Wind speed forecasting system based on deep learning

Abstract

The invention discloses a wind speed forecasting system based on deep learning, and belongs to the technical field of weather forecasting. The system comprises a coupled mode framework which integrates a triple nested architecture of an atmospheric mode WRF, an ocean mode ROMS and a sea wave mode SWAN, and transmits variables through a model coupling tool MCT; the deep learning roughness parameterization module is integrated in an MYNN boundary layer scheme of the atmospheric mode WRF and calculates the sea surface roughness in real time; the full-wave spectrum integration module is embedded into the ocean mode ROMS, receives wave spectrum data provided by the ocean mode SWAN and calculates a Stokes drift profile; and the intelligent correction module is used for carrying out spatial interpolation on the wind speed data output by the mode, correcting errors by combining a LightGBM model, and outputting a final forecast product. According to the method, the deep neural network is embedded into the WRF boundary layer model kernel, the sea surface roughness parameter is updated in real time, the wave evolution process is dynamically responded, and the space-time adaptive capacity of wind field forecasting under the complex sea condition is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and particularly to a wind speed forecasting system based on deep learning. Background Art

[0002] The near-surface wind field forecasting is of great significance for air pollution control, ocean engineering and social activity services. At present, due to problems such as initial field deviation, insufficient resolution and imperfect physical parameterization in numerical weather forecasting models, there are significant deviations in near-surface wind field forecasting.

[0003] At present, traditional sea surface roughness parameterization schemes rely on linear regression or simple non-linear relationships, making it difficult to capture the complex coupling effects between atmospheric and wave parameters, resulting in large deviations in sea surface roughness estimation. In addition, in existing wave-current coupling models, the calculation of Stokes drift is mostly based on the single-spectrum method, ignoring the contribution of the full wave spectrum, resulting in an underestimation of the shear strength in the ocean boundary layer. Although some scholars have tried to optimize the parameterization scheme by combining field observation data, their generalization ability and accuracy are still limited by the limitations of traditional methods. Therefore, there is an urgent need for a high-precision wind field forecasting technical solution that combines deep learning and the full wave spectrum method to solve the above technical problems. Summary of the Invention

[0004] The purpose of the present invention is to provide a wind speed forecasting system based on deep learning, which improves the sea surface roughness parameterization scheme by introducing deep learning algorithms, optimizes the calculation of Stokes drift by combining the full wave spectrum method, and uses a machine learning model to intelligently correct the output of the numerical model, so as to reduce the wind field forecasting error and solve the problems raised in the above background art.

[0005] To solve the above technical problems, the present invention provides a wind speed forecasting system based on deep learning, including a coupling model framework, a deep learning roughness parameterization module, a full wave spectrum integration module and an intelligent correction module.

[0006] Coupling model framework: Integrate a triple nested architecture: the atmospheric model WRF (9km), the ocean model ROMS (3km), and the wave model SWAN (1km), and transfer variables through the model coupling tool MCT.

[0007] The variables include wind stress, sea surface temperature SST, and wave spectrum data.

[0008] Improve the multi-scale coupling accuracy and achieve the full coupling interaction of the atmosphere-ocean-wave. Solve the problem of insufficient resolution of traditional models by gradually reducing the scale, and reduce the initial field deviation.

[0009] Deep learning roughness parameterization module: Integrated into the MYNN boundary layer scheme of the atmospheric model WRF for real-time calculation of sea surface roughness.

[0010] The deep learning roughness parameterization module uses a deep neural network (DNN), with an input layer (7 parameters: wind speed WSPD, wind direction DIRwnd, friction velocity u*, significant wave height Hs, wave speed Cp, wave crest direction DIRwav, and the angle between the wave crest direction and the wind direction DIRwav_wnd), 3 fully connected hidden layers (16→32→16 neurons, activated by the hyperbolic tangent function Tanh), and a single output layer.

[0011] The training data is generated based on multi-source observational data and numerical model simulations. The integration method is to embed the WRF-MYNN boundary layer scheme code and dynamically update the air-sea fluxes. The output layer is a single neuron that outputs the predicted value of the sea surface roughness.

[0012] The deep learning roughness parameterization module dynamically updates the calculation of air-sea interface fluxes in real time by modifying the MYNN boundary layer scheme code of the atmospheric model WRF.

[0013] Deep learning sea surface roughness parameterization: uses a deep neural network (DNN) to fuse multi-source parameters such as wind speed and waves, analyzes the non-linear relationship between the atmosphere and waves, and the root mean square error (RMSE) of the test set is reduced by more than 50% compared to the traditional scheme.

[0014] By capturing the non-linear coupling relationship between wave parameters and atmospheric conditions, it improves the calculation accuracy of sea surface roughness, replaces the traditional linear regression scheme, and reduces the wind field prediction bias caused by parameterization errors.

[0015] The full wave spectrum integration module: embeds the ocean model ROMS and receives the wave spectrum data provided by the wave model SWAN to calculate the Stokes drift profile.

[0016] First, calculate the Stokes drift spectrum of the discrete frequency-direction spectrum through the SWAN model. Second, couple the MCT to transfer data to ROMS and calculate the vertical Stokes drift profile layer by layer. Finally, for waves above the cut-off wave number, use the deep sea dispersion relation formula to process.

[0017] The specific implementation steps include:

[0018] S1. Through the SWAN model, use the radial integration formula 1 to calculate the Stokes drift spectrum of the discrete frequency-direction spectrum:

[0019]

[0020] Among them, the dynamic spectral density N(ω,θ) = E(ω,θ) / ω, which is calculated and generated by the SWAN model. E(ω,θ) is the two-dimensional wave spectrum, which gives the energy distribution of the waves at the angular frequency ω and the propagation direction θ, k is the wave number, Vθ is the calculated azimuth angle interval, and N θ is the number of azimuth angle equal parts.

[0021] Full-spectrum Stokes drift calculation: Based on the discrete frequency-direction spectrum integration and high-frequency tail spectrum optimization, it completely quantifies the influence of the wave spectrum on the vertical shear of the ocean boundary layer and improves the physical accuracy of the drift intensity calculation.

[0022] S2. Transfer the Stokes drift spectrum generated by formula (1) to the ROMS model through the MCT coupler, and calculate the Stokes drift profile layer by layer using formula (2) at the vertical calculation grid depth of ROMS:

[0023]

[0024] where z is the model height, h is the water depth, u ss is the Stokes drift spectrum, Vz is the depth difference between two model layers, N f0 is the critical frequency of high and low frequency waves, N f is the cut-off frequency.

[0025] S3. For waves with a wave number k greater than the cut-off wave number k c = ω 2 / g, when the deep-sea dispersion relationship is satisfied, use formulas (3) and (4) to calculate:

[0026]

[0027] where g is the acceleration of gravity, a ± = -2(z ± Vz / 2)k c , and erfc is the complementary error function.

[0028] Avoid the spectral limitation of the single-spectrum method through full-spectrum integration, improve the integrity of Stokes drift calculation, accurately quantify the wave-induced vertical shear effect in the ocean boundary layer, and improve the wave-mass transfer coupling process.

[0029] Intelligent correction module: Perform spatial interpolation on the wind speed data output by the model, correct the error in combination with the LightGBM model, and output the final forecast product. The implementation steps include:

[0030] S1. Interpolate the grid point wind speed output by the model to the measured stations to construct an error data set.

[0031] S2. Divide the training set, validation set, and test set according to the time and space dimensions.

[0032] S3. Optimize the hyperparameters α and β of the hybrid loss function through grid search.

[0033] The mathematical expression of the hybrid loss function is:

[0034] Loss(y pred ,y true ,L,H) = αMSE(ypred , y true ) + β[ReLU(L - y pred ) 2 + ReLU(y pred - H) 2 (5)

[0035] Among them, ypred is the wind speed value predicted by the numerical model, ytrue is the true observed wind speed value, L and H are respectively the lower and upper limits of the wind speed interval corresponding to the true level, and α and β are hyperparameters.

[0036] The hyperparameters α and β are optimized by grid search. The loss function consists of two parts:

[0037] The mean squared error term MSE: αMSE(y pred , y true ), which is used to ensure that the predicted wind speed is close to the true value;

[0038] The interval penalty term: β[ReLU(L - y pred ) 2 + ReLU(y pred - H) 2 , which is used to impose a penalty when the predicted value exceeds the wind speed interval [L, H] corresponding to the true level, forcing the predicted value to fall within the correct level; by adjusting the weight of β, the boundary constraint is strengthened.

[0039] Hybrid constraint intelligent correction: Design a loss function (Equation 5) that combines the mean squared error (MSE) and the interval penalty term. By dynamically weighting (β > α), force the predicted value to match the true wind speed level interval, reducing the RMSE of the 10 - meter - height wind speed forecast from 2.1777 to 2.0081, taking into account both numerical accuracy and level accuracy.

[0040] Combine the physical model and statistical learning to reduce the systematic bias of the model. The hybrid loss function balances numerical accuracy and level constraints, improving the usability of the forecast results.

[0041] Through the integration of high - resolution coupling of numerical models, deep - learning parameterization substitution, full - wave spectrum integration algorithms, and machine - learning post - processing correction, a closed - loop optimization system is formed, significantly improving the spatio - temporal accuracy and physical consistency of the near - surface wind field forecast.

[0042] Compared with the existing technology, the beneficial effects achieved by the present invention are:

[0043] Breaking through the limitations of traditional parameterization: Traditional sea surface roughness schemes rely on linear regression or simple non-linear models, making it difficult to characterize the complex interactions between the ocean and the atmosphere. This scheme captures the non-linear correlations of multiple parameters such as wind speed and waves through a deep neural network, enabling high-dynamic sea-air flux calculations and fundamentally solving the problem of insufficient generalization of traditional methods.

[0044] Analysis of the contribution of the full spectrum of waves: Existing Stokes drift calculations are mostly based on the single-spectrum method, ignoring the cumulative effect of the full-wave spectrum energy distribution on boundary layer shear. This scheme reconstructs the vertical drift profile through discrete frequency-direction spectrum integration, fully reflecting the impact of wave energy distribution on ocean mass transfer and avoiding the systematic underestimation of drift intensity by traditional methods.

[0045] Multi-mode tightly coupled interaction: Traditional wave-current coupling models often produce errors due to mismatched resolutions or delays in the transfer of physical quantities. This scheme achieves high-frequency synchronous interaction between the atmosphere, ocean, and wave models through a triple-nested architecture (WRF-ROMS-SWAN), ensuring real-time two-way feedback of key parameters such as wind stress and wave spectra and enhancing the overall coordination of the system.

[0046] Physical-data dual-driven correction: Existing error correction methods are mostly limited to pure statistical models and are prone to deviating from physical laws. This scheme proposes a hybrid loss function that integrates numerical accuracy constraints and hierarchical boundary physical rules in the LightGBM model, absorbing observational data information while maintaining the consistency between the forecast results and the true meteorological grades through interval penalty terms, achieving physically interpretable intelligent correction.

[0047] Dynamic adaptive boundary optimization: Compared with traditional fixed parameterization schemes, this technology embeds a deep neural network into the WRF boundary layer model kernel, updates the sea surface roughness parameters in real time, dynamically responds to the wave evolution process, and significantly improves the spatio-temporal adaptability of wind field forecasts in complex sea conditions. Description of the Drawings

[0048] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0049] Figure 1 is a coupling mode system framework diagram of a wind speed forecasting system based on deep learning according to the present invention.

[0050] Figure 2 is a schematic diagram of the triple-nested area of a wind speed forecasting system based on deep learning according to the present invention.

[0051] Figure 3 is a deep neural network structure diagram of a wind speed forecasting system based on deep learning according to the present invention.

[0052] Figure 4 This is a scatter plot for comparing sea surface roughness predictions of a wind speed prediction system based on deep learning according to the present invention.

[0053] Figure 5 This is a flowchart for intelligent correction of wind speed of a wind speed prediction system based on deep learning according to the present invention.

[0054] Figure 6 This is a distribution map of wind speed RMSE of a wind speed prediction system based on deep learning according to the present invention. Specific embodiments

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0056] The present invention provides a wind speed prediction system based on deep learning, including a coupled model framework, a deep learning roughness parameterization module, a full-spectrum integration module, and an intelligent correction module.

[0057] Please refer to Figure 1 , the coupled model system framework diagram shows the data interaction between WRF, ROMS, and SWAN.

[0058] Coupled model integrated triple-nested architecture: the atmospheric model WRF (9 km), the ocean model ROMS (3 km), and the wave model SWAN (1 km), transfer variables through the model coupling tool MCT.

[0059] The variables include wind stress, sea surface temperature SST, and wave spectrum data.

[0060] Please refer to Figure 2 , a schematic diagram of the coastal triple-nested area (resolution 9 km, 3 km, 1 km).

[0061] Improve the multi-scale coupling accuracy and achieve full-coupling interaction of the atmosphere-ocean-wave. Solve the problem of insufficient resolution of traditional models by gradually downscaling and reduce the initial field deviation.

[0062] Deep learning roughness parameterization module: integrated into the MYNN boundary layer scheme of the atmospheric model WRF, used to calculate the sea surface roughness in real time.

[0063] Please refer to Figure 3 , the deep neural network structure diagram, the input layer contains 7 features, and the output is the sea surface roughness.

[0064] The deep learning roughness parameterization module uses a deep neural network (DNN), with an input layer (7 parameters: wind speed WSPD, wind direction DIRwnd, friction velocity u*, significant wave height Hs, wave speed Cp, wave crest direction DIRwav, and the angle between the wave crest direction and the wind direction DIRwav_wnd), 3 fully connected hidden layers (16→32→16 neurons, activated by the hyperbolic tangent function Tanh), and a single output layer.

[0065] The training data is generated based on multi-source observational data and numerical model simulations. The integration method is to embed the WRF-MYNN boundary layer scheme code and dynamically update the air-sea fluxes. The output layer is a single neuron that outputs the predicted value of the sea surface roughness.

[0066] Please refer to Figure 4 , the scatter plot of the sea surface roughness prediction comparison between the traditional scheme and the scheme of the present invention, showing that the new scheme is significantly closer to the observed values.

[0067] The deep learning roughness parameterization module dynamically updates the calculation of the air-sea interface fluxes in real time by modifying the MYNN boundary layer scheme code of the atmospheric model WRF.

[0068] Deep learning sea surface roughness parameterization: Using a deep neural network (DNN) to fuse multi-source parameters such as wind speed and waves, analyze the non-linear relationship between the atmosphere and waves, and the root mean square error (RMSE) of the test set is reduced by more than 50% compared with the traditional scheme.

[0069] By capturing the non-linear coupling relationship between wave parameters and atmospheric conditions, improving the calculation accuracy of the sea surface roughness, replacing the traditional linear regression scheme, and reducing the wind field prediction deviation caused by parameterization errors.

[0070] Full wave spectrum integration module: Embedded in the ocean model ROMS, receiving wave spectrum data provided by the wave model SWAN to calculate the Stokes drift profile.

[0071] First, calculate the Stokes drift spectrum of the discrete frequency-direction spectrum through the SWAN model. Secondly, couple the MCT to transfer data to ROMS and calculate the vertical Stokes drift profile layer by layer. Finally, for waves above the cut-off wave number, use the deep sea dispersion relation formula to process.

[0072] The specific implementation steps include:

[0073] S1. Through the SWAN model, use the radial integration formula 1 to calculate the Stokes drift spectrum of the discrete frequency-direction spectrum:

[0074]

[0075] Among them, the dynamic spectral density N(ω,θ) = E(ω,θ) / ω, which is generated by the SWAN model. E(ω,θ) is the two-dimensional wave spectrum, which gives the energy distribution of the waves at the angular frequency ω and the propagation direction θ. k is the wave number, Vθ is the calculated azimuth angle interval, Nθ is the number of equal azimuth angles.

[0076] Full-wave spectral Stokes drift calculation: Based on the discrete frequency-direction spectrum integration and the high-frequency tail spectrum optimization, the influence of the complete wave spectrum on the vertical shear of the ocean boundary layer is quantified, and the physical accuracy of the drift intensity calculation is improved.

[0077] S2. Transfer the Stokes drift spectrum generated by formula 1 to the ROMS model through the MCT coupler, and calculate the Stokes drift profile layer by layer with formula 2 at the vertical calculation grid depth of ROMS:

[0078]

[0079] Among them, z is the model height, h is the water depth, u ss is the Stokes drift spectrum, Vz is the depth difference between two model layers, N f0 is the critical frequency of high and low frequency waves, N f is the cut-off frequency.

[0080] S3. For waves with a wave number k greater than the cut-off wave number k c = ω 2 / g, when the deep-sea dispersion relationship is satisfied, use formulas 3 and 4 for calculation:

[0081]

[0082] Among them, g is the acceleration of gravity, a ± = -2(z ± Vz / 2)k c , and erfc is the complementary error function.

[0083] Avoid the spectral limitation of the single-spectrum method through full-wave spectral integration, improve the integrity of the Stokes drift calculation, accurately quantify the wave-induced vertical shear effect in the ocean boundary layer, and improve the wave-mass transfer coupling process.

[0084] Intelligent correction module: Perform spatial interpolation on the wind speed data output by the model, correct the error in combination with the LightGBM model, and output the final forecast product. The implementation steps include:

[0085] S1. Interpolate the grid point wind speed output by the model to the measured stations to construct an error data set.

[0086] S2. Divide the training set, validation set and test set according to the time and space dimensions.

[0087] S3. Optimize the hyperparameters α and β of the hybrid loss function through grid search.

[0088] Please refer to Figure 5 , the intelligent correction flowchart of wind speed, including data interpolation, error training and model application steps.

[0089] The mathematical expression of the hybrid loss function is:

[0090] Loss(y pred ,y true ,L,H) = αMSE(y pred ,y true ) + β[ReLU(L - y pred ) 2 + ReLU(y pred - H) 2 (5)

[0091] Where ypred is the wind speed value predicted by the numerical model, ytrue is the true observed wind speed value, L and H are the lower and upper limits of the wind speed interval corresponding to the true level respectively, and α and β are hyperparameters.

[0092] The hyperparameters α and β are optimized through grid search, and the loss function consists of two parts:

[0093] The mean square error term MSE: αMSE(y pred ,y true ), which is used to ensure that the predicted wind speed is close to the true value;

[0094] The interval penalty term: β[ReLU(L - y pred ) 2 + ReLU(y pred - H) 2 , which is used to impose a penalty when the predicted value exceeds the wind speed interval [L, H] corresponding to the true level, forcing the predicted value to fall within the correct level; by adjusting the weight of β, the boundary constraint is strengthened.

[0095] Hybrid constraint intelligent correction: Design a loss function (Formula 5) that combines the mean square error (MSE) and the interval penalty term. By using dynamic weights (β > α), force the predicted value to match the true wind speed level interval, reducing the RMSE of the 10-meter height wind speed forecast from 2.1777 to 2.0081, taking into account both numerical accuracy and level accuracy.

[0096] Please refer to Figure 6 , the RMSE distribution map of the 10-meter wind speed in the offshore area, comparing the error improvement effects of the traditional scheme and the new scheme.

[0097] Combining physical models and statistical learning to reduce the systematic bias of the model. The hybrid loss function balances numerical accuracy and level constraints, improving the usability of the forecast results.

[0098] Through the integration of high-resolution coupling of numerical models, replacement of deep learning parameterization, full-wave spectrum integration algorithm, and machine learning post-processing correction, a closed-loop optimization system is formed, significantly improving the spatio-temporal accuracy and physical consistency of near-surface wind field forecasts.

[0099] Example 1:

[0100] 1. Data preparation and preprocessing:

[0101] Use NCEP GFS data as initial and boundary conditions, HYCOM / NCODA provides the ocean initial field, and WW3 provides the wave boundary conditions.

[0102] Perform quality control on the ASIT dataset, normalize it after removing outliers, and use it to train the deep neural network.

[0103] 2. Model coupling and operation:

[0104] The WRF model configures the WSM6 microphysical scheme, RRTM long-wave radiation scheme, RUC land surface scheme, and MYNN boundary layer scheme, with 58 vertical layers.

[0105] For the boundary conditions of temperature, salinity, and 3D velocity field, use the radiation-nudging condition, adopt the Chapman boundary condition for the free surface, and use the Flather scheme for the vertically averaged velocity.

[0106] The mixing scheme adopts the Generic Length Scale scheme. SWAN calculates wave growth through the Komen formula.

[0107] The full-wave spectrum method calculates the Stokes drift by integrating the discrete frequency-direction spectrum (Formula 2), and uses the deep-sea approximation formula (Formulas 3-4) for the high-frequency part.

[0108] 3. Implementation of intelligent correction:

[0109] Interpolate the model-predicted wind speed to the measured stations, construct an error dataset, and divide it into a training set, a validation set, and a test set.

[0110] Use the LightGBM model to optimize error prediction, and introduce Formula 5 as the loss function:

[0111] Set the wind speed grade interval according to the Beaufort wind scale division, and optimize α and β through grid search (e.g., α = 1.0, β = 2.0).

[0112] Monitor the misjudgment rate of wind speed grades on the validation set to ensure that the CER is reduced by at least 20% in strong wind scenarios (e.g., y true ≥15m).

[0113] 4. Verification and Output: Compare the CCMP data to evaluate the improvement effect of RMSE, and generate the final revised wind speed forecast product.

[0114] It should be noted that in this document, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0115] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent substitution on some of the technical features. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A wind speed forecasting system based on deep learning, characterized by: The system includes: a coupled mode framework, a deep learning roughness parameterization module, a full spectrum integration module and an intelligent correction module; Coupled model framework: A triple nested architecture integrating the atmospheric model WRF, the ocean model ROMS, and the wave model SWAN, with variables transferred through the model coupling tool MCT; Deep Learning Roughness Parameterization Module: MYNN boundary layer scheme integrated into the atmospheric model WRF for real-time calculation of sea surface roughness; Full spectrum integration module: embedded in the ocean model ROMS, receives the wave spectrum data provided by the wave model SWAN to calculate the Stokes drift profile; Intelligent correction module: performs spatial interpolation on the wind speed data output by the model, combines the LightGBM model to correct the error, and outputs the final forecast product.

2. A wind speed forecasting system based on deep learning according to claim 1, characterized in that: The resolutions of the triple nested architecture are 9 km, 3 km and 1 km respectively, and the variables include wind stress, sea surface temperature SST, and wave spectrum data.

3. A wind speed forecasting system based on deep learning according to claim 1, characterized in that: The deep learning roughness parameterization module adopts a deep neural network DNN, and the network structure is: training data source, input layer, hidden layer and output layer.

4. A wind speed forecasting system based on deep learning according to claim 3, characterized in that: The input layer contains characteristic parameters such as wind speed WSPD, wind direction DIRwnd, friction speed u*, significant wave height Hs, wave speed Cp, crest direction DIRwav and the angle DIRwav_wnd between the crest direction and wind direction; the hidden layer consists of three fully connected layers, the number of neurons is 16, 32, and 16 respectively, and the activation function is hyperbolic tangent Tanh; the output layer is a single neuron, which outputs the predicted value of sea surface roughness.

5. The wind speed forecasting system based on deep learning according to claim 1, characterized in that: The deep learning roughness parameterization module modifies the MYNN boundary layer scheme code of the atmospheric model WRF to dynamically update the air-sea interface flux calculation in real time.

6. A wind speed forecasting system based on deep learning according to claim 1, characterized in that: The full spectrum integration module implementation steps include: S1. Calculate the Stokes drift spectrum of the discrete frequency-direction spectrum using radial integration formula 1 through the SWAN mode: Among them, the dynamic spectrum density N(ω,θ) = E(ω,θ) / ω is calculated and generated by the SWAN mode; E(ω,θ) is a two-dimensional wave spectrum, which gives the energy distribution of the wave at the angular frequency ω and the propagation direction θ, k is the wave number, Vθ is the calculation azimuth interval, N θ is equal parts of azimuth; S2. The Stokes drift spectrum generated by Formula 1 is transferred to the ROMS mode through the MCT coupler, and the Stokes drift profile is calculated hierarchically using Formula 2 at the vertical calculation grid depth of ROMS: Where z is the model height, h is the water depth, and u ss is the Stokes drift spectrum, Vz is the depth difference between the two mode layers, N f0 is the critical frequency of high and low frequency waves, N f is the cut-off frequency; S3, for the wave number greater than the cutoff k c =ω 2 / g waves, when the deep sea dispersion relation is satisfied, use formulas 3 and 4 to calculate: Where g is the acceleration due to gravity, a ± =-2(z±Vz / 2)k c , erfc is the residual function.

7. The wind speed forecasting system based on deep learning according to claim 1, characterized in that: The intelligent correction module implementation steps include: S1, interpolate the grid point wind speed output by the model to the measured site to construct an error data set; S2, divide the training set, validation set and test set according to the time and space dimensions; S3. Optimize the hyperparameters α and β of the hybrid loss function through grid search.

8. A wind speed forecasting system based on deep learning according to claim 7, characterized in that: The mathematical expression of the hybrid loss function is: Loss(and pred ,and true ,L,H)=αMSE(y pred ,and true )+β[ReLU(Ly pred ) 2 +ReLU(and pred -H) 2 ](5) Among them, y pred is the wind speed value predicted by the numerical model, y true is the actual observed wind speed value, L and H are the lower and upper limits of the wind speed range corresponding to the actual level, and α and β are hyperparameters.

9. A wind speed forecasting system based on deep learning according to claim 8, characterized in that: The hyperparameters α and β in the hybrid loss function are optimized by grid search. The loss function consists of two parts: Mean square error MSE: αMSE(y pred ,y true ), used to ensure that the predicted wind speed is close to the true value; Interval penalty term: β[ReLU(Ly pred ) 2 +ReLU(y pred -H) 2 ] is used to impose a penalty when the predicted value exceeds the wind speed interval [L,H] corresponding to the true level, forcing the predicted value to fall within the correct level; by adjusting the weight of β, the boundary constraint is strengthened.

Images