A short-term and imminent multi-meteorological element forecasting method based on a deep neural network

Abstract

The present invention provides a short-term and imminent multi-meteorological element forecasting method based on a deep neural network, including the steps of: obtaining historical precipitation data and historical pressure, temperature, humidity and wind data of the area to be measured; establishing a kNN Barnes factor map according to the station dictionary and uniform station dictionary of meteorological element categories; using the knn proximity algorithm to sample and connect stations to construct a graph to obtain adjacency graph feature data; preprocessing the adjacency graph feature data and then inputting it into the UNet network to output a feature map of the number of categories; calculating the physical properties and change process of the atmosphere according to the historical pressure, temperature, humidity and wind data, and using the knn algorithm to construct a graph connection for stations according to the honeycomb hexagonal structure to obtain a graph feature set; matching the graph feature set with samples and substituting it into a predictor through multiple networks to obtain a multi-element time series forecast of the station set. This method can obtain a multi-element time series forecast of the station set and can be used for operational weather forecasting.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological element forecasting, and particularly relates to a short-term and multi-meteorological element forecasting method based on a deep neural network. Background Technique

[0002] Weather forecasting involves various fields such as meteorology, agriculture, transportation, tourism, and scientific research, and is an important means to ensure national economic and national defense construction. Currently, meteorological forecasting methods are mainly divided into numerical meteorological forecasting and numerical-driven methods. The former relies on large computers to solve the atmospheric physical model to obtain the forecasting results, while the latter makes predictions on the future development of the atmosphere through statistical or machine learning methods; after a long period of development, the numerical forecasting method has become the current mainstream meteorological forecasting method. However, due to the coordination problem of the physical model, there are forecasting deviations in some cases. In addition, due to the long time required for large-scale numerical calculations, it faces the disadvantages of high cost and slow response in the forecasting of sudden weather and short-term weather.

[0003] Compared with traditional numerical weather forecasting, the advantage of deep learning is that it can automatically learn from a large amount of data through algorithms, thereby extracting the internal characteristics of the data and the relevant laws contained therein, and does not require very professional prior knowledge in the meteorological field. After the deep learning model is trained, its prediction results are almost real-time. Therefore, compared with the numerical forecasting method, it is more suitable for medium- and short-term and sudden meteorological forecasting, and has the characteristics of low cost and rapid response. However, due to the chaos of the meteorological model, the long-term forecasting accuracy of deep learning is not as good as numerical calculation. Summary of the Invention

[0004] In view of this, one of the purposes of the present invention is to provide a short-term and multi-meteorological element forecasting method based on a deep neural network, which can complete the forecasting of multi-element time series of a station set.

[0005] To achieve the above purpose, the technical solution of the present invention is: a short-term and multi-meteorological element forecasting method based on a deep neural network, characterized by including the following steps:

[0006] S1: Obtain the historical precipitation data and historical pressure, temperature, humidity, and wind data of the area to be measured, and group the historical precipitation data and the historical pressure, temperature, humidity, and wind data according to a preset time series segment to obtain the climate maximum value, minimum value, and average value;

[0007] S2: According to the station dictionary and uniform station dictionary of different meteorological element categories, establish a kNN Barnes factor graph, and use the knn proximity algorithm to sample and connect the data of the kNN Barnes factor graph to construct a graph to obtain the adjacency graph feature data;

[0008] S3: Preprocess the adjacency graph feature data obtained in S2 and then input it into the UNet model and the residual network for downsampling first. Through convolutions of different degrees, deep features are learned, and then through upsampling, it is restored to the original image size, and finally a feature map with the number of categories is output;

[0009] S4: Calculate the dynamic meteorology of the physical properties and change processes of the atmosphere according to the historical pressure, temperature, humidity, and wind data, and connect the feature maps obtained in S3 according to the honeycomb hexagonal structure to obtain a graph feature set;

[0010] S5: Perform sample matching on the graph feature set obtained in S4 and then substitute it into the time series graph network model, and finally obtain the multi-element time series prediction results of the station set.

[0011] Furthermore, the station dictionary of meteorological element categories includes: visibility stations, national stations, pressure stations, humidity stations, wind direction and speed stations, temperature stations, and rainfall stations.

[0012] Furthermore, the steps of using the knn proximity algorithm to sample the kNN Barnes factor graph specifically include:

[0013] Samples where the main station has precipitation characteristics and the adjacent station has precipitation characteristics;

[0014] Samples where the main station has precipitation characteristics and the adjacent station has no precipitation characteristics;

[0015] Samples where the main station has no precipitation characteristics and the adjacent station has precipitation characteristics;

[0016] Samples where the main station or the adjacent station has precipitation characteristics within a preset first time period;

[0017] For space, all samples of national stations and minute assessment stations;

[0018] For time, hourly data at the exact hour is preferred.

[0019] Furthermore, the steps of preprocessing the adjacency graph feature data obtained in S2 specifically include:

[0020] First, perform feature normalization on the adjacency graph to ensure that the input and output feature values of the model are consistent. Input the obtained adjacency graph feature data into the UNet model and the residual network, first perform downsampling, through convolutions of different degrees, learn deeper image features, and then through upsampling, restore it to the original image size, and finally output a feature map with the same number of categories.

[0021] Compared with the prior art, the present invention has the following advantages:

[0022] The present invention discloses a short-term and imminent multi-meteorological element forecasting method based on a deep neural network. The historical data of the physical characteristics of meteorological stations is prepared by using the classical atmospheric dynamics equations and meteorological physical quantity calculations. With the goal of forecasting from the characteristic time series to the label time series, components such as deep learning residual networks, multi-resolution networks, recurrent networks, and time-distributed linear networks are used to design a general deep weather forecasting structure. The general forecasting structure is migrated to different seasons and hours for training and iteration to obtain a set of deep weather forecasting models. The real-time observation data is fused and processed and substituted into the forecasting model to obtain the multi-element time series forecasting of the station set, which can be used for the weather forecasting of operational operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a flowchart of a short-term and imminent multi-meteorological element forecasting method based on a deep neural network of the present invention;

[0025] Figure 2 It is a relationship diagram between a station and its neighboring stations in the present invention;

[0026] Figure 3 It is a schematic diagram of a knn adjacency graph in the present invention;

[0027] Figure 4 It is a UNet network structure diagram of the present invention;

[0028] Figure 5 It is a time series graph network model of the present invention;

[0029] Figure 6 It is a schematic diagram of the construction principle of the time series graph network model of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] In order to make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0031] The embodiments are provided to better illustrate the present invention, but the content of the present invention is not limited to the embodiments. Therefore, those skilled in the art can make non-essential improvements and adjustments to the implementation solutions based on the above invention content, which still fall within the protection scope of the present invention.

[0032] It should be noted that the subscripts of each representative formula in this embodiment are only used for distinction without special meaning if there is no special explanation.

[0033] Embodiment 1

[0034] In this embodiment, a short-term and multi-meteorological element forecasting method based on a deep neural network is disclosed. Referring to Figure 1 the flowchart, the method includes the following steps:

[0035] S1: Obtain the historical precipitation data and historical pressure, temperature, humidity, and wind data of the area to be measured, and group the historical precipitation data and historical pressure, temperature, humidity, and wind data according to a preset time series segment to obtain the climate maximum value, minimum value, and average value;

[0036] In this embodiment, first, historical precipitation data and historical pressure, temperature, humidity, and wind data of the area to be measured are obtained to form a feature set. The above data can be obtained from climate background, surface climate, satellite data, radar data, GPS data, etc. Then, these data are preprocessed for the first time. The preprocessing process can be atmospheric dynamic processing, WeChat meteorological processing, etc. Finally, the required historical precipitation data and historical pressure, temperature, humidity, and wind data of the area to be measured are obtained: data items in CIMISS hourly surface data of China, surface minute precipitation data of the area to be measured, and surface minute pressure, temperature, humidity, and wind data of the area to be measured: Station_Id_d (station number), Station_levl (station level), Lat (latitude), Lon (longitude), Alti (station altitude), PRS (pressure), PRS_Sea (sea-level pressure), TEM (temperature / air temperature), DPT (dew point temperature), RHU (relative humidity), VAP (vapor pressure), PRE (precipitation), WIN_D_Avg_2mi (wind direction), WIN_S_Avg_2mi (wind speed), GST (surface temperature), VIS_HOR_10MI (horizontal visibility), and the data time range is 2016 - 2019. Then, these data are grouped according to a preset time series segment, such as grouping by every pentad (0 - 2] hours, (2 - 4] hours, (4 - 6] hours, (6 - 8] hours, (8 - 10] hours, (10 - 12] hours, (12 - 14] hours, (14 - 16] hours, (16 - 18] hours, (18 - 20] hours, (20 - 22] hours, (22 - 24] hours, and the climate maximum, minimum, and average values are calculated. The purpose of calculating the maximum and minimum values is to understand the vibration state and evolution trend of climate elements, and the purpose of calculating the average value of climate elements is to verify whether the average value can represent the average situation of climate elements in this period.

[0037] S2: Establish a kNN Barnes factor map according to the station dictionaries of different meteorological element categories and the uniform station dictionary;

[0038] The kNN Barnes factor map is a neighboring station factor map established using the knn nearest neighbor algorithm model and the Barnes Gaussian function. It can complete the data filling of spatial stations through the acyclic belief propagation iterative algorithm.

[0039] In this step, according to the station dictionaries of different element categories and the uniform station dictionary, an element station dictionary U i (i = 1..7) and the kNN Barnes factor map between the uniform station dictionaries V, where U i Corresponds to seven meteorological element dictionaries: visibility station, national station, pressure station, humidity station, wind direction and speed station, air temperature station, and rainfall station respectively. The relationship diagram between the station and the neighboring station is as Figure 2As shown, R is the maximum distance between site ★ and k (k = 40) adjacent sites, and r is the distance between site ★ and a certain site. Both R and r are solved using the metric function formula:

[0040] In the geodetic coordinate system, the distance between any two points m, n ∈ M, m = (x1, y1), n = (x2, y2) can be expressed by the metric r(m, n) as:

[0041]

[0042] where x i , y i respectively represent the longitude and latitude positions of the points;

[0043] Then calculate the weight w of each adjacent station relative to point ★:

[0044]

[0045] S3: Use the knn nearest neighbor algorithm to perform data sampling on the kNN Barnes factor graph and connect the kNN Barnes factor graph to obtain adjacency graph feature data;

[0046] The principle of data sampling in this step is based on the kNN Barnes factor graph principle. Specifically, use the knn nearest neighbor algorithm. If most of the k nearest samples of a sample in the feature space belong to a certain category, then this sample also belongs to this category and has the characteristics of the samples in this category. The schematic diagram of the adjacency graph is as Figure 3 shown;

[0047] In a specific embodiment, the data sampling process is to construct a kNN Barnes factor graph connection for all visibility stations, national stations, barometric pressure stations, humidity stations, wind direction and speed stations, air temperature stations, and rainfall stations in the area to be measured, and use the knn nearest neighbor algorithm for sampling. The sampling follows the following principles:

[0048] Samples where the main station has precipitation characteristics and the adjacent station has precipitation characteristics;

[0049] Samples where the main station has precipitation characteristics and the adjacent station has no precipitation characteristics;

[0050] Samples where the main station has no precipitation characteristics and the adjacent station has precipitation characteristics;

[0051] Samples where the main station or the adjacent station has precipitation characteristics within a preset first time period;

[0052] Space, all samples of national stations and minute assessment stations;

[0053] Time, preferential for hourly data at the exact hour.

[0054] S4: Preprocess the adjacency graph feature data obtained in S3 and then input it into the UNet model and the residual network for downsampling first. Through convolutions of different degrees, deeper image features are learned, and then it is upsampled back to the original image size, with transposed convolution used for upsampling. Finally, a feature map with the number of classes is output;

[0055] In this example, since the pattern dominated by numerical weather prediction remains the mainstream of weather forecasting, but numerical weather prediction attempts to accurately simulate the physical characteristics and changing processes of the atmosphere. In terms of parameter estimation, it is difficult to propose a unified parameter estimation model. In this embodiment, a UNet model (UNet, a variant of the fully convolutional neural network) plus a residual network are introduced to attempt to estimate the parameter set of the physical characteristics and changing processes of the atmosphere, which is implemented using an attention UNet model. The model is divided into two parts. The first is the dynamic meteorology calculation of the physical characteristics and changing processes of the atmosphere; the second is the attention UNet model plus the residual network; UNet is an efficient convolutional neural network architecture, and its structure diagram can be referred to Figure 4 Specifically:

[0056] UNet uses the encoding (Encode) / decoding (Decode) structure of the neural network, which is vividly represented by the capital letter U. The encoding (Encode) / decoding (Decode) ends with appropriate scales are connected by residual network components to achieve a result similar to that of a complex connection model with fewer parameters. The UNet network can be simply regarded as first downsampling, learning deep features through convolutions of different degrees, then upsampling back to the original image size, with transposed convolution used for upsampling, and finally outputting a feature map with the number of classes;

[0057] S5: Calculate the dynamic meteorology of the physical characteristics and changing processes of the atmosphere based on historical pressure, temperature, humidity, and wind data, and connect the feature maps obtained in S4 according to the honeycomb-like hexagonal structure to obtain a graph feature set;

[0058] In this step, first complete the dynamic meteorology calculation of the physical characteristics and changing processes of the atmosphere. Specifically, calculate based on the historical precipitation data and historical pressure, temperature, humidity, and wind data obtained in step S1:

[0059] Calculation of the density ρ of moist air, where the pressure is p, E is energy, and T k represents the thermodynamic temperature of moist air, kg is kilogram, and m is meter:

[0060]

[0061] Calculation of the climate background fields of the u, v, w factors:

[0062] Calculation of the angular velocity ω of the Earth's rotation:

[0063] ω = 7.292×10-5 s -1 ;

[0064] Calculate the geostrophic parameter f:

[0065] f = 2ωsinθ;

[0066] where θ is the geographical latitude (unit: degree).

[0067] Take the derivative of θ in the above formula to obtain the derivative of the geostrophic parameter f with respect to θ

[0068]

[0069] Substitute the formula f and ω into the following system of equations to find u, v, and w, where u, v, and w represent the velocity components in the x, y, and z directions respectively, the pressure is p, g is the acceleration due to gravity, and t represents the time series.

[0070]

[0071]

[0072] Calculation of the advection equation F. Substitute v t+1 The obtained result into the following formula to calculate the advection equation F:

[0073]

[0074] Substitute the advection air temperature T into F t+1 to obtain the calculation equation of the advection air temperature T:

[0075]

[0076] Water vapor equation M v , and calculate the water vapor flux accordingly, where ρ is the air density, q is the specific humidity, V represents the wind speed magnitude, the pressure is p, and ΔlΔz represents the plane area of the plane orthogonal to the wind direction.

[0077]

[0078]

[0079] Brunt - Vaisala internal gravity wave. The Brunt - Vaisala frequency (N) is an important parameter representing the stability of the atmospheric stratification. In practical applications, N is usually used 2 to represent, where θ is the potential temperature and g is the acceleration due to gravity:

[0080]

[0081] Gravity internal waves are waves generated by vertical disturbances due to the action of gravity in a stably stratified atmosphere. N 2 >0 is a necessary condition for the existence of gravity internal waves. Substituting N 2 into the following two equations can obtain the circular frequency ω and wave speed c of gravity internal waves:

[0082]

[0083]

[0084] where k and n are the wave numbers in the x and z directions respectively, and c l is the adiabatic sound speed. The propagation speed of gravity internal waves is mainly determined by N 2 , and the Brunt-Vaisala frequency is the limiting frequency of gravity internal waves.

[0085] The governing equation of Kelvin-Helmholtz is the Euler equation. The two-dimensional conservative form of the Euler equation is where the conserved quantity U, and the fluxes E, F are respectively expressed as:

[0086]

[0087] where ρ is the density of the fluid, u is the velocity of the fluid in the x direction, v is the velocity of the fluid in the y direction, p is the pressure of the fluid, represents the energy per unit mass-energy, C represents the velocity, ρe represents the energy intensity, and T is the temperature of the fluid.

[0088] The calculation formula for Kelvin-Helmholtz stability, where k is the wave number of the plane wave and g is the acceleration due to gravity. When KH > 0, it represents stability; when KH = 0, it represents neutrality; when KH < 0, it represents instability.

[0089]

[0090] where ρ1, ρ2, are respectively the density of fluid 1, the density of fluid 2, the velocity of fluid 1 in the x direction, and the velocity of fluid 2 in the x direction.

[0091] The calculation formula for Richardson instability, T is the absolute temperature, V represents the magnitude of the wind speed, g is the acceleration due to gravity, z represents the height, and u represents the wind speed.

[0092]

[0093]

[0094]

[0095] Among them, θ2, θ1, u2, and u1 respectively correspond to the potential temperature and wind speed at heights z2 and z1.

[0096] After the above calculations, data feature analysis is carried out to analyze the change range. The following Table 1 shows the change range of physical quantities of the required data:

[0097] Table 1 Change Range of Data Physical Quantities

[0098]

[0099] Since the value ranges of each physical quantity are not the same, if the feature analysis is directly carried out according to the value range in Table 1, the result is not very accurate and it is not convenient for analysis. Therefore, it is necessary to perform feature normalization processing on the data so that each feature is in the same numerical order of magnitude, and then carry out the analysis. The normalization processing is carried out according to the parameters in Table 2, and the minimum value, maximum value, etc. in step S1 are processed according to the method in the table:

[0100] Table 2 Data Item Normalization Processing

[0101]

[0102]

[0103]

[0104] Then, the features of the physical characteristics and change process of the atmosphere are processed, and according to the honeycomb hexagonal structure, the knn algorithm is used to construct a graph connection for all stations in the area to be measured, realizing the conversion of the meteorological physical characteristics of the stations to the hexagonal graph structure of the stations. The overall process from the data preprocessing to the intermediate model passed through and then to the final result time series forecast set is as Figure 1 shown;

[0105] S6: After sample matching the graph feature set obtained in S5, it is input into the time series graph network model, and finally the multi-element time series forecast results of the station set are obtained.

[0106] In this step, referring to Figure 5 the schematic diagram of the structure of the time series graph network model in Figure 6 , the design of the time series graph network includes the physical basis and the design basis. The physical basis of the time series network is built according to the short-term and imminent forecast physical model, and the construction of the time series network basis is based on the knn adjacency graph and physical cause analysis.

[0107] In this embodiment, after obtaining the graph feature set data, the corresponding activation function PReLU is used to process the data, and then the BiLSTM model is used to perform the Enconder (encoder) operation on the data, that is, to memorize and understand the information, and refine the information into a vector. After the processing is completed, it goes through a PReLU function and the Deconder (decoder) operation of the BiLSTM (bidirectional long short-term memory network) model, starts to recall and apply this information, and decodes it into the required form after processing, and finally obtains the required data form through a PReLU function, that is, the final time series prediction result data.

[0108] In the method of this embodiment, according to the time and space principles, in terms of time, there are 24 hours in a day, and a model is established every 2 hours; there are 30 days in a month, and a model is established every 5 days as a pentad; there are 12 months in a year, and a corresponding model is established for each month. In terms of space, the setting is the national scope. Therefore, the total number of models is: 12 * 6 * 12; in the method of this embodiment, the input is a preset time series segment and a preset first time period sequence. For example, in a specific embodiment, the input is two hours, with each 10 minutes as a sequence, and there are a total of 12 time series. The output of this method is also two hours, with a total of 12 time series.

[0109] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the purpose of the present invention and the scope protected by the claims. These all fall within the protection scope of the present invention.

Claims

1. A short-term and imminent multi-meteorological element forecasting method based on a deep neural network, characterized in that, It includes the following steps: S1: Obtain the historical precipitation data and historical pressure, temperature, humidity, and wind data of the area to be measured, and group the historical precipitation data and the historical pressure, temperature, humidity, and wind data according to a preset time series segment to obtain the climate maximum value, minimum value, and average value; S2: Establish a kNN Barnes factor map based on the station dictionaries and uniform station dictionaries of different meteorological element categories. Use the knn proximity algorithm to perform data sampling on the kNN Barnes factor map to connect the stations to construct a graph and obtain the adjacency graph feature data; Establish a kNN Barnes factor map based on the station dictionaries and uniform station dictionaries of different meteorological element categories, including: Based on the station dictionaries of different feature categories and the uniform station dictionary, establish the feature station dictionary U i (i = 1..7) and the kNN Barnes factor graph between the uniform station dictionary V, where U i respectively correspond to seven meteorological feature dictionaries: visibility station, national station, barometric pressure station, humidity station, wind direction and speed station, air temperature station, and rainfall station; The distance between any two points m, n ∈ M, m = (x1, y1), n = (x2, y2) in the geodetic coordinate system can be expressed by the metric r(m, n) as: where x i , y i represent the longitude and latitude positions of the point, respectively; Calculate the weight w of each station: where R is the farthest distance between the station and k neighboring stations; Based on the kNN Barnes factor map, use the knn proximity algorithm to perform data sampling on the kNN Barnes factor map to connect the stations to construct a graph and obtain the adjacency graph feature data, including: The data sampling process is to construct and connect the kNN Barnes factor map for all visibility stations, national stations, pressure stations, humidity stations, wind direction and speed stations, temperature stations, and rainfall stations in the area to be measured, and use the knn proximity algorithm for sampling. The sampling follows the following principles: Samples where the main station has precipitation characteristics and the adjacent station has precipitation characteristics; Samples where the main station has precipitation characteristics and the adjacent station has no precipitation characteristics; Samples where the main station has no precipitation characteristics and the adjacent station has precipitation characteristics; Samples where the main station or the adjacent station has precipitation characteristics within a preset first time period; All samples of national stations and minute assessment stations in space; In terms of time, the hourly data at the exact hour is preferred; S3: After preprocessing the adjacency graph feature data obtained in S2, input it into the UNet model and the residual network to perform downsampling to learn deep features, then upsample it back to the original image size, and finally output the feature map with the number of categories; After preprocessing the adjacency graph feature data obtained in S2, input it into the UNet model and the residual network to perform downsampling to learn deep features, then upsample it back to the original image size, and finally output the feature map with the number of categories, including: Introduce the UNet model plus the residual network to attempt to estimate the parameter set of the physical characteristics and change process of the atmosphere. It is implemented using the attention UNet model. The model is divided into two parts. The first is the dynamic meteorological calculation of the physical characteristics and change process of the atmosphere; the second is the attention UNet model plus the residual network; UNet is an efficient convolutional neural network architecture; UNet undergoes convolution to different degrees to learn deep features, then upsamples it back to the original image size. The upsampling is implemented using transposed convolution, and finally outputs the feature map with the number of categories; S4: Calculate the dynamic meteorology of the physical characteristics and change process of the atmosphere based on the historical pressure, temperature, humidity, and wind data, and connect the feature maps obtained in S3 according to the honeycomb hexagonal structure to obtain the graph feature set; Dynamical meteorology for calculating the physical properties and change processes of the atmosphere based on historical pressure, temperature, humidity, and wind data, and connecting the feature maps obtained in S3 according to a honeycomb hexagonal structure to obtain a graph feature set, including: Perform the following calculations based on the historical precipitation data and historical pressure, temperature, humidity, and wind data of the area to be measured obtained in step S1: Calculate the density ρ of moist air: where the pressure is p, E is the energy, and T k represents the thermodynamic temperature of moist air, kg is kilogram, and m is meter; Calculate the climatic background fields of the u, v, and w factors: Calculate the angular velocity ω of the Earth's rotation: ω = 7.292×10 -5 s -1 ; Calculate the geostrophic parameter f: f = 2ωsinθ; where θ is the geographical latitude; Differentiate with respect to θ to obtain the derivative of the geostrophic parameter f with respect to θ Substitute f and ω into the following system of equations to find u, v, and w: where u, v, and w represent the velocity components in the x, y, and z directions respectively, the pressure is p, g is the acceleration due to gravity, and t represents the time series; Calculation of the advection equation F: Substitute v t+1 The obtained result into the following formula to calculate the advection equation F: Substitute the advection air temperature T into F t+1 to obtain the calculation equation for the advection air temperature T: Calculate the water vapor equation M using the following formula v : where ρ is the air density, q is the specific humidity, V represents the magnitude of the wind speed, the pressure is p, and ΔlΔz represents the planar area of the plane orthogonal to the wind direction; Calculate the Brunt-Vaisala frequency N using the following formula 2 : where θ is the potential temperature and g is the acceleration due to gravity; Substitute N 2 into the following two equations to obtain the circular frequency ω and wave speed c of internal gravity waves: where k and n are the wave numbers in the x- and z-directions, respectively, and c l is the adiabatic sound speed; The governing equations of Kelvin-Helmholtz are the Euler equations. The two-dimensional conservative form of the Euler equations is where the conserved quantity U, and the fluxes E, F are expressed as follows: where ρ is the density of the fluid, u is the fluid velocity in the x-direction, v is the fluid velocity in the y-direction, p is the fluid pressure, E represents the energy per unit mass-energy, C represents the velocity, ρe represents the energy intensity, and T is the temperature of the fluid; The calculation formula for the Kelvin-Helmholtz stability is as follows: where k is the wave number of the plane wave, g is the acceleration due to gravity. When KH > 0, it represents stability; when KH = 0, it represents neutrality; when KH < 0, it represents instability. ρ1, ρ2, are respectively the density of fluid 1, the density of fluid 2, the velocity of fluid 1 in the x-direction, and the velocity of fluid 2 in the x-direction; The calculation formula for the Richardson instability is as follows: Among them, T is the absolute temperature, V represents the wind speed magnitude, g is the acceleration due to gravity, z represents the height, u represents the wind speed, and θ2, θ1, u2, and u1 respectively correspond to the potential temperature and wind speed at heights z2 and z1. After the above calculations, perform data feature analysis, analyze the change range, and obtain the value range of each physical quantity: Perform feature normalization processing on the value range of each physical quantity to obtain normalized data: Based on the normalized data, use the knn algorithm to construct graph connections for all stations in the area to be measured according to the honeycomb hexagonal structure, and realize the conversion of the meteorological physical characteristics of the stations to the hexagonal graph structure of the stations; S5: Perform sample matching on the graph feature set obtained in S4 and substitute it into the time series graph network model to finally obtain the multi-element time series prediction results of the station set; Perform sample matching on the graph feature set obtained in S4 and substitute it into the time series graph network model to finally obtain the multi-element time series prediction results of the station set, including: After obtaining the graph feature set data, use the corresponding activation function PReLU to process the data, then perform an encoder operation on the data through the BiLSTM model, and extract information to form a vector. Then, after passing through a PReLU function and a decoder operation of the bidirectional long short-term memory network model, start to recall and use this information, process and decode it into the required form, and finally obtain the final time series prediction result data through a PReLU function.

Images