Troposphere four-dimensional wind field prediction method based on attention mechanism

By constructing a four-dimensional wind field prediction model using a multi-head sparse attention computation method based on the attention mechanism, the accuracy and speed issues in tropospheric wind field prediction are solved, achieving high-precision and fast wind field prediction results.

CN119026728APending Publication Date: 2024-11-26SHANGHAI JIAOTONG UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411022250.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing technologies for predicting tropospheric wind fields suffer from poor prediction accuracy, difficulty in obtaining wind information dependence, and slow model convergence speed, especially in long-sequence prediction and large-scale spatiotemporal prediction.

Method used

A four-dimensional wind field prediction model is constructed by adopting a multi-head sparse attention computation method based on the attention mechanism. The wind information is standardized by initialization matrix and normalization function, and the time position code is calculated by using sine function. The model is then combined with sparse attention computation and multi-head attention layer for fully connected computation to achieve high-precision and fast wind field prediction.

Benefits of technology

It achieves high-precision and rapid prediction of tropospheric wind fields, solves the problems of poor prediction accuracy and slow convergence speed of the model, and improves the efficiency of feature extraction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026728A_ABST
    Figure CN119026728A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of troposphere four-dimensional wind field prediction, and provides a troposphere four-dimensional wind field prediction method based on an attention mechanism. Comprising the steps of preset wind information initialization and standardization, first time position code extraction, feature map calculation, historical wind information collection, historical wind information weighting, second time position code extraction, attention matrix calculation and full connection calculation. According to the method, a four-dimensional wind field prediction model is constructed through a multi-head sparse attention calculation method, and the prediction method which is high in precision, simple in feature extraction and rapid in model calculation is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of tropospheric four-dimensional wind field prediction, and in particular to a tropospheric four-dimensional wind field prediction method based on an attention mechanism. Background Art

[0002] Due to the uncertainty and temporal mutation of the tropospheric wind field, the prediction of the tropospheric wind field has the following characteristics: First, with the development of neural networks, the tropospheric wind field prediction has a good effect in short-term prediction tasks, but it is difficult to predict long sequences and guide the long-term operation of medium and low altitude aerostats; secondly, wind field prediction is mostly used for electric energy prediction of wind power systems. The predicted wind field range is small, and the tasks performed by medium and low altitude aerostats are often large in range. Therefore, how to predict a large range of time and space is also one of the bottlenecks of wind field prediction. In order to achieve long-sequence prediction of large-scale wind fields, it is necessary to perceive and model the wind field, and use supercomputers to deduce physical models. However, this method requires a lot of computing costs and does not have sufficient promotion value for tasks such as civil detection and industrial monitoring. Based on the above wind field prediction problems, generative long short-term memory networks and long short-term memory networks with shared weights can effectively improve the accuracy of long-term wind field prediction by mining long-term dependencies and trends from limited available data samples, but this will inevitably increase the amount of calculation. Different from the traditional time series prediction task, the tropospheric wind field prediction for medium and low altitude aerostats is a four-dimensional prediction task, that is, the time series prediction of the wind field in three-dimensional space is carried out simultaneously. The training samples for prediction are not a repetitive time period, but wind field data is collected before the aerostat mission, and the operation of the aerostat is guided by predicting the wind field conditions in the next few days. Therefore, the tropospheric wind field prediction is a multivariate time series prediction task for strongly correlated data. The particle swarm optimization algorithm is an effective method for finding the optimal correlation, which effectively avoids the local optimal scenario and provides a solution for the rapid prediction of wind fields for near-space low-speed aircraft missions. Although the particle swarm optimization algorithm provides the possibility of global search, since the particle swarm optimization algorithm only uses the group optimal and individual optimal information during the iteration process, it cannot guarantee convergence to the global optimal point.

[0003] In summary, the existing technology has the problems of poor prediction accuracy, difficulty in obtaining wind information dependence, and slow model convergence speed. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a four-dimensional wind field prediction method for the troposphere based on an attention mechanism. Through a multi-head sparse attention calculation method, a four-dimensional wind field prediction model is constructed to achieve a prediction method with high precision, simple feature extraction and fast model calculation.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A four-dimensional tropospheric wind field prediction method based on attention mechanism, comprising:

[0007] The preset wind information in the database is initialized and normalized using the initialization matrix and normalization function to obtain the encoding input and graph attention weights;

[0008] Calculating the first time position code of the preset wind information by using a sine function;

[0009] Using a sparse attention calculation formula, performing attention calculation on the encoded input in chronological order according to the first time position code to obtain a feature map;

[0010] Collect historical wind information of the coordinate points to be predicted;

[0011] Calculate the product of the historical wind information and the graph attention weight to obtain weighted wind information;

[0012] Calculating a second time position code of the weighted wind information using a sine function;

[0013] Using a sparse attention calculation formula, performing attention calculation on the weighted wind information in chronological order according to the second time position code to obtain an attention matrix;

[0014] The attention matrix and the feature map are input into a preset multi-head attention layer for full connection calculation to obtain the target prediction result.

[0015] Preferably, the sparse attention calculation formula is: e ij is the importance of the feature of node j to node i; α ij is the attention coefficient of the feature of node j to node i; N i is the total number of nodes; LeakyReLU is the activation function.

[0016] Preferably, the calculation formula of the characteristic map is: C is the number of attention mechanism heads; is the attention coefficient of the feature of node j to node i; is the attention weight matrix in the time dimension; is the weight matrix in the time dimension.

[0017] Preferably, the calculation formula of the graph attention weight is: Q is the query matrix; K is the key matrix; V is the actual value matrix; d is the vector length; A(Q, K, V) is the graph attention weight; Softmax() is the normalized exponential function.

[0018] The present invention discloses the following technical effects:

[0019] The present invention provides a method for predicting four-dimensional tropospheric wind field based on an attention mechanism. Through a multi-head sparse attention calculation method, the problems of poor model prediction accuracy, difficult feature extraction, and slow convergence speed are solved, and high-precision and rapid prediction of the four-dimensional wind field is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. 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 paying creative labor.

[0021] Figure 1 A schematic diagram of a four-dimensional tropospheric wind field prediction process based on an attention mechanism provided in an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of a tropospheric four-dimensional wind field prediction module provided in an embodiment of the present invention;

[0023] Figure 3 An attention weight map provided by an embodiment of the present invention;

[0024] Figure 4 A comparison chart of the actual value and the predicted value provided by the embodiment of the present invention;

[0025] Figure 5 This is a diagram of experimental test results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] The purpose of the present invention is to provide a four-dimensional wind field prediction method for the troposphere based on an attention mechanism. Through a multi-head sparse attention calculation method, a four-dimensional wind field prediction model is constructed to achieve a prediction method with high precision, simple feature extraction and fast model calculation.

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Figure 1 A schematic diagram of a four-dimensional tropospheric wind field prediction process based on an attention mechanism provided in an embodiment of the present invention, Figure 2 A schematic diagram of a tropospheric four-dimensional wind field prediction module provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a tropospheric four-dimensional wind field prediction method based on an attention mechanism, comprising:

[0030] Step 100: Use the initialization matrix and normalization function to respectively initialize the preset wind information d in the database. en Initialize and standardize to get the encoded input d' en and graph attention weight att(W);

[0031] Step 200: Calculate the first time position code of the preset wind information using a sine function;

[0032] Step 300: using a sparse attention calculation formula, performing attention calculation on the coded input in chronological order according to the first time position code to obtain a feature map;

[0033] Step 400: Collect historical wind information d of the coordinate point to be predicted de ;

[0034] Step 500: Calculate the product of the historical wind information and the graph attention weight to obtain the weighted wind information d' de ;

[0035] Step 600: Calculate the second time position code of the weighted wind information using a sine function;

[0036] Step 700: using a sparse attention calculation formula, performing attention calculation on the weighted wind information in chronological order according to the second time position code to obtain an attention matrix;

[0037] Step 800: Input the attention matrix and the feature map into the preset multi-head attention layer for full connection calculation to obtain the target prediction result d pre .

[0038] Furthermore, initialization includes: selecting and cutting historical wind information according to a specified length. In this embodiment, it is necessary to collect wind information for 4 days, 11 heights, and 2 directions. Therefore, the wind information of each coordinate point needs to be processed into the form of (1, 96, 22).

[0039] Specifically, the implementation process of this embodiment is actually divided into two steps. The first step is to fix the time and position information, and only train the attention weights for the wind information at different heights and angles at the same coordinate point to obtain the attention feature map. After this step is completed, each height and angle is converted into new wind information through weighting (which can be understood as being mapped to a new coordinate system), and then the time series prediction of the time dimension is performed. The new wind information obtained in the previous step is obtained by the sparse attention mechanism according to time to obtain the attention feature map of the time dimension, that is, the information at different time points is combined with the weights of the information at other time points. The two feature maps are feature maps of the spatial dimension and the time dimension, respectively.

[0040] Specifically, the location information is the result of normalizing the longitude and latitude information of the data collection point so that all location codes are between (-1, 1); in this embodiment, the experimental location range is east longitude [86°, 92°] and north latitude [43°, 46°], and the location code of the coordinates (X, Y) is Time coding is to encode the month, day, and hour information in a sinusoidal manner. For example, if the data is measured at month a, day b, and hour c, the time coding of that moment is

[0041] Furthermore, the sparse attention calculation formula is: e ij is the importance of the feature of node j to node i; α ij is the attention coefficient of the feature of node j to node i; N i is the total number of nodes; LeakyReLU is the activation function.

[0042] Optionally, LeakyReLU(x)=max(0,x)+α*min(0,x), and in this embodiment, α=0.01.

[0043] Specifically, the calculation formula of the feature map is: C is the number of attention mechanism heads; is the attention coefficient of the feature of node j to node i; is the attention weight matrix in the time dimension; is the weight matrix in the time dimension.

[0044] Specifically, α ij is the attention coefficient of the feature of node j to node i, which is the weight result of the graph attention mechanism. is the input time series, W dis the attention weight matrix in the time dimension; multiple heads refer to different consideration angles. In this embodiment, we set C different initial weight values ​​to obtain C results through training, and finally obtain the average of the results. Therefore, the graph attention coefficient of the cth head is The weight matrix in the time dimension is

[0045] Furthermore, the calculation formula of the graph attention weight is: Q is the query matrix; K is the key matrix; V is the actual value matrix; d is the vector length; A(Q, K, V) is the graph attention weight; Softmax() is the normalized exponential function.

[0046] Furthermore, A(Q, K, V) is the graph attention weight, that is, the weight distribution of different vectors participating in other vectors. Assuming that there are N-dimensional wind information in total (N=22 in this embodiment), the purpose of the graph attention weight is to consider the influence of other height and angle wind information on the wind information of this dimension when calculating a certain wind information. Therefore, A(Q, K, V) saves an N*N weight matrix, which corresponds to the correlation between the N-dimensional parameters and themselves and other dimensions, to calculate a weighted wind information value. Softmax() is a normalized exponential function, and its purpose is to map the obtained results to the interval of 0-1 in an exponential manner, and to further increase the discrimination of the weights through exponential calculation. The part with smaller weight will be closer to 0, reducing the error caused by it, which is beneficial to subsequent calculations.

[0047] Specifically, this embodiment predicts long-term four-dimensional wind field information based on the historical time series of the three-dimensional wind field and a small amount of prediction reference values. Wind field information is a collective concept. The tropospheric wind field information of a specific geographic coordinate point includes the wind speed and wind direction data at different heights of the coordinate point. In the actual measurement process, the wind speed and direction data are quantified into U, V, and W components. The vertical component W is difficult to obtain, so the wind information at a certain coordinate and a certain height is usually composed of U and V components. The U and V components represent the horizontal wind direction and wind speed information of the point. For the same geographic coordinate point, there is a certain correlation between the U and V components at different heights. In order to explore the quantitative relationship between the correlation between wind information at different heights, the wind information of the coordinate point is used. Indicates that is the wind information at the i-th height, and the wind information after feature extraction is The weight matrix is ​​set to W d .

[0048] Furthermore, wind information at different heights constitutes a set of wind information at the coordinate point. There is a strong correlation between the U-direction and V-direction parameters at the same height, and the parameter values ​​at adjacent heights also affect each other and share similar properties. In order to capture spatial dependencies, this embodiment adopts the graph attention mechanism (GAT) method, transforms and propagates information as the first layer, and obtains the association weights between different information nodes.

[0049] Among them, the attention coefficients of different nodes are:

[0050]

[0051] e ij Indicates the importance of the feature of node j to node i. ij To normalize:

[0052]

[0053] refer to Figure 3 ,Since the correlation of wind field information often exists in adjacent nodes, the adjacency matrix is ​​used to select the topological neighborhood in the graph to calculate the attention coefficient. This embodiment uses the LeakyReLU function as the activation function, and the function generation expression is as follows:

[0054]

[0055] The calculated results are used to find the final output characteristics of each node. In order to improve the accuracy of the calculation and avoid accidental errors, the calculation is performed in a multi-headed manner, and the average value of the output results is used as the final output result:

[0056]

[0057] Where C represents the number of attention mechanism heads.

[0058] Furthermore, after the graph attention mechanism processes the data, the wind farm information obtains the dependencies in the spatial dimension. However, time series prediction also requires obtaining the dependencies in the time dimension of the sequence, which requires further use of the attention mechanism. The traditional attention mechanism requires training three target matrices: the query matrix Q, the key matrix K, and the actual value matrix V. For the wind farm data scenario, the calculation formulas for the three matrices are:

[0059]

[0060] Among them, the weight matrix W Q , W K , W VThis is the target training matrix. The attention mechanism pairs the query matrix corresponding to the input at each moment with the key matrix at all moments, and adjusts the attention weight through the actual value matrix to achieve attention allocation encoding. The attention calculation method of the three matrices is as follows:

[0061]

[0062] Where d represents the length of the vector. Assume that the i-th row in Q, K, and V is q i ,k i ,v i , then the attention of the i-th row is:

[0063]

[0064] Since the traditional attention mechanism has a relatively cumbersome operation mechanism, it will cause a lot of redundancy when performing attention calculation on the data. Therefore, a sparse attention calculation method is adopted:

[0065]

[0066] in It represents the sparse Q, that is, compared with the traditional Q, this matrix only uses the parameter values ​​that have a greater impact on the result.

[0067] Specifically, based on the above analysis and deduction, this embodiment proposes a WGIN model that combines the graph attention mechanism and the sparse attention mechanism for four-dimensional wind field prediction. The overall framework structure of the WGIN model is as follows: Figure 2 First, the measured values ​​of the wind field data are respectively passed into the graph attention module as part of the encoded wind information data and the decoded wind information data to be predicted. The weights between the nodes of the wind information graph are learned through the attention mechanism, and the attention weight coefficients are obtained to perform weighted processing of the wind information data.

[0068] Furthermore, the specific functions of each part are as follows:

[0069] 1) Attention weight layer and softmax layer:

[0070] Initialize the weight matrix W Q , W K , W V Matrix (usually all 0 or all 1 matrix), then each wind information X is weighted as the query matrix Q, key matrix K and actual value matrix V, and the value of the weighted wind information is calculated through the three vectors, that is, the output value of the attention layer. The specific calculation method is as follows:

[0071]

[0072] First, the Q matrix is ​​multiplied by the K matrix, that is, the inner product of the q vector and the k vector in the matrix is ​​calculated, and the calculation result is normalized according to the length d of the vector. Then, it is mapped to the interval of 0-1 through softmax, and the V matrix is ​​weighted to obtain the attention layer output result A(Q, K, V) of the wind information.

[0073] 2) Graph Attention Weight and Feature Map:

[0074] The feature map is the attention weight map in the multi-head attention mechanism. The figure above shows the feature map of the first head when the time length is 48. Each point in the feature map represents the weight value corresponding to the point at other time points. The feature map or attention weight with the minimum loss is obtained by training the training set and stored in the system in preparation for the decoding process.

[0075] 3) Fully connected layer:

[0076] The fully connected layer is a linear regression (single-layer neural network). The output result A(Q, K, V) of the attention layer is used as the weighted result of the wind information as the input of the neural network, the actual wind information to be predicted is used as the output of the neural network, and the weight parameters of this layer are trained.

[0077] Furthermore, when it is necessary to predict the wind information of a certain coordinate point at several future time points, it is first necessary to collect the wind information values ​​of the coordinate point over a period of time in the past, and then initialize the wind information to be predicted as an all-0 matrix to form the input wind information. Then, through the graph attention mechanism, the wind information is weighted in a highly dimensional manner to obtain weighted wind information, and the time position information encoding of the time point and the coordinate point is added. After the time position information of the weighted wind information is encoded, it is necessary to perform calculations on the masked multi-head sparse attention layer and the multi-head attention layer. First, for the input of the first time point of the wind information, the attention calculation result corresponding to the wind information at this point is calculated through the weights in the feature map (that is, the input of the fully connected layer). Then, for the input of the second time point of the wind information, first, it is necessary to calculate the attention value of the information and the first time point, and then use the result to calculate the attention calculation result corresponding to the wind information at this point through the weight in the feature map (that is, each time point is masked, the subsequent wind information is covered as an unknown quantity, and only the wind information at this time point and all time points before this time point are calculated to calculate the self-attention). Similarly, the output result of the attention layer corresponding to the wind information at each time point is calculated as the input of the fully connected layer. Then, the month, day, and hour features of the time information are normalized to form time coding data, and the position coordinates of the wind information are normalized as position coding data, which are used together with the weighted wind information as the input of the encoding and decoding model. Finally, the input comprehensive wind information data is trained through a multi-head sparse attention layer to obtain an attention feature map. And through the corresponding relationship of the feature map, the unknown wind information is decoded and predicted in combination with the known wind information data.

[0078] Optionally, the test data of this embodiment is derived from the wind direction and wind speed data measured by a certain atmospheric physics research institute in 2021. The data measurement interval is north latitude [43°, 46°] east longitude [86°, 92°], with a total of 29161 coordinate points, and data collection is performed once every hour. The collection height interval is [1000hpa, 100hpa], with a total of 11 heights. This embodiment extracts data from June, July, August, and September in the data set, and randomly selects eight consecutive days of data from each month to form a wind field data set sample. The wind field data of the first four days of each group of samples is used as known information, and part of the real data in the last four days of data (randomly extracting 1% to 20% of the coordinate points) is regarded as the calculation result of the supercomputer, and the time series is trained, and the airship performs mission scenarios to simulate the wind field information of other coordinate points in the area. Long sequence prediction.

[0079] Furthermore, this embodiment uses two commonly used indicators of time series prediction to evaluate the prediction effectiveness of the model constructed in this embodiment and other models:

[0080] 1) Mean Square Error (MSE):

[0081]

[0082] Assume that the sample has a total of N sets of vectors {y1,y2,...,y N}, each set of vectors has M elements in is the true value of a single element, for its predicted value.

[0083] 2) Mean Absolute Error (MAE):

[0084]

[0085] MSE and MAE are measures of the error between the actual value and the predicted value. The smaller the value of MSE and MAE, the better the prediction performance.

[0086] Furthermore, this study uses three time series prediction methods for comparison, including LSTMa, Informer, and MTSMixer, to demonstrate the performance of the proposed model. LSTMa: As a variant of LSTM, LSTMa is widely used in the field of machine translation and has better long-sequence prediction capabilities than traditional LSTM methods. Informer: Based on the sparse attention mechanism, it has excellent performance in long-sequence prediction. MTSMixer: The attention mechanism is optimized to have better long-sequence prediction capabilities.

[0087] Specifically, the test batch size of this embodiment is unified as 32, the epoch of each training is 10 (epoch represents the number of complete traversals of the data set during model training), and the learning rate is 1e -4 ,The wind speed data and position coding are both normalized by the mean, and each group of samples are normalized separately, and the time coding is considered in terms of month, day, and hour.

[0088] refer to Figure 4 , June data set, at an altitude of 1000hpa, the comparison result of the predicted value of the U-direction wind field information of a random coordinate point and the true value of the point is shown in the figure. Therefore, this embodiment has a good effect in the long-series time prediction of the troposphere.

[0089] Specifically, Table 1 summarizes the evaluation results of all methods for time prediction of different lengths in four data sets. In the comparative experiment, 10% of the coordinate point data is used as the training set, 10% of the data is used as the validation set, and 80% of the data is used as the test set. In order to reduce the error caused by chance, the average value of the experimental results of different samples is taken.

[0090] Table 1

[0091]

[0092]

[0093] Furthermore, the following conclusions can be drawn from Table 1:

[0094] 1) GWIN has a significant performance improvement over the Informer method in all time forecasts. Due to the particularity of the wind field environment, there is a large difference in the degree of correlation between wind information at different heights. According to the traditional Informer method, data processing will assign too much weight to wind information with large height differences, reducing the prediction effect.

[0095] 2) The MTSMixer algorithm abandons the structure of the attention mechanism to seek better long-sequence prediction capabilities, and has certain advantages in the long-sequence MAE index. Since this method seeks to explore the periodicity of time series laws, the algorithm starts from the data itself and achieves good results, but when the data has unknown periodicity, the effect still needs to be improved. The GWIN method starts more from the wind field structure itself, improves the correlation between known data and unknown data, and thus more effectively finds the periodic laws of time series.

[0096] 3) The LSTMa algorithm is inferior to the other three time series algorithms in all indicators. LSTM and its derivative algorithms are based on neural networks and also take a long time to train. From the data, we can see that compared with short- and medium-term time series prediction, the GWIN algorithm has a more obvious advantage in medium- and long-term time series prediction.

[0097] Specifically, the GWIN method and the Informer method are compared in Table 1, and experimental tests are conducted on training samples and training time length. In order to further explore the ultimate performance of the model proposed in this paper and explore the influence of each module in the model, the weight part, time coding part, and position coding part of the GWIN model are removed respectively, and 1%, 2%, 5%, 10%, and 20% of the samples are used as training sets to test the samples.

[0098] Further, through Figure 5The following conclusions can be drawn: whether to add a graph attention mechanism module to the model has the greatest impact on the experimental results, and time encoding also has a great impact on the experimental results. The position encoding module can improve the accuracy of prediction, but the impact on the experiment is small. As the number of training data samples decreases, the improvement in experimental results brought by position encoding will increase slightly. The reason for the above phenomenon is that the smaller the sample data, the more difficult it is to extract regularity and periodicity information from the training data, so the similarity of the coordinate point position will have a greater guiding significance for the law.

[0099] The beneficial effects of the present invention are as follows:

[0100] The present invention constructs a four-dimensional wind field prediction model through a multi-head sparse attention calculation method, realizing a prediction method with high precision, simple feature extraction and fast model calculation.

[0101] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0102] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A four-dimensional tropospheric wind field prediction method based on attention mechanism, It is characterized in that include: The preset wind information in the database is initialized and normalized using the initialization weight matrix and the normalized exponential function to obtain the encoding input and graph attention weights; Calculating the first time position code of the preset wind information by using a sine function; Using a sparse attention calculation formula, performing attention calculation on the encoded input in chronological order according to the first time position code to obtain a feature map; Collect historical wind information of the coordinate points to be predicted; Calculate the product of the historical wind information and the graph attention weight to obtain weighted wind information; Calculating a second time position code of the weighted wind information using a sine function; Using a sparse attention calculation formula, performing attention calculation on the weighted wind information in chronological order according to the second time position code to obtain an attention matrix; The attention matrix and the feature map are input into a preset multi-head attention layer for full connection calculation to obtain the target prediction result.

2. According to claim 1, a method for predicting tropospheric four-dimensional wind field based on attention mechanism, It is characterized in that The sparse attention calculation formula is: e ij is the importance of the feature of node j to node i; α ij is the attention coefficient of the feature of node j to node i; N i is the total number of nodes; LeakyReLU is the activation function.

3. According to claim 1, a method for predicting tropospheric four-dimensional wind field based on attention mechanism, It is characterized in that The calculation formula of the feature map is: C is the number of attention mechanism heads; is the attention coefficient of the feature of node j to node i; W d c is the attention weight matrix in the time dimension; is the weight matrix in the time dimension.

4. According to claim 1, a method for predicting tropospheric four-dimensional wind field based on attention mechanism, It is characterized in that The calculation formula of the graph attention weight is: Q is the query matrix; K is the key matrix; V is the actual value matrix; d is the vector length; A(Q, K, V) is the graph attention weight; Softmax() is the normalized exponential function.