Adaptive Correction Hydrodynamic Numerical Prediction Method Based on Observational Data

Abstract

The adaptive correction hydrodynamic numerical prediction method, which integrates observational data, belongs to the field of marine hydrodynamic numerical (such as tidal level and current velocity) prediction. To address the issue of improving the accuracy of hydrodynamic model predictions of tidal levels and currents, the key points are: the hydrodynamic mathematical model outputs prior hydrodynamic simulation results for the prediction point based on the open boundary conditions output by the deep learning network model; it calculates the similarity and weighted similarity between any prediction point in the selected area and its nearest neighbor observation point on any selected feature in the environmental field data of the prediction point in the selected area; and based on the prior state, weighted similarity, measured hydrodynamic results from the nearest neighbor observation point of any prediction point, and the Kalman gain matrix, it calculates the posterior state of any prediction point, which is used in the posterior hydrodynamic simulation results for that prediction point.

Description

Technical Field

[0001] This invention belongs to the field of marine hydrodynamic numerical prediction, such as tidal level prediction. It is an adaptive correction hydrodynamic numerical prediction method that integrates observation data and its implementation. Background Technology

[0002] The Age of Exploration ushered in the modern era of global civilization, with coastal cities undertaking numerous marine resource development projects, including port construction, coastal protection, and aquaculture. The construction and operation of these projects require a thorough understanding of the surrounding sea conditions, including ocean dynamics, sediment, and water quality. Research methods primarily include field observation, physical model testing, and numerical simulation, with computer simulation combined with long-term observational data comparison becoming the mainstream approach. During this process, many excellent marine numerical models have been developed and improved, such as the international models ROMS, POM, FVCOM, Delft3D, ECOM, and MIKEZERO, and the Chinese models GPUOM, HydroInfo, and MaCOM1.0.

[0003] The errors in the numerical model mainly stem from two aspects. First, there are uncertainties in the initial and boundary conditions (UIBC). The initial field and boundary constraints of the model are insufficient to accurately characterize the dynamic processes in complex marine areas, leading to deviations in the simulation of elements such as tidal levels, current movements, and dynamic changes in topography within the region. Second, there is insufficient description of key physical processes (IRKPP). Existing theories still lack complete mathematical representations of fluid turbulence, sediment transport and deposition processes, and interactions of water components, which limits the simulation accuracy of the model. Summary of the Invention

[0004] To address the issue of improving the accuracy of hydrodynamic models in predicting hydrodynamic data such as tide levels and currents based on environmental data, this application presents an adaptive correction method for numerical hydrodynamic prediction that integrates observational data in some embodiments, including...

[0005] The first deep learning network model outputs hydrodynamic simulation results for all marked points on the open boundary line of the selected region, based on the astronomical tide and environmental field data of the marked points in the selected region, and is used for open boundary conditions.

[0006] The hydrodynamic mathematical model outputs prior hydrodynamic simulation results for the selected area based on the open boundary conditions and the environmental field data of the selected area to be predicted.

[0007] The hydrodynamic mathematical model calculates the similarity of any point to be predicted in the selected area with its nearest neighbor observation point on any selected feature in the prior hydrodynamic simulation results data of the point to be predicted in the selected area and the environmental field data of the point to be predicted in the selected area.

[0008] The hydrodynamic mathematical model calculates the weighted similarity of any point to be predicted with its nearest neighbor observation point in the prior hydrodynamic simulation results data of the point to be predicted in the selected area, as well as the environmental field data of the point to be predicted in the selected area, based on the similarity on any selected feature and the weight of the selected feature.

[0009] The hydrodynamic mathematical model calculates the Kalman gain matrix based on the prior state and the propagation prediction error covariance matrix;

[0010] The hydrodynamic mathematical model calculates the posterior state of any point to be predicted based on the prior state, the weighted similarity between any point to be predicted and its nearest neighbor observation point, the measured hydrodynamic results data of the nearest neighbor observation point of any point to be predicted, and the Kalman gain matrix, and uses the posterior hydrodynamic simulation results data of any point to be predicted.

[0011] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the similarity between any point to be predicted in a selected area and its nearest neighbor observation point on any selected feature in the prior hydrodynamic simulation results data of the point to be predicted in the selected area and the environmental field data of the point to be predicted in the selected area is calculated, including...

[0012] Calculate the coefficient of determination of any point to be predicted in the selected region and its nearest neighbor observation points on the selection characteristics of the prior hydrodynamic simulation results of the point to be predicted in the selected region, as well as the environmental field data of the point to be predicted in the selected region. As shown in the following formula:

[0013]

[0014] In the formula, The coefficient of determination is represented by the coefficient of determination. Indicates the observation point number The value of each selected feature. Indicates the point to be predicted. The value of each selected feature. Indicates the number of features selected. Select the mean of the features for the observation points;

[0015] Will The truncation is mapped to [0,1]. The similarity between any point to be predicted in the selected region and its nearest neighbor observation point on any selected feature in the prior hydrodynamic simulation results data of the point to be predicted in the selected region and the environmental field data of the point to be predicted in the selected region is calculated. As shown in the following formula:

[0016]

[0017] In the formula, This represents the similarity on the k-th selected feature;

[0018] The weighted similarity between any point to be predicted and its nearest neighbor observation point in the prior hydrodynamic simulation results data of the point to be predicted in the selected area, and the selected features in the environmental field data of the point to be predicted in the selected area, is calculated by the following formula:

[0019]

[0020] In the formula, S represents the weighted similarity, and K represents the number of selected features. This represents the weight of the selected feature.

[0021] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the hydrodynamic mathematical model calculates the prior states in the Kalman gain matrix based on the prior states and the propagation prediction error covariance matrix. As shown in the following formula:

[0022]

[0023] In the formula, the subscript Indicates the first Time step, Represents prior values, Representing the hydrodynamic mathematical model, Represents the posterior value;

[0024] Among them, the propagation prediction error covariance matrix As shown in the following formula:

[0025]

[0026] In the formula, This represents the propagation forecast error covariance matrix. Indicates the number of members in the set. This represents the value of the i-th member of the forecast set. The mean of the members of the forecast set;

[0027] Wherein, the Kalman gain matrix As shown in the following formula:

[0028]

[0029] In the formula, Represents the observation matrix. This represents the variance matrix of the observation error.

[0030] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the hydrodynamic mathematical model calculates the posterior state of any point to be predicted based on the prior state, the weighted similarity between any point to be predicted and its nearest neighbor observation points, the measured hydrodynamic results data of the nearest neighbor observation points of any point to be predicted, and the Kalman gain matrix. This posterior state is used for the posterior hydrodynamic simulation results data of any point to be predicted, as shown in the following formula:

[0031]

[0032] In the formula, Indicates the posterior state. Indicates weighted similarity. Represents the Kalman gain matrix. This represents the measured hydrodynamic results data. Represents the observation matrix. This represents the prior state.

[0033] According to the adaptive correction hydrodynamic numerical prediction method for fusing observation data in some embodiments of this application, the first deep learning network model is obtained as follows: the first deep learning network model is constructed and trained, the input of which includes astronomical tide and environmental field data of marked points on the open boundary line of the selected region, and the output includes measured hydrodynamic result data of marked points on the open boundary line of the selected region.

[0034] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the first deep learning network model includes a temporal convolutional network, wherein the model hyperparameters include a historical time step of 24h and hidden layers [256,256,256,256,128,128,128,128].

[0035] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the selection of features and their weights is obtained based on the following method:

[0036] Construct and train a second deep learning network model. The input of its training dataset includes prior hydrodynamic simulation results data of observation points in the selected area, as well as environmental field data of observation points in the selected area. The output includes measured hydrodynamic results data of observation points in the selected area.

[0037] The second deep learning network model outputs the predicted hydrodynamic simulation results for the selected area based on the hydrodynamic simulation results of the selected area to be predicted and the environmental field data of the selected area to be predicted.

[0038] Based on the characteristics of the hydrodynamic simulation results data of the selected area to be predicted points, the environmental field data of the selected area to be predicted points, and the predicted hydrodynamic simulation results data of the selected area to be predicted points, the weight of any feature of the hydrodynamic simulation results data of the selected area to be predicted points and the environmental field data of the selected area to be predicted points is obtained.

[0039] Select the top few features with the largest weights among any given features as selection features, and normalize the weights of the selection features to obtain the weight of any selection feature.

[0040] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the weights of any feature of the hydrodynamic simulation result data of the selected area to be predicted point and the environmental field data of the selected area to be predicted point are obtained through SHAP analysis.

[0041] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the second deep learning network model includes a temporal convolutional network, wherein the model hyperparameters include a historical time step of 24h and hidden layers [256,256,256,256,128,128,128,128].

[0042] According to the adaptive correction hydrodynamic numerical prediction method based on fused observation data in some embodiments of this application, the points to be predicted include grid cells and / or nodes of the selected region to be predicted;

[0043] The hydrodynamic simulation results include tidal level and / or tidal current data;

[0044] The characteristics of the prior hydrodynamic simulation results include the prior hydrodynamic simulation results, and the environmental field data include the 10-meter meridional wind component, the 10-meter zonal wind component, the 2-meter dew point temperature, the 2-meter temperature, the mean sea level pressure, the surface pressure, evaporation, surface runoff, groundwater runoff, and total precipitation.

[0045] Beneficial effects:

[0046] This invention proposes a dual-path fusion enhancement framework: G-UIBC-2C (GPUOM-Uncertainties in Initia and Boundary Conditions-to Current) can reduce initial input errors; secondly, EnKF (Ensemble Kalman Filter) is coupled in GPUOM-HD to realize dynamic fusion of observation data and mechanistic models, and G-EnKF-AL (GPUOM–EnKF-Adaptive Localization) is designed to generate adaptive localization coefficients offline, thereby simultaneously improving forecast accuracy and computational efficiency.

[0047] The adaptive correction method for hydrodynamic numerical prediction based on fused observation data of this invention involves coupling EnKF within the hydrodynamic mathematical model GPUOM-HD. This coupling uses the actual observed tidal and tidal current data from the nearest neighbor observation points to correct the prior tidal and tidal current prediction data output by the hydrodynamic mathematical model based on open boundary conditions and environmental field data. This achieves dynamic fusion of observation data and simulated data from the mechanistic model. This correction is implemented in EnKF using a weighted similarity method considering feature weights, resulting in more accurate posterior values ​​for the corrected tidal and tidal current predictions, thus improving prediction accuracy. Furthermore, the iteration based on EnKF can quickly predict the corrected posterior values ​​across multiple time steps, improving computational efficiency.

[0048] Among them, the G-UIBC-2C model provides high-fidelity initial / boundary conditions. By using astronomical tide and environmental field observation data, as well as the corresponding real tide levels, the model is trained. Based on the astronomical tide and environmental field data, the model can directly, accurately, and quickly predict the real tide levels of all marked points on the open boundary line of the selected area, which is used for the open boundary conditions. This is the role of the first deep learning model of this invention.

[0049] The G-UIBC-2C model provides data assimilation optimization: by using prior tide level and environmental field observation data of selected area observation points, as well as the corresponding measured tide level of the selected area observation points, the model is trained. The model directly, accurately, and quickly predicts the actual tide level of the selected area to be predicted based on the prior tide level and environmental field data. The predicted actual tide level is then analyzed by SHAP to obtain the weights assigned based on the influence of each feature in the prior tide level and environmental field on the tide level and tidal current, thereby determining important features (selected features) and their weights. Based on the feature type and weight, the G-EnKF-AL method supports similarity analysis between spatial points and specific stations to improve assimilation accuracy.

[0050] In this invention, G-UIBC-2C is used to generate high-fidelity initial and boundary conditions; G-EnKF-AL fuses observational data to improve the accuracy of the mechanistic hydrodynamic model. Through systematic evaluation of the coupling effect between the above methods and GPUOM, the following experimental conclusions are obtained:

[0051] (1). Optimization and Application Results of the G-UIBC-2C Model: The hyperparameter scheme P4 (B01 / B04 dataset + 24h historical step size + TCN hidden layer [256,256,256,256,128,128,128,128]) performed best in tidal simulation, with R² values ​​of 95.07% and 92.91% for LT1 and LT2 stations, respectively, representing improvements of 2.45% and 6.32% compared to the traditional MIKE astronomical tide scheme. This improvement stems from the enhanced simulation accuracy under normal meteorological conditions. Further significant improvements in accuracy are expected under extreme scenarios such as storm surges, which our research team will explore in more detail later.

[0052] (2). SHAP Analysis: The global bar chart shows that the SHAP values ​​of meteorological factors (d2m-0h, sp-24h, t2m-24h) reach 0.05, proving that they cannot be ignored in long-term tide level prediction. This verifies the feasibility of providing a reference for parameter calibration of the GPUOM-HD model through SHAP analysis.

[0053] (3) Adaptability of the G-EnKF-AL assimilation strategy: Assimilation at all locations and throughout all time periods significantly improved accuracy, with the MAE of flow velocity in LC1-LC5 decreasing by 77.88%-92.38% and the MAE of flow direction decreasing by 58.72%-86.05%. A partial assimilation strategy (assimilating LC1 / LC3-LC5, validating LC2) still achieved a 25.00% reduction in the MAE of flow velocity in LC2 and a 35.34% reduction in the MAE of flow direction, validating the feasibility of sparse observation. Assimilation at all locations and during partial time periods significantly improved accuracy, with the MAE of flow velocity in LC1-LC5 decreasing by 21.52%-63.96% and the MAE of flow direction decreasing by -143.99%-42.45%.

[0054] The methods proposed in this paper, such as G-UIBC-2C and G-EnKF-AL, are not only applicable to the hydrodynamic simulations shown in the examples, but can also be applied to scenarios involving sediment and water quality to improve simulation accuracy. Furthermore, these methods can be coupled with GPUOM to serve marine digital twin systems, enabling high-fidelity simulations of typical marine phenomena; they can both recreate their dynamic evolution processes and make predictions. Attached Figure Description

[0055] Figure 1 The following are RNN and LSTM units: (a) RNN; (b) LSTM.

[0056] Figure 2It is a TCN residual block structure.

[0057] Figure 3 This is a flowchart of the present invention.

[0058] Figure 4 The grid diagram and measured points for the Bohai Sea case study are shown.

[0059] Figure 5 This is a water depth map for a calculation example in the Bohai Sea (negative values ​​represent water depth).

[0060] Figure 6 The optimal solution for the G-UIBC-2C Bohai Sea simulation is used to simulate the LT1-LT2 tide levels on GPUOM-HD.

[0061] Figure 7 The simulation results of LC1-LC5 flow velocity and direction using the MIKE astronomical tide prediction scheme and the optimal G-UIBC-2C scheme on GPUOM-HD are shown below: a) LC1 flow velocity comparison; b) LC1 flow direction comparison; c) LC2 flow velocity comparison; d) LC2 flow direction comparison; e) LC3 flow velocity comparison; f) LC3 flow direction comparison; g) LC4 flow velocity comparison; h) LC4 flow direction comparison; i) LC5 flow velocity comparison; j) LC5 flow direction comparison.

[0062] Figure 8 To illustrate the simulation results of flow rate and direction in LC1-LC5 of the G-EnKF-AL series solution on GPUOM-HD, a) shows the flow rate comparison for LC1; b) shows the flow direction comparison for LC1; c) shows the flow rate comparison for LC2; d) shows the flow direction comparison for LC2; e) shows the flow rate comparison for LC3; f) shows the flow direction comparison for LC3; g) shows the flow rate comparison for LC4; h) shows the flow direction comparison for LC4; i) shows the flow rate comparison for LC5; and j) shows the flow direction comparison for LC5.

[0063] Figure 9 Comparison of tidal levels in LT1-LT2 for datasets B01 / B04 and B01-B04.

[0064] Figure 10 This is a global bar chart (2000, left B01, right B04). Detailed Implementation

[0065] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments, as shown in the flowchart of the present invention.

[0066] This disclosure discloses an adaptive correction hydrodynamic numerical prediction method that integrates observation data to optimize the predicted hydrodynamic simulation results data, wherein the hydrodynamic simulation results data includes tidal level and / or tidal current data, and this embodiment takes tidal level as an example.

[0067] Terminology Explanation:

[0068] (1) Astronomical Tide (AT): Astronomical tide refers to the periodic rise and fall of sea level caused by the gravitational interaction between the moon and the sun.

[0069] (2) Characteristics in environmental field data include:

[0070] u10: Meridional wind component over 10 meters; v10: Zonal wind component over 10 meters; d2m: Dew point temperature over 2 meters;

[0071] t2m: Temperature at 2 meters; msl: Mean sea level pressure; sp: Surface pressure;

[0072] e: evaporation; sro: surface runoff; ssro: groundwater runoff; tp: total precipitation.

[0073] Hydrodynamic results data: including tidal level and / or tidal current data.

[0074] Selected region open boundary line Figure 4 The line connecting B01 and B04.

[0075] The observation point has relevant measurements of astronomical tides, environmental fields, tide levels and / or tidal currents, which are denoted as LC in this invention.

[0076] The points to be predicted include grid cells and / or nodes of the selected area to be predicted, i.e., points of tidal level and / or tidal flow data to be predicted.

[0077] The model involved in this embodiment is described below:

[0078] G-UIBC-2C (GPUOM - Uncertainties in Initia and Boundary Conditions - to Current) is a deep learning network model, including a first deep learning network model and a second deep learning network model. This invention, based on experiments, selects TCN from Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN) for the deep learning network model and determines the optimal hyperparameters of the network (see experimental examples below).

[0079] GPUOM-HD is a hydrodynamic mathematical model that couples an Ensemble Kalman Filter (EnKF) with offline-generated adaptive localization coefficients G-EnKF-AL (GPUOM – EnKF - AdaptiveLocalization).

[0080] The continuity equation (Equation (1.4)) and momentum equation (Equations (1.5)-(1.6)) of the hydrodynamic mathematical model GPUOM-HD are as follows:

[0081] (1.4)

[0082] (1.5)

[0083] (1.6)

[0084] In the formula:

[0085] Indicates water level; Represents coordinates in Cartesian space. Represents Cartesian space Directional coordinates Represents Cartesian space The coordinates of the direction, z represents the Cartesian space. Coordinates of direction;

[0086] ( , ( ) represents the average vertical velocity in Cartesian coordinates. Represents the Cartesian coordinate system The average vertical velocity in the direction of flow. Represents the Cartesian coordinate system The average vertical velocity in the direction; , ) represents the square of the average vertical velocity;

[0087] The water depth below zero surface; Total water depth; Represents coordinates in Cartesian space. Represents Cartesian space Directional coordinates Represents Cartesian space Coordinates of direction;

[0088] Represents the vertical coordinate, in Location ,exist Location ;

[0089] Indicates the evaporation flux rate; Indicates the flux rate of rainfall; Indicates the density of water; Indicates the inflow flux rate; Indicates the area of ​​the control volume; Indicates the Coriolis force coefficient; To represent gravitational acceleration; To represent the reference density of water; For time; The vertical eddy viscosity coefficient;

[0090] It is the horizontal viscous diffusion term, calculated according to equation (1.7); Calculate according to formula (1.8); It is a resistance term, calculated according to the corresponding flow obstruction parameterization scheme.

[0091] (1.7)

[0092] (1.8)

[0093] Considering The water surface gradient term can be expressed by equations (1.9)-(1.10), where and For the bottom slope source term:

[0094] (1.9)

[0095] (1.10)

[0096] In equation (1.7) is Horizontal turbulent viscosity coefficient. In this model, it can be given as a constant or calculated using the Smagorinsky rotational parametric model, i.e., equation (1.11):

[0097] (1.11)

[0098] In the formula, For constant parameters, Let MCE be the area of ​​the momentum control element.

[0099] Based on the above, the adaptive correction hydrodynamic numerical prediction method of the present invention, which integrates observation data, includes the following steps:

[0100] Step 1. The first deep learning model of G-UIBC-2C is used to predict the true tide level in the engineering area based on astronomical tide and environmental field data. Specifically, the model outputs the tide level at the marked points on the open boundary line of the selected area based on the astronomical tide and environmental field data of the points marked on the open boundary line of the selected area, which is used as the open boundary condition. Among them, the features in the environmental field data include u10: 10-meter meridional wind component, v10: 10-meter zonal wind component, d2m: 2-meter dew point temperature, t2m: 2-meter temperature, msl: mean sea level pressure, sp: surface pressure, e: evaporation, sro: surface runoff, ssro: groundwater runoff, and tp: total precipitation.

[0101] Step 2. The GPUOM-HD hydrodynamic mathematical model outputs the prior tide level of the selected area to be predicted point based on the open boundary conditions, environmental field data of the selected area to be predicted point, hydrodynamic example grid, water depth and parameter information (as shown in Table 2).

[0102] Step 3. For any point to be predicted, the assimilation point of the nearest neighbor observation point is used, where the predicted point includes grid cells and / or nodes of the selected region to be predicted. Calculate the prediction value for each point to be predicted. With observation point Euclidean distance For each point to be predicted Match the nearest observation point To form a dataset During subsequent assimilation, the assimilation point uses the observation value of the observation point with the closest Euclidean distance. If the distance is less than or equal to the cutoff radius, assimilation is performed.

[0103] Step 4. The hydrodynamic mathematical model calculates the similarity of any point to be predicted in the selected area with its nearest neighbor observation point on any selected feature in the prior hydrodynamic simulation results data of the point to be predicted in the selected area and the environmental field data of the point to be predicted in the selected area.

[0104] In step 4, feature acquisition includes the following steps:

[0105] Step 4.1. Using the prior tide level from Step 2 ( Features, such as u10, v10, e, etc. in the environmental field, are aligned with the timestamps and then missing values ​​are processed using cubic spline interpolation.

[0106] Step 4.2. Using 11 features, including the prior tide level, u10, v10, and e, as inputs and the measured tide level as output, train the sub-model using the G-UIBC-2C optimal hyperparameter scheme. Perform SHAP analysis to obtain the weights of each input feature. Select the K features with the largest weights as the final variables to consider in G-EnKF-AL.

[0107] Among them, the SHAP (ShaPley Additive Explanations) method, inspired by cooperative game theory and proposed by Lundberg et al., is an interpretable machine learning framework based on Shapley values. This method achieves accurate evaluation of the importance of each input feature in the model by quantifying the marginal contribution of individual features in multi-feature collaborative scenarios. Its core lies in the theory of Shapley value calculation—this theory was initially used to solve problems of fair resource allocation (such as profit distribution, cost sharing, and measurement of cooperative contribution), and has now been extended to feature attribution analysis.

[0108] The core computation of the SHAP method can be represented as:

[0109] (1.9)

[0110] In the formula, Representative prediction function Input Sample Features Regarding relative importance, For the set of all features; For features not included The set. The calculation process involves three steps:

[0111] Subset modeling: for features containing Feature subset Training Model and in removing features Post-trained model .

[0112] Marginal effect calculation: Inferring sample subsets using the trained model respectively. and The output is calculated, and its average prediction error is calculated. .

[0113] Expectation value integration: Traversing features Find the marginal contribution across all permutations and combinations and take its expected value.

[0114] The SHAP method combines global and local interpretability: it reveals the distribution of feature importance in a dataset at the global level and analyzes the prediction attribution of a single observation sample at the local level. Therefore, SHAP is widely used for feature selection and model interpretability enhancement in fields such as road surface recognition, disease diagnosis, hyperspectral inversion, and saltwater intrusion prediction.

[0115] Step 4.3. Normalize the weights of the K features with the largest weights to a sum of 1, resulting in the new weights:

[0116]

[0117] In the formula, This indicates the first result obtained from SHAP analysis. The weights of the 11 features (in total). Indicates the first New weights for each feature. The corresponding features are called selection features.

[0118] In step 4, the similarity between any point to be predicted in the selected area and its nearest neighbor observation point on any selected feature in the prior hydrodynamic simulation results data of the point to be predicted in the selected area, and the environmental field data of the point to be predicted in the selected area, is calculated, including...

[0119] Step 4.4. G-EnKF-AL uses R² to evaluate the similarity between the assimilation points and the observation points on each variable:

[0120] (1.2)

[0121] In the formula, R2 represents, Indicates the observation point number The value of each selected feature. Indicates the point to be predicted. The value of each selected feature. Indicates the number of features selected. Choose the mean of the features for the observation points.

[0122] Step 4.5. Truncate the R² values ​​and map them to [0,1] to obtain the variables among the K features with the largest weights. (No. Similarity of features :

[0123] (1.3)

[0124] In the formula, This represents the similarity on the k-th selected feature;

[0125] Step 5. Based on the similarity on any selected feature and the weight of the selected feature, calculate the weighted similarity on the selected features between any point to be predicted and its nearest neighbor observation points in the prior hydrodynamic simulation results data of the point to be predicted in the selected area, and in the environmental field data of the point to be predicted in the selected area, as shown in the following formula:

[0126]

[0127] In the formula, S represents the weighted similarity, and K represents the number of selected features. This represents the weight of the selected feature.

[0128] Step 6. Using the weighted similarity S as a correction factor, dynamically fuse the observed data with the mechanistic model through EnKF, and generate adaptive localization coefficients offline by G-EnKF-AL (GPUOM – EnKF - Adaptive Localization), including...

[0129] Create a prior state

[0130] (3.1)

[0131] In the formula, the subscript Indicates the first Time step, Represents prior values, Representing the hydrodynamic mathematical model, Represents the posterior value;

[0132] Calculate the propagation prediction error covariance matrix

[0133] (3.2)

[0134] In the formula, This is an approximate ensemble forecast error covariance matrix. The number of members in the set. For the i-th member of the forecast set, This is the mean of the forecast set.

[0135] Calculate the Kalman gain matrix

[0136] (3.3)

[0137] In the formula, Represents the observation matrix. Represents the observation error variance matrix;

[0138] The hydrodynamic mathematical model calculates the posterior state of any point to be predicted based on the prior state, the weighted similarity between any point to be predicted and its nearest neighbor observation points, the measured hydrodynamic results data of the nearest neighbor observation points of any point to be predicted, and the Kalman gain matrix. This posterior state is then used to calculate the posterior state based on the posterior hydrodynamic simulation results data of the point to be predicted.

[0139] (3.4)

[0140] In the formula, Indicates the posterior state. Indicates weighted similarity. Represents the Kalman gain matrix. This represents the measured hydrodynamic results data. Represents the observation matrix. This represents the prior state.

[0141] In the prediction phase of this invention, GPUOM is used to generate the prior state for the next time step, and then an ensemble estimate of the forecast error covariance is generated. In the update phase, new observations are used to correct the prior state, making the generated posterior state closer to the true state. In this invention, to enable the digital twin system and the operational forecasting platform to perform flexible and targeted assimilation operations on multi-source heterogeneous observation indicators, the model encapsulates the core function of "Ensemble Kalman Filter (EnKF) and its improved algorithm". This function provides a unified variable interface, supporting on-demand calls based on the type and number of observation indicators. This design enhances the model's scalability and ensures seamless integration with subsequent functional modules to be developed. Moreover, the matrix operations involved in formulas (3.2) to (3.4) are dimensionality-reduced and optimized into theoretically equivalent scalar operations within the model, improving computational efficiency while maintaining accuracy.

[0142] (1) Analysis of Deep Learning Model and Hyperparameter Optimal Scheme: Astronomical Tide (AT) Observation Data: To verify the accuracy of the model constructed, trained, and applied in this paper, corresponding tide level and tidal current observation data were obtained in the Bohai Sea. The specific geographical location and time series range of each observation station are detailed. The research area of ​​the Bohai Sea hydrodynamic example in this invention is the sea area west of the line connecting Dalian, Liaoning and Yantai, Shandong. The example grid ( Figure 4 ), water depth ( Figure 5 The parameter values ​​are based on previous research by the team, and the specific parameters are shown in Table 3. The measured locations, data types, and time ranges are shown in Table 2. Figure 4 .

[0143] Table 1. Measured locations, data types, and time ranges.

[0144]

[0145] Note: The time format is YYYY year, MM month, DD day, HH hour (Beijing time); data for 1984, 1985, and 1996 are missing for Qinhuangdao, and data for 1983 and 1995 are missing for Tanggu; the estimated original tidal reference levels for Tiger Beach, Qinhuangdao, Tanggu, Yantai, and LT1 and LT2 are 219.58, 86.57, 262.22, and 214.41 cm and 0.03 cm and 0.04 cm below the mean sea level, respectively; the actual observation interval is 1 hour, and the time series data interpolation method is cubic spline interpolation.

[0146] Table 2. Grid, water depth, and parameter information for the Bohai Sea hydrodynamic example.

[0147]

[0148] Environmental field data: This paper uses the fifth-generation global atmospheric reanalysis dataset ERA5 released by the European Centre for Medium-Range Weather Forecasts (ECMWF). This dataset covers the period from 1940 to the present. Ten parameters were selected, including the 10-meter meridional wind component, the 10-meter zonal wind component, and the 2-meter dew point temperature. The spatiotemporal attributes of the data include: temporal resolution (1h), spatial resolution (... ), Time span (covering all periods involved in the aforementioned analysis in this article).

[0149] Table 3. Environmental Field Data (Complete and Abbreviations)

[0150]

[0151] Note: 1 Indicates in 1 The area was evenly covered with 1 Deep water.

[0152] This invention selects a deep learning network model based on the following approach: The G-UIBC-2C model aims at two key aspects of the hydrodynamic mathematical model (GPUOM-HD):

[0153] (1) High-fidelity initial / boundary conditions: By using astronomical tide and environmental field observation data, as well as the corresponding real tide levels, the model is trained. The model directly, accurately and quickly predicts the real tide levels of all marked points on the open boundary line of the selected area based on the astronomical tide and environmental field data. This is used for the open boundary conditions, which is the role of the first deep learning model of this invention.

[0154] (2) Data assimilation optimization: By using the prior tide level and environmental field observation data of the selected area observation points, as well as the corresponding measured tide level of the selected area observation points, the model is trained. The model can directly, accurately and quickly predict the true tide level of the selected area to be predicted based on the prior tide level and environmental field data. The predicted true tide level is then analyzed by SHAP to obtain the weights assigned to the influence of each feature in the prior tide level and environmental field on the tide level, thereby determining the important features (selected features) and their weights. Based on the feature type and weight, the similarity analysis between spatial points and specific stations is supported in the G-EnKF-AL method to improve the assimilation accuracy. This is the role of the second deep learning model of this invention.

[0155] In the candidate model of this invention, Long Short-Term Memory (LSTM) network and Temporal Convolutional Network (TCN), LSTM can effectively capture and utilize long-range dependencies in time-series data. It is widely used in tasks related to sequence data learning, such as speech recognition, machine translation, natural language processing, and time series prediction.

[0156] LSTM cells such as Figure 1 As shown in (b), it has three types of gating: input, forget, and output gates, respectively. , and express.

[0157] Input gate: controls the current input data. Inflow into memory unit How many are there?

[0158] (1.1)

[0159] Forget Gate: Controls which information is retained and which is forgotten.

[0160] (1.2)

[0161] (1.3)

[0162] Output gate: controls the memory unit. For the current output value The impact.

[0163] (1.4)

[0164] (1.5)

[0165] In the formula, This is the weight matrix of the network; It is the bias vector; The sigmoid function maps a real value to an interval. ; , and are respectively time steps At that time, the hidden unit contains the input vector, output vector, and memory unit.

[0166] In the candidate models of this invention, Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN), the Temporal Convolutional Network (TCN) combines the inherent parallel computing advantages of CNNs with the long-term dependency modeling capabilities of RNNs, making it particularly suitable for temporal tasks. TCN employs various ingenious processing methods, such as causal convolution, dilated convolution, residual connections, and random deactivation. Its residual block structure is as follows: Figure 2 As shown. Causal Convolutional: Taking time-series data as an example, in standard convolution operations, the output value is based on its past, present, and future input values; causal convolution, by constraining the convolution operation, makes the output value depend only on its current and historical input values, thus strictly avoiding the leakage of future information. Dilated Convolution: Dilated convolution increases the receptive field by skipping some input units during convolution. This operation enables TCN to effectively capture a wider range of contextual information in the input data. Residual Connections: Residual connections directly add the output of one layer to the input of subsequent deeper layers. This operation significantly alleviates the gradient vanishing problem during the training of deep networks, improving the efficiency and stability of model training. Dropout: Dropout is a regularization technique that reduces the interdependence between neurons. During the training phase, the output values ​​(activation values) of some neurons are randomly set to zero with a probability of dropout_rate, but the weights obtained during training are not deleted or modified. During the testing phase, all neurons are activated, but their output values ​​are multiplied by the retention probability of that neuron during training (i.e., 1-dropout_rate).

[0167] Based on the two models mentioned above, this invention uses two datasets (the B01 / B04 subset and the complete B01-B04 dataset) from the B01-B04 sites to train the G-UIBC-2C model in the Bohai hydrodynamic case study. This model integrates the LSTM (Long Short-Term Memory) and TCN (Temporal Convolutional Network) time prediction algorithms.

[0168] The first deep learning network model's training set input includes astronomical tide and environmental field data for marked points along the open boundary line of the selected region, and its output includes measured hydrodynamic results for all marked points along the open boundary line of the selected region. The second deep learning network model, based on its input hydrodynamic simulation results for the points to be predicted in the selected region, and the environmental field data for those points, outputs predicted hydrodynamic simulation results for the points to be predicted in the selected region.

[0169] Based on the selected historical time step and input layer features, and combined with the preliminary hyperparameter tuning results, this invention studies and constructs six different model configurations.

[0170] Specifically as follows:

[0171] (1) 5-layer structure (input layer; LSTM hidden layer [256,128,64]; output layer), -24 hours, based on B01 / B04 dataset.

[0172] (2) 5-layer structure (input layer; LSTM hidden layer [512,256,128]; output layer), -24 hours, based on B01 / B04 dataset.

[0173] (3) 10-layer structure (input layer; TCN hidden layer [256,256,128,128,64,64,32,32]; output layer), -24 hours, based on B01 / B04 dataset.

[0174] (4) 10-layer structure (input layer; TCN hidden layer [256,256,256,256,128,128,128,128]; output layer), -24 hours, based on B01 / B04 dataset.

[0175] (5) 10-layer structure (input layer; TCN hidden layer [512,512,256,256,128,128,64,64]; output layer), -48 hours, based on B01 / B04 dataset.

[0176] (6) 10-layer structure (input layer; TCN hidden layer [512,512,512,512,256,256,256,256]; output layer), -48 hours, based on B01 / B04 dataset.

[0177] Therefore, this invention selects the B01 / B04 dataset, designs six hyperparameter schemes for training, and evaluates their training effects on B01 / B04. Furthermore, using the 2016 MIKE astronomical tide forecast and meteorological data as input, the above six schemes are used to predict the "true tide level". The tide levels generated by the MIKE scheme and the "true tide levels" generated by the six G-UIBC-2C schemes are used as open boundary conditions for the GPUOM model in the Bohai Sea example, respectively, to conduct hydrodynamic simulations in 2016, comparing the tidal simulation application effects of each scheme at the LT1-LT2 stations (Table 5).

[0178] A comparison of the performance of various G-UIBC-2C schemes at LT1-LT2 was conducted. Six hyperparameter schemes were designed for training, and their training performance on B01 / B04 was evaluated. Furthermore, using the 2016 MIKE astronomical tide forecast and meteorological data as input, the above six schemes were used to predict the "true tide level." The tide levels generated by the MIKE scheme and the "true tide levels" generated by the six G-UIBC-2C schemes were used as open boundary conditions for the GPUOM model in a Bohai Sea example, and hydrodynamic simulations were conducted in 2016 to compare the tidal simulation application effects of each scheme at LT1-LT2 stations.

[0179] Table 4 G-UIBC-2C_LSTM Parameter Settings

[0180]

[0181] Note: To improve the model's generalization ability and eliminate the bias of the tide gauge station's datum, the model output Y is standardized using the statistics of the astronomical tide component in the input X.

[0182] Model Training and Validation: Evaluation Metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination. Its mathematical definition and statistical significance are as follows:

[0183] (1).MAE

[0184] (1.6)

[0185] In the formula, For the first Measured values ​​of each sample These are the predicted values ​​from the corresponding model; is the sample size. MAE represents the arithmetic mean of the absolute values ​​of the prediction errors.

[0186] (2).RMSE

[0187] (1.7)

[0188] RMSE assigns higher weights to larger errors through squaring, making it more sensitive to outliers. In terms of dimensional consistency, RMSE maintains the same units as the target variable, allowing for a direct quantification of the prediction error.

[0189] (3). Calculation

[0190] (1.8)

[0191] In the formula, This represents the mean of the measured samples. This represents the proportion of variance explained by the model to the total variance, and its theoretical range is... When the value approaches 1, it indicates that the model has excellent explanatory power; if a negative value appears, it means that the model's prediction accuracy is lower than the simple mean model benchmark, reflecting significant model failure.

[0192] Table 5. MIKE Astronomical Tide, Six Training Schemes, and Application Results

[0193]

[0194] Note: P0 represents the MIEK astronomical tide forecast; P1-P6 represent the six schemes; 1-MAE represents the MAE of LT1; T represents the performance of the training results in the dataset.

[0195] As shown in Table 5, T-MAE, T-RMSE, and T-R² demonstrate the model training performance, with P2 representing the hyperparameter scheme with the best training effect. 1-MAE, 1-RMSE, and 1-R² demonstrate the simulation performance of LT1, with P4 representing the hyperparameter scheme with the best simulation performance. Similarly, 2-MAE, 2-RMSE, and 2-R² demonstrate the simulation performance of LT2, with P4 again representing the best simulation performance. This indicates that the MAR, RMSE, and R² metrics during G-UIBC-2C training are not entirely positively correlated with their application performance in the GPUOM-HD model. This suggests that when designing G-UIBC-2C and similar time-series prediction models, one should not only pursue training metric optimization through complex architectures or hyperparameter tuning, but also focus on the model's performance in specific marine scenarios.

[0196] T-MAE, T-RMSE, and T-R² demonstrate the model training results, with P2 representing the hyperparameter scheme that yields the best training performance. 1-MAE, 1-RMSE, and 1-R² demonstrate the simulation results for LT1, while 2-MAE, 2-RMSE, and 2-R² demonstrate the simulation results for LT2.

[0197] Based on the implementation results of the G-UIBC-2C scheme, Scheme 4 was selected to train on the B01 / B04 and B01-B04 datasets respectively, and the application effect of the two sets of data in LT1-LT2 was evaluated.

[0198] Table 6 shows the application effects of datasets B01 / B04 and B01-B04 in LT1-LT2.

[0199]

[0200] Figure 6 The results of tidal level simulations at LT1-LT2 stations using the optimal scheme from the G-UIBC-2C Bohai Sea example are presented on GPUOM-HD. The results show that, compared to the MIKE scheme, the open boundary conditions of most G-UIBC-2C schemes improve the accuracy of tidal simulations. Scheme 4 achieves tidal simulation accuracies (R²) of 95.07% and 92.91% for LT1 and LT2 stations, respectively, representing improvements of 2.45% and 6.32% compared to the MIKE scheme.

[0201] Figure 9 The table shows the tidal level comparison between LT1 and LT2 for datasets B01 / B04 and B01-B04, based on Table 6 and... Figure 1 The results showed that the B01-B04 dataset failed to improve the application effect of LT1-LT2. The optimal solution was finally determined to be: B01 / B04 dataset combined with a historical time step of 24h and TCN hidden layer [256,256,256,256,128,128,128,128].

[0202] (2) Evaluation and analysis of the optimal solution of G-UIBC-2C in LC1-LC5:

[0203] Using the 2021 MIKE astronomical tide forecast and meteorological data as input, the optimal scheme was used to predict the "true tide level". The tide level generated by the MIKE scheme and the "true tide level" generated by the optimal scheme were used as open boundary conditions to carry out tidal current simulation in 2021, and the application effects of the two schemes on the tidal current simulation at LC1-LC5 stations were compared (Table 8).

[0204] Table 7 Comparison of current velocities (m / s) between the MIKE astronomical tide forecasting scheme and the G-UIBC-2C optimal scheme in LC1-LC5.

[0205]

[0206] Table 8 Comparison of flow direction (°) between the MIKE astronomical tide forecasting scheme and the G-UIBC-2C optimal scheme in LC1-LC5

[0207]

[0208] Note: MIKE refers to the MIKE astronomical tide forecasting scheme; G-UIBC-2C refers to the hyperparameter scheme mentioned in P4 above. Improv.(%) = (MAE) MIKE -MAE G-UIBC-2C )*100% / MAEMIKE. The flow direction difference is min([abs(os),360-abs(os)], where o is the observed value and s is the simulated value. The model evaluation uses the flow data from 15:00 on September 18, 2021 to 15:00 on September 19, 2021.

[0209] The results show that, compared with the MIKE scheme, the optimal scheme improves the accuracy of flow velocity simulation by an average of 7.42% (maximum improvement of 42.88%) and flow direction simulation by an average of 0.05% (maximum improvement of 12.00%) for LC1-LC5 stations.

[0210] (3) SHAP Analysis of the Optimal Scheme of G-UIBC-2C: Through SHAP analysis of the optimal scheme, the influence of astronomical tides and meteorological factors on the actual tide level can be quantified, providing a reference for the parameter calibration of the GPUOM model. Taking the Bohai Sea in 2016 as an example: Figure 6 The global bar chart shows that, apart from astronomical tides, d2m_-0h, sp_-24h, and t2m_-24h are the key factors for accurate tide level forecasting, with SHAP values ​​of 0.05, 0.04, and 0.04, respectively. This indicates that even on a 20-year timescale, the influence of meteorological factors on the tide level in the Bohai Sea is still significant.

[0211] (4) Application effect of G-EnKF-AL at all locations and all time periods: The GPUOM-HD prior values ​​obtained by the optimal scheme and meteorological data are used as inputs, and the measured tidal current values ​​are used as outputs. The optimal hyperparameter scheme is used to train the model and SHAP analysis is performed. This method can quantify the influence of prior values ​​and meteorological factors on the actual tidal current, and provide a basis for the weight calculation in G-EnKF-AL. Since the measured tidal current period is relatively short (25 hours), the similarity between the assimilation point and the measured point is directly quantified using the prior U and V components. LC1-LC5 are selected as the assimilation and verification stations of GPUOM-HD. The accuracy of coupling G-EnKF-AL and GPUOM-HD is verified using the measured tidal current data from 15:00 on September 18 to 15:00 on September 19, 2021. The parameter configuration is as follows: the set member N_f is 100, the mean of the prior and measurement errors is set to 0, the variance is 0.001, and the random seed is 250517.

[0212] Table 9 Comparison of Flow Rate (m / s) for Application of G-EnKF-AL Series Solutions

[0213]

[0214] Table 10 Comparison of Application and Flow Direction of G-EnKF-AL Series Solutions (°)

[0215]

[0216] Note: G-EnKF-AL(NO), G-EnKF-AL(FP,FT), G-EnKF-AL(PP,FT), and G-EnKF-AL(PP,PT) represent: no application of the G-EnKF-AL algorithm, application at all locations throughout the entire time period, application at some locations throughout the entire time period, and application at all locations at some time periods, respectively. Improv. (%, 1 / 0) refers to the percentage performance improvement of G-EnKF-AL(FP,FT) compared to G-EnKF-AL(NO) (and so on). Model evaluation uses measured power flow data from 15:00 on September 18th to 15:00 on September 19th, 2021.

[0217] Table 9-10 shows the comparison results of flow velocity and direction at LC1-LC5 stations with and without G-EnKF-AL application across all locations and time periods. Compared to the model without G-EnKF-AL coupling, after G-EnKF-AL assimilation across all locations and time periods, the simulated MAEs for flow velocity at LC1-LC5 stations are 0.0096, 0.0092, 0.0072, 0.0079, and 0.0094 m / s, respectively, and the simulated MAEs for flow direction are 6.7002°, 6.9455°, 12.3357°, 4.6916°, and 7.5950°, respectively. Figure 8 (G-EnKF-AL-1).

[0218] (5) Analysis of the full-time application effect of G-EnKF-AL at some locations: Based on the hydrodynamic characteristics of stations LC1-LC5 during the ebb and flow of tides, LC1 and LC3-LC5 were selected as the assimilation and pre-validation stations for the GPUOM model, and LC2 was selected as the final validation station. Compared with the model without G-EnKF-AL coupling, after full-time assimilation at some locations using G-EnKF-AL, the simulated MAE of current velocity at station LC2 decreased by 25.00%, and the simulated MAE of current direction decreased by 35.34%. Figure 8 (G-EnKF-AL-2).

[0219] (6) Analysis of the application effect of G-EnKF-AL in some time periods at all locations: LC1-LC5 were selected as the assimilation and verification stations of the GPUOM model. The measured tidal flow data from 15:00 on September 18 to 9:00 on September 19, 2021 were used to test the effect of the G-EnKF-AL coupled model with GPUOM on improving the accuracy of tidal flow forecast.

[0220] Table 11 Comparison of G-EnKF-AL application at all locations and during specific time periods - flow velocity (m / s)

[0221]

[0222] Table 12 Comparison of G-EnKF-AL application status at all locations and during specific time periods - flow direction (°)

[0223]

[0224] Note: The model evaluation uses power flow data from 10:00 to 15:00 on September 19, 2021.

[0225] like Figure 8 As shown in Tables 11-12 of G-EnKF-AL-3, compared with the baseline model without G-EnKF-AL coupling, after assimilation at all points and all time periods, the average MAE of the simulated flow velocity at stations LC1-LC5 decreased by 43.96% (with a maximum decrease of 63.96%), and the average MAE of the simulated flow direction increased by -9.03% (with a maximum decrease of 42.45%).

[0226] Based on the above embodiments, this application also provides a computer storage medium storing a computer program, which, when executed by a computer, causes the computer to perform the methods provided in the above embodiments.

[0227] The storage medium can be any available medium that a computer can access. For example, but not limited to, a computer-readable medium can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer.

[0228] Based on the above embodiments, this application also provides a chip for reading a computer program stored in a memory to implement the method provided in the above embodiments.

[0229] Based on the above embodiments, this application provides a computer program product that implements the methods provided in the above embodiments when the computer program product is run on an electronic device.

[0230] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0231] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0232] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0233] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

Claims

1. An adaptive correction hydrodynamic numerical prediction method that fuses observational data, characterized in that, comprising The first deep learning network model outputs the hydrodynamic simulation result data of all the marked points on the open boundary connecting line of the selected region according to the astronomical tide and environmental field data of the selected region on the open boundary connecting line of the marked points, which is used as the open boundary condition; The hydrodynamic mathematical model outputs the prior hydrodynamic simulation result data of the selected region to be predicted points according to the input of the open boundary condition and the environmental field data of the selected region to be predicted points; The hydrodynamic mathematical model calculates a decision coefficient on a selected feature in the prior hydrodynamic simulation result data of any to-be-predicted point in the selected area and the observation point nearest to the to-be-predicted point in the selected area, and in the environmental field data of the to-be-predicted point in the selected area is shown by the following formula: wherein, denotes the determination coefficient, denotes the value of the i-th selected feature of the observation point, denotes the value of the i-th selected feature of the observation point, denotes the value of the i-th selected feature of the observation point, denotes the value of the i-th selected feature of the observation point, denotes the number of selected features, is the mean value of the selected features of the observation point; The The similarity between any selected feature of the prior hydrodynamic simulation result data of the selected region to be predicted and the environmental field data of the nearest neighbor observation point of the selected region to be predicted is calculated is shown by the following formula: wherein represents the similarity on the kth selected feature; The hydrodynamic mathematical model calculates the weighted similarity S of the selected feature between the selected region to be predicted point and its nearest neighbor observation point in the prior hydrodynamic simulation result data of the selected region to be predicted points and the environmental field data of the selected region to be predicted points according to the similarity on the selected feature and the weight of the selected feature, which is shown as follows: In the formula, S represents a weighted similarity, K represents the number of selected features, represents the weight of the selected feature; The hydrodynamic mathematical model calculates the Kalman gain matrix according to the prior state and the propagation prediction error covariance matrix; The hydrodynamic mathematical model calculates the posterior state of the selected region to be predicted point according to the prior state, the weighted similarity of the selected region to be predicted point and its nearest neighbor observation point, the measured hydrodynamic result data of the nearest neighbor observation point of the selected region to be predicted point, and the Kalman gain matrix, which is used for the posterior hydrodynamic simulation result data of the selected region to be predicted point.

2. The adaptive correction hydrodynamic numerical prediction method of fused observational data according to claim 1, characterized in that, Wherein, The water power mathematical model calculates the prior state in the Kalman gain matrix according to the prior state and a propagation prediction error covariance matrix is shown by the following formula: In the formula, subscript represents the time step, represents the prior value, represents the hydrodynamic mathematical model, represents the posterior value; where the propagation forecast error covariance matrix is given by: wherein denotes the propagated forecast error covariance matrix, denotes the number of ensemble members, denotes the value of the i-th member of the forecast ensemble, is the mean of the forecast ensemble members; wherein the Kalman gain matrix is given by the following equation: wherein denotes the observation matrix, denotes the observation error variance matrix.

3. The adaptive correction hydrodynamic numerical prediction method of fused observational data according to claim 2, wherein, Wherein, The hydrodynamic mathematical model calculates the posterior state of the selected region to be predicted point according to the prior state, the weighted similarity of the selected region to be predicted point and its nearest neighbor observation point, the measured hydrodynamic result data of the nearest neighbor observation point of the selected region to be predicted point, and the Kalman gain matrix, which is used for the posterior hydrodynamic simulation result data of the selected region to be predicted point, which is shown as follows: wherein represents a posterior state, represents a weighted similarity, represents a Kalman gain matrix, represents a measured hydrodynamic result data, represents an observation matrix, represents a prior state.

4. The adaptive correction hydrodynamic numerical prediction method of fused observational data according to claim 1, wherein, Wherein, The first deep learning network model is obtained based on the following way: constructing and training the first deep learning network model, the input of the training set of which includes the astronomical tide and environmental field data of the marked points on the open boundary connecting line of the selected region, and the output includes the measured hydrodynamic result data of the marked points on the open boundary connecting line of the selected region.

5. The adaptive correction hydrodynamic numerical prediction method of fusing observation data according to claim 4, characterized in that, Wherein, The first deep learning network model includes a time domain convolution network, wherein the model hyperparameters include a historical time step 24h, and a hidden layer [256, 256, 256, 256, 128, 128, 128, 128].

6. The adaptive correction hydrodynamic numerical prediction method of fusing observation data according to claim 1, wherein, Wherein, The selected feature and its weight are obtained based on the following way: constructing and training the second deep learning network model, the input of the training data set of which includes the prior hydrodynamic simulation result data of the selected region observation points and the environmental field data of the selected region observation points, and the output includes the measured hydrodynamic result data of the selected region observation points; The second deep learning network model outputs the predicted hydrodynamic simulation result data of the selected region to be predicted points according to the input of the hydrodynamic simulation result data of the selected region to be predicted points and the environmental field data of the selected region to be predicted points; According to the water dynamic simulation result data of the selected region to be predicted point, the characteristics of the environmental field data of the selected region to be predicted point, and the predicted water dynamic simulation result data of the selected region to be predicted point, the weight of any feature of the water dynamic simulation result data of the selected region to be predicted point and the environmental field data of the selected region to be predicted point is obtained. Select the first several features with the maximum weight in the weight of any feature as the selected feature, and normalize the weight of the selected feature to obtain the weight of any selected feature.

7. The adaptive correction hydrodynamic numerical prediction method of fusing observation data according to claim 6, wherein, Wherein, Through SHAP analysis, the weight of any feature of the water dynamic simulation result data of the selected region to be predicted point and the environmental field data of the selected region to be predicted point is obtained.

8. The adaptive correction hydrodynamic numerical prediction method of fusing observation data according to claim 6, wherein, Wherein, The second deep learning network model includes a time domain convolution network, wherein the model hyperparameters include a historical time step of 24h and a hidden layer of [256, 256, 256, 256, 128, 128, 128, 128].

9. The adaptive correction hydrodynamic numerical prediction method of fusing observation data according to claim 1, wherein, Wherein, The to-be-predicted point includes a grid cell and / or a node of the selected region to be predicted. Wherein, the water dynamic simulation result data includes tidal level and / or tidal flow data. Wherein, the characteristics of the prior water dynamic simulation result data include prior water dynamic simulation results, and the environmental field data includes 10-meter zonal wind component, 10-meter meridional wind component, 2-meter dew point temperature, 2-meter temperature, mean sea level pressure, surface pressure, evaporation, surface runoff, groundwater runoff, and total precipitation.

Images