A method for predicting arctic atmosphere coupling by automatically fusing sea ice concentration and thickness

Abstract

The application provides an automatic Arctic atmosphere coupling prediction method of fusing sea ice density and thickness, and belongs to the technical field of meteorology. First, GFS background field data are acquired and corrected by applying an ice-sea heat exchange correction equation, then sea ice density and thickness information are extracted and physical consistency verification is carried out based on a thermodynamic equilibrium equation. Then, multi-element fusion is carried out through an Arctic ice-sea assimilation model, sea ice surface temperature is calculated by applying a polar sea ice heat flux equation, and a sea ice density thickness boundary layer interaction equation set is established to describe the modulation effect of sea ice on the boundary layer. The core step is to apply a feature fusion model to deeply fuse various sea ice features, to adaptively adjust the weight based on the ice-air interface balance index, to generate an optimized model initial field and boundary condition, and finally to set a parameter scheme to execute an integral program to obtain a prediction result, so that the technical problem of low prediction accuracy caused by the physical inconsistency of sea ice information is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of meteorology, and in particular, relates to an automatic Arctic atmospheric coupling prediction method fusing sea ice concentration and thickness. BACKGROUND

[0002] Arctic numerical weather prediction is an important field of meteorological science research, and is of great significance for polar scientific research, route planning and climate change research. Traditional Arctic weather prediction is mainly based on the background field data provided by the global forecast system (GFS), combined with limited observation data for assimilation before inputting into the numerical model. The existing technology usually inputs the sea ice concentration as the boundary condition into the model, and simulates the influence of sea ice on the atmospheric boundary layer through a simplified sea ice parameterization scheme.

[0003] However, the traditional technology has obvious defects in processing sea ice information. On the one hand, the existing method usually only considers the sea ice concentration information, and ignores the key role of sea ice thickness, which leads to the inability of the model to accurately describe the sea ice heat capacity and its modulation effect on the atmospheric boundary layer; on the other hand, even if the sea ice concentration and thickness information are introduced at the same time, the physical consistency between them cannot be guaranteed, and they are often provided by different data sources and lack effective physical constraint mechanisms.

[0004] This technical problem of insufficient physical consistency between sea ice concentration and thickness information leads to significant simulation deviation of the Arctic atmospheric boundary layer structure, especially in the rapidly changing sea ice area, the calculation error of the boundary layer turbulent flux accumulates, which further affects the simulation accuracy of the large-scale circulation, and finally causes the insufficient accuracy of the Arctic regional weather prediction, which cannot meet the high-precision demand of polar scientific research and route planning. SUMMARY

[0005] Therefore, the present application provides an automatic Arctic atmospheric coupling prediction method fusing sea ice concentration and thickness, which can solve the technical problem of low accuracy of Arctic atmospheric coupling prediction caused by insufficient physical consistency between sea ice concentration and thickness information in the prior art.

[0006] The present application is implemented as follows: The present application provides an automatic Arctic atmospheric coupling prediction method fusing sea ice concentration and thickness, which includes obtaining and correcting global GFS background field data, extracting and verifying sea ice concentration and thickness information, performing multi-element fusion to generate a background field, calculating sea ice surface temperature by applying a polar sea ice heat flux equation, calculating boundary layer structure by establishing a sea ice concentration and thickness boundary layer interaction equation set, performing deep fusion of sea ice concentration, sea ice thickness, sea ice surface temperature and sea ice modulated boundary layer structure by applying a feature fusion model, adaptively adjusting the weight of each element in the output vector of the feature fusion model based on the ice-air interface balance index, generating an optimized model initial field and boundary condition, executing an integral program to obtain a prediction result and performing post-processing.

[0007] The step of obtaining global GFS background field data and correcting, specifically, obtaining global GFS background field data, the GFS background field data containing 41 layers in the vertical direction, a time interval of 3 hours, a single prediction duration of 168 hours, and applying an ice-sea heat exchange correction equation to preliminarily correct the boundary layer parameters of the sea ice region in the GFS background field data.

[0008] The step of extracting and verifying sea ice density and sea ice thickness information, specifically, extracting sea ice density and sea ice thickness information from the GFS background field data, verifying the physical consistency of the sea ice density and the sea ice thickness information based on the sea ice thermodynamic equilibrium equation, and rewriting the verified sea ice density and sea ice thickness information into an intermediate file recognizable by a model pre-processing program.

[0009] The step of performing multi-element fusion to generate a background field, specifically, setting a model pre-processing parameter file, executing a terrain processing program, a format rewriting program, and an interpolation program, and performing multi-element fusion of meteorological elements, sea ice density, and sea ice thickness information through an Arctic ice-sea assimilation model to generate a model background field with higher physical consistency.

[0010] The step of applying a feature fusion model to deeply fuse sea ice density, sea ice thickness, sea ice surface temperature, and sea ice modulated boundary layer structure, specifically, applying a feature fusion model to deeply fuse sea ice density, sea ice thickness, sea ice surface temperature, and sea ice modulated boundary layer structure, and adaptively adjusting the weight of each element in the output vector of the feature fusion model based on the ice-air interface balance index.

[0011] The step of executing an integration program to obtain prediction results and performing post-processing, specifically, setting physical parameterization schemes and sea ice parameterization schemes, executing an integration program to obtain future 7-day weather prediction results using the model initial field and the boundary conditions, and evaluating the uncertainty influence of sea ice density and sea ice thickness on the weather prediction results through ensemble prediction technology.

[0012] The step of executing an integration program to obtain prediction results and performing post-processing, further including post-processing the weather prediction results, calculating prediction skill scores by comparing historical observation data, and generating standardized prediction products according to the application requirements of the Arctic region.

[0013] The GFS background field data refers to grid-based prediction data generated by the Global Forecast System, containing atmospheric, oceanic, terrestrial, and sea ice multi-layer elements, and the GFS background field data provides initial information for subsequent processing.

[0014] The ice-air interface balance index is used to evaluate the balance state of heat, momentum and water vapor exchange in the interaction process between sea ice and atmosphere, and the ice-air interface balance index is obtained by calculating the product of the sea ice density and the sea ice thickness divided by the surface net radiation flux. When the ice-air interface balance index is less than 0.3, it indicates that the exchange process is weak, when the ice-air interface balance index is greater than 0.7, it indicates that the exchange process is strong, and when the ice-air interface balance index is between the two, it indicates that the exchange process is moderate.

[0015] The feature fusion model is a deep learning model based on a multi-head self-attention mechanism and a spatio-temporal graph convolution, which is used to optimize the fusion of the sea ice density, the sea ice thickness, the sea ice surface temperature and the sea ice modulated boundary layer structure.

[0016] The present application realizes the deep fusion of sea ice information and atmospheric prediction through a multi-step systematic process. The method first performs physical consistency verification on the sea ice density and thickness information based on the sea ice thermodynamic balance equation, and then performs multi-element fusion on the verified sea ice information and meteorological elements through the Arctic ice-ocean assimilation model to generate a mode background field with higher physical consistency. By establishing a sea ice density and thickness boundary layer interaction equation set, the present application realizes accurate description of the sea ice modulated boundary layer structure, and applies a feature fusion model to deep fusion of the sea ice density, thickness, surface temperature and boundary layer structure, and adjusts the feature weight adaptively based on the ice-air interface balance index, thereby solving the problem of insufficient physical consistency of sea ice information in traditional methods. The feature fusion model considers both prediction accuracy and physical consistency constraints during the training process, ensuring that the fusion result not only conforms to the physical law but also has high precision characteristics. The present application successfully solves the technical problem of low accuracy of atmospheric coupling prediction in the Arctic region caused by insufficient physical consistency of sea ice density and thickness information, and provides a more reliable method for weather prediction in the Arctic region. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the method of the present application.

[0018] Figure 2 The Taylor diagram of the prediction elements in Example 2.

[0019] Figure 3 The prediction result chart of each element in the Arctic region on August 11, 2021 in Example 2, including four subgraphs, (a) 2m temperature prediction chart, (b) 2m dew point temperature prediction chart, (c) sea level pressure prediction chart, (d) 10m wind speed prediction chart. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0021] As Figure 1 shown is a flow chart of an automatic Arctic atmosphere coupling prediction method provided by the application, the method comprises the following steps:

[0022] S01, obtain GFS background field data from a global weather center server through an API interface, including multiple layers of air pressure from the ground to the upper air, and apply a Arctic sea ice heat exchange correction equation to preliminarily correct the boundary layer parameters of the sea ice region, and improve the description of the polar boundary layer parameters;

[0023] S02, extract sea ice density and thickness information from the GFS background field data, and perform physical consistency verification based on the sea ice thermodynamic equilibrium equation. When the sea ice thickness deviates too much from the theoretical value, adjust it through an iterative method. Rewrite the verified data into an intermediate file that can be recognized by the model preprocessing program;

[0024] S03, set the model preprocessing parameter file, set the projection mode and resolution suitable for the Arctic region, execute the terrain processing program, format rewriting program and interpolation program, and generate a higher physically consistent model background field by using the Arctic ice sea assimilation model based on four-dimensional variational assimilation technology;

[0025] S04, configure the sea ice thermodynamic condition parameters and the sea-air momentum exchange coefficient in the model parameters, calculate the sea ice surface temperature by applying the polar sea ice heat flux equation, and determine the ice-air interface temperature state by comprehensively considering various radiation and heat flux processes;

[0026] S05, establish a sea ice density and thickness boundary layer interaction equation set, consider the modulation effect of sea ice on the atmospheric boundary layer structure, calculate the sea ice surface roughness, thermal stability parameter, boundary layer turbulence flux, boundary layer height and temperature and humidity gradient, and generate a sea ice modulated boundary layer structure;

[0027] S06, apply a feature fusion model to deeply fuse sea ice density, thickness, surface temperature and boundary layer structure, adaptively adjust the feature weight based on the ice-air interface balance index, optimize the multi-source feature fusion process, and generate high-quality model initial field and boundary conditions;

[0028] S07, set the physical parameterization scheme and the sea ice parameterization scheme, use the initial field and boundary conditions generated in the previous step to execute the integral program to obtain the weather prediction result, and evaluate the uncertainty influence of the sea ice parameters on the prediction result through the ensemble prediction technology;

[0029] S08, post-process the weather prediction result, including error correction and product generation, calculate the prediction skill score by comparing historical observation data, and generate standardized prediction products according to the application requirements of the Arctic region.

[0030] wherein the GFS background field data refers to the global forecast system generated grid-based forecast data containing atmospheric, ocean, land surface and sea ice multi-layer elements, and the GFS background field data provides initial information for subsequent processing.

[0031] wherein the sea ice concentration refers to the percentage of the area covered by sea ice in a unit sea area, the value range is 0 to 1, 0 represents no sea ice, and 1 represents complete coverage by sea ice, and the sea ice concentration affects the sea-air heat exchange process.

[0032] wherein the sea ice thickness refers to the vertical distance from the bottom to the top of the sea ice, the unit is meter, and the sea ice thickness affects the thermal inertia and conduction characteristics of the sea ice.

[0033] wherein the sea ice heat exchange correction equation is used to correct the boundary layer parameter deviation of the standard atmospheric model under the condition of polar sea ice, the input includes the standard calculated boundary layer turbulent flux, the sea ice influence factor, the sea ice concentration and the sea ice thickness, and the output is the corrected boundary layer turbulent flux, and the corrected boundary layer turbulent flux is used to optimize the initial conditions of the model.

[0034] wherein the sea ice thermodynamic equilibrium equation is a basic equation describing the sea ice heat change and phase change process, the input parameters include the sea ice thermal conductivity, the sea ice salinity, the sea ice surface temperature, the sea ice bottom temperature and the sea ice internal temperature profile, and the output parameters are the sea ice thickness change rate and the internal temperature distribution state, and the sea ice thickness change rate and the internal temperature distribution state are used to verify the physical consistency of the sea ice concentration and the sea ice thickness.

[0035] wherein the polar sea ice heat flux equation is used to calculate the sea ice surface temperature and the ice-air interface energy exchange process, and accurately describes the thermal equilibrium state of the Arctic sea ice region. The equation is based on the principle of energy balance, and comprehensively considers the solar shortwave radiation, atmospheric longwave radiation, sea ice surface emission radiation, and heat, latent heat and conduction heat exchange and other physical processes. The input includes sea ice albedo, downward shortwave radiation flux, sea ice emissivity, downward longwave radiation flux, Stefan-Boltzmann constant, near-surface wind speed, near-surface air temperature, near-surface specific humidity, saturated specific humidity, sea ice thermal conductivity, sea water freezing temperature and sea ice thickness, etc. Physical quantities. The output is the sea ice surface temperature, which is a key physical parameter obtained by solving the energy balance equation, and directly affects the ice-air interface heat exchange process.

[0036] The sea ice density thickness boundary layer interaction equation set is used to describe the modulation effect of sea ice on the atmospheric boundary layer structure, establish a quantitative relationship between sea ice characteristics and atmospheric boundary layer parameters, and realize sea ice-atmosphere coupling simulation. The equation set considers the influence of ice ridge structure on surface roughness and the regulation of atmospheric thermal stability on turbulent exchange process. The input includes sea ice thickness, sea ice density, gravitational acceleration, virtual potential temperature and its vertical difference, height difference, wind speed difference, air density, drag coefficient, near-surface wind speed, air constant-pressure specific heat capacity, sensible heat exchange coefficient and near-surface air temperature. The output is sea ice surface roughness, boundary layer height, temperature gradient and humidity gradient distribution in the boundary layer, boundary layer turbulent flux, including momentum flux, sensible heat flux and latent heat flux, which are the basic data for describing the sea ice-atmosphere interaction process and constructing the boundary layer parameterization scheme.

[0037] The ice-air interface balance index is used to evaluate the balance state of heat, momentum and water vapor exchange in the sea ice and atmosphere interaction process. The ice-air interface balance index is obtained by calculating the product of the sea ice density and the sea ice thickness divided by the surface net radiation flux. When the ice-air interface balance index is less than 0.3, it indicates that the exchange process is weak, and when it is greater than 0.7, it indicates that the exchange process is strong, and when it is between the two, it indicates that the exchange process is moderate. The ice-air interface balance index is used to adjust the weight of each element in the feature fusion model output vector.

[0038] The feature fusion model is a deep learning model based on multi-head self-attention mechanism and spatio-temporal graph convolution, which is used to optimize the fusion of sea ice density, sea ice thickness, sea ice surface temperature and sea ice modulated boundary layer structure. The specific structure of the feature fusion model is that the input layer receives the feature data of sea ice density, sea ice thickness, sea ice surface temperature and sea ice modulated boundary layer structure, captures the correlation between different variables through the multi-head self-attention layer, then extracts the physical relationship pattern in the local region through the spatio-temporal graph convolution layer, then integrates the feature representation of different spatial scales through the multi-scale feature fusion layer, and finally outputs the optimized feature representation through the full connection layer. The whole network adopts residual connection structure to ensure stable gradient propagation and complete feature information preservation.

[0039] The steps of establishing the training data set of the feature fusion model include collecting historical Arctic detection satellite data and in-situ observation data, normalizing according to spatial resolution and temporal resolution, extracting key features such as sea ice density, sea ice thickness, sea ice surface temperature, near-surface air temperature and near-surface wind speed, constructing spatio-temporal sequence sample pairs, and filtering out unreasonable data according to the principle of physical conservation. Finally, a training data set containing multi-scale spatio-temporal features is obtained, wherein the validation set and the test set account for 20% and 10% of the total data amount, respectively.

[0040] The step of training the feature fusion model specifically comprises: firstly initializing network parameters using pre-trained physical knowledge guided weight distribution; then optimizing the network parameters using a small batch stochastic gradient descent method; considering prediction accuracy and physical consistency constraints in a loss function; using a cosine annealing strategy for a learning rate; dynamically adjusting weight distribution of a feature fusion layer according to the ice-air interface balance index during the training process to enhance the expression capability of an abnormal area of the index; and avoiding overfitting through an early stopping method, so as to finally obtain a model with both physical interpretability and prediction accuracy.

[0041] The near-surface wind speed is a horizontal wind speed measured at a height of 10 meters from the ground, and is in units of meters per second.

[0042] The surface net radiation flux is the difference between solar radiation received by the ground and long-wave radiation emitted by the ground, and is in units of watts per square meter.

[0043] The snow cover thickness is the thickness of the snow layer covering the sea ice surface, and is in units of meters.

[0044] The model initial field is a starting state field for the start of integral calculation of the index value mode, and contains physical quantities such as wind, temperature, humidity, and pressure of each layer of the atmosphere.

[0045] The boundary condition is a side boundary and bottom boundary condition required during the integral process of the index value mode.

[0046] The weather forecast result is a prediction of future weather conditions obtained by mode integral calculation, including elements such as temperature, humidity, wind direction and speed, air pressure, and precipitation.

[0047] The Arctic sea ice assimilation model is used to perform multi-source data fusion on meteorological elements, sea ice density, and sea ice thickness information, and to realize collaborative updating of multiple variables and multiple scales based on four-dimensional variational assimilation technology.

[0048] The near-surface air temperature is the atmospheric temperature measured at a height of 2 meters from the ground, and is in units of degrees Celsius.

[0049] The relative humidity is the percentage of water vapor content in the air to the saturated water vapor content at the same temperature.

[0050] wherein the downward shortwave radiation flux refers to the solar radiation energy reaching the ground surface, and the unit is watt per square meter, and the downward shortwave radiation flux is obtained from the GFS background field data.

[0051] wherein the downward longwave radiation flux refers to the longwave radiation energy emitted downward by the atmosphere to reach the ground surface, and the unit is watt per square meter, and the downward longwave radiation flux is obtained from the GFS background field data.

[0052] The specific implementation of the above steps is described in detail below. The specific implementation of step S01 is to obtain GFS background field data from a global weather center server through an API interface, which covers the global geographical range, contains 41 non-uniformly distributed pressure layers from the ground surface to about 0.01 hPa in the vertical direction, has a horizontal resolution of 0.25°x0.25°, a time resolution of 3 hours, and a single prediction duration of 168 hours. After obtaining the data, the ice-sea heat exchange correction equation is applied to preliminarily correct the boundary layer parameters of the sea ice region. The correction equation is based on the Monin-Obukhov similarity theory, considers the influence of sea ice characteristics on the boundary layer stability and flux exchange, introduces the sea ice density and thickness as important parameters, and the correction coefficient is 0.85 when the sea ice density is greater than 0.7, 0.92 when the sea ice density is between 0.3 and 0.7, and 0.98 when the sea ice density is less than 0.3. The purpose of this step is to obtain high-quality background field data and preliminarily improve the description of the polar boundary layer parameters, laying a foundation for subsequent processing.

[0053] The specific implementation of step S02 is to extract the sea ice density and sea ice thickness information from the GFS background field data, and the bilinear interpolation algorithm is used to ensure spatial continuity. Then, the extracted sea ice density and thickness information are physically consistent checked based on the sea ice thermodynamic equilibrium equation. The checking process first calculates the theoretical sea ice thickness value under the thermodynamic equilibrium condition, and when the relative deviation of the actual sea ice thickness from the theoretical value exceeds 20%, the sea ice thickness is adjusted by an iterative method to meet the thermodynamic constraint. The coordination of sea ice density and sea ice thickness is also tested, and for the newly formed thin ice region, the sea ice density and thickness should satisfy the positive correlation, that is, when the sea ice thickness is less than 0.3 m, the sea ice density should not exceed 0.85. The checked sea ice density and thickness data are rewritten into an intermediate file recognizable by the model preprocessing program in binary format, and a lossless compression algorithm is used to reduce the storage space. The purpose of this step is to ensure the physical consistency of the sea ice parameters and avoid non-physical solutions in the model integration process.

[0054] The specific implementation of step S03 is to set the mode pre-processing parameter file, including key parameters such as mode resolution, vertical layer number, projection mode and calculation domain range. For the polar region, the polar stereographic projection is used to reduce the grid deformation, the horizontal resolution is set to 15 km, the vertical direction is set to 57 layers, and the top layer height is set to 10 hPa. When the terrain processing program is executed, the Gaussian filter is used to smooth the terrain, and the filter radius is 4 grid point distances, so as to eliminate numerical noise. The format rewriting program converts various input data into a mode-specific binary format, and the interpolation program uses a conformal interpolation algorithm to ensure the conservation of physical quantities in the spatial interpolation process. The Arctic sea ice assimilation model uses four-dimensional variational assimilation technology, introduces a background error covariance matrix to describe the correlation between different variables, and the covariance structure is calculated by the NMC method. Considering the particularity of the sea ice region, the flow correlation length scale adjustment algorithm is introduced, so that the assimilation process can better maintain the frontal characteristics of the sea ice edge. This step generates a physical consistency higher mode background field through multi-element fusion, and provides high-quality initial conditions for subsequent mode integration.

[0055] The specific implementation of step S04 is to configure the sea ice thermodynamic condition parameters and the sea air momentum exchange coefficient in the mode parameters. The sea ice thermal conductivity is set to 2.0 W / (m·K), and the sea ice albedo is dynamically adjusted according to the sea ice thickness, which is between 0.5 and 0.7 when the thickness is less than 0.5 m, and between 0.7 and 0.85 when the thickness is greater than 0.5 m. The sea air momentum exchange coefficient uses the modified Charnock relationship, and the roughness length parameter is 0.0012 when the sea ice density is greater than 0.6, and 0.0018 when the sea ice density is less than 0.6. The polar sea ice heat flux equation is used to calculate the sea ice surface temperature. The equation is based on the energy balance principle, considering solar radiation, atmospheric longwave radiation, sea ice surface emission radiation, and heat and latent heat exchange, etc. The sea ice surface temperature is obtained by solving the nonlinear equation set by Newton iteration method. When the sea ice thickness is less than 0.1 m, the heat penetration coefficient 0.3 is introduced to consider the influence of sea water heat flux penetrating through sea ice; when the sea ice thickness is greater than 1.5 m, the vertical stratification is not less than 5 layers to accurately describe the internal temperature profile. The purpose of this step is to accurately describe the sea ice thickness and atmospheric heat exchange process, and to provide key input for boundary layer parameterization. The Charnock relationship is a dimensionless relationship in fluid mechanics, which is used to describe the characteristics of liquid surface waves. It was proposed by British fluid mechanicist John Charnock in 1955, mainly used in the fields of marine meteorology and ocean engineering.

[0056] The specific implementation of step S05 is to establish a sea ice density thickness boundary layer interaction equation set based on the modified boundary layer theory, considering the modulation effect of sea ice on the atmospheric boundary layer structure. First, the sea ice surface roughness is calculated, the roughness length is related to the sea ice density, ice ridge height and distribution, the ice ridge height is associated with the sea ice thickness through an empirical relationship, and the typical value is between 0.01 m and 0.05 m. Then the thermal stability parameter is calculated, and the Richardson number is used to represent, when the Richardson number is greater than 0.25, the boundary layer is in a strong stable state, and when the Richardson number is less than 0, it is in an unstable state. The turbulent flux is calculated based on the stability corrected K theory, and the boundary layer height is determined by the critical Richardson number method, which is usually 100 m to 300 m on the sea ice stable boundary layer. The temperature gradient and humidity gradient are calculated using the discrete difference method, and the vertical resolution is increased at the bottom of the boundary layer, with the first layer thickness of 5 m. The sea ice modulated boundary layer structure generated in this step contains key parameters such as boundary layer height, temperature gradient and humidity gradient, which are used as input for the subsequent feature fusion model, with the purpose of accurately describing the influence of sea ice on the boundary layer process.

[0057] The specific implementation of step S06 is to apply a feature fusion model to deeply fuse sea ice density, sea ice thickness, sea ice surface temperature and sea ice modulated boundary layer structure. The feature fusion model first normalizes the input features using the Z-score standardization method, then captures the spatio-temporal correlation between different variables through a multi-head self-attention mechanism, with 8 attention heads and a hidden layer dimension of 256. The spatio-temporal graph convolution layer uses a three-layer stacking structure, with a convolution kernel size of 3x3, and the output channel number of each layer is 128, 64 and 32, and the activation function uses LeakyReLU. The multi-scale feature fusion layer integrates feature representations of different spatial scales through a pyramid pooling network, with pooling scales of 1x1, 2x2, 4x4 and 8x8. Based on the ice-air interface balance index, the weight of each element in the feature fusion model output vector is adjusted adaptively, when the index is less than 0.3, the sea ice density feature weight is increased by 30%; when the index is greater than 0.7, the sea ice surface temperature feature weight is increased by 25%; when the index is between 0.3 and 0.7, the weights of each feature remain balanced. This step optimizes the multi-source feature fusion process through deep learning method, generates a more physically consistent model initial field and boundary condition, and provides high-quality input for model integration.

[0058] The specific implementation of step S07 is to set the physical parameterization scheme and the sea ice parameterization scheme. The physical parameterization scheme selection includes that the radiation scheme adopts the RRTMG long-short wave scheme, the boundary layer scheme adopts the YSU scheme, the cumulus convection scheme adopts the Kain-Fritsch scheme, and the microphysical scheme adopts the WSM6 scheme. The sea ice parameterization scheme selection includes that the heat conduction adopts a multi-layer nonlinear diffusion equation, the phase transition process considers the influence of salinity, and the dynamics process adopts an elastoplastic rheology method. The integral program is executed by using the model initial field and boundary conditions generated in step S06, the time step is set to 45 seconds, the integration time is 168 hours, and the output frequency is 1 hour. The ensemble prediction technique is used to evaluate the uncertainty influence of sea ice density and sea ice thickness on the meteorological prediction result, the number of ensemble members is 21, and the ensemble members are generated by perturbing the sea ice density by ±10% and the sea ice thickness by ±20%. The ensemble prediction result describes the prediction uncertainty by a probability distribution function, and when the dispersion of the ensemble exceeds a preset threshold, the system automatically marks it as a high uncertainty area, and the threshold is set to be greater than 1.5 times the standard deviation of the climate average. The purpose of this step is to obtain the meteorological prediction result for the next 7 days and quantify the prediction uncertainty. Specifically, the RRTMG long-short wave radiation scheme is a fast and accurate atmospheric radiation transfer model, which calculates the radiation flux and heating rate of the atmospheric layer by considering the absorption spectrum lines of various gas components and the effect of aerosols, balancing the calculation accuracy and efficiency; the YSU boundary layer scheme is a non-local K-theory closure mode, which explicitly processes the turbulent mixing and top entrainment process in the atmospheric boundary layer, and improves the simulation of heat and water vapor exchange between the night boundary layer and the free atmosphere; the Kain-Fritsch cumulus convection scheme is based on the Lagrangian mass flux framework, which simulates the convective updraft and downdraft, considers the cloud-environment mixing and precipitation process, and is particularly suitable for mesoscale grid precipitation simulation; the WSM6 microphysical scheme is a complex single-moment parameterization scheme, which tracks the generation, transformation and sedimentation of six types of water condensate (water vapor, cloud water, rainwater, cloud ice, snow and graupel), and can more completely describe the microphysical processes of cold, warm clouds and mixed-phase clouds.

[0059] The specific implementation of step S08 is post-processing of the weather forecast results, including system error correction, downscaling processing, and product generation. The system error correction adopts a quantile mapping method, and an error correction model is constructed based on the statistical relationship between historical forecasts and observations, with a training sample length of not less than 365 days. The forecast skill score is calculated by comparing historical observation data, and the scoring indicators include root mean square error, bias, correlation coefficient, and forecast skill score. The threshold value of the forecast skill score in the Arctic region requires that the 72-hour forecast skill of the 500 hPa potential height field be not less than 0.7, and the 48-hour forecast skill of the ground temperature be not less than 0.65. According to the application requirements of the Arctic region, standardized forecast products are generated, including routine meteorological element forecast maps, sea ice density change forecast maps, extreme weather event probability forecast maps, and sailing index forecast maps. The product resolution is dynamically adjusted according to user requirements, with a standard resolution of 0.1°x0.1°, and the product format supports multiple formats such as GRIB2, NetCDF, and GeoTIFF. The purpose of this step is to convert the original forecast results into user-friendly forecast products and evaluate the forecast quality through skill score.

[0060] The detailed structure of the feature fusion model adopts a deep neural network framework, specifically including the following main components: the input layer receives normalized sea ice density, sea ice thickness, sea ice surface temperature, and sea ice modulated boundary layer structure feature data, with a feature dimension of [batch size, time step, spatial height, spatial width, feature channel number]. The multi-head self-attention layer contains 8 attention heads, each of which independently learns the association patterns between different features. The input features are first linearly transformed to generate query, key, and value matrices, then the attention weights are calculated, and finally the outputs of each head are spliced and linearly transformed to obtain the output features. The spatio-temporal graph convolution layer is composed of 3 layers of stacked graph convolution networks, each layer uses a combination of spatial convolution and temporal convolution, with a spatial convolution kernel size of 3x3 and a temporal convolution kernel size of 3. Batch normalization and LeakyReLU activation functions are used between layers. The multi-scale feature fusion layer integrates information of different spatial scales through a pyramid pooling network, performs 1x1, 2x2, 4x4, and 8x8 pooling operations respectively, then upsamples the pooling results back to the original resolution and splices them. The fully connected layer maps the fused features to the output space, adopts a residual connection structure to ensure stable gradient propagation and complete feature information preservation, and the final output layer dimension is consistent with the initial field and boundary condition format. The entire network uses a batch size of 64, an initial learning rate of 0.001, applies a cosine annealing strategy, and trains for 300 rounds during training.

[0061] The detailed steps for establishing the training data set of the feature fusion model include: first, collecting Arctic exploration satellite data from multiple data sources, including the sea ice concentration product provided by AMSR2, with a resolution of 6.25 km; the sea ice thickness product jointly inverted by CryoSat-2 and SMOS, with a resolution of 25 km; and the sea ice surface temperature product provided by MODIS, with a resolution of 1 km. At the same time, in-situ data observed by polar research stations, buoys and ships are collected, including boundary layer temperature profile, humidity profile and wind speed profile, etc. The collected data are resampled according to the spatial resolution and unified to 15 km grid, and the time resolution is unified to daily scale, and the bilinear interpolation method is used to keep the conservation of physical quantities. From the processed data, key features such as sea ice concentration, sea ice thickness, sea ice surface temperature, near-surface air temperature and near-surface wind speed are extracted, and sliding window samples with a time window of 30 days are constructed, with the input being the previous 27-day features and the output being the next 3-day features. According to the principle of physical conservation, unreasonable data are screened and removed, including checking whether the sea ice concentration is within the range of 0 to 1, whether the sea ice thickness is a positive value and the concentration in the thin ice area does not exceed the threshold, whether the surface temperature meets the thermodynamic constraints, etc. The samples that do not meet the conditions are removed. Finally, a training data set containing multi-scale spatio-temporal features is obtained, with a total sample size of not less than 10,000, of which the validation set and the test set account for 20% and 10% of the total data amount respectively, ensuring that the model has good generalization ability.

[0062] It should be noted that the mode pre-processing program receives the GFS background field data obtained through the API interface call in step S01. These data cover the global range, contain 41 layers of non-uniformly distributed pressure layers in the vertical direction, have a horizontal resolution of 0.25°x0.25°, and have a time resolution of 3 hours. The program first applies an ice-ocean heat exchange correction equation to preliminarily correct the boundary layer parameters in the sea ice area. Different correction coefficients (0.85, 0.92, 0.98) are applied according to the sea ice concentration interval (>0.7, 0.3-0.7, <0.3). In step S02, the program extracts the sea ice concentration and sea ice thickness information from the GFS background field, and performs physical consistency verification through the sea ice thermodynamic equilibrium equation. When the actual sea ice thickness deviates from the theoretical value by more than 20%, it is adjusted through iteration to meet the thermodynamic constraints; for the newly formed thin ice area (thickness <0.3 m), the sea ice concentration is limited to not more than 0.85. The verified data are rewritten into a binary format intermediate file for subsequent processing.

[0063] It should be noted that the terrain processing program is responsible for processing the terrain data required by the model, which is an important part of the model pre-processing in step S03. For the Arctic region, the program uses polar stereographic projection to reduce grid deformation, and sets the horizontal resolution to 15 km. During the terrain processing, the program applies Gaussian filtering to smooth the terrain, with a filter radius of 4 grid points, to eliminate small-scale terrain features that may cause numerical noise. This processing ensures that the terrain data meets the stability requirements of the numerical model, avoiding computational instability caused by excessive terrain gradients. The smoothed terrain field output by the terrain processing program will directly affect the boundary layer parameterization and surface process simulation in the model, which is particularly important for the simulation of the Arctic sea ice region.

[0064] It should be noted that the format rewriting program converts various input data into a binary format specific to the model, ensuring consistency between the data structure and the internal representation of the model. The interpolation program uses a conformal interpolation algorithm to ensure the conservation of physical quantities during spatial interpolation, which is crucial for accurately describing the physical characteristics of the sea ice edge region. The Arctic sea ice assimilation model uses four-dimensional variational assimilation technology, introducing a background error covariance matrix to describe the correlation between different variables, and the covariance structure is calculated by the NMC method. At the same time, considering the special nature of the sea ice region, the program introduces a flow correlation length scale adjustment algorithm to better maintain the frontal characteristics of the sea ice edge during the assimilation process. The coordinated work of these programs ensures the fusion processing of multi-source data and generates a model background field with higher physical consistency.

[0065] It should be noted that the integration program uses the initial field and boundary conditions generated in step S06, based on the set physical parameterization scheme (the RRTMG long and short wave scheme is used for radiation scheme, the YSU scheme (modified) is used for boundary layer scheme, the Kain-Fritsch scheme is used for cumulus convection scheme, and the WSM6 scheme is used for microphysical scheme) and sea ice parameterization scheme (multi-layer nonlinear diffusion equation is used for heat conduction, salinity effect is considered for phase change process, and elastoplastic rheology method is used for dynamics process). The time step of the integration program is set to 45 seconds, the integration time is 168 hours (7 days), and the output frequency is 1 hour. The program also implements ensemble prediction technology, generates 21 ensemble members by perturbing the sea ice density by ±10% and the sea ice thickness by ±20%, to evaluate the uncertainty of sea ice parameters on the meteorological prediction results. The ensemble prediction results describe the prediction uncertainty through the probability distribution function, and when the ensemble dispersion exceeds the preset threshold (the standard deviation is greater than 1.5 times the climate average), the system automatically marks it as a high uncertainty region.

[0066] Specifically, the principle of the present application is that the technical principle of the present application is based on the physical mechanism of the sea ice-atmosphere coupling system and the multi-source data fusion theory. The core is to build a complete sea ice information processing chain and a sea ice-atmosphere interaction model, and to realize the whole process optimization from data acquisition, physical consistency verification to multi-element fusion.

[0067] Firstly, the present application corrects the boundary layer parameters of the sea ice region in the GFS background field data by the ice-sea heat exchange correction equation, solves the systematic deviation of the standard atmospheric model in the polar region. The key innovation is to apply the sea ice thermodynamic equilibrium equation to physically consistent verification of the sea ice density and thickness information, to ensure that the two key parameters meet the basic laws of thermodynamics. This step makes up for the physical contradiction caused by the inconsistent source of sea ice density and thickness information in the traditional method, and lays a solid foundation for subsequent processing.

[0068] Secondly, the present application establishes the polar sea ice heat flux equation and the sea ice density and thickness boundary layer interaction equation set, which are used to calculate the sea ice surface temperature and the sea ice modulation boundary layer structure respectively. The two equation sets fully consider the comprehensive influence of multiple physical factors such as sea ice density, thickness, near-surface wind speed, surface net radiation flux and snow cover thickness, and can accurately describe the energy and momentum exchange process of the sea ice-atmosphere interface, thereby providing more physically realistic boundary conditions.

[0069] The most innovative is the feature fusion model proposed by the present application, which is based on multi-head self-attention mechanism and spatiotemporal graph convolution, and can learn the complex nonlinear relationship between sea ice density, thickness, surface temperature and boundary layer structure. The model introduces the ice-air interface balance index as an adaptive weight adjustment mechanism, solving the problem of insufficient model generalization ability under different sea ice conditions. In the training process, the loss function considers both prediction accuracy and physical consistency constraints, ensuring that the model output not only conforms to the historical observation law but also does not violate the basic physical laws, thereby theoretically guaranteeing the physical consistency of sea ice information and the reliability of the prediction results.

[0070] In summary, the present application establishes a complete physical link of sea ice density, thickness and atmosphere coupling through systematic sea ice information processing and deep learning driven feature fusion, effectively solves the problem of insufficient physical consistency of sea ice information in traditional methods, and provides theoretical and technical support for improving the accuracy of weather prediction in the Arctic region.

[0071] A specific embodiment 1 of the present application is provided below, and the specific implementation manner of each step in the embodiment 1 is described in detail as follows.

[0072] The specific implementation of step S01 is to obtain GFS background field data from a global weather center server through an API interface call. The data covers a global geographical range, contains 41 non-uniformly distributed pressure layers from the ground to about 0.01 hPa in the vertical direction, has a horizontal resolution of 0.25°x0.25°, a time resolution of 3 hours, and a single prediction length of 168 hours. After obtaining the data, the boundary layer parameters of the sea ice region are preliminarily corrected by using an ice-sea heat exchange correction equation. The ice-sea heat exchange correction equation is specifically represented as follows:

[0073] ;

[0074] ;

[0075] In the formula, is the corrected boundary layer turbulent flux, with the unit of ; is the standard calculated boundary layer turbulent flux, with the unit of ; is a correction coefficient, dimensionless; is a sea ice influence factor, with a value range of 0.05-0.15, dimensionless; is a sea ice density, dimensionless, with a value range of 0-1; is a sea ice thickness, with the unit of m. The purpose of this step is to obtain high-quality background field data and preliminarily improve the description of the polar boundary layer parameters, laying a foundation for subsequent processing.

[0076] The specific implementation of step S02 is to extract sea ice density and sea ice thickness information from the GFS background field data, and the extraction adopts a bilinear interpolation algorithm to ensure spatial continuity. Then, the extracted sea ice density and thickness information are physically consistent verified based on a sea ice thermodynamic equilibrium equation. The sea ice thermodynamic equilibrium equation is specifically represented as follows:

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] In the formula, is a sea ice density, with a value of 917 ; is a sea ice specific heat capacity, with a value of 2067 ; is a sea ice internal temperature, with the unit of K; is time, with the unit of s; is the vertical coordinate, unit: m; is the sea ice thermal conductivity, unit: ; is the sea ice thickness, unit: m; is the sea ice bottom heat flux, unit: ; is the sea ice top heat flux, unit: ; is the sea ice latent heat, value: 3.35 x J / kg; is the pure ice thermal conductivity, value: 2.03 ; is the salinity influence coefficient, value: 0.13; is the sea ice salinity, unit: ‰; is the sea ice surface temperature, unit: K; is the sea ice freezing temperature, unit: K.

[0082] Based on the above equations, the physical consistency check is carried out, when the relative deviation of actual sea ice thickness and theoretical value exceeds 20%, the sea ice thickness is adjusted by iteration method to meet the thermodynamic constraint. The coordination of sea ice density and sea ice thickness is also tested, for the newly born thin ice area, the relationship is defined as:

[0083] ;

[0084] wherein, is the maximum reasonable value of sea ice density in thin ice area. The checked sea ice density and thickness data are rewritten into intermediate files in binary format which can be recognized by the model pre-processing program. The purpose of this step is to ensure the physical consistency of sea ice parameters and avoid non-physical solutions in the model integration process.

[0085] The specific implementation of step S03 is to set the model pre-processing parameter file, including key parameters such as model resolution, vertical layer number, projection method and calculation domain range. For the Arctic region, polar stereographic projection is used to reduce grid distortion, the horizontal resolution is set to 15 km, the vertical direction is 57 layers, and the top layer height is 10 hPa. When executing the terrain processing program, Gaussian filter is used to smooth the terrain, and the filter radius is 4 grid point distances to eliminate numerical noise. The format rewriting program converts various input data into model-specific binary format, and the interpolation program uses conformal interpolation algorithm to ensure the conservation of physical quantities in spatial interpolation process. The Arctic ice-ocean assimilation model uses four-dimensional variational assimilation technology, the basic expression is as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] wherein, is a cost function; is an analysis field state vector; is a background field state vector; is a background error covariance matrix; is an observation operator; is an observation vector; is an observation error covariance matrix; is a flow-dependent length scale, with a unit of km; is a reference length scale, with a value of 50 km; is an adjustment coefficient, with a value of 2.5; is a sea ice concentration gradient, is a sea ice concentration gradient maximum. This step generates a model background field with higher physical consistency through multi-element fusion, thereby providing high-quality initial conditions for subsequent model integration.

[0090] The specific implementation of step S04 is to configure sea ice thermodynamic condition parameters and sea-air momentum exchange coefficients in the model parameters. The sea ice thermodynamic parameters include a sea ice thermal conductivity set to 2.0 , and a sea ice albedo dynamically adjusted according to the sea ice thickness, with a specific expression as follows:

[0091] ;

[0092] wherein, is a sea ice albedo, dimensionless; is a sea ice thickness, with a unit of m.

[0093] The sea-air momentum exchange coefficient adopts a modified Charnock relationship, with an expression as follows:

[0094] ;

[0095] wherein, is a roughness length parameter, with a unit of m; is a sea ice concentration, dimensionless.

[0096] The sea ice surface temperature is calculated by applying a polar sea ice heat flux equation, with a specific expression as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] wherein, is the sea ice albedo, dimensionless; is the downward shortwave radiation flux, with the unit of ; is the sea ice emissivity, with the value of 0.97, dimensionless; is the downward longwave radiation flux, with the unit of ; is the Stefan-Boltzmann constant, with the value of 5.67x ; is the sea ice surface temperature, with the unit of K; is the sensible heat flux, with the unit of ; is the latent heat flux, with the unit of ; is the conductive heat flux, with the unit of ; is the air density, with the unit of ; is the air constant-pressure specific heat capacity, with the value of 1004 ; is the sensible heat exchange coefficient, dimensionless; is the near-surface wind speed, with the unit of m / s; is the near-surface air temperature, with the unit of K; is the water vaporization latent heat, with the value of 2.5x J / kg; is the latent heat exchange coefficient, dimensionless; is the near-surface specific humidity, dimensionless; is the saturated specific humidity, dimensionless; is the sea ice thermal conductivity, with the unit of ; is the sea water freezing temperature, with the unit of K; is the sea ice thickness, with the unit of m; is the thermal penetration coefficient, with the value of 0.3 when the sea ice thickness is less than 0.1 m, and 0 otherwise; is the attenuation coefficient, with the value of 7, dimensionless. The purpose of this step is to accurately describe the sea ice thickness and atmospheric heat exchange process, and to provide key inputs for boundary layer parameterization.

[0102] The specific implementation of step S05 is to establish a sea ice density thickness boundary layer interaction equation set, which is based on the modified boundary layer theory and considers the modulation effect of sea ice on the atmospheric boundary layer structure. First, the sea ice surface roughness is calculated, and the specific expression is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] where, is the sea ice surface roughness, with unit of m; is the ice ridge coverage ratio, dimensionless; is the ice ridge height, with unit of m; is the reference ice ridge height, with value of 1 m; is the sea ice thickness, with unit of m.

[0107] Then the thermal stability parameter is calculated, with expression as:

[0108] ;

[0109] where, is the Richardson number, dimensionless; is the gravity acceleration, with value of 9.8 ; is the virtual potential temperature, with unit of K; is the virtual potential temperature difference, with unit of K; is the height difference, with unit of m; is the wind speed difference, with unit of m / s.

[0110] Based on the above stability, the boundary layer turbulent flux is calculated:

[0111] ;

[0112] ;

[0113] ;

[0114] where, is the momentum flux, with unit of ; is the air density, with unit of ; is the drag coefficient, dimensionless; is the near-surface wind speed, with unit of m / s; is the stability correction function, dimensionless; is the sensible heat flux, with unit of ; is the air constant-pressure specific heat capacity, with unit of ; is the sensible heat exchange coefficient, dimensionless; is the near-surface air temperature, with unit of K; is the sea ice surface temperature, with unit of K.

[0115] The boundary layer height is determined by the critical Richardson number method:

[0116] ;

[0117] Where, is the boundary layer height, in m; is the height of the jth floor, in meters; For height Richardson number at ; is the critical Richardson number, which is 0.25.

[0118] The temperature gradient and humidity gradient are calculated using the discrete difference method:

[0119] ;

[0120] ;

[0121] Where, is the temperature gradient, in K / m; is the temperature difference, in K; is the height difference, in m; is the humidity gradient, in kg / (kg·m); The sea ice modulated boundary layer structure generated in this step contains key parameters such as boundary layer height, temperature gradient, and humidity gradient. It is used as input for the subsequent feature fusion model to accurately describe the impact of sea ice on boundary layer processes.

[0122] The specific implementation of step S06 is to use the feature fusion model to deeply fuse sea ice density, sea ice thickness, sea ice surface temperature and sea ice modulated boundary layer structure. The feature fusion model first normalizes the input features using the Z-score normalization method:

[0123] ;

[0124] Where, is the normalized feature; is the original feature; is the characteristic mean; is the characteristic standard deviation.

[0125] The weight of each element in the feature fusion model output vector is adaptively adjusted based on the ice-air interface balance index. The ice-air interface balance index is calculated as follows:

[0126] ;

[0127] Where, is the sea ice surface temperature, unit is K; ; is the sea ice concentration, dimensionless; is the sea ice thickness, unit is m; is the surface net radiation flux, unit is W / m2; .

[0128] The weight adjustment method is:

[0129] ;

[0130] In the formula, is the adjusted weight; is the original weight; is the feature index, 1 represents the sea ice concentration, and 3 represents the sea ice surface temperature. This step optimizes the multi-source feature fusion process through a deep learning method, generates a mode initial field and boundary condition with higher physical consistency, and provides high-quality input for mode integration.

[0131] The specific implementation of step S07 is to set a physical parameterization scheme and a sea ice parameterization scheme. The physical parameterization scheme selection includes that the radiation scheme adopts the RRTMG long-short wave scheme, the boundary layer scheme adopts the modified YSU scheme, the cumulus convection scheme adopts the Kain-Fritsch scheme, and the microphysical scheme adopts the WSM6 scheme. The sea ice parameterization scheme selection includes that the heat conduction adopts a multi-layer nonlinear diffusion equation, the phase change process considers the influence of salinity, and the dynamics process adopts an elastoplastic rheology method. The integration program is executed by using the mode initial field and boundary condition generated in step S06, the time step is set to 45 seconds, the integration time is 168 hours, and the output frequency is 1 hour. The ensemble prediction technology is used to evaluate the uncertainty influence of the sea ice concentration and the sea ice thickness on the meteorological prediction result, and the ensemble member disturbance mode is: ; ; ; ; In the formula, is the sea ice concentration of the jth ensemble member; is the sea ice concentration of the control prediction; is the sea ice concentration disturbance coefficient; is the sea ice thickness of the jth ensemble member; is the sea ice thickness of the control prediction; is the sea ice thickness disturbance coefficient; , which represents a normal distribution. The ensemble prediction result describes the prediction uncertainty through a probability distribution function, and the high uncertainty area determination standard is: ;

[0132] In the formula, is the standard deviation of the ensemble prediction result; is the standard deviation of the climate mean. The purpose of this step is to obtain the weather forecast results for the next 7 days and quantify the forecast uncertainty.

[0133] The specific implementation of step S08 is to post-process the weather forecast results, including system error correction, downscaling processing, and product generation. The system error correction uses the quantile mapping method, and the specific expression is: ;

[0134] In the formula, is the corrected forecast result; is the inverse function of the observation probability distribution; is the forecast probability distribution function; is the original forecast result. The forecast skill score is calculated by comparing historical observation data, and the scoring indicators include root mean square error, bias, correlation coefficient, and forecast skill score, and the calculation method is as follows: ; ; ; ; In the formula, is the root mean square error; is the forecast value; is the observation value; is the sample size; is the bias; is the correlation coefficient; is the forecast mean; is the observation mean; is the skill score; is the reference forecast. The threshold of the Arctic region forecast skill score requires that the 72-hour forecast skill of the 500 hPa potential height field be not less than 0.7, and the 48-hour forecast skill of the surface temperature be not less than 0.65. According to the application requirements of the Arctic region, standardized forecast products are generated, including routine weather element forecast charts, sea ice concentration change forecast charts, extreme weather event probability forecast charts, and sailing index forecast charts. The purpose of this step is to convert the original forecast results into user-friendly forecast products and evaluate the forecast quality through skill scores.

[0135] The detailed structure of the feature fusion model adopts a deep neural network framework, and specifically includes the following main components: the input layer receives normalized sea ice density, sea ice thickness, sea ice surface temperature, and sea ice boundary layer structure feature data. The detailed steps for establishing the training data set of the feature fusion model include: first, collecting Arctic satellite data from multiple data sources, including sea ice density products provided by AMSR2, with a resolution of 6.25 km; sea ice thickness products jointly inverted by CryoSat-2 and SMOS, with a resolution of 25 km; and sea ice surface temperature products provided by MODIS, with a resolution of 1 km. At the same time, collecting in-situ data from polar research stations, buoys, and ships, including boundary layer temperature profiles, humidity profiles, and wind speed profiles. The collected data is resampled according to the spatial resolution and unified to 15 km grid points, and the time resolution is unified to daily scale, and the bilinear interpolation method is used to maintain the conservation of physical quantities.

[0136] From the processed data, key features such as sea ice density, sea ice thickness, sea ice surface temperature, near-surface air temperature, and near-surface wind speed are extracted, and a sliding window sample pair with a time window of 30 days is constructed, with the input being the previous 27-day features and the output being the next 3-day features. According to the principle of physical conservation, unreasonable data is screened out, including checking whether the sea ice density is within the range of 0 to 1, whether the sea ice thickness is positive and the density in the thin ice area does not exceed the threshold, and whether the surface temperature meets the thermodynamic constraints. Samples that do not meet the conditions are removed. Finally, a training data set containing multi-scale spatio-temporal features is obtained, with a total sample size of not less than 10,000, of which the validation set and the test set account for 20% and 10% of the total data volume, respectively.

[0137] Optionally, for the sea ice thermodynamic equilibrium equation, the sea ice salinity is taken as 12‰ in newly formed ice and 4‰ in multi-year ice. In terms of boundary conditions, the sea ice surface temperature is affected by atmospheric forcing, and the sea ice bottom temperature is taken as the seawater freezing temperature, which is about -1.8℃ for seawater with a salinity of 35‰. The physical consistency verification process first solves the sea ice thermodynamic equilibrium equation to obtain the theoretical sea ice thickness value, and then compares it with the actual sea ice thickness. When the relative deviation exceeds the threshold of 20%, the sea ice thickness is adjusted by an iterative method. In the iterative process, the adjustment amplitude is 50% of the deviation at each step, the maximum number of iterations is 10, and the convergence criterion is that the relative deviation is less than 5%.

[0138] Optionally, in the four-dimensional variational assimilation technique, in the area with large gradient, the length scale is small, which helps to maintain the frontal characteristics of the sea ice edge; in the area with small gradient, the length scale is large, which is conducive to the large-scale spread of information. The solution of the four-dimensional variational assimilation adopts the conjugate gradient method, the maximum iteration number is 100, and the convergence criterion is that the gradient norm is reduced to 0.01 of the initial value or the gradient change is less than 0.1% for 3 consecutive iterations.

[0139] Optionally, the polar sea ice heat flux equation is based on the energy balance principle, and considers solar radiation, atmospheric long-wave radiation, sea ice surface emission radiation, and sensible heat, latent heat and conduction heat exchange and other physical processes. The left side of the equation represents the energy balance of the sea ice surface, and each item is: represents the solar shortwave radiation absorbed by the sea ice, represents the atmospheric long-wave radiation absorbed by the sea ice, represents the long-wave radiation emitted by the sea ice surface to the outside, is a sensible heat exchange coefficient, which is 0.0013 under neutral conditions, is a latent heat exchange coefficient, which is 0.0013 under neutral conditions. This equation is solved by Newton iteration method, and the initial guess value is , and the iteration process is:

[0140] ;

[0141] ;

[0142] ;

[0143] In the formula, is the surface temperature of the th iteration, is the energy balance residual, is the derivative of the residual with respect to the surface temperature. The iteration termination condition is or the maximum iteration number is reached for 20 times.

[0144] Optionally, in the sea ice density thickness boundary layer interaction equation set: the boundary layer height is determined by the critical Richardson number method, that is, searching from the ground upwards, when the Richardson number of a certain layer is greater than or equal to the critical Richardson number (taking the value of 0.25), the height of this layer is the boundary layer height . The temperature gradient and the humidity gradient are calculated by the discrete difference method, and the vertical resolution is encrypted at the bottom of the boundary layer during calculation, the thickness of the first layer is 5m, which gradually increases with height, and the maximum does not exceed 50m.

[0145] Optionally, the ice-air interface balance index is when the index When it is less than 0.3, it indicates that the exchange process is weak, and the weight of the enhanced sea ice density characteristic increases by 30%; when the index When it is greater than 0.7, it indicates that the exchange process is strong, and the weight of the enhanced sea ice surface temperature characteristic is increased by 25%; when the index When the index is between 0.3 and 0.7, the weights of the various features remain balanced. The threshold setting of the ice-air interface balance index is determined based on historical data analysis. Statistics show that areas with an index less than 0.3 generally correspond to thin ice or ice edge areas, where changes in sea ice density are more sensitive to atmospheric influences; areas with an index greater than 0.7 generally correspond to thick ice areas, where changes in sea ice surface temperature have a more significant impact on the atmosphere.

[0146] Optionally, the terrain processing program can be expressed mathematically as follows:

[0147] Gaussian filter smoothing:

[0148] ;

[0149] ;

[0150] Where, is the height of the smoothed terrain; is the original terrain height; is the Gaussian filter kernel; is the filter radius, which is equal to the distance between 4 grid points; is the standard deviation of the Gaussian distribution, which is .

[0151] Polar Stereographic Coordinate Transformation: ; Where, is the rectangular coordinate of the projection plane; is the radius of the Earth; is latitude; is the longitude; is the central longitude.

[0152] Optionally, the mathematical expressions of the format rewriting procedure and the interpolation procedure are as follows:

[0153] Shape-preserving interpolation algorithm: ; ; Where, The interpolation result of the target point; For the The value of a known point; is the weight coefficient; is the distance from the target point to the known point.

[0154] The four-dimensional variational assimilation technique has been described previously and will not be described in detail here.

[0155] Optionally, the integral program is mathematically expressed as follows:

[0156] Momentum equation (horizontal direction): ;

[0157] Continuity equation: ;

[0158] Thermodynamic equation: ;

[0159] Water vapor conservation equation: ;

[0160] In the formula, is the horizontal wind speed vector; is time; is the vertical velocity; is the vertical coordinate; is the air density; is the air pressure; is the Coriolis parameter; is the vertical unit vector; is the friction; is the temperature; is the specific volume; is the constant-pressure specific heat; is the heating rate; is the specific humidity; is the water vapor source and sink term.

[0161] Optionally, the integral program also includes ensemble prediction perturbations and prediction uncertainty determination criteria:

[0162] Optionally, the integral program also includes a heat conduction multi-layer nonlinear diffusion equation in sea ice parameterization:

[0163] ;

[0164] ;

[0165] Optionally, the integral program also includes a dynamic equation of the elastoplastic rheology method:

[0166] ;

[0167] ;

[0168] In the formula, is the sea ice temperature; is the sea ice thermal conductivity; is the sea ice salinity; is the pure ice thermal conductivity; is a salinity influence coefficient; is a temperature dependence coefficient; is a freezing temperature; is a sea ice velocity vector; is a sea ice thickness; is a stress tensor; is an atmospheric stress; is an ocean stress; is a Coriolis parameter; is a gravitational acceleration; is a sea surface height; is a shear viscosity; is a bulk viscosity; is a rate of strain tensor; is a trace of the rate of strain tensor; is a unit tensor; is a sea ice internal strength.

[0169] It should be noted that the four schemes in step S07 are prior art, in order to more comprehensively describe the embodiment, the mathematical description of the four schemes is provided below.

[0170] The radiation flux calculation is based on a fast radiation transfer model, and the basic equation is as follows:

[0171] ;

[0172] In the formula, is the radiation intensity with a wavelength of , and the unit is ; is an optical path, and the unit is ; is an extinction coefficient, and the unit is ; is an emission coefficient, and the unit is .

[0173] The atmospheric shortwave radiation transfer equation can be expressed as:

[0174] ;

[0175] In the formula, is an optical thickness, dimensionless; is a cosine of zenith angle, dimensionless; is an azimuth angle, and the unit is ; is a source function, and the unit is .

[0176] The source function considers two parts of scattering and thermal emission:

[0177] ;

[0178] where, is the single-scattering albedo, dimensionless, with a value range of ; is the scattering phase function, dimensionless; is the Planck function, with the same unit as ; is the temperature, with the unit of .

[0179] The Planck function is expressed as:

[0180] ;

[0181] where, is the Planck constant, with a value of ; is the speed of light, with a value of ; is the Boltzmann constant, with a value of ; is the wavelength, with the unit of ; is the temperature, with the unit of .

[0182] In the RRTMG scheme, the radiation calculation adopts the Correlated-k Distribution method, which calculates the full-spectrum radiation transfer through the grouping summation method:

[0183] ;

[0184] where, is the total radiation flux, with the unit of ; is the weight coefficient of the th group, dimensionless; is the radiation flux of the th group, with the unit of ; is the number of groups, with 14 groups for shortwave and 16 groups for longwave in the RRTMG scheme.

[0185] In the modified YSU scheme, the vertical distribution expression of the turbulent flux is:

[0186] ;

[0187] where, is the turbulent flux of the physical quantity , with the unit depending on the nature of ; is the turbulent diffusion coefficient, with the unit of ; is the vertical gradient of the physical quantity ; is the non-local term, with the same unit as .

[0188] The turbulent diffusion coefficient is calculated as:

[0189] ;

[0190] where, is the dimensionless coefficient, with the value of 0.4 for momentum, and 0.4 for heat and water vapor; is the mixing length, with the unit of ; is the turbulent kinetic energy, with the unit of ; is the stability correction function, dimensionless.

[0191] The non-local term is expressed as:

[0192] ;

[0193] where, is the non-local term coefficient, dimensionless, with the value range of ; is the surface turbulent flux, with the same unit as ; is the boundary layer height, with the unit of ; is the height, with the unit of ; is the power index, dimensionless, with the value of .

[0194] To adapt to the sea ice environment, the modified YSU scheme uses the following formula to calculate the boundary layer height in the sea ice area:

[0195] ;

[0196] where, is the friction velocity, with the unit of ; is the Coriolis parameter, with the unit of ; is the Monin-Obukhov length, with the unit of ; is the constant, dimensionless, with the value of 0.4 on sea ice and 0.7 on the sea surface.

[0197] The friction velocity is calculated as:

[0198] ;

[0199] where, is the drag coefficient, dimensionless; is the wind speed at the reference height, with the unit of .

[0200] The Monin-Obukhov length is calculated as:

[0201] ;

[0202] where, is the virtual potential temperature, with the unit of ; is the Karman constant, dimensionless, with the value of 0.4; is the gravitational acceleration, with the unit of , with the value of ; is the virtual potential temperature flux, with the unit of .

[0203] The stability correction function is expressed as:

[0204] ;

[0205] where, is the Richardson number, dimensionless, with the calculation formula as:

[0206] ;

[0207] where, is the virtual potential temperature difference, with the unit of ; is the height difference, with the unit of ; and are the east-west and north-south wind speed differences, with the unit of .

[0208] In the Kain-Fritsch scheme, the cloud base mass flux calculation expression is:

[0209] ;

[0210] where, is the cloud base mass flux, with the unit of ; is the cloud base air density, with the unit of ; is the cloud base vertical velocity, with the unit of ; is the cloud base area proportion, dimensionless, with the value range of .

[0211] The parameterization expression of the vertical velocity of cloud base is:

[0212] ;

[0213] In the formula, is the convective available potential energy, and the unit is ; is a proportional coefficient, and the unit is , and the value is .

[0214] The calculation formula of CAPE is:

[0215] ;

[0216] In the formula, is the free convection height, and the unit is ; is the equilibrium height, and the unit is ; is the gravity acceleration, and the unit is ; is the virtual temperature of the ascending air mass, and the unit is ; is the virtual temperature of the environment, and the unit is ; is the vertical height infinitesimal, and the unit is .

[0217] The calculation formula of the convection adjustment timescale is:

[0218] ;

[0219] In the formula, is the adjustment timescale, and the unit is ; is the convective available potential energy, and the unit is ; is the CAPE consumption rate, and the unit is , and the calculation formula is:

[0220] ;

[0221] The thermodynamic equation of the ascending air flow in the cloud is:

[0222] ;

[0223] ;

[0224] In the formula, is the potential temperature in the cloud, and the unit is ; H is the height, unit is m; ; L is the latent heat of water vaporization, unit is J / kg; ; Cp is the specific heat at constant pressure, unit is J / (kg·K); ; Exner is the Exner function, dimensionless; ; E is the entrainment ratio, unit is kg / kg; ; M is the updraft mass flux, unit is kg / (m2·s); ; Θe is the environmental potential temperature, unit is K; ; T is the total water ratio, dimensionless; ;

[0225] The parameterization expression of the entrainment ratio is:

[0226] ;

[0227] In the formula, α is the proportional coefficient, dimensionless, and the value is 0.03; R is the cloud radius, unit is m. .

[0228] In the WSM6 scheme, the basic kinetic equation of water phase change is:

[0229] ;

[0230] ;

[0231] ;

[0232] ;

[0233] ;

[0234] ;

[0235] In the formula, q is the water vapor ratio, dimensionless; qci is the cloud water ratio, dimensionless; qri is the rainwater ratio, dimensionless; qci is the cloud ice ratio, dimensionless; qs is the snow ratio, dimensionless; qg is the graupel ratio, dimensionless; t is the time, unit is s; ; N is the cloud water condensation rate, unit is s-1; ; is the water vapor deposition rate, with unit of ; is the rainwater condensation rate, with unit of ; is the evaporation rate, with unit of ; is the melting rate, with unit of ; is the auto-conversion rate, with unit of ; is the collision rate, with unit of ; is the rainwater collection cloud ice rate, with unit of ; is the rainwater freezing into snow rate, with unit of ; is the rainwater freezing into graupel rate, with unit of ; is the snow collection cloud ice rate, with unit of ; is the advection term, with unit of .

[0236] The cloud water condensation rate is calculated by the following formula:

[0237] ;

[0238] wherein, is the saturated water vapor ratio, dimensionless; is the latent heat of water vaporization, with unit of ; is the specific heat at constant pressure, with unit of ; is the water vapor gas constant, with unit of ; is the temperature, with unit of ; is the time step, with unit of .

[0239] The auto-conversion rate (cloud water to rainwater) is calculated by the following formula:

[0240] ;

[0241] wherein, is the conversion rate coefficient, with unit of , and the value is ; is the critical cloud water content, dimensionless, and the value is ; is the Heaviside step function, dimensionless.

[0242] The collision rate (raindrop collecting cloud droplet) is calculated by the following formula:

[0243] ;

[0244] wherein, is the collision coefficient, with the unit of , and the value is .

[0245] The parameterized expression of the falling speed is:

[0246] ;

[0247] ;

[0248] ;

[0249] wherein, , , are the falling speeds of raindrops, snow and graupel, respectively, with the unit of ; is the air density, with the unit of ; , , are the diameters of raindrops, snow and graupel, respectively, with the unit of ; , , are coefficients, with the unit of , and the values are , , ; , , are coefficients, with the unit of , , ; , , are power indexes, with the unit of , , .

[0250] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: the embodiment details the physical consistency verification of sea ice density and thickness information, the multi-element fusion of the Arctic ice sea assimilation model, the configuration of sea ice thermodynamic condition parameters, and the deep fusion processing of the feature fusion model, thereby significantly improving the accuracy of the Arctic regional weather forecast. The forecast method comprises a data download program, a sea ice thickness preprocessing program, a data preprocessing program, a mode calculation program, a post-processing program and an overall control program.

[0251] In this embodiment, the researchers selected the route of the 12th Chinese Arctic Scientific Expedition north of 60°N (12 UTC on July 21, 2021 to 23 UTC on September 15, 2021) for testing and verification. First, step S01 is implemented by a data download program to download global GFS background field data from the National Environmental Prediction Center server using the wget download function of the Linux server. The data has 41 layers in the vertical direction, a horizontal resolution of 0.25°x0.25°, a time interval of 3 hours, and a single prediction duration of 168 hours. The program provides data download detection and restart functions, which automatically restart the download process when an interruption is detected to ensure data integrity. Program running information is saved in the download_err file. As shown in Table 1:

[0252] Table 1 Main parameters of GFS background field data

[0253]

[0254] After the data is downloaded, the sea ice thickness preprocessing program implements step S02 to extract sea ice density and sea ice thickness information from the GFS background field and perform physical consistency verification based on the sea ice thermodynamic equilibrium equation. First, the wgrib2 software is used to extract and convert the sea ice thickness in the GFS background field to the NETCDF format, and then the NCL script is used to read the processed sea ice thickness data and rewrite it as an intermediate file recognizable by WPS. In the physical consistency verification process, when the sea ice thickness is less than 0.3 m, the corresponding sea ice density is limited to not more than 0.85, which meets the physical characteristics of the newly formed thin ice area; when the relative deviation of the actual sea ice thickness and the thermodynamic equilibrium theoretical value exceeds 20%, the sea ice thickness is adjusted by an iterative method to meet the thermodynamic constraints. The verified sea ice density and thickness data are rewritten in binary format to an intermediate file recognizable by the model preprocessing program, and a lossless compression algorithm is used to reduce storage space in the specific implementation. In order to make the results of the sea ice thickness preprocessing program consistent with the time name of other binary files, the preprocessing program is designed with a time loop function to ensure that the generated intermediate file information can be viewed using the rd_intermediate.exe program provided by WPS.

[0255] The data preprocessing program implements step S03, and the meteorological elements, sea ice density and sea ice thickness information are fused by the Arctic Ocean assimilation model to generate a more physically consistent model background field. According to the operation requirements of the Polar WRF model, the namelist.wps file is set. For the Arctic region, the polar stereographic projection is used to reduce the grid deformation, the horizontal resolution is set to 15 km, the vertical layering is 57 layers, and the top layer height is 10 hPa. The terrain processing program geogrid.exe, the format rewriting program ungrb.exe and the interpolation program metgrid.exe are executed in turn. When executing the terrain processing program, Gaussian filtering is used to smooth the terrain, and the filtering radius is 4 grid point distances to eliminate numerical noise; the format rewriting program converts various input data into a special binary format for the model; the interpolation program uses a conformal interpolation algorithm to ensure the conservation of physical quantities in the spatial interpolation process. Before executing metgrid.exe, the sea ice thickness information needs to be added in METGRID.TBL, and the ICEDEPTH option is added in fg_name in namelist.wps. The Arctic Ocean assimilation model uses four-dimensional variational assimilation technology, introduces a background error covariance matrix to describe the correlation between different variables, and the covariance structure is calculated by the NMC method. Considering the particularity of the sea ice area, a flow correlation length scale adjustment algorithm is introduced to ensure that the assimilation process can better maintain the frontal characteristics of the sea ice edge. The results of the data preprocessing are shown in Table 2:

[0256] Table 2 Main parameters and result files of data preprocessing

[0257]

[0258] The mode calculation program implements steps S04 to S07, in sequence, to perform sea ice thermodynamic condition parameter configuration, sea ice modulated boundary layer structure calculation, feature fusion model deep fusion processing, and mode integral calculation. First, the background field data is converted into the initial field required by the WRF mode, i.e., the initial field wrfinput, the side boundary wrfbdy, and the bottom boundary wrflowinp, by using the real.exe. Since the real.exe of the Polar WRF version has an interpolation bug, the real.exe program of the WRFV3.9.1 version is used for processing. In the mode parameters, the sea ice thermodynamic condition parameters and the sea air momentum exchange coefficient are configured: the sea ice thermal conductivity is set to 2.0 W / (m·K), the sea ice albedo is dynamically adjusted according to the thickness, and when the thickness is less than 0.5 m, the value is 0.5-0.7, and when the thickness is greater than 0.5 m, the value is 0.7-0.85; the sea air momentum exchange coefficient uses the modified Charnock relationship, and the roughness length parameter is set to 0.0012 when the sea ice density is greater than 0.6, and is set to 0.0018 when the sea ice density is less than 0.6. The sea ice surface temperature is calculated by using the polar sea ice heat flux equation. The equation is based on the energy balance principle, considers multiple physical processes such as solar radiation, atmospheric longwave radiation, sea ice surface emission radiation, and sensible and latent heat exchange, and solves the nonlinear equation set by using the Newton iteration method.

[0259] The sea ice density thickness boundary layer interaction equation set is established to calculate the sea ice modulated boundary layer structure. The equation set is based on the modified boundary layer theory and considers the modulation effect of sea ice on the atmospheric boundary layer. First, the sea ice surface roughness is calculated, and then the thermal stability parameter is calculated. The Richardson number is used to represent the stability state, and the K theory based on the stability correction is used to calculate the turbulent flux. The feature fusion model is used for deep fusion of the sea ice density, the sea ice thickness, the sea ice surface temperature, and the sea ice modulated boundary layer structure. The weight of each element in the output vector of the feature fusion model is adaptively adjusted based on the ice-air interface balance index. When the index is less than 0.3, the weight of the sea ice density feature is increased by 30%; when the index is greater than 0.7, the weight of the sea ice surface temperature feature is increased by 25%. The specific configuration of the physical parameterization scheme and the sea ice parameterization scheme is shown in Table 3:

[0260] Table 3 Key parameter settings of the mode calculation program

[0261]

[0262] The post-processing program implements step S08 to post-process the weather forecast results, calculate the forecast skill score by comparing historical observation data, and generate standardized forecast products according to the application requirements of the Arctic region. The ARWpost software is used to interpolate and format the wrfout file obtained by model integration, and convert it to a binary format readable by GrADs. Then, a Fortran program is used to convert these binary files to standard NetCDF format. The system error correction uses the quantile mapping method, and the error correction model is constructed based on the statistical relationship between historical forecasts and observations. The training sample length is not less than 365 days. The forecast skill score is calculated by comparing historical observation data, and the scoring indicators include root mean square error, bias, correlation coefficient, and forecast skill score. The threshold of the Arctic region forecast skill score requires that the 72-hour forecast skill of the 500 hPa potential height field be not less than 0.7, and the 48-hour forecast skill of the ground temperature be not less than 0.65. The final generated forecast products include 2-meter temperature, 2-meter dew point temperature, sea level pressure, and 10-meter wind speed, etc. The forecast maps of conventional meteorological elements, as well as the sea ice density change forecast map, extreme weather event probability forecast map, and navigation index forecast map. The product resolution is 0.1°x0.1°, and supports multiple formats such as NetCDF, GrADS, and GeoTIFF.

[0263] The entire forecast process is connected through the overall control program, and each program is executed according to the steps of S01 to S08. The running status of each program can be viewed in real time through the msg.run file. To evaluate the forecast effect of this method, the route north of 60°N of the 12th Chinese Arctic Scientific Expedition (2021 July 21 12UTC-September 15 23UTC) was selected for testing and comparison. The model forecast variables include wind, temperature, relative humidity, pressure, precipitation, etc. The single forecast duration is 7 days. The initial field, side boundary and bottom boundary data are updated every 24 hours, and finally 57 groups of forecast products are obtained. The first 24 hours of each product are extracted and spliced to form a 24-hour forecast result. In this way, 48-hour and 72-hour forecast products are formed. Researchers also drew Taylor diagrams of 24-hour forecast elements, which can display correlation coefficient, standard deviation ratio and root mean square error three statistical indicators at the same time, so as to more comprehensively evaluate the performance of model prediction. For example Figure 2As shown in FIG. 24, in the 24-hour forecast, the model has the best prediction effect on the sea level pressure, the prediction levels of 2-meter temperature and 2-meter dew point temperature are close, and the prediction effect of 10-meter wind speed is the worst. This is directly related to the optimization in steps S04 to S06 for the sea ice surface temperature and the boundary layer structure, indicating the superiority of the method in the simulation of thermodynamic processes. The researchers designed two sets of comparative experiments: the baseline experiment (Exp-A) does not consider the fusion of sea ice density and sea ice thickness, and the improved experiment (Exp-B) uses the method of the present application to fuse the sea ice density and the sea ice thickness. After fusing the sea ice density and the sea ice thickness distribution, the prediction effect of the model on the 2-meter temperature and the 2-meter dew point temperature is obviously improved, the root mean square error of the 2-meter temperature in 24h~72h is reduced by 5.2~10.3% compared with Exp-A, and the root mean square error of the 2-meter dew point temperature in 24h~72h is reduced by 8.6~14.7% compared with Exp-A; the prediction errors of the sea level pressure and the 10-meter wind speed change little, and the improvement is less than 1.2%. At the same time, it can be seen that with the increase of the prediction time, the sea level pressure error accumulates rapidly, and the 72-hour prediction error is 3.4 times of the 24-hour prediction error; the 2-meter temperature and the 2-meter dew point error gradually increase; the 10-meter wind speed error changes little with the increase of the prediction time.

[0264] Finally, the GrADs software is used to draw the results to obtain the prediction results of each element in the Arctic region as shown in FIG. 24. Figure 3

[0265] ​This result shows that the application effectively improves the prediction ability of the boundary layer temperature field by fusing sea ice density and thickness, especially by the multi-element fusion of the Arctic ice-ocean assimilation model and the deep fusion processing of the feature fusion model. The traditional Arctic weather prediction method mainly has three problems: first, the influence of sea ice density and sea ice thickness is ignored, which leads to inaccurate description of boundary layer parameterization and heat exchange process; second, the complex relationship between sea ice thickness and atmospheric heat exchange is not considered, and the coupling effect of sea ice surface temperature and thickness is ignored; third, there is a lack of physical consistency checking mechanism, which may lead to physical contradictions between sea ice parameters. The core technical innovations of the application include: first, in step S02, the physical consistency of sea ice density and thickness information is checked based on the sea ice thermodynamic equilibrium equation to ensure the physical rationality of the data; second, in step S03, the multi-element fusion of meteorological elements, sea ice density and sea ice thickness information is realized through the Arctic ice-ocean assimilation model; third, in step S04, the sea ice surface temperature is accurately calculated by applying the polar sea ice heat flux equation to effectively describe the sea ice thickness and atmospheric heat exchange process; fourth, in step S05, the sea ice density thickness boundary layer interaction equation set is established to calculate the sea ice modulated boundary layer structure; fifth, in step S06, the multi-source data is deeply fused by applying the feature fusion model, and the feature fusion model output weight is adaptively adjusted based on the ice-air interface balance index. These innovative technologies are verified in the practical application of the twelfth Arctic scientific expedition, which significantly improves the prediction accuracy of near-surface air temperature and dew point temperature, and the root mean square error of 2-meter air temperature and 2-meter dew point temperature 72-hour prediction is reduced by 10.3% and 14.7%, respectively.

[0266] The application successfully solves the technical problems of insufficient multi-source data fusion and inaccurate sea ice-atmosphere interaction process description in traditional methods, and provides a more reliable weather prediction basis for Arctic scientific expedition and polar navigation.

[0267] It should be noted that the variables involved in the application are explained in detail as shown in Tables 4, 5 and 6.

[0268] Table 4 Variable explanation table (first part)

[0269]

[0270] Table 5 Variable explanation table (second part)

[0271]

[0272] Table 6 Variable explanation table (third part)

[0273]

[0274] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. An Arctic atmospheric coupled forecasting method that automatically integrates sea ice density and thickness, characterized by: The method includes obtaining and correcting global GFS background field data, extracting and verifying sea ice density and sea ice thickness information, performing multi-factor fusion to generate a background field, applying the polar sea ice heat flux equation to calculate the sea ice surface temperature, establishing a sea ice density thickness boundary layer interaction equation group to calculate the boundary layer structure, applying a feature fusion model to deeply fuse sea ice density, sea ice thickness, sea ice surface temperature and sea ice modulated boundary layer structure, adaptively adjusting the weight of each element in the feature fusion model output vector based on the ice-air interface balance index, generating an optimized model initial field and boundary conditions, and executing an integration program to obtain a forecast result; the step of extracting and verifying sea ice density and sea ice thickness information specifically comprises extracting sea ice density and sea ice thickness information from the GFS background field data, performing physical consistency verification on the sea ice density and sea ice thickness information based on the sea ice thermodynamic balance equation, and rewriting the verified sea ice density and sea ice thickness information into an intermediate file recognizable by a model pre-processing program; The polar sea ice heat flux equation is specifically expressed as follows: ； ； ； ； Where, is the sea ice albedo, is the downward shortwave radiation flux, is the sea ice emissivity, is the downward longwave radiation flux, is the Stefan-Boltzmann constant, is the sea ice surface temperature, is the sensible heat flux, is the latent heat flux, is the conductive heat flux, is the air density, is the specific heat capacity of air at constant pressure, is the sensible heat exchange coefficient, is the near-surface wind speed, is the surface air temperature, is the latent heat of water vaporization, is the latent heat exchange coefficient, is the near-surface specific humidity, is the saturated specific humidity, is the sea ice thermal conductivity, is the freezing temperature of seawater, is the sea ice thickness, is the thermal conductivity coefficient, is the attenuation coefficient; The calculation of the boundary layer structure specifically involves: considering the modulation effect of sea ice on the atmospheric boundary layer structure, calculating the sea ice surface roughness, thermal stability parameters, boundary layer turbulent flux, boundary layer height and temperature and humidity gradients, and generating the sea ice modulated boundary layer structure; The weights of the elements in the feature fusion model output vector are adaptively adjusted based on the ice-air interface balance index. Specifically, when the ice-air interface balance index is less than 0.3, the weight of the sea ice density feature is increased by 30%; when the ice-air interface balance index is greater than 0.7, the weight of the sea ice surface temperature feature is increased by 25%. The ice-air interface balance index is obtained by calculating the product of the sea ice density and the sea ice thickness and dividing it by the surface net radiation flux, where the surface net radiation flux refers to the difference between the solar radiation received by the surface and the long-wave radiation emitted by the surface.

2. The method according to claim 1, characterized in that The GFS background field data contains 41 layers in the vertical direction, with a time interval of 3 hours and a single forecast duration of 168 hours. The ice-sea heat exchange correction equation is used to perform a preliminary correction on the boundary layer parameters of the sea ice area in the GFS background field data.

3. The method according to claim 2, characterized in that The step of executing multi-factor fusion to generate a background field specifically includes setting a model pre-processing parameter file, executing a terrain processing program, a format rewriting program, and an interpolation program, and fusing meteorological elements, the sea ice density, and the sea ice thickness information through an Arctic ice-sea assimilation model to generate a model background field with higher physical consistency.

4. The method according to claim 3, characterized in that The deep fusion specifically applies a feature fusion model to deeply fuse the sea ice density, the sea ice thickness, the sea ice surface temperature and the sea ice modulated boundary layer structure, and adaptively adjusts the weight of each element in the output vector of the feature fusion model based on the ice-air interface balance index.

5. The method according to claim 4, characterized in that The forecast results are weather forecast results for the next 7 days.

6. The method according to claim 3, characterized in that It also includes post-processing of weather forecast results, calculating forecast skill scores by comparing them with historical observation data, and generating standardized forecast products based on application needs in the Arctic region.

7. The method according to claim 6, characterized in that The GFS background field data refers to the gridded forecast data generated by the global forecast system that includes multiple layer elements of the atmosphere, ocean, land surface and sea ice.

Images