Meteorological element gridding processing method based on multi-air-pressure-layer data fusion

Abstract

The invention provides a meteorological element gridding processing method based on multi-air-pressure-layer data fusion, and belongs to the technical field of meteorological data processing and deep learning. The method comprises the following steps: S1, carrying out preprocessing, quality control and standard unified operation on data; s2, carrying out space gridding conversion; s3, carrying out multi-layer data fusion operation on the meteorological data and carrying out interlayer physical consistency constraint; s4, performing feature enhancement operation on the fused meteorological elements, and performing normalization operation; and S5, merging the feature channels in the meteorological elements after feature enhancement to form a multi-dimensional tensor. The method has the beneficial effects that the physical significance of meteorological data is reserved by cooperatively processing data of a plurality of barometric layers, designing a self-adaptive gridding mapping algorithm, mining a coupling relationship among different barometric layers and performing data fusion based on a meteorological principle, so that the meteorological data adapts to an input format of a deep learning model; and the meteorological prediction precision and the generalization ability of the model are improved.

Description

Technical Field

[0001] This invention belongs to the fields of meteorological data processing and deep learning technology, and proposes a gridded processing method for meteorological elements based on multi-barometric layer data fusion. This method solves the problems of data structure mismatch, poor spatial interpolation accuracy, and insufficient utilization of vertical information in traditional meteorological data processing methods in the field of deep learning through collaborative processing of multi-barometric layer data, adaptive gridded mapping, and feature fusion algorithms. Background Technology

[0002] Weather forecasting is an indispensable scientific and technological tool in modern society, profoundly impacting various industries such as agriculture, aviation, transportation, and energy. Accurate weather forecasts can anticipate extreme weather events, safeguarding people's lives and property and improving the efficiency of natural resource utilization. However, with the intensification of climate change, weather forecasting faces increasingly complex challenges, such as the massive volume of meteorological data, the complexity of its spatial and temporal dimensions, and the variability of nonlinear relationships.

[0003] Traditional weather forecasting methods are primarily based on numerical weather prediction (NWP) models. These methods rely on a combination of numerous physical equations and meteorological observation data to make short-term or long-term weather predictions through numerical simulation. Although these models have made significant progress in prediction accuracy, their complexity still poses considerable challenges in data processing, computational costs, and model optimization. Furthermore, traditional methods also suffer from insufficient utilization of the spatial correlation and vertical structure information of meteorological elements, affecting the effective capture of high-dimensional weather changes.

[0004] In recent years, deep learning technology has made groundbreaking progress in fields such as image recognition and natural language processing. Its powerful data processing and modeling capabilities have provided new ideas and methods for weather forecasting. Deep learning models, especially convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM), can automatically learn complex spatial and temporal patterns from large-scale meteorological data, and therefore their application prospects in the field of weather forecasting are widely regarded as promising.

[0005] The main advantages of deep learning in weather forecasting are: Automatic feature extraction and learning: Deep learning models can automatically extract useful features from raw meteorological data without the need for manual design of complex feature engineering, making them more efficient than traditional methods in high-dimensional data processing.

[0006] Modeling capability of nonlinear relationships: Traditional weather forecasting methods rely on linear models, which make it difficult to capture complex nonlinear relationships, while deep learning can effectively model complex nonlinear relationships between meteorological elements and improve forecast accuracy.

[0007] High-efficiency computing power: With the help of modern computing platforms, deep learning can be trained and inferred efficiently on large-scale datasets, thus meeting the challenges brought by big data.

[0008] Meteorology is a three-dimensional dynamic system, and meteorological data from different pressure layers reflect the state of the atmosphere at different levels. In recent years, with the advancement of meteorological observation technology, more and more meteorological stations are able to provide data from multiple pressure layers, such as the 500 hPa, 850 hPa, and 1000 hPa layers. This data provides important information for a deeper understanding and forecasting of atmospheric processes. However, how to effectively integrate data from different pressure layers has become a major challenge in meteorological data processing. Summary of the Invention

[0009] The main objective of this invention is to provide a method for gridded processing of meteorological elements based on multi-barometric layer data fusion, thereby solving the problems mentioned in the background art.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for gridded processing of meteorological elements based on multi-pressure layer data fusion, comprising the following steps: S1. Perform data preprocessing, quality control, and standardization operations; S2. Perform spatial gridding transformation on meteorological data; S3. Perform multi-layer data fusion on meteorological data and impose physical consistency constraints between layers; S4. Perform feature enhancement and normalization operations on the fused meteorological elements; S5. Merge the feature channels in the enhanced meteorological elements to form a multidimensional tensor.

[0011] Furthermore, the detailed process of step S1 is as follows: S101. Select typical pressure layers as input layers, including: 500 hPa, 850 hPa and 1000 hPa; , , These represent the temperature values ​​of the corresponding pressure layers, in that order. , , These represent the air pressure values ​​of the corresponding pressure layers, in that order. , , The wind speeds at the corresponding layers are represented sequentially. Each meteorological element is organized using station number, latitude and longitude coordinates, and observation timestamp as indexes, and stored uniformly for subsequent processing; S102. Observational data that exceeds the limits of physical common sense or the range of equipment shall be discarded; missing data shall be filled in. All data quality processing steps are logged to ensure process traceability; S103. Align all data to a unified time base and spatial reference system; for time alignment, select a unified time step and perform time resampling on the original data; For non-integer observation data, complete the data accordingly; Regarding spatial coordinates, if the original data uses non-WGS84 coordinates, it needs to be converted to the EPSG:4326 standard latitude and longitude coordinate system.

[0012] Furthermore, the detailed process of step S2 is as follows: S201. For the prediction area, specify the target point. Centered on a regular spatial grid structure, each grid is called a grid cell; The spacing between grid cells ranges from 10 to 50 kilometers and can be adjusted as needed. The center of each grid cell is called the grid point. ; S202. Perform spatial analysis on grid points and observation stations, specifically: Select the grid points to be analyzed For grid points Calculate the distance between it and all observation points to determine the range threshold. Only observation points whose distance is within the range threshold are retained as candidate sets; For each pair of "grid point – observation station" combinations, historical time-series data can be used to calculate the Pearson correlation coefficient: ; in, For grid points With observation point The Pearson correlation coefficient between them For grid points Historical observations For observation point The observed values, Describing covariance, Indicates standard deviation; Determine a correlation threshold When there is At that time, the observation point Place it into the interpolation neighborhood; Organize the observation point information in the interpolation neighborhood of the grid point, with the format: grid point-observation point-distance-correlation; S203, Grid points The observation points in the interpolation neighborhood are represented in vector form: The corresponding observed value is Then grid points The observed values ​​can be expressed as: ; in, For grid points The observed values, For the first The weight of each observation point For the first Observed values ​​at each observation point; S204. Introduce a terrain correction factor to correct the observed values: ; in, For grid points Corrected observations For grid points Terrain correction factor.

[0013] Furthermore, the detailed process of step S3 is as follows: S301. The observed values ​​of the 500 hPa, 850 hPa, and 1000 hPa pressure layers are expressed as follows: , and Their local variances are respectively , and Their fusion weights are respectively , and The fusion weight of the pressure layer is: ; S302. Perform weighted fusion based on the fusion weights: ; in, This represents a comprehensive set of meteorological element values. S303. Use the ReLU function to process the integrated meteorological element values: ; in, This is a comprehensive value processed from meteorological elements; S304. Introduce physical consistency constraints to constrain the integrated meteorological element processing values, and obtain constrained meteorological element processing values. Physical consistency is obtained, including temperature decrease constraints and wind speed vertical gradient constraints; Temperature decrease constraint refers to the decrease in temperature with altitude, and the temperature difference. With height difference The relationship is: ; in, For dry adiabatic lapse rate; Vertical gradient constraint of wind speed, wind speed difference With height difference The relationship is: ; in, This is the proportionality coefficient.

[0014] Furthermore, the detailed process in step S4 is as follows: Calculate the spatial gradient, including the spatial gradient along the longitude direction and the spatial gradient along the latitude direction; Spatial gradient along the longitude direction: ; Spatial gradient along the latitudinal direction: ; in, Represents grid points time step The constrained meteorological element processing values ​​are as follows: and These are the grid spacings in the longitude and latitude directions, respectively. Its gradient magnitude is expressed as: ; Its gradient direction can be represented as: ; in, The gradient direction; The formula for calculating the time derivative is: ; Next, auxiliary information is added to further enhance the expressive power of meteorological data; auxiliary information includes: topographic data, seasonal or periodic features; topographic data includes: slope and altitude, periodic features include: intraday period calculated by hour and seasonal period calculated by day; Meteorological elements are standardized and normalized.

[0015] Furthermore, the format of the multidimensional tensor is: ; in, and It refers to the number of rows and columns of the grid. The number of channels includes raw meteorological elements, spatial gradient, temporal derivative, and auxiliary information.

[0016] Furthermore, in step S102, the strategy for filling in missing values ​​is as follows: Different processing strategies are adopted according to the missing proportion: when the missing proportion is low, bilinear interpolation is used for imputation. When the missing proportion is high, a machine learning-based temporal imputation method can be introduced for repair. All data quality processing steps are logged to ensure process traceability.

[0017] Furthermore, the first Weight of each observation point It is necessary to determine the relationship between the grid points and the first... Distance between observation points Make corrections: ; in, It is a small positive number. For distance attenuation control parameters, For the first The level to which each observation point belongs, For the first The historical mean square error of each observation point , This is an adjustable coefficient.

[0018] Furthermore, the formula for the intraday cycle is: ; Among them, time Using the time of day, in hours, to represent the first hour of the day. Hour; The formula for the seasonal cycle is: ; Among them, time Using a year as the unit, the number of days in a day represents the first day of the year. sky.

[0019] Furthermore, in step S2, the distance between the grid point and the observation point can be Euclidean distance or spherical distance; When the distance is less than 300km, use Euclidean distance: ; in, , These are grid points and the The x-coordinates of each observation point; , These are grid points and the The ordinates of each observation point; When the distance is greater than 300 km or is a spherical region, the Haversine formula is used to calculate the great circle distance on a sphere: ; in, For the Earth's radius, , Grid points and the Longitude of each observation point , Grid points and the The latitude of each observation point.

[0020] Beneficial effects: This invention employs an innovative adaptive gridded mapping algorithm through the collaborative processing of data from multiple pressure layers. This algorithm fully explores the coupling relationships between different pressure layers and performs data fusion based on meteorological principles. This method not only preserves the physical meaning of meteorological data but also adapts it to the input format of deep learning models, thereby improving the accuracy of weather forecasts and the generalization ability of the models. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a graph showing the data source and the results of time / space alignment; Figure 3 This is a schematic diagram of spatial neighborhood and weight calculation; Figure 4 This is a schematic diagram of the temperature field after topographic contour lines and lapse rate correction; Figure 5 This is a diagram illustrating the effect of vertical physical consistency constraints. Figure 6 This is a schematic diagram of feature enhancement and tensor construction; Figure 7 This is a comparison chart of the effects of this method with other methods. Detailed Implementation

[0022] Example 1 like Figure 1 As shown, a method for gridding meteorological elements based on multi-barometric layer data fusion includes the following steps: S1. Perform data preprocessing, quality control, and standardization operations; The detailed process is as follows: S101. Select typical pressure layers as input layers, including: 500 hPa, 850 hPa and 1000 hPa; , , These represent the temperature values ​​of the corresponding pressure layers, in that order. , , These represent the air pressure values ​​of the corresponding pressure layers, in that order. , , The wind speeds at the corresponding layers are represented sequentially. Each meteorological element is organized using station number, latitude and longitude coordinates, and observation timestamp as indexes, and stored uniformly for subsequent processing; S102. Observational data that exceeds the physical common sense or the range of the equipment shall be discarded; missing data shall be filled in. All data quality processing steps are logged to ensure process traceability; Different processing strategies are adopted according to the missing percentage: when the missing percentage is less than 5%, bilinear interpolation is used for imputation. The process of bilinear interpolation is as follows: Given the four vertices of a rectangle in a two-dimensional plane, , , , The third item Given the value of this point, find the corresponding point. The value of ; First, interpolate along the x-axis to obtain two points. , : ; ; Interpolation along the y-axis: ; When the missing percentage is greater than or equal to 5%, historical statistical value interpolation can be used for repair. The process is as follows: For a certain data point, such as the missing data at 14:30 on April 15, 2023, it can be filled in using data from the same time before 2023. Specifically, the median method is used, that is, the median of the above data is taken as the filling value for the missing data, which can ensure the stability of the data. All data quality processing steps are logged to ensure process traceability; S103. Align all data to a unified time base and spatial reference system. For time alignment, select a unified time step; in the preferred scheme, the time step can be 1 hour, and perform time resampling processing on the original data. For non-integer observation data, use linear interpolation to fill in the gaps. For example, if the previous value is 3, the next value is 6, and two values ​​are missing in between, then: ; Fill in the two missing values, 4 and 5, in sequence; Regarding spatial coordinates, if the original data uses non-WGS84 coordinates, it needs to be converted to the EPSG:4326 standard latitude and longitude coordinate system.

[0023] S2. Spatial gridding transformation of meteorological data; the detailed process is as follows: S201. For the prediction area, specify the target point. Centered on a regular spatial grid structure, each grid is called a grid cell; The spacing between grid cells ranges from 10 to 50 kilometers and can be adjusted as needed. The center of each grid cell is called the grid point. ; S202. Perform spatial analysis on grid points and observation stations, specifically: Select the grid points to be analyzed For grid points Calculate the distance between it and all observation points to determine the range threshold. Only observation points whose distance is within the range threshold are retained as candidate sets; For each pair of "grid point – observation station" combinations, historical time-series data can be used to calculate the Pearson correlation coefficient: ; in, For grid points With observation point The Pearson correlation coefficient between them For grid points Historical observations For observation point The observed values, Describing covariance, Indicates standard deviation; Determine a correlation threshold When there is At that time, the observation point Place it into the interpolation neighborhood; Organize the observation point information in the interpolation neighborhood of the grid point, with the format: grid point-observation point-distance-correlation; S203, Grid points The observation points in the interpolation neighborhood are represented in vector form: The corresponding observed value is Then grid points The observed values ​​can be expressed as: ; in, For grid points The observed values, For the first The weight of each observation point For the first Observed values ​​at each observation point; for It is necessary to determine the relationship between the grid points and the first... Distance between observation points Make corrections: ; in, It is a small positive number. For distance attenuation control parameters, For the first The level to which each observation point belongs, For the first The historical mean square error of each observation point , This is an adjustable coefficient; And distance It can be either Euclidean distance or spherical distance; When the distance is less than 300km, use Euclidean distance: ; in, , These are grid points and the The x-coordinates of each observation point; , These are grid points and the The ordinates of each observation point; When the distance is greater than 300 km or is a spherical region, the Haversine formula is used to calculate the great circle distance on a sphere: ; in, For the Earth's radius, , Grid points and the Longitude of each observation point , Grid points and the Latitude of each observation point S204. Introduce a terrain correction factor to correct the observed values: ; in, For grid points Corrected observations For grid points The terrain correction factor, which is usually related to the elevation of the grid point, can be estimated using the vertical lapse rate model. The specific process is as follows: Regarding temperature, we have: ; in, This is the adiabatic lapse rate. For grid points altitude, The elevation of the reference point Temperature at the reference point; Regarding wind speed, we have: ; in, This refers to the wind speed topography coefficient; S3. Perform multi-layer data fusion on meteorological data and impose physical consistency constraints between layers; the detailed process is as follows: S301. The observed values ​​of the 500 hPa, 850 hPa, and 1000 hPa pressure layers are expressed as follows: , and Their local variances are respectively , and Their fusion weights are respectively , and The fusion weight of the pressure layer is: ; S302. Perform weighted fusion based on the fusion weights: ; in, This represents a comprehensive set of meteorological element values. S303. Use the ReLU function to process the integrated meteorological element values: ; in, This is a comprehensive value processed from meteorological elements; S304. Introduce physical consistency constraints to constrain the integrated meteorological element processing values, and obtain constrained meteorological element processing values. Physical consistency is obtained, including temperature decrease constraints and wind speed vertical gradient constraints; Temperature decrease constraint refers to the decrease in temperature with altitude, and the temperature difference. With height difference The relationship is: ; in, For dry adiabatic lapse rate; Vertical gradient constraint of wind speed, wind speed difference With height difference The relationship is: ; in, This is the proportionality coefficient.

[0024] S4. Perform feature enhancement and normalization operations on the fused meteorological elements; the detailed process is as follows: Calculate the spatial gradient, including the spatial gradient along the longitude direction and the spatial gradient along the latitude direction; Spatial gradient along the longitude direction: ; Spatial gradient along the latitudinal direction: ; in, Represents grid points time step The constrained meteorological element processing values ​​are as follows: and These are the grid spacings in the longitude and latitude directions, respectively. Its gradient magnitude is expressed as: ; Its gradient direction can be represented as: ; in, The gradient direction; The formula for calculating the time derivative is: ; Next, auxiliary information is added to further enhance the expressive power of meteorological data; auxiliary information includes: topographic data, seasonal or periodic features; topographic data includes: slope and altitude, periodic features include: intraday period calculated by hour and seasonal period calculated by day; The formula for intraday cycles is: ; Among them, time Using the time of day, in hours, to represent the first hour of the day. Hour; The formula for the seasonal cycle is: ; Among them, time Using a year as the unit, the number of days in a day represents the first day of the year. sky Meteorological elements are standardized and normalized. The standardization process is as follows: ; in, For grid points The standardized observations and These are the mean and standard deviation of the meteorological element, respectively. Next, the standardized values ​​are normalized to map the data to the [0,1] interval: ; in, For grid points The normalized observations, and These are grid points The maximum and minimum values ​​of the standardized observations.

[0025] S5. Merge the feature channels in the enhanced meteorological elements to form a multidimensional tensor; the format of the multidimensional tensor is: ; in, and It refers to the number of rows and columns of the grid. The number of channels includes raw meteorological elements, spatial gradient, temporal derivative, and auxiliary information.

[0026] Example 2 To further illustrate the advantages of the method of this invention, in conjunction with Example 1, a method for gridded processing of meteorological elements based on multi-pressure layer data fusion is described. After unifying time and coordinates, it combines spatial neighborhood weights, topographic lapse rate correction, and vertical physical consistency constraints to fuse multi-layer elements such as 1000 / 850 / 500 hPa according to inverse variance weights. Feature enhancement is then used to construct a [T, M, N, C] tensor to improve estimation accuracy and stability. Comparative and ablation experiments verify that this method achieves significant improvements over conventional IDW, Kriging, and single-layer input in terms of mean absolute error (MAE).

[0027] like Figure 2 As shown, observation stations from data sources A and B are used based on the eastern edge of the plateau (100.0°–101.2°E, 29.8°–30.7°N) with a step size of 1 hour. In the figure, "×" represents the original location of source A, "+" represents the original location of source B, and solid circles represent stations aligned to a unified grid and time step. Dashed lines indicate the mapping from the original coordinates / time to the alignment result. After completing multi-source spatiotemporal alignment, stations from different sources are aligned consistently in both location and time, significantly reducing the systematic bias caused by mismatches and laying a solid foundation for subsequent gridding and fusion.

[0028] like Figure 3 As shown, under the same area / time setting, weights are applied around the grid points: black squares represent target grid points, red dots represent observation stations, and black radial lines are the lines connecting the grid points to each station; black markings next to the stations indicate the distances. d i With weight w i . Grid points Choose the radius centered on r K observation stations within and calculate the distance d 1… d k Normalized weights w1…w are obtained by inverse distance weighting. k The closer the distance, the greater the weight; if necessary, combine the relevance threshold ρ* to reduce the weight of low-relevance sites or remove them.

[0029] like Figure 4 As shown, the contour lines represent the elevation H distribution, and the base map is the temperature field corrected for the lapse rate (the lapse rate is...). (0.0065 °C / m). After correction, the temperature in high-altitude areas is relatively lower and the temperature in low-lying areas is higher, making the element field consistent with the topographic prior and reducing false signals from the slope.

[0030] like Figure 5As shown, the original temperature profile exhibits an inversion at approximately 1.5 km; after applying consistency constraints, the inversion is suppressed, and the profile better conforms to the physical expectation of a dry adiabatic linearity of 6.5 K / km, thereby reducing the unreasonable vertical jumps introduced by interlayer fusion.

[0031] like Figure 6 As shown, the original elements temperature T, humidity RH, east-west wind field U, and north-south wind field V are compared with the derived features. , Topographic information (DEM) and seasonal data are stitched together in the channel dimension to form a multi-channel feature tensor of size [T, M, N, C], which serves as the input for subsequent modeling and inference. Among the derived features mentioned above, Meteorological elements The gradient reflects its rate of spatial change; This indicates the rate of change of meteorological elements over time.

[0032] like Figure 7 As shown, under the same region and sample division, the MAE values ​​of IDW, Kriging, single-layer input, and the method of this invention are 1.85, 1.72, 1.60, and 1.18, respectively. This invention significantly improves overall accuracy by introducing "multi-layer fusion + terrain correction + vertical consistency constraint + feature enhancement".

[0033] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for gridded processing of meteorological elements based on multi-pressure layer data fusion, characterized in that: Includes the following steps: S1. Perform data preprocessing, quality control, and standardization operations; S2. Perform spatial gridding transformation on meteorological data; S3. Perform multi-layer data fusion on meteorological data and impose physical consistency constraints between layers; S4. Perform feature enhancement and normalization operations on the fused meteorological elements; S5. Merge the feature channels in the enhanced meteorological elements to form a multidimensional tensor.

2. The meteorological element gridding processing method based on multi-pressure layer data fusion according to claim 1, characterized in that: The detailed process of step S1 is as follows: S101. Select typical pressure layers as input layers, including: 500 hPa, 850 hPa and 1000 hPa; , , These represent the temperature values ​​of the corresponding pressure layers, in that order. , , These represent the air pressure values ​​of the corresponding pressure layers, in that order. , , The wind speeds at the corresponding layers are represented sequentially. Each meteorological element is organized using station number, latitude and longitude coordinates, and observation timestamp as indexes, and stored uniformly for subsequent processing; S102. Observational data that exceeds the limits of physical common sense or the range of equipment shall be discarded; missing data shall be filled in. All data quality processing steps are logged to ensure process traceability; S103. Align all data to a unified time base and spatial reference system; for time alignment, select a unified time step and perform time resampling on the original data; For non-integer observation data, complete the data accordingly; In terms of spatial coordinates, the WGS84 geographic coordinate system is used, and its official identifier is EPSG:4326. If the original data does not use the WGS84 geographic coordinate system, or uses other reference / projection coordinate systems, it should be uniformly converted to EPSG:4326.

3. The meteorological element gridding processing method based on multi-pressure layer data fusion according to claim 1, characterized in that: The detailed process of step S2 is as follows: S201. For the prediction area, specify the target point. Centered on a regular spatial grid structure, each grid is called a grid cell; The spacing between grid cells ranges from 10 to 50 kilometers and can be adjusted as needed. The center of each grid cell is called the grid point. ; S202. Perform spatial analysis on grid points and observation stations, specifically: Select the grid points to be analyzed For grid points Calculate the distance between it and all observation points to determine the range threshold. Only observation points whose distance is within the range threshold are retained as candidate sets; For each "grid point – observation station" combination, historical time-series data can be used to calculate the Pearson correlation coefficient: （1）； in, For grid points With observation point The Pearson correlation coefficient between them For grid points Historical observations For observation point The observed values, Describing covariance, Indicates standard deviation; Determine a correlation threshold When there is At that time, the observation point Place it into the interpolation neighborhood; Organize the observation point information in the interpolation neighborhood of the grid point, with the format: grid point-observation point-distance-correlation; S203, Grid points The observation points in the interpolation neighborhood are represented in vector form: The corresponding observed value is Then grid points The observed values ​​can be expressed as: （2）； in, For grid points The observed values, For the first The weight of each observation point For the first Observed values ​​at each observation point; S204. Introduce a terrain correction factor to correct the observed values: （3）； in, For grid points Corrected observations For grid points Terrain correction factor.

4. The method for gridded processing of meteorological elements based on multi-pressure layer data fusion according to claim 1, characterized in that: The detailed process of step S3 is as follows: S301. The observed values ​​of the 500 hPa, 850 hPa, and 1000 hPa pressure layers are expressed as follows: , and Their local variances are respectively , and Their fusion weights are respectively , and The fusion weight of the pressure layer is: （4）； S302. Perform weighted fusion based on the fusion weights: （5）； in, This represents a comprehensive set of meteorological element values. S303. Use the ReLU function to process the integrated meteorological element values: （6）； in, This is a comprehensive value processed from meteorological elements; S304. Introduce physical consistency constraints to constrain the integrated meteorological element processing values, and obtain constrained meteorological element processing values. Physical consistency is obtained, including temperature decrease constraints and wind speed vertical gradient constraints; Temperature decrease constraint refers to the decrease in temperature with altitude, and the temperature difference. With height difference The relationship is: （7）； in, For dry adiabatic lapse rate; Vertical gradient constraint of wind speed, wind speed difference With height difference The relationship is: （8）； in, This is the proportionality coefficient.

5. The meteorological element gridding processing method based on multi-pressure layer data fusion according to claim 4, characterized in that: The detailed process in step S4 is as follows: Calculate the spatial gradient, including the spatial gradient along the longitude direction and the spatial gradient along the latitude direction; Spatial gradient along the longitude direction: （9）； Spatial gradient along the latitudinal direction: （10）； in, Represents grid points time step The constrained meteorological element processing values ​​are as follows: and These are the grid spacings in the longitude and latitude directions, respectively. Its gradient magnitude is expressed as: （11）； Its gradient direction can be represented as: （12）； in, The gradient direction; The formula for calculating the time derivative is: （13）； Next, auxiliary information is added to further enhance the expressive power of meteorological data; auxiliary information includes: topographic data, seasonal or periodic features; topographic data includes: slope and altitude, periodic features include: intraday period calculated by hour and seasonal period calculated by day; Meteorological elements are standardized and normalized.

6. The method for gridded processing of meteorological elements based on multi-pressure layer data fusion according to claim 1, characterized in that: The format of the multidimensional tensor is: （14）； in, and It refers to the number of rows and columns of the grid. The number of channels includes raw meteorological elements, spatial gradient, temporal derivative, and auxiliary information.

7. The method for gridded processing of meteorological elements based on multi-pressure layer data fusion according to claim 2, characterized in that: In step S102, the strategy for filling in missing values ​​is as follows: Different processing strategies are adopted according to the missing proportion: when the missing proportion is low, bilinear interpolation is used for imputation. When the missing proportion is high, a machine learning-based temporal imputation method can be introduced for repair. All data quality processing steps are logged to ensure process traceability.

8. The method for gridded processing of meteorological elements based on multi-pressure layer data fusion according to claim 3, characterized in that: The first Weight of each observation point It is necessary to determine the relationship between the grid points and the first... Distance between observation points Make corrections: （15）； in, It is a small positive number. For distance attenuation control parameters, For the first The level to which each observation point belongs, For the first The historical mean square error of each observation point , This is an adjustable coefficient.

9. The method for gridded processing of meteorological elements based on multi-pressure layer data fusion according to claim 5, characterized in that: The formula for the intraday cycle is: （16）； Among them, time Using the time of day, in hours, to represent the first hour of the day. Hour; The formula for the seasonal cycle is: （17）； Among them, time Using a year as the unit, the number of days in a day represents the first day of the year. sky.

10. The meteorological element gridding processing method based on multi-pressure layer data fusion according to claim 7, characterized in that: In step S2, the distance between the grid point and the observation point can be Euclidean distance or spherical distance. When the distance is less than 300km, use Euclidean distance: （18）； in, , These are grid points and the The x-coordinates of each observation point; , These are grid points and the The ordinates of each observation point; When the distance is greater than 300 km or is a spherical region, the Haversine formula is used to calculate the great circle distance on a sphere: （19）； in, For the Earth's radius, , Grid points and the Longitude of each observation point , Grid points and the The latitude of each observation point.

Images