Tropical cyclone boundary layer parameterization method and system based on deep learning
Parameterized the tropical cyclone boundary layer by deep learning-based methods, the problem of uncertainty in turbulent transmission coefficients and insufficient interpretation ability of small-scale processes in traditional methods is solved, and more efficient data processing and prediction are achieved, and computing efficiency is improved. .
Patent Information
- Application Number
- CN202410850036.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-27
- Publication Date
- 2024-11-08
AI Technical Summary
In the application of traditional boundary layer parameterization methods, there is uncertainty in the turbulent transmission coefficient and insufficient interpretation ability of small-scale processes in the application of tropical cyclone boundary layers.
Using deep learning-based methods, we perform numerical simulation of meteorological topography data, obtain cyclone simulation data, and obtain turbulence integration flux through interpolation and integration, build a turbulence forecast model, and suppress the influence of extreme values through nonlinear transformation and normalization. Improve the simulation capabilities of the model.
The problem of uncertainty in turbulent transmission coefficients and insufficient interpretation ability of small-scale processes in traditional methods is solved, the model's simulation ability of actual situations is improved, and faster data processing and prediction are achieved, and the computing efficiency is improved.
Smart Images

Figure CN118917169A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of meteorological numerical simulation and calculation, and in particular to a tropical cyclone boundary layer parameterization method and system based on deep learning. Background Art
[0002] A tropical cyclone is a warm-core low-pressure vortex system formed on the warm ocean surface in tropical or subtropical regions. It is characterized by strong suddenness and great destructive power. Coastal areas often suffer from natural disasters such as strong winds, heavy rains and sea tides due to the landing of tropical cyclones, causing huge economic losses and casualties. However, tropical cyclones can also bring abundant precipitation to arid areas. Therefore, in-depth research on the formation mechanism of tropical cyclones is of great practical significance.
[0003] The tropical cyclone boundary layer plays an important role in the intensity evolution of tropical cyclones. In the theory of tropical cyclone development, the second type of conditional instability and cooperative reinforcement theory point out that the tropical cyclone boundary layer consumes kinetic energy, enhances water vapor accumulation, and provides unstable energy for the development of the cyclone. At the same time, the key physical processes in the development of tropical cyclones - wind-induced surface heat flux exchange and positive feedback between sea surface wind and water vapor flux - also mainly occur in the boundary layer. However, limited by the current computing power, mesoscale numerical models can only resolve processes at the kilometer level. Although the existing boundary layer parameterization methods perform well in reproducing turbulence and vertical mass flux diffusion in the convective boundary layer, they have significant defects when applied to the tropical cyclone boundary layer. For example, different boundary layer parameterization methods lead to uncertainties in the turbulent transport coefficient, which can explain the high sensitivity of the development, intensity and small-scale structure of tropical cyclones to the boundary layer parameterization. In addition, the boundary layer parameterization cannot explain the contribution of some small-scale processes to the turbulent flux, such as roll vortices and eyewall vorticity maximum, although these contributions are not negligible in tropical cyclones. Summary of the invention
[0004] In view of the problems existing in the existing traditional boundary layer parameterization method, the present invention is proposed.
[0005] Therefore, the problem to be solved by the present invention is that the application of the traditional boundary layer parameterization method to the tropical cyclone boundary layer has the uncertainty of the turbulent transmission coefficient and the insufficient ability to explain small-scale processes.
[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0007] In a first aspect, an embodiment of the present invention provides a tropical cyclone boundary layer parameterization method based on deep learning, which comprises:
[0008] Performing numerical simulation on meteorological and topographic data to obtain cyclone simulation data; the meteorological and topographic data includes static topographic data and gridded meteorological data;
[0009] Overlaying the cyclone simulation data onto a plurality of specific horizontal planes of the tropical cyclone boundary layer to obtain interpolation data;
[0010] Extracting and integrating the flux in the interpolation data to obtain a turbulence integrated flux, and processing the turbulence integrated flux;
[0011] A turbulence prediction model is constructed, and the processed turbulence integrated flux is input into the turbulence prediction model to obtain a parameterized result of the tropical cyclone boundary layer flux.
[0012] As a preferred solution of the tropical cyclone boundary layer parameterization method based on deep learning of the present invention, the acquisition of the interpolation data includes:
[0013] Obtain the distance between irregular grid points and regular grid points, and determine the nearest neighbor of each grid point on the regular grid;
[0014] The influence weight of each irregular grid point on the interpolation point on the regular grid is calculated based on the distance and correlation information of the nearest neighbor point;
[0015] Traversing each grid point on the regular grid, performing high-order interpolation calculation on each grid point using the data of its nearest neighbor point and the influence weight, to obtain the function value of the point to be interpolated;
[0016] The calculation of the function value of the interpolation point is as follows:
[0017] Q(x,y)=(1-α)(1-β)Q 00 +α(1-β)Q 10 +(1-α)βQ 01 +αβQ 11
[0018] Among them, Q 00 , Q 10 , Q 01 and Q 11 are the four adjacent grid points around the regular grid point, α and β are the interpolation weights in the horizontal and vertical directions, which are determined by the relative distances between the target point and the adjacent points, respectively.
[0019] As a preferred solution of the tropical cyclone boundary layer parameterization method based on deep learning described in the present invention, the flux in the interpolation data is extracted by determining the turbulent disturbance part in each small block and the average variable of each timestamp, and the flux in the interpolation data is obtained by subtracting the turbulent disturbance part in each small block from the average variable.
[0020] As a preferred solution of the tropical cyclone boundary layer parameterization method based on deep learning of the present invention, wherein: the flux extracted from the interpolation data also includes the surface flux, the surface flux is filtered and integrated with the analyzed turbulent flux;
[0021] The surface flux filtering process includes:
[0022] Comprehensively process each time point and spatial location data and determine the optimal estimate of the current time point;
[0023] Determine the weighted average of neighboring points based on the processed spatial position data;
[0024] Combine the results of comprehensive processing to determine the final spatiotemporal consistency data.
[0025] As a preferred embodiment of the tropical cyclone boundary layer parameterization method based on deep learning of the present invention, the weighted average value of the calculated neighboring points is as follows:
[0026]
[0027] in, is the filtered value, Q(i+m,j+n) is the value of the neighboring point, and w(m,n) is the weight coefficient.
[0028] As a preferred embodiment of the tropical cyclone boundary layer parameterization method based on deep learning of the present invention, the nonlinear transformation of the turbulent integrated flux includes:
[0029] Determining a mean and a standard deviation of the turbulent flux data, determining a center point of the turbulent flux data based on the mean, and evaluating a dispersion of the data based on the standard deviation;
[0030] Obtaining a standardized deviation of each element in the turbulent flux data, and determining a turbulent flux extreme value index according to the standardized deviation;
[0031] The turbulence integrated flux data are passed through a nonlinear logarithmic transformation function to adjust the distribution shape of the data;
[0032] The calculation of the turbulent flux extreme index is shown in the following formula:
[0033]
[0034] Where M is the number of rows, H is the number of columns, each element turflux(j,k) represents the turbulent flux value in the j-th row and k-th column, σ is the standard deviation, and μ is the mean.
[0035] As a preferred embodiment of the tropical cyclone boundary layer parameterization method based on deep learning of the present invention, the turbulence prediction model is as follows:
[0036]
[0037] Among them, F is the parameterized model of deep learning, U, V, W are the three components of velocity, θ is the potential temperature, Q is the water vapor mixing ratio, Qc is the cloud water mixing ratio, is the turbulent output flux.
[0038] In a second aspect, an embodiment of the present invention provides a tropical cyclone boundary layer parameterization system based on deep learning, which includes:
[0039] The data acquisition module is used to obtain static terrain data and gridded NCEP reanalysis meteorological data, conduct large eddy simulation on idealized tropical cyclones, reproduce the evolution process of tropical cyclones, and obtain cyclone simulation data;
[0040] Data interpolation module, used to interpolate ideal tropical cyclone simulation data to a specific horizontal plane of the boundary layer to ensure sufficient spatial coverage of meteorological data;
[0041] The flux integration module is used to obtain directly resolved turbulent fluxes and resolved variables from the interpolated data, extract the surface flux from the interpolated data, and introduce a consistent constraint filtering process to integrate it with the resolved turbulent flux to obtain the turbulent surface integrated flux, obtain the sub-grid scale turbulent flux from the interpolated data, and integrate it with the turbulent surface integrated flux to obtain the final turbulent integrated flux;
[0042] The model building and forecasting module is used to build a turbulence forecasting model and obtain the parameterization results of tropical cyclone boundary layer flux through training.
[0043] In a third aspect, an embodiment of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the processor executes the computer program, it implements any step of the above-mentioned deep learning-based tropical cyclone boundary layer parameterization method.
[0044] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the above-mentioned deep learning-based tropical cyclone boundary layer parameterization method is implemented.
[0045] The beneficial effect of the present invention is to use a deep learning model to explain complex boundary layer processes, solving the problems of uncertainty in turbulent transport coefficients and insufficient ability to explain small-scale processes in the application of traditional boundary layer parameterization methods in tropical cyclone boundary layers. The turbulent integrated flux is transformed nonlinearly, and the influence of extreme values is suppressed through nonlinear transformation and normalization processing, making the turbulent flux data smoother and more stable, thereby improving the model's ability to simulate actual conditions. The constructed VT-BL self-attention mechanism realizes parallel computing, which can process and predict data more quickly than traditional parameterization methods, providing higher computing efficiency in weather research and forecasting models. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:
[0047] Figure 1 Flowchart of the tropical cyclone boundary layer parameterization method based on deep learning.
[0048] Figure 2 Schematic diagram of the WRF_LES simulation area based on the deep learning-based tropical cyclone boundary layer parameterization method.
[0049] Figure 3 The distribution diagram of turbulent flux at altitudes of 50m, 387m, and 906m in large eddy simulation of the tropical cyclone boundary layer parameterization method based on deep learning.
[0050] Figure 4 Schematic diagram of the overall structure of the VT-BL network model for tropical cyclone boundary layer parameterization method based on deep learning.
[0051] Figure 5 Schematic diagram of the structure of the multi-head adaptive attention module in the VT-BL network model of the tropical cyclone boundary layer parameterization method based on deep learning.
[0052] Figure 6 Scatter plot of turbulent flux forecast output by the VT-BL network model based on deep learning-based tropical cyclone boundary layer parameterization method.
[0053] Figure 7 A comparison of the spatial distribution of turbulent flux at 251m using YSU, large eddy simulation, and VT-BL for tropical cyclone boundary layer parameterization methods based on deep learning. DETAILED DESCRIPTION
[0054] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.
[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0056] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.
[0057] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.
[0058] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0059] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0060] Example 1
[0061] Reference Figure 1 to Figure 7, which is the first embodiment of the present invention, and provides a tropical cyclone boundary layer parameterization method based on deep learning, comprising:
[0062] Obtain static terrain data and gridded NCEP reanalysis meteorological data and perform data preprocessing.
[0063] The static terrain data are digital elevation models and land cover type data, which are used to simulate the energy and water exchange process of the surface. The gridded meteorological data uses NCEP reanalysis data with a data accuracy of 0.25°*0.25° and a time resolution of every 6 hours.
[0064] Furthermore, the idealized tropical cyclone was simulated using the numerical simulation method WRF to reproduce the evolution of the tropical cyclone and obtain cyclone simulation data;
[0065] LES simulations of tropical cyclones were performed using WRFv3.9.1.1. Four nested domains were set up with their centers coinciding, named Domain 1 to Domain 4. The side lengths of these domains are 2000, 1200, 750, and 200 km, respectively, with horizontal resolutions of 12.5, 2.5, 0.5, and 0.1 km, and time steps of 25, 5, 1, and 0.2 seconds, respectively. Domain 1 adopts periodic boundary conditions involving 69 vertical layers, of which 15 are below 1200 m. The vertical grid length gradually increases from tens of meters in the boundary layer to about 900 m near the top of the model, which is set at 22 km. Rayleigh damping is applied within the upper 5 km to dissipate reflected gravity waves. The sea surface temperature is fixed at 28 °C, and the initial environmental profile uses the average sounding data from Jordan.
[0066] The initial vortex follows the analytical vortex of Rotunno and Emanuel, and the vortex parameters are the default values in WRF. The outer radius of the vortex is 412.5 km, the maximum wind speed radius is 82.5 km, the maximum wind speed and the vortex depth are 15 m / s and 20 km, respectively. Lin microphysical parameterization is used in all four domains, the longwave radiation parameterization adopts the Goddard and Dudhia scheme, and the shortwave radiation parameterization adopts the Goddard and Dudhia scheme. The modified Monin-Obukhov scheme is used for the surface layer, and the Kain-Fritsch cumulus convection parameterization is only applied to domain 1. The YSU-BL scheme is used for domains 1 and 2. In LES, domains 3 and 4 use the 1.5-order turbulent kinetic energy closure scheme as the SGS parameterization. In order to improve numerical stability, the 6th-order numerical diffusion coefficient is set to 0.2.
[0067] Furthermore, the ideal tropical cyclone simulation data is interpolated to 14 specific horizontal planes of the boundary layer using the weighted bilinear interpolation method to ensure sufficient spatial coverage of meteorological data. Finally, the interpolation results are written into the output file. The ideal tropical cyclone simulation data includes six variables: three components of wind speed, potential temperature, water vapor mixing ratio, and cloud water mixing ratio.
[0068] For example, in one possible embodiment, the cyclone simulation data is interpolated to a specific level of the boundary layer as follows:
[0069] The nearest neighbor search algorithm is used. In the process of searching for the nearest neighbor, the Euclidean distance is used to measure the distance between the irregular grid points and the regular grid points. The nearest neighbor of each grid point on the regular grid is determined based on the calculated distance.
[0070] According to the distance and correlation information of the neighboring points, the weighted least squares method is used to calculate the influence weight of each irregular grid point on the point to be interpolated on the regular grid. Each grid point on the regular grid is traversed. For each grid point, the data of its neighboring points and the weight coefficient are used to perform high-order interpolation calculation to obtain the function value of the point to be interpolated.
[0071] Divide the space into multiple regular grid cells, each grid cell contains one or more irregular grid points, assign the irregular grid points to corresponding grid cells according to their spatial positions, construct a grid index, and for each grid cell, record the index and coordinate information of the irregular grid points contained therein;
[0072] The weighted bilinear interpolation method uses the values of four adjacent grid points for interpolation, and determines the value of the target point through the interpolation weights α and β. The improved bilinear interpolation calculation formula is:
[0073] Q(x,y)=(1-α)(1-β)Q 00 +α(1-β)Q 10 +(1-α)βQ 01 +αβQ 11
[0074] Among them, Q 00 , Q 10 , Q 01 and Q 11 are the four adjacent grid points around the regular grid point, α and β are the interpolation weights in the horizontal and vertical directions, which are determined by the relative distances between the target point and the adjacent points, respectively.
[0075] Furthermore, directly resolved turbulent fluxes and analytical variables are obtained from the interpolated data;
[0076] By specifying the input path and output path, as well as the average scaling, the source of the data and the storage location after processing are determined. Then, by looping through the hours and minutes, the meridional wind speed (u), zonal wind speed (v), vertical wind speed (w), potential temperature (th), water vapor mixing ratio (qv), and cloud water mixing ratio (qc) are read from the interpolated high-resolution variables.
[0077] For each timestamp, the average variable is calculated, that is, the grid is divided into small blocks, and the average value of each small block is calculated to obtain the average horizontal wind speed (u avg ), average vertical wind speed (v avg ), average vertical wind speed (w avg ), average potential temperature (th avg ), average water vapor mixing ratio (qv avg ) and the average cloud water mixing ratio (qc avg ).
[0078] In order to calculate the resolved turbulent flux, the grid is divided into small blocks, and the turbulent disturbance part in each small block is calculated, that is, the value in each small block is subtracted from the average value. Then, these turbulent disturbance parts are used to calculate the resolved turbulent flux, including vertical turbulent momentum flux (uw, vw), potential temperature turbulent flux (thw) and water vapor mixing ratio turbulent flux (qvw). The vertical turbulent momentum flux is calculated as:
[0079]
[0080] Where w′ and u′ represent the turbulent disturbance part of the vertical wind speed and the meridional wind speed, respectively. and It represents the average value of vertical wind speed and meridional wind speed;
[0081] The calculation of potential temperature turbulent flux is:
[0082]
[0083] Where θ′ represents the turbulent disturbance part of the potential temperature, Indicates the average value of potential temperature;
[0084] The calculation of the turbulent flux of water vapor mixing ratio is:
[0085]
[0086] Where q′ represents the turbulent disturbance part of the water vapor mixing ratio, Represents the average value of the water vapor mixing ratio.
[0087] Furthermore, the surface flux is extracted from the interpolated data, and a filtering process with consistency constraints is introduced to integrate it with the resolved turbulent flux to obtain the turbulent surface integrated flux;
[0088] Furthermore, surface fluxes are extracted from the WRF-LES output, including the turbulent stress between horizontal and vertical wind speeds (uw sf 、vw sf ), turbulent heat flux (thw sf ) and the turbulent flux of water vapor mixing ratio (qvw sf ).
[0089] First, define the data path and output path, as well as the gas constant (r_d) and specific heat capacity (cp) of dry air used to calculate the surface flux. Loop through the timestamps and read the required surface flux variables, including friction velocity (UST), surface sensible heat flux (HFX), surface water vapor flux (QFX), surface static pressure (PSFC), temperature (T), water vapor mixing ratio (QVAPOR), horizontal wind speed (U), and vertical wind speed (V).
[0090] According to the calculation formula of surface flux, the turbulent stress between horizontal velocity and vertical velocity, turbulent heat flux and turbulent flux of water vapor mixing ratio are calculated. The specific calculation formula is as follows:
[0091] Calculation of air density (ρ):
[0092]
[0093] Calculation of specific heat correction term (Cpm):
[0094] Cpm=C p (1+0.8q)
[0095] Horizontal and vertical turbulence stresses (uw sf 、vw sf ) calculation:
[0096]
[0097] Turbulent heat flux (thw sf ) calculation:
[0098]
[0099] Turbulent flux of water vapor mixing ratio (qvw sf ) calculation:
[0100]
[0101] After calculating the surface flux, the surface flux data is subjected to spatiotemporal joint filtering, and the spatiotemporal relationship is used to optimize the data. Through filtering, the consistency and continuity of the surface flux data in time and space are ensured, and possible noise and outliers are eliminated.
[0102] For example, in one feasible embodiment, the filtering process is as follows:
[0103] For the data at each time point and spatial position, a joint filtering algorithm is applied, and the Kalman filter is used for temporal consistency processing. In the temporal filtering process, the Kalman filter performs a comprehensive processing of the data at the past time point and the data at the current time point through the prediction-update process, and calculates the optimal estimate value at the current time point:
[0104]
[0105] in, is the updated state estimate, is the predicted state estimate, K k is the Kalman gain, z k is the observation value, H is the observation matrix. In the spatial filtering process, a Gaussian filter is used to smooth the data at each spatial position and calculate the weighted average of the neighboring points:
[0106]
[0107] in, is the filtered value, Q(i+m,j+n) is the value of the neighboring point, and w(m,n) is the weight coefficient.
[0108] Combining the results of time filtering and space filtering, the final spatiotemporal consistency data is calculated by weighted averaging to ensure the smoothness and continuity of the data in time and space:
[0109]
[0110] Among them, α is the weighting coefficient, is the time filtering result, is the spatial filtering result.
[0111] Furthermore, the sub-grid scale turbulent flux is obtained from the interpolated data and integrated with the turbulent surface integrated flux to obtain the final turbulent integrated flux.
[0112] Starting from the surface fluxes and subgrid-scale trends, the subgrid-scale fluxes at the full level are obtained by interpolating at the half level. Using the surface fluxes and subgrid-scale trends, the subgrid-scale fluxes at the full level are inferred using the trapezoidal rule. Next, the average subgrid-scale fluxes are calculated to be added to the resolved fluxes to obtain the final turbulent fluxes, thereby obtaining more accurate turbulent fluxes.
[0113] Preferably, sub-grid scale turbulent fluxes are obtained from the interpolated data.
[0114] For example, in one feasible embodiment, as follows:
[0115] Keeping the original spatial structure and physical characteristics, four neighboring points on the target half-level are selected and cubic interpolation is performed on the half-level.
[0116] Using the data of surface fluxes and subgrid-scale trends, the subgrid-scale fluxes on the complete level are inferred using the trapezoidal rule. For each level, the area of the trapezoid between two adjacent levels is calculated:
[0117]
[0118] Among them, Q i and Q j are the flux values of two adjacent levels, and Δz is the height difference between the levels.
[0119] After obtaining the subgrid-scale flux at the complete level, the average processing is performed. After calculating the average subgrid-scale flux, the subgrid-scale flux is merged with the turbulent surface integrated flux through the weighted average method to obtain the final turbulent integrated flux.
[0120] Furthermore, the turbulent integrated flux is transformed nonlinearly and normalized
[0121] For example, in one feasible embodiment, as follows:
[0122] First, calculate the mean of the entire turbulent flux data and find the center point of the data as follows:
[0123]
[0124] Where M is the number of rows, H is the number of columns, and each element turflux(j,k) represents the turbulent flux value of the jth row and kth column;
[0125] Then calculate the standard deviation of the entire turbulent flux data to evaluate the dispersion of the data. The formula is as follows:
[0126]
[0127] For each element turflux(j,k) in the turbulent flux data, calculate its standardized deviation, that is, the deviation of the value from the mean divided by the standard deviation, to obtain the degree of deviation of the value from the mean, sum all standardized deviations, and then divide by the total number of elements to obtain the turbulent flux extreme value index ETFI, the formula is as follows:
[0128]
[0129] The turbulence integrated flux data is transformed through a nonlinear logarithmic transformation function to adjust the distribution of the data so that the extreme values are suppressed while maintaining the overall trend and sign of the data. The nonlinear logarithmic transformation function is as follows:
[0130] T log (j,k)=log(turflux(j,k)+1)
[0131] Among them, T log (j, k) is the turbulent flux value after logarithmic transformation, turflux(j, k) is the original turbulent flux value, and 1 is added to avoid negative and zero value problems in logarithmic operations;
[0132] After the nonlinear logarithmic transformation, in order to further alleviate the imbalanced nature of the data, the data is normalized, and the formula is as follows:
[0133]
[0134] Among them, T min and T max are the minimum and maximum values of the transformed data respectively;
[0135] The normalized data is adjusted in combination with the turbulent flux extreme index to further smooth the influence of extreme values and make the data more uniform and stable. The formula is as follows:
[0136]
[0137] Furthermore, a turbulence forecast model VT-BL was constructed, and the parameterization results of tropical cyclone boundary layer flux were obtained through training.
[0138] For example, in one feasible embodiment, as follows:
[0139] VT-BL parameterization can be abstracted as follows:
[0140]
[0141] F is a parameterized model for deep learning. The input or output vectors are concatenated as a two-dimensional matrix with aligned lengths. The length of the vectors is 14, corresponding to the 14 levels selected below 1500 meters. The input resolution variables are the three components of velocity (U, V, W), potential temperature (Θ), water vapor mixing ratio (Q), and cloud water mixing ratio (Qc).
[0142] The turbulent output flux is
[0143] The tropical cyclone boundary layer has obvious vertical structural characteristics, and the turbulence characteristics and variable interaction relationships at different altitudes are different. For example, the altitude layers close to the ground (such as layers 1 and 2) are mainly affected by surface friction, while higher layers (such as layers 13 and 14) are mainly affected by atmospheric stability and convection. Therefore, each altitude layer requires an independent mechanism to learn and adjust its internal interaction relationships. The adaptive attention mechanism can capture the complex interactions of input variables at different altitudes.
[0144] Traditional attention mechanisms generally use globally shared weight parameters, which cannot fully capture the differences between different height layers. In order to enhance the expressiveness and flexibility of the model, an independent learnable parameter matrix W is designed for each height layer to adjust the attention weight of the layer. A multi-layer Transformer structure is used, and each layer includes a multi-head self-attention mechanism and a feedforward neural network layer to perform deep feature extraction and representation learning on the input sequence.
[0145] In each Transformer layer, the self-attention mechanism is used to capture long-distance dependencies in the input sequence. The model pays attention to the input data at different positions to better capture the correlation between the data. In each Transformer layer, a feedforward neural network layer is included to perform nonlinear transformation and dimensional mapping on the features, dividing the input sequence data into blocks. Each block contains a certain amount of input data, reducing the dimension of the input data and improving the processing efficiency of the model. For each block, the Vit module is applied to extract spatial information.
[0146] The Vit module maps the input block into a continuous low-dimensional space, and uses the self-attention mechanism to capture the spatial relationship within and between blocks. The features of all blocks processed by the Vit module are fused to obtain global spatial information. Residual connections and layer normalization are introduced inside the encoder and decoder to alleviate the gradient vanishing problem in deep networks, accelerate model training, improve model stability, and enhance the prediction ability of turbulent flux.
[0147] Preferably, the VT-BL model introduces an adaptive attention mechanism to adaptively adjust the attention weights to capture the interaction relationship between input variables at different height levels.
[0148] For example, in a feasible embodiment, the specific steps of the adaptive attention mechanism are as follows:
[0149] Each height layer randomly initializes an independent learnable parameter matrix W with a size of 6*6, which is used to adjust the attention weight of the layer. The 14 height layers correspond to 14 W i (i=1,2,…,14);
[0150] The input matrix X is used as the input of the adaptive attention mechanism, and the query vector Q, key vector K and value vector V are obtained by three independent linear transformations;
[0151] Calculate the dot product between the query vector Q and the key vector K and pass it through W i Adjust to get the adaptive attention weight α i , the formula is as follows:
[0152] α i =softmax(f(Q i ,K i ,W i ))
[0153] Among them, Q i represents the value of the query vector at the i-th altitude layer, K i represents the value of the key vector at the i-th height layer, and the weight generation function f(Q i ,K i ,W i ) is realized through linear transformation and is defined as follows:
[0154]
[0155] Among them, d is the scaling factor, which is generally set to the dimension of the key vector to balance the attention distribution;
[0156] Use the calculated adaptive attention weight αi to the value vector V i Perform weighted summation to obtain a weighted output vector.
[0157] Furthermore, this embodiment also provides a tropical cyclone boundary layer parameterization system based on deep learning, including:
[0158] The data acquisition module is used to obtain static terrain data and gridded NCEP reanalysis meteorological data, conduct large eddy simulation on idealized tropical cyclones, reproduce the evolution process of tropical cyclones, and obtain cyclone simulation data;
[0159] Data interpolation module, used to interpolate ideal tropical cyclone simulation data to a specific horizontal plane of the boundary layer to ensure sufficient spatial coverage of meteorological data;
[0160] The flux integration module is used to obtain directly resolved turbulent fluxes and resolved variables from the interpolated data, extract the surface flux from the interpolated data, and introduce a consistent constraint filtering process to integrate it with the resolved turbulent flux to obtain the turbulent surface integrated flux, obtain the sub-grid scale turbulent flux from the interpolated data, and integrate it with the turbulent surface integrated flux to obtain the final turbulent integrated flux;
[0161] The model building and forecasting module is used to build a turbulence forecasting model and obtain the parameterization results of tropical cyclone boundary layer flux through training.
[0162] This embodiment also provides a computer device, which is suitable for the tropical cyclone boundary layer parameterization method based on deep learning, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the tropical cyclone boundary layer parameterization method based on deep learning proposed in the above embodiment.
[0163] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.
[0164] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the tropical cyclone boundary layer parameterization method based on deep learning as proposed in the above embodiment.
[0165] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.
[0166] Example 2
[0167] Reference Figure 2 to Figure 7 , which is the second embodiment of the present invention, provides a tropical cyclone boundary layer parameterization method based on deep learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.
[0168] The FNL data of the historical global reanalysis data of the NCEP Center for Environmental Prediction of the United States was used as the initial field and boundary conditions, and the WRFv3.9.1 mesoscale numerical model was used to simulate the large eddy of tropical cyclones. Four nested domains were set, and their centers coincided, named domain d01 to domain d04. The side lengths of these domains were 1500, 850, 550 and 100 kilometers, respectively, with horizontal resolutions of 9, 3, 1 and 0.2 kilometers, respectively, and time steps of 45, 15, 5 and 1 seconds, respectively.
[0169] The boundary conditions used in domain 1 involve 45 vertical layers, of which 14 are below 1500 m. The vertical grid length gradually increases from tens of meters in the boundary layer to about 900 m near the top of the model, which is set at 22 km. Rayleigh damping is applied within the upper 5 km to dissipate reflected gravity waves. The sea surface temperature is fixed at 28 °C, and the initial environmental profile uses the average sounding data from Jordan. The four domains all use the WSM5 microphysical parameterization. In the calculation process, cloud ice and cloud water are considered to be the same category, so this scheme is a simple ice scheme. When diagnosing the number concentration of ice, the WSM5 scheme is not based on the temperature of ice, but focuses on the mass content of ice. This scheme is an improved version of the NCEP5 scheme. The longwave radiation parameterization uses the RRTM scheme. The longwave processes caused by the optical thickness of clouds, carbon dioxide, ozone, and water vapor are expressed using a pre-processed comparison table. The shortwave radiation parameterization uses the Dudhia scheme. The changes in solar radiation flux include factors such as cloud reflection and absorption, and dry air scattering.
[0170] The modified Monin-Obukhov scheme is used for the surface layer. The Kain-Fritsch cumulus convection parameterization is only applied to domain 1, which includes relatively rough microphysical processes, water vapor entrainment and entrainment, and rising and sinking motions into a simple cloud model. The YSU boundary layer parameterization scheme is used for domains d01 and d02. In LES, the 1.5-order turbulent kinetic energy closure scheme is used as the SGS parameterization for domains d03 and d04. In order to improve numerical stability, the 6th-order numerical diffusion coefficient is set to 0.2.
[0171] The specific experimental data are shown in the following table:
[0172] Table 1 Experimental data table
[0173]
[0174] It can be seen from the above table that the scheme of the present invention shows excellent results in all simulation indicators.
[0175] Specifically:
[0176] The absolute errors of average wind speed, potential temperature and water vapor mixing ratio are the smallest, which are 1.95m / s, 0.63K and 0.0014 respectively, which are 39.3%, 45.2% and 56.3% lower than those of existing technologies, significantly improving the simulation accuracy of these key variables.
[0177] The absolute errors of turbulent momentum flux, potential temperature flux and water vapor flux are also the smallest, which are 0.18m^2 / s^2, 0.07K*m / s and 0.0005 respectively, which are 60%, 66.7% and 72.2% lower than the existing technology, greatly improving the accuracy of turbulence output.
[0178] The computational efficiency is optimal, with each time step taking only 125 seconds, which is 31% higher than the existing technology, reflecting the innovative contribution of the invention in improving computing performance.
[0179] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A tropical cyclone boundary layer parameterization method based on deep learning, characterized by: include, Performing numerical simulation on meteorological and topographic data to obtain cyclone simulation data; the meteorological and topographic data includes static topographic data and gridded meteorological data; Overlaying the cyclone simulation data onto a plurality of specific horizontal planes of the tropical cyclone boundary layer to obtain interpolation data; Extracting and integrating the flux in the interpolation data to obtain a turbulence integrated flux, and processing the turbulence integrated flux; A turbulence prediction model is constructed, and the processed turbulence integrated flux is input into the turbulence prediction model to obtain a parameterized result of the tropical cyclone boundary layer flux.
2. The tropical cyclone boundary layer parameterization method based on deep learning according to claim 1, characterized in that: The acquisition of the interpolation data includes: Obtain the distance between irregular grid points and regular grid points, and determine the nearest neighbor of each grid point on the regular grid; The influence weight of each irregular grid point on the interpolation point on the regular grid is calculated based on the distance and correlation information of the nearest neighbor point; Traversing each grid point on the regular grid, performing high-order interpolation calculation on each grid point using the data of its nearest neighbor point and the influence weight, to obtain the function value of the point to be interpolated; The calculation of the function value of the interpolation point is as follows: Q(x,y)=(1-a)(1-b)Q 00 +α(1-β)Q 10 +(1-α)βQ 01 +αβQ 11 Among them, Q 00 , Q 10 , Q 01 and Q 11 are the four adjacent grid points around the regular grid point, α and β are the interpolation weights in the horizontal and vertical directions, which are determined by the relative distances between the target point and the adjacent points, respectively.
3. The tropical cyclone boundary layer parameterization method based on deep learning according to claim 2, characterized in that: The flux in the interpolated data is extracted by determining the turbulent disturbance part in each small block and the average variable at each time stamp, and the flux in the interpolated data is obtained by subtracting the turbulent disturbance part in each small block from the average variable.
4. The tropical cyclone boundary layer parameterization method based on deep learning according to any one of claims 1 or 3, characterized in that: Extracting flux from the interpolation data also includes surface flux, filtering the surface flux and integrating it with the analyzed turbulent flux; The surface flux filtering process includes: Comprehensively process each time point and spatial location data and determine the optimal estimate of the current time point; Determine the weighted average of neighboring points based on the processed spatial position data; Combine the results of comprehensive processing to determine the final spatiotemporal consistency data.
5. The tropical cyclone boundary layer parameterization method based on deep learning according to claim 4, characterized in that: The weighted average of the neighboring points is calculated as follows: in, is the filtered value, Q(i+m,j+n) is the value of the neighboring point, and w(m,n) is the weight coefficient.
6. The tropical cyclone boundary layer parameterization method based on deep learning according to claim 1, characterized in that: The nonlinear transformation of the turbulent integrated flux includes, Determining a mean and a standard deviation of the turbulent flux data, determining a center point of the turbulent flux data based on the mean, and evaluating a dispersion of the data based on the standard deviation; Obtaining a standardized deviation of each element in the turbulent flux data, and determining a turbulent flux extreme value index according to the standardized deviation; The turbulence integrated flux data are passed through a nonlinear logarithmic transformation function to adjust the distribution shape of the data; The calculation of the turbulent flux extreme index is shown in the following formula: Where M is the number of rows, H is the number of columns, each element turflux(j,k) represents the turbulent flux value in the j-th row and k-th column, σ is the standard deviation, and μ is the mean.
7. The tropical cyclone boundary layer parameterization method based on deep learning according to claim 6, characterized in that: The turbulence prediction model is shown in the following formula: Among them, F is the parameterized model of deep learning, U, V, W are the three components of velocity, θ is the potential temperature, Q is the water vapor mixing ratio, Qc is the cloud water mixing ratio, is the turbulent output flux.
8. A tropical cyclone boundary layer parameterization system based on deep learning, based on the tropical cyclone boundary layer parameterization method based on deep learning according to any one of claims 1 to 7, characterized in that: include, The data acquisition module is used to obtain static terrain data and gridded NCEP reanalysis meteorological data, conduct large eddy simulation on idealized tropical cyclones, reproduce the evolution process of tropical cyclones, and obtain cyclone simulation data; Data interpolation module, used to interpolate ideal tropical cyclone simulation data to a specific horizontal plane of the boundary layer to ensure sufficient spatial coverage of meteorological data; The flux integration module is used to obtain directly resolved turbulent fluxes and resolved variables from the interpolated data, extract the surface flux from the interpolated data, and introduce a consistent constraint filtering process to integrate it with the resolved turbulent flux to obtain the turbulent surface integrated flux, obtain the sub-grid scale turbulent flux from the interpolated data, and integrate it with the turbulent surface integrated flux to obtain the final turbulent integrated flux; The model building and forecasting module is used to build a turbulence forecasting model and obtain the parameterization results of tropical cyclone boundary layer flux through training.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the tropical cyclone boundary layer parameterization method based on deep learning described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the tropical cyclone boundary layer parameterization method based on deep learning according to any one of claims 1 to 7 are implemented.
Citation Information
Cited By
TCBL wind field high-resolution LES simulation method facing gray area parameterization, medium and program product
CN119272530A
Method for improving TCBL grey area scale perception similarity function based on LES simulation optimization, medium and program product
CN119272533A
LES ultrahigh-resolution TC large airflow field numerical simulation method and system based on airship radar and down-cast sounding data assimilation driving
CN121351709A
LES super-high-resolution tc atmospheric flow field numerical simulation method and system
CN121351709B
