural network branch, the long short-term memory neural network branch, and the position encoding branch, and the remaining water level is predicted by forward propagation. Calculate the loss function value between the predicted residual water level and the actual residual water level label; Based on the loss function value, the parameters of the network structure are optimized using the backpropagation algorithm to obtain a trained hybrid neural network model.

6. The method for calculating the residual water level of a virtual tide gauge station as described in claim 1, characterized in that, The preprocessing of the multi-source data to obtain a spatiotemporal data cube and residual water level data from each reference tide gauge station includes: Based on the real-time measured total water level data and the real-time astronomical tide forecast data, the initial residual water level data of each reference tide gauge station is calculated. The initial residual water level data, the real-time climate reanalysis data, and the pre-stored seabed topography data are spatiotemporally aligned to obtain the residual water level data of each reference tide gauge station, as well as the spatiotemporally aligned climate reanalysis data and seabed topography data. The spatiotemporal data cube is obtained by overlaying the spatiotemporal aligned climate reanalysis data and seabed topography data along the channel dimension.

7. The method for calculating the residual water level of a virtual tide gauge station as described in claim 6, characterized in that, The step of spatiotemporally aligning the initial residual water level data, the real-time climate reanalysis data, and the pre-stored seabed topography data includes: The initial residual water level data of each reference tide gauge station will be uniformly resampled to a preset time interval; The real-time climate reanalysis data is resampled to the preset time interval and interpolated to a preset spatial grid. The seabed topography data is cropped and sampled into the preset spatial grid.

8. A virtual tide gauge residual water level calculation system, characterized in that, include: The data acquisition module is used to acquire real-time multi-source data from multiple reference tide gauge stations, as well as the spatial location information of each reference tide gauge station and the target virtual tide gauge station; wherein, the real-time multi-source data includes real-time measured total water level data of each reference tide gauge station, real-time astronomical tide forecast data, real-time climate reanalysis data, and pre-stored seabed topography data; The data preprocessing module is used to preprocess the multi-source data to obtain a spatiotemporal data cube and residual water level data of each reference tide gauge station. The residual water level estimation module is used to estimate the residual water level based on the spatial location information, the spatiotemporal data cube, and the residual water level data, using a hybrid neural network model to obtain the estimated residual water level sequence of the target virtual tide gauge station; wherein, the hybrid neural network model includes a convolutional neural network branch, a long short-term memory neural network branch and a position encoding branch, a feature fusion layer and a decoding estimation layer.

9. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the virtual tide gauge residual water level estimation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the virtual tide gauge residual water level estimation method as described in any one of claims 1 to 7.

Images