A high-precision wave field prediction method and system based on expert and space gating

Abstract

This invention discloses a high-precision wave field prediction method and system based on expert and spatial gating, relating to the field of nearshore wave element prediction technology. The prediction method includes the following steps: Step 1: Acquire multi-year low spatial resolution wind field data and wave data for the target sea area; Step 2: Obtain high spatial resolution wave element field for the corresponding year in the target sea area; Step 3: Construct a full dataset; Step 4: Construct a full data benchmark model and a forecast model; Step 5: Perform operational forecasting based on the forecast model trained in Step 4, with the low spatial resolution wind field data for the forecast period as the input to the model. The method of this invention can achieve rapid prediction of high-precision wave fields solely through wind field data, and solves the problem of "significant overall seasonal differences in wind fields, while exhibiting strong non-uniformity in local areas."

Description

Technical Field

[0001] This invention relates to the field of nearshore wave element prediction technology, and in particular to a high-precision wave field prediction method and system based on expert and spatial gating. Background Technology

[0002] Nearshore wave parameters have a significant impact on the safety of coastal infrastructure, port loading and unloading and port entry and exit scheduling, offshore wind power operation and maintenance, and nearshore disaster prevention and mitigation. In practical engineering applications, wave parameters such as effective wave height, mean wave direction, and mean wave period with high spatiotemporal resolution are usually required to accurately assess key indicators such as the stress on structures, ship motion, and static stability of the harbor basin under extreme sea conditions.

[0003] Currently, obtaining high spatial resolution wave features in complex nearshore areas mainly relies on long-term numerical simulations using shallow-water wave numerical models (such as the SWAN model or the ROMS–SWAN coupled model) on nested grids. These methods typically use low spatial resolution wind fields and offshore wave features as driving and boundary conditions, achieving a refined description of physical processes such as wave refraction, shallowing, and diffraction by densifying the grid in nearshore areas. However, nested numerical models are computationally extremely expensive, placing stringent demands on CPU resources and time, making it difficult to meet the operational needs of large-scale, multi-year, and real-time updates. This is particularly true in disaster early warning and engineering scheme evaluation, where rapid response is crucial, traditional numerical simulation methods often struggle to provide timely results. Furthermore, existing operational nearshore wave forecasts usually rely on both wind field and offshore wave inputs. Since both wind field and wave forecasts inherently contain uncertainties, the more input features there are, the longer the error propagation chain becomes. Moreover, simultaneous "high-precision consistency" of wind field and wave boundaries at the same time is uncommon in engineering practice, easily leading to deviations or insufficient stability in high-precision nearshore wave field results. Therefore, there is an urgent need for a fast prediction method that relies solely on wind fields during the operational phase and reduces dependence on wave-type inputs and error coupling.

[0004] On the other hand, nearshore wave spatial distribution and sea state morphology exhibit significant seasonal differences. Different seasons show different dominant regions and extreme value distribution characteristics due to the superposition of wind field structure, the influence of incoming waves from the open sea, and topographic effects. Given that the wind field exhibits "significant overall seasonal differences within a certain range, but strong local heterogeneity," using a single model to uniformly model the entire year's samples easily leads to trade-offs between different seasonal sea states, resulting in limited fitting ability in complex nearshore topographic areas and extreme sea states. Furthermore, existing methods mostly remain at the level of single-model downscaling or single-process implementation, lacking a system-level solution oriented towards business operations that can explicitly model seasonal differences and achieve spatial adaptive fusion, thus limiting its widespread application in port areas, offshore engineering, and coastal disaster prevention and mitigation.

[0005] In view of this, this invention is hereby proposed. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a high-precision wave field prediction method and system based on expert and spatial gating. This method can achieve rapid prediction of high-precision wave fields using only wind field data and solves the problem of wind fields exhibiting "significant overall seasonal differences and strong local non-uniformity".

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A high-precision wave field prediction method based on expert and spatial gating includes the following steps:

[0009] Step 1: Determine the target sea area and obtain coastline / land boundary and water depth data; then obtain multi-year low spatial resolution wind field data and wave data of the target sea area from publicly available meteorological reanalysis databases;

[0010] Step 2: Using the low spatial resolution wind field data and wave data of the target sea area obtained in Step 1, perform simulation calculations using a shallow water wave numerical model to obtain the high spatial resolution wave element field of the target sea area for the corresponding year.

[0011] Step 3: Perform time alignment and spatial consistency processing on the multi-year low spatial resolution wind field data of the target sea area obtained in Step 1 and the corresponding year high spatial resolution wave element field of the target sea area obtained in Step 2, and use a unified mask to mark the land area and missing measurement points to construct the full dataset.

[0012] Step 4: Construct the full-data baseline model and forecast model; the full-data baseline model takes low-spatial-resolution wind field data as input and outputs high-spatial-resolution wave element field, which is used for transfer training of the seasonal expert model; the forecast model includes parallel seasonal expert models and spatially gated models; there are four seasonal expert models, corresponding to spring, summer, autumn, and winter, which take low-spatial-resolution wind field data as input and output high-spatial-resolution wave element field; the spatially gated model takes low-spatial-resolution wind field data as input and outputs a weighted graph of the outputs of the seasonal expert models. The weighted graph has the same resolution as the outputs of the seasonal expert models and is used to weight the outputs of the seasonal expert models as the overall output of the forecast model;

[0013] Step 5: Perform operational forecasts based on the forecast model trained in Step 4. The input to the model is low spatial resolution wind field data for the forecast period.

[0014] Furthermore, in step 1, the low spatial resolution wind field data includes at least two orthogonal components u of the wind speed at a height of 10 m. 10 and v 10In step 2, the high spatial resolution wave element field includes at least the significant wave height Hs(t); in step 3, the samples in the constructed full dataset include:

[0015] ;

[0016] };

[0017] In the formula, X(t,n,m) is the low spatial resolution wind field data, which is the input signal, t is the time index point, and n and m are the spatial index points in the low spatial resolution grid; Y(t,i,j) is the high spatial resolution wave element field, which is the monitoring signal, and i and j are the spatial index points in the high spatial resolution grid.

[0018] Furthermore, in step 4, the full-data baseline model, the seasonal expert model, and the spatial gating model are trained through the following steps:

[0019] Step A4.1: Train a full-data baseline model using the full dataset;

[0020] Step A4.2: Divide the entire dataset into four subsets according to time: spring, summer, autumn, and winter; use the four subsets of spring, summer, autumn, and winter to train the full dataset baseline model trained in step A4.1, and obtain the four seasonal expert models.

[0021] Step A4.3: Train the forecast model using the full dataset. During training, freeze the parameters of the expert models for four quarters and train only the parameters of the spatial gating model.

[0022] Furthermore, in step 4, the spatial gating model outputs four spatial grid weight maps corresponding to the expert models for each season, as shown in the following formula:

[0023] ;

[0024] ;

[0025] ;

[0026] In the formula, t is the time index point, i and j are the spatial index points, and k is the seasonal expert model index point.

[0027] Furthermore, in step 4, the formula for the output of the prediction model is as follows:

[0028] ;

[0029] In the formula, The output results are for the expert models for the four seasons.

[0030] Furthermore, during the training of the full-data baseline model, seasonal expert model, and spatial gating model in step 4, the loss function is calculated through the following steps:

[0031] Step B4.1: Based on the high spatial resolution wave element field of the target sea area obtained in Step 2, statistically analyze the data of each spatial grid point for each year to obtain the 90th quantile threshold map, as shown in the following formula:

[0032] ;

[0033] In the formula, t is the time index point, and i and j are the spatial index points;

[0034] Step B4.2: Calculate the sea state weighting coefficient based on the 90th percentile threshold map, using the following formula:

[0035] ;

[0036] Step B4.3: Calculate the loss using the following formula:

[0037] ;

[0038] In the formula, The mask is divided into valid points for ocean and invalid points for land.

[0039] Furthermore, in step 4, the full data benchmark model and the seasonal expert model include a feature extraction sub-network and a spatial reconstruction sub-network; the feature extraction sub-network is the encoding stage, with a transform network structure, used to extract multi-scale spatial features from low spatial resolution wind field data; the spatial reconstruction sub-network is the decoding stage, with a convolutional neural network structure, used to generate wave element prediction fields on the target high-precision grid; the spatial gating model is a convolutional neural network.

[0040] To achieve the above objectives, the present invention also employs the following technical solution:

[0041] A high-precision wave field prediction system based on hybrid expert and spatial gating, characterized in that it includes a prediction model trained by any one of the above-mentioned high-precision wave field prediction methods based on expert and spatial gating proposed in this invention.

[0042] Compared with the prior art, the beneficial effects of this invention are as follows:

[0043] (1) Reduce input dependence and error coupling: The operational phase only uses low spatial resolution wind field data as input, without relying on wave observation, wave reanalysis or wave forecast model output, reducing the error superposition caused by the introduction of multiple source elements and improving forecast stability.

[0044] (2) Explicit modeling of seasonal differences: By using expert models for four seasons to learn the differences in seasonal sea conditions, we can avoid the loss of accuracy caused by the compromise of a single model on the annual sample and improve the generalization ability under different seasons.

[0045] (3) Spatial adaptive fusion: The spatial gating network outputs a grid-level weight map to realize the adaptive allocation of expert contributions for different regions in each season, thereby enhancing the prediction accuracy under complex nearshore terrain and spatial non-uniform distribution.

[0046] (4) High computational efficiency and easy business integration: The computational cost in the inference stage is significantly lower than that of traditional nested numerical simulation, which can meet the needs of high-frequency updates and rapid response, and is suitable for integration with port area, marine engineering and coastal disaster prevention and mitigation business systems. Attached Figure Description

[0047] Figure 1 This is a high-precision wave field prediction method based on expert and spatial gating;

[0048] Figure 2 A comparison chart of predictions from the full-data baseline model and the seasonal expert model;

[0049] Figure 3 This is a comparison chart of the prediction results of the forecasting model and the monitoring signal. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0051] Example 1:

[0052] A high-precision wave field prediction method based on expert and spatial gating, such as Figure 1 As shown, it includes the following steps:

[0053] Step 1: Determine the target sea area and obtain coastline / land boundary and water depth data; then obtain multi-year low spatial resolution wind field data and wave data of the target sea area from publicly available meteorological reanalysis databases.

[0054] In this embodiment, multi-year low spatial resolution wind field data of the target sea area can be obtained through the reanalysis data path ERA5; the years to be obtained can be determined as needed, such as 2015-2023.

[0055] In this embodiment, in step 1, the low spatial resolution wind field data includes at least two orthogonal components u of the wind speed at a height of 10 m. 10 and v 10 .

[0056] Step 2: Using the low spatial resolution wind field data and wave data of the target sea area obtained in Step 1, perform simulation calculations using a shallow water wave numerical model to obtain the high spatial resolution wave element field of the target sea area for the corresponding year.

[0057] In this embodiment, the shallow water wave numerical model can use the SWAN model; when simulating the shallow water wave numerical model, wind field data is used to create the driving file of the numerical model, and wave field data is used to create the boundary file and initial field file of the numerical model.

[0058] In this embodiment, in step 2, the high spatial resolution wave element field includes at least the effective wave height Hs(t).

[0059] Step 3: Perform time alignment and spatial consistency processing on the multi-year low spatial resolution wind field data of the target sea area obtained in Step 1 and the corresponding year high spatial resolution wave element field of the target sea area obtained in Step 2, and use a unified mask label to mark the land area and missing measurement points to construct the full dataset.

[0060] In this embodiment, in step 3, the samples in the constructed full dataset include:

[0061] ;

[0062] };

[0063] In the formula, X(t,n,m) is the wind field data with low spatial resolution (e.g., 48×48), which is the input signal, t is the time index point, and n and m are the spatial index points in the low spatial resolution grid; Y(t,i,j) is the wave element field with high spatial resolution (e.g., 366×364), which is the monitoring signal, and i and j are the spatial index points in the high spatial resolution grid.

[0064] In this embodiment, time correspondence is performed with an hourly time resolution; and a unified mask is used to mark land areas and missing measurement points to ensure that invalid points can be removed during subsequent loss function calculation.

[0065] In this embodiment, steps 1 to 3 are the offline sample construction stage. The above wave data and the simulation calculation of the shallow water wave numerical model are only used for the construction of offline samples. In the subsequent operation stage, it is not necessary to access any wave observation, wave reanalysis or wave model output. Only wind field data is needed to complete the rapid forecast.

[0066] Step 4: Construct a full-data baseline model and a forecast model; the full-data baseline model takes low-spatial-resolution wind field data as input and outputs high-spatial-resolution wave element fields, which are used for transfer training of the seasonal expert model; the forecast model includes parallel seasonal expert models and spatially gated models; there are four seasonal expert models, corresponding to spring, summer, autumn, and winter, which take low-spatial-resolution wind field data as input and output high-spatial-resolution wave element fields; the spatially gated model takes low-spatial-resolution wind field data as input and outputs a weighted graph of the outputs of the seasonal expert models. The weighted graph has the same resolution as the outputs of the seasonal expert models and is used to weight the outputs of the seasonal expert models as the overall output of the forecast model.

[0067] In this embodiment, step 4 involves training the full-data baseline model, the seasonal expert model, and the spatial gating model through the following steps:

[0068] Step A4.1: Train a full-data benchmark model using the full dataset.

[0069] Step A4.2: Divide the entire dataset into four subsets according to time: spring, summer, autumn, and winter; use the four subsets of spring, summer, autumn, and winter to train the full dataset baseline model trained in step A4.1, and obtain the four seasonal expert models.

[0070] In this embodiment, after steps 4.1 and 4.2, the seasonal expert model, based on its ability to predict high spatial resolution wave element fields from low spatial resolution wind field data, focuses on predicting wave element fields from spring, summer, autumn, and winter wind field data, respectively.

[0071] Step A4.3: Train the forecast model using the full dataset. During training, freeze the parameters of the expert models for four quarters and train only the parameters of the spatial gating model.

[0072] In this embodiment, in step 4, the spatial gating model outputs four spatial grid weight maps corresponding to the expert models for each season, as shown in the following formula:

[0073] ;

[0074] ;

[0075] ;

[0076] In the formula, t is the time index point, i and j are the spatial index points, and k is the seasonal expert model index point.

[0077] In this embodiment, the formula for the output of the prediction model in step 4 is as follows:

[0078] ;

[0079] In the formula, The output results are for the expert models for the four seasons.

[0080] In this embodiment, during the training of the full-data baseline model, seasonal expert model, and spatial gating model in step 4, the loss function is calculated through the following steps:

[0081] Step B4.1: Based on the high spatial resolution wave element field of the target sea area obtained in Step 2, statistically analyze the data of each spatial grid point for each year to obtain the 90th quantile threshold map, as shown in the following formula:

[0082] ;

[0083] In the formula, t is the time index point, and i and j are the spatial index points.

[0084] Step B4.2: Calculate the sea state weighting coefficient based on the 90th percentile threshold map, using the following formula:

[0085] .

[0086] Step B4.3: Calculate the loss using the following formula:

[0087] ;

[0088] In the formula, The mask is divided into valid points for ocean and invalid points for land.

[0089] In this embodiment, considering the significant differences in wave amplitude at different locations in the target sea area (larger in the open sea and smaller nearshore), using a single global threshold would cause the model to overemphasize the open sea and ignore the nearshore. Therefore, step B4.1 calculates the 90th percentile threshold map, and step B4.2 calculates the weighting coefficients, applying a weight factor greater than 1 to spatial grid points where the actual wave elements exceed the threshold map. This enhances the model's ability to fit the relative extremes of each region (especially high-wave events nearshore) without introducing additional wave input, while ensuring that land and invalid points do not participate in gradient calculation through masking.

[0090] In this embodiment, in step 4, the full data benchmark model and the seasonal expert model include a feature extraction sub-network and a spatial reconstruction sub-network; wherein the feature extraction sub-network is the encoding stage, with a transform network structure, used to extract multi-scale spatial features from low spatial resolution wind field data; the spatial reconstruction sub-network is the decoding stage, with a convolutional neural network structure, used to generate wave element prediction fields on the target high-precision grid; the spatial gating model is a convolutional neural network.

[0091] Step 5: Perform operational forecasts based on the forecast model trained in Step 4. The input to the model is low spatial resolution wind field data for the forecast period.

[0092] In this embodiment, step 5 inputs low spatial resolution wind field data, such as u from weather forecast output. 10 and v 10 The significant wave height is calculated and output through the forecast model.

[0093] This embodiment presents a high-precision wave field prediction method based on expert and spatial gating, using low spatial resolution wind field data (u 10 v 10 Using wave data as the sole business input, the system constructs four seasonal expert models and introduces a spatial gating network to achieve adaptive allocation of expert contributions for different seasons and spatial regions. This allows for the rapid generation of high-precision wave element fields for the target sea area without the need to access wave observation, wave reanalysis, or wave numerical model outputs, thereby improving the prediction accuracy and stability under complex spatial distributions and extreme sea conditions.

[0094] Specifically, the high-precision wave field prediction method based on experts and spatial gating in this embodiment has the following effects: (1) Reduced input dependence and error coupling: The business stage only uses low spatial resolution wind field data as input, without relying on wave observation, wave reanalysis or wave forecasting model output, reducing the error superposition caused by the introduction of multiple source elements and improving prediction stability; (2) Explicit modeling of seasonal differences: By learning the seasonal sea state differences through four seasonal expert models, the accuracy loss caused by the compromise of a single model on the annual sample is avoided, and the generalization ability under different seasons is improved; (3) Spatial adaptive fusion: The spatial gating network outputs a grid-level weight map to realize the adaptive allocation of expert contributions for different regions in each season, and enhance the prediction precision under complex nearshore terrain and spatial non-uniform distribution; (4) High computational efficiency and easy business integration: The computational cost in the inference stage is significantly lower than that of traditional nested numerical simulation, which can meet the needs of high-frequency updates and rapid response, and is suitable for the integration of port area, marine engineering and coastal disaster prevention and mitigation business systems.

[0095] To verify the effectiveness of the scheme in this embodiment, the Bohai Sea was used as the target sea area for verification. Data from 2015 to 2022 was obtained for training. The samples from 2015 to 2022 were merged into a training pool and then randomly divided into an internal training set and an internal validation set for training the full data benchmark model and the forecast model. The samples from 2023 were used as the final test set and did not participate in any model parameter updates or learning rate scheduling.

[0096] For example, the prediction comparison between the full data baseline model and the seasonal expert model Figure 2 As shown in the figure, RRmse represents the relative root mean square error of the prediction results of each seasonal model in the four-season subset dataset. Figure 2 It can be seen that there are significant differences between seasons. The RRmse is highest in autumn, followed by summer, then spring and winter, which means that discussing the seasonality of data distribution is necessary for machine learning. It can also be observed that the values ​​in the diagonal (bottom left to top right) are the lowest in each column. Taking spring as an example, the values ​​in the first column from bottom to top are 0.278, 0.280, 0.280, and 0.283, indicating that in the spring sample, the model's prediction relative root mean square error is 0.278, while the prediction errors decrease from summer to autumn to winter. Compared to the baseline model in the first row, the expert prediction errors are further improved in summer and winter, further reflecting the impact of data seasonality on model training. In summary, the heatmap results show that seasonal experts have the lowest prediction errors in their respective seasons, indicating that the seasonal model has learned seasonal-specific characteristics in this season.

[0097] At longitude 37.40°, latitude 122.73°, from May 6th to June 17th, a comparison is made between the prediction results of the forecast model and the monitoring signals. Figure 3 As shown, according to Figure 3 It can be seen that the prediction results of the forecast model represented by red and the supervision signal represented by black have achieved phase matching under multiple fluctuations, and the numerical prediction has also achieved accurate prediction. The calculated root mean square error is 0.1517m.

[0098] In summary, it can be seen that seasonal differences are an issue that cannot be ignored in wave prediction. Seasonal expert models built to address seasonal differences can learn seasonal differences better on the basis of benchmark models. Furthermore, after being combined with gated models, the forecast model can learn wave characteristics from the wind field to make high-precision predictions, which greatly reduces the need for model forecasting.

[0099] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-precision wave field prediction method based on expert and spatial gating, characterized in that, Includes the following steps: Step 1: Determine the target sea area and obtain coastline / land boundary and water depth data; then obtain multi-year low spatial resolution wind field data and wave data of the target sea area from publicly available meteorological reanalysis databases; Step 2: Using the low spatial resolution wind field data and wave data of the target sea area obtained in Step 1, perform simulation calculations using a shallow water wave numerical model to obtain the high spatial resolution wave element field of the target sea area for the corresponding year; the high spatial resolution wave element field includes at least the significant wave height Hs(t). Step 3: Perform time alignment and spatial consistency processing on the multi-year low spatial resolution wind field data of the target sea area obtained in Step 1 and the corresponding year high spatial resolution wave element field of the target sea area obtained in Step 2, and use a unified mask to mark the land area and missing measurement points to construct the full dataset. Step 4: Construct the full-data baseline model and forecast model; the full-data baseline model takes low-spatial-resolution wind field data as input and outputs high-spatial-resolution wave element fields, used for transfer training of the seasonal expert model; the forecast model includes parallel seasonal expert models and spatially gated models; there are four seasonal expert models, corresponding to spring, summer, autumn, and winter, taking low-spatial-resolution wind field data as input and outputting high-spatial-resolution wave element fields; the spatially gated model takes low-spatial-resolution wind field data as input and outputs a weighted graph of the seasonal expert model outputs, with the weighted graph having the same resolution as the seasonal expert model outputs, used to weight the seasonal expert model outputs as the overall output of the forecast model; the loss function is calculated during the training of the full-data baseline model, seasonal expert models, and spatially gated models through the following steps: Step B4.1: Based on the high spatial resolution wave element field of the target sea area obtained in Step 2, statistically analyze the data of each spatial grid point for each year to obtain the 90th quantile threshold map, as shown in the following formula: ； In the formula, t is the time index point, and i and j are the spatial index points; Step B4.2: Calculate the sea state weighting coefficient based on the 90th percentile threshold map, using the following formula: ； Step B4.3: Calculate the loss using the following formula: ； In the formula, The mask is divided into valid points for ocean and invalid points for land. Y(t,i,j) is the prediction model; Y(t,i,j) is the high spatial resolution wave element field, t is the supervision signal, i and j are the spatial index points in the high spatial resolution grid; Step 5: Perform operational forecasts based on the forecast model trained in Step 4. The input to the model is low spatial resolution wind field data for the forecast period.

2. The high-precision wave field prediction method based on expert and spatial gating according to claim 1, characterized in that, In step 1, the low spatial resolution wind field data includes at least two orthogonal components u of the wind speed at a height of 10 m. 10 and v 10 In step 3, the samples in the constructed full dataset include: ； }； In the formula, X(t,n,m) is the low spatial resolution wind field data, which is the input signal, t is the time index point, and n and m are the spatial index points in the low spatial resolution grid; Y(t,i,j) is the high spatial resolution wave element field, which is the monitoring signal, and i and j are the spatial index points in the high spatial resolution grid.

3. The high-precision wave field prediction method based on expert and spatial gating according to claim 1, characterized in that, In step 4, the full-data baseline model, seasonal expert model, and spatial gating model are trained through the following steps: Step A4.1: Train a full-data baseline model using the full dataset; Step A4.2: Divide the entire dataset into four subsets according to time: spring, summer, autumn, and winter; The full-data baseline model trained in step A4.1 is trained using four subsets of spring, summer, autumn and winter respectively, resulting in four seasonal expert models; Step A4.3: Train the forecast model using the full dataset. During training, freeze the parameters of the expert models for four quarters and train only the parameters of the spatial gating model.

4. The high-precision wave field prediction method based on expert and spatial gating according to claim 1, characterized in that, In step 4, the spatial gating model outputs four spatial grid weight maps corresponding to the expert models for each season, as shown in the following formula: ； ； ； In the formula, t is the time index point, i and j are the spatial index points, and k is the seasonal expert model index point.

5. The high-precision wave field prediction method based on expert and spatial gating according to claim 4, characterized in that, In step 4, the formula for the output of the prediction model is as follows: ； In the formula, The output results are for the expert models for the four seasons.

6. The high-precision wave field prediction method based on expert and spatial gating according to claim 1, characterized in that, In step 4, the full data benchmark model and the seasonal expert model include a feature extraction subnetwork and a spatial reconstruction subnetwork. The feature extraction subnetwork is the encoding stage, with a transform network structure, used to extract multi-scale spatial features from low spatial resolution wind field data. The spatial reconstruction subnetwork is the decoding stage, with a convolutional neural network structure, used to generate wave element prediction fields on the target high-precision grid. The spatial gating model is a convolutional neural network.

7. A high-precision wave field prediction system based on hybrid expert and spatial gating, characterized in that, The prediction model trained using a high-precision wave field prediction method based on expert and spatial gating, as described in any one of claims 1 to 6.

Images