TEC space-time prediction method based on fusion of feature analysis and self-attention mechanism

Abstract

The invention discloses a TEC space-time prediction method based on fusion of feature analysis and a self-attention mechanism, and the method comprises the steps: screening strong effective features through analyzing the correlation between TEC historical data and each feature parameter, calculating the attention weight between data points on each time sequence through employing a self-attention coding mechanism, and carrying out the weighted summation to obtain the attention output. In an ionosphere total electron content prediction task, firstly, for sun and geomagnetic activity characteristic parameters, the correlation between each parameter and a TEC sequence in the same time period and a TEC sequence after a day interval is evaluated, strong correlation characteristics are screened, TEC sequences of adjacent grid points are used as characteristic quantities, multiple spatio-temporal characteristic matrixes of different time scales in a global range are established, and the total electron content of the ionosphere is predicted. And encoding the input feature information according to a self-attention encoding mechanism to obtain an attention output result, and decoding in a decoding layer through a linear model to obtain a TEC predicted value at a future moment. According to the method, efficient extraction of spatio-temporal information can be completed, and higher-precision prediction of the TEC is realized through a self-attention mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically is a TEC spatiotemporal prediction method integrating feature analysis with a self-attention mechanism. Background Art

[0002] As the requirements for navigation and positioning performance of the global satellite navigation system continue to increase, the impact of space weather disturbances on satellite signal transmission has gradually become an important factor that cannot be ignored in affecting signal quality. Since satellite signals will be affected to varying degrees when passing through the magnetosphere, ionosphere and atmosphere, especially the ionospheric effect has different intensities of interference on the transmission path and energy of navigation signals. Among them, the charged particles filled in the ionospheric space will cause satellite signals of different frequencies to produce different degrees of signal refraction. This phenomenon will cause a delay in signal transmission time, resulting in changes in the time it takes for the signal to reach the ground receiver. The resulting time error also reduces the accuracy of navigation and positioning, and in severe cases, it can reach meter-level errors. This type of ionospheric delay phenomenon is closely related to the ionospheric TEC in the columnar area per unit cross-sectional area on the satellite's path through the ionosphere, and has a greater impact on navigation and positioning performance.

[0003] In order to solve the interference of complex ionospheric environment on satellite navigation system in practical applications and reduce the pseudo-range error caused by ionospheric delay as much as possible, the dual-frequency difference method can be used to eliminate the ionospheric delay error, achieving an ionospheric delay error correction effect of more than 90%, but its correction effect is affected by the observation baseline distance and is only applicable to dual-frequency or multi-frequency receivers. Single-frequency receivers can estimate the ionospheric delay through traditional empirical models, mainly including the International Reference Ionosphere Model (IRI), Klobuchar model, NeQuick model and related improved models. Such methods are simple and efficient, but the correction accuracy is limited, and it is difficult to reflect the long time series and spatial characteristics of TEC. In addition, the existing ionospheric delay error estimation can be achieved by predicting TEC values. Typical neural network models such as Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU) and related improved methods mainly improve the model's TEC prediction performance by better capturing TEC time series information. However, due to the structure of the model itself, the above methods are difficult to obtain TEC information of longer time series or the serial operation structure makes their operation efficiency limited. Therefore, making full use of TEC spatiotemporal information and designing a reasonable prediction model according to data characteristics will help further improve TEC prediction accuracy and efficiency, which is of great significance for maintaining the stable operation of satellite navigation positioning.

[0004] Compared with the existing technology, the technical differences are as follows:

[0005] Technical comparison with patent CN111814398B "A graph-based method for predicting surface solar radiation by integrating spatiotemporal attention"

[0006] Patent CN111814398B constructs a SWR subnet of the actual measurement site for the geographical location of the surface solar radiation to be predicted. The extracted feature data is the measured solar radiation of each subgrid point, which is directly used as the feature parameters of the first P time steps to predict the solar radiation of the next Q time sub-step grid points and the points to be measured. In our TEC prediction task, a set of feature vectors is constructed for each point to be predicted worldwide. In addition to the historical data of the current point TEC, its parameters also include related influencing parameters such as the solar activity index and the geomagnetic activity index. The historical TEC data around the point to be measured is extracted to capture its spatial correlation, and the TEC time periodicity is converted into multi-time scale feature parameters. Before model encoding, feature analysis and screening are also completed, that is, the above features are correlated with the TEC historical data of the current period and after 24 hours, and effective features are screened to achieve enrichment of feature parameters from multiple angles, fully extract spatiotemporal information, optimize feature parameters, improve data quality, and thus improve model efficiency and accuracy.

[0007] Technical comparison with patent CN114492978B "Global ionospheric TEC prediction method and system based on improved spatiotemporal prediction model"

[0008] Patent CN114492978B converts global TEC grid data into a heat map TEC map, corresponds longitude and latitude to image height and width, and four features to the number of channels of the image. The sample is global TEC data from 2020 to 2023. We use all grid points in the global range to construct feature vectors, whose elements include historical TEC data of the current point and surrounding points, more solar and geomagnetic activity related parameters, and multi-time scale feature parameters. In the data preprocessing stage, the correlation analysis of the above features with the current TEC sequence and the TEC prediction sequence after one day is used to screen high-quality feature information to optimize feature parameters and improve data quality. At the same time, the samples we use are global TEC historical data, solar and geomagnetic activity data from 1998 to 2018. The data covers the 23rd and 24th solar activity cycles, which is more conducive to extracting rich TEC feature information and constructing a more generalized prediction model.

[0009] Patent CN114492978B uses the attention mechanism to dynamically adjust the output weight of the LSTM network. This method calculates the attention between the input sequence and the decoder, and requires the decoder's state to interact with the input sequence to generate the output, and calculates step by step. We use the Self-attention mechanism, which does not require an external decoder part and emphasizes self-attention within the input sequence. Each position can achieve self-attention by calculating the self-attention score, which can handle the long-distance dependency of the TEC sequence in the modeling process and has stronger parallel computing capabilities. Summary of the invention

[0010] In view of the above problems, the present invention proposes a TEC spatiotemporal prediction method that integrates feature analysis and self-attention mechanism. The method combines solar and geomagnetic activity indexes and TEC historical data, and constructs TEC spatiotemporal feature parameters on this basis. The correlation between TEC historical data and each feature parameter is first analyzed and features with high correlation are selected. Then, the input feature vector is encoded using the Transformer model self-attention encoding mechanism, thereby extracting TEC spatiotemporal feature information, and finally decoding the linear model at the decoding layer to obtain the TEC prediction value at the future moment. The method can complete the efficient extraction of spatiotemporal information and achieve more accurate TEC prediction through the self-attention mechanism.

[0011] To achieve the above object, the technical solution adopted by the present invention is:

[0012] The TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism includes the following steps:

[0013] (1) Extract solar and geomagnetic activity parameters, and use the TEC historical data of the current grid point and the last four grid points as spatial characteristic parameters. According to the characteristics of the solar activity cycle, construct indexes with year, season, month, and hour as time scales as temporal characteristic parameters.

[0014] (2) Perform correlation analysis on the solar and geomagnetic activity characteristic parameters and time characteristic parameters in the above period of time with the TEC historical data of the same period of time and the TEC historical data after one day, and select characteristic parameters with strong correlation. The above selected strongly correlated characteristic parameters, the TEC historical data of the current point and the last four points are used as feature vectors, and the TEC value of the current point in the next day to be predicted is used as sample data; wherein the characteristic parameters in each sample data are the historical data of the previous n days, and the value to be predicted is the TEC value of the n+1th day;

[0015] (3) In units of hours, the Transformer model self-attention encoding mechanism is used to encode the feature vectors of T = 24*n time points in the previous n days, and the query vector, key vector and value vector of each feature parameter in each set of feature vectors are calculated. The self-attention weight vector of the position corresponding to each feature vector is calculated based on the query vector and the key vector. Finally, for each feature vector, the weighted sum between the weight vector and the value vector is calculated, and then the self-attention encoding process is completed through forward feedback and layer normalization to obtain the self-attention score matrix. The output dimension is N×T×S, where S is the number of feature elements and N is the number of samples.

[0016] (4) The output of the self-attention encoding layer is used as the input information of the decoding layer. The input data is converted into 24 TEC prediction values ​​to be predicted by linear transformation to obtain a vector containing 24 elements.

[0017] As a further improvement of the present invention, step (1) specifically comprises:

[0018] (1.1) Extract the solar activity index SSN and F10.7, the geomagnetic activity intensity index Kp, ap, AE and Dst, and the TEC historical data of the current grid point and the last four points as characteristic parameters;

[0019] (1.2) Taking 11 years as the cycle, the years of the 23rd and 24th solar activity cycles are marked as FY = {1, 2, ..., 11, 1, 2, ..., 11}, and the rest of the years are deduced in the same way. Then, the following cycle conversion formula is used to convert FY into annual characteristic parameters FYS and FYC:

[0020]

[0021] (1.3) Taking the total number of days in a year T as the period, where T = 365 in a normal year and T = 366 in a leap year, the annual cumulative day DOY is constructed into characteristic parameters FDS and FDC according to the sine and cosine formulas:

[0022]

[0023] Where, DOY=1,2,...,T;

[0024] (1.4) With a 24-hour period, the hourly HOD of a day is constructed into characteristic parameters FHS and FHC according to the sine and cosine formulas:

[0025]

[0026] Where, HOD=0,1,2,...,23;

[0027] (1.5) Take the historical TEC data of a grid point in a low-latitude area during the period of strong solar and geomagnetic activity, and calculate the correlation with the solar and geomagnetic characteristic parameters SSN, F10.7, Kp, ap, AE and Dst, and the time scale characteristic parameters FYS, FYC, FDS, FDC, FHS and FHC during the same period; at the same time, take the historical TEC data of the same time length after one day, and calculate the correlation with the above characteristic parameters, and select the characteristic parameters with the highest correlation;

[0028] (1.6) Extract the current grid point TEC0 and the nearest four grid points TEC1, TEC2, TEC3 and TEC4 as spatial correlation feature parameters.

[0029] As a further improvement of the present invention, step (2) specifically comprises:

[0030] (2.1) Assume that the feature vector formed by the filtered feature parameters is as follows. In units of hours, the feature parameters of the i-th hour in the previous n days are constructed as the feature vector f i :

[0031] f i ={SSN i ,F10.7 i ,Kp i ,ap i ,…,TEC 0,i ,TEC 1,i ,TEC 2,i TEC 3,i ,TEC 4,i}

[0032] Wherein, i=1,2,...,24*n;

[0033] (2.2) All the characteristic parameters of the previous n days are constructed into a characteristic vector F = {f1,f2,...,f n};

[0034] (2.3) The predicted TEC value for the 24 hours of the previous day corresponding to the feature vector of the previous day is constructed as the predicted label vector Y: Y = {TEC0, TEC1, TEC2, ..., TEC 23}.

[0035] As a further improvement of the present invention, step (3) specifically comprises:

[0036] (3.1) Substitute the eigenvectors f1, f2, ...f in the eigenvector F i ...f N , where i = 1, 2, ..., N, N = 24*n parallel input, and the self-attention mechanism is used to calculate the query vector q corresponding to the input vectori , key vector k i Sum value vector v i :

[0037]

[0038] (3.2) Calculate the input vector f for each position i The weight of the query vector in . Taking the feature vector f1 as an example, calculate the attention weight value of position j, j=1,2,...,N for f1:

[0039]

[0040] Where, d k is the key matrix dimension;

[0041] (3.3) Calculate the feature information f extracted by the self-attention mechanism for the feature vector f1 1,Attention :

[0042]

[0043] (3.4) Similarly, for the input vectors f2,f3,...,f N , extract the corresponding feature information according to steps (2.2) and (2.3), and combine them to get the single-head self-attention score matrix F Attention :

[0044] F Attention ={f 1,Attention ,f 2,Attention ,...,f N,Attention}

[0045] (3.5) Based on the calculation of the single-head self-attention mechanism in step (3.4), a multi-head self-attention mechanism is used to convert all feature vectors in step (3.1) into the query matrix Q = {q1, q2, ..., q 120}, key matrix K = {k1, k2, ..., k 120} and value matrix V = {v1,v2,...,v 120} is divided into h equal parts, and h groups of query matrices Q are obtained m , key matrix K m Sum value matrix V m (m=1,2,..,h);

[0046] (3.6) Using the initialized linear change matrix and Map the above matrix to different spaces and execute steps (2.1) to (2.4) to obtain h groups of self-attention score matrices:

[0047]

[0048] Among them, head m The dimension is T×(S / h);

[0049] (3.7) Merge h groups of attention score matrices and initialize the linear transformation matrix W O Get the multi-head self-attention score matrix F with dimension T×S Attention :

[0050] Head=[head1,head2,...,head h ]

[0051] F Attention =Head·W O

[0052] (3.8) The above self-attention score matrix F Attention The self-attention encoding process is then completed through forward feedback and layer normalization to obtain the self-attention encoding output matrix Z with a dimension of N×T×S, where S is the number of feature elements and N is the number of samples, that is, the model predicts the sample size batch_size for each batch.

[0053] As a further improvement of the present invention, step (4) specifically comprises:

[0054] (4.1) The output matrix Z of each batch of encoding part obtained by step (3.8) is used as the input matrix of the decoding part, and it is converted into the prediction result TEC using linear transformation pred :

[0055]

[0056] Among them, W L and b L Preset initial values ​​for linear transformation. For a single sample, Z is a vector of 1×(T*S), W L is a linear transformation matrix of (T*S)×24, b L is a 1×24 vector, and the TEC obtained from the above formula is pred That is, the TEC predicted value for 24 hours in the n+1th day obtained by the prediction model proposed in the present invention based on the historical data of the previous n days.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention proposes a TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism. In addition to extracting solar and geomagnetic activity indexes closely related to TEC changes as feature parameters, the method evaluates the correlation between each parameter and the TEC sequence of the same period and the TEC of the period after one day, selects strong correlation features, and constructs the TEC sequence of inverse distance adjacent grid points considering spatial correlation as feature quantity. Combined with TEC historical data, multiple spatiotemporal feature matrices of different time scales in the global scope are established to improve the richness of sample data. Then, in the encoding stage of the Transformer model, the self-attention mechanism is used to extract features of TEC spatiotemporal information, which fully captures the time periodicity and spatial correlation of the TEC long time series. Finally, a linear transformation is used in the decoding layer, and the self-attention output matrix is ​​used as the input of the decoding layer to obtain the TEC prediction value for the next 24 hours. The method adopts the parallel operation structure of feature screening and model, which greatly improves the model training and prediction speed, and has unique advantages for processing solar activity related data with a cycle of 11 years. At the same time, the model's multi-head self-attention mechanism can map feature information to spaces of different dimensions, achieving full extraction of TEC feature data, which helps to improve the learning effect of the model. Compared with the existing prediction models, it can effectively improve the TEC prediction accuracy in different regions, which is of great significance to reducing the positioning error of the satellite navigation system caused by ionospheric disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of TEC spatiotemporal prediction by integrating feature analysis and self-attention mechanism of the present invention;

[0060] Figure 2 It is the Transformer self-attention encoding TEC prediction flow chart of the spatiotemporal information fusion of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0062] This embodiment discloses a TEC spatiotemporal prediction method that integrates feature analysis and self-attention mechanism, such as Figure 1 and Figure 2 As shown, including:

[0063] Step 1: Extract solar and geomagnetic activity parameters, and use the TEC historical data of the current grid point and the last four grid points as spatial characteristic parameters. According to the characteristics of the solar activity cycle, construct indexes with year, season, month, and hour as time scales as time characteristic parameters.

[0064] The specific steps include:

[0065] (1.1) Extract the solar activity index SSN and F10.7, the geomagnetic activity intensity index Kp, ap, AE and Dst, and the TEC historical data of the current grid point and the last four points as characteristic parameters;

[0066] (1.2) Taking 11 years as the cycle, mark the years of the 23rd and 24th solar activity cycles (1997 to 2018) as FY = {1, 2, ..., 11, 1, 2, ..., 11}, and the rest of the years are similar. Then, the following cycle conversion formula is used to convert FY into annual characteristic parameters FYS and FYC:

[0067]

[0068] (1.3) Taking the total number of days in a year T as the period, where T = 365 in a normal year and T = 366 in a leap year, the annual cumulative day DOY is constructed into characteristic parameters FDS and FDC according to the sine and cosine formulas:

[0069]

[0070] Where, DOY=1,2,...,T;

[0071] (1.4) With a 24-hour period, the hourly HOD of a day is constructed into characteristic parameters FHS and FHC according to the sine and cosine formulas:

[0072]

[0073] Where, HOD=0,1,2,...,23;

[0074] (1.5) Take the historical TEC data of a grid point in a low-latitude area during the period of strong solar and geomagnetic activity, and calculate the correlation with the solar and geomagnetic characteristic parameters SSN, F10.7, Kp, ap, AE and Dst, and the time scale characteristic parameters FYS, FYC, FDS, FDC, FHS and FHC during the same period; at the same time, take the historical TEC data of the same time length after one day, and calculate the correlation with the above characteristic parameters, and select the characteristic parameters with the highest correlation;

[0075] (1.6) Extract the current grid point TEC0 and the nearest four grid points TEC1, TEC2, TEC3 and TEC4 as spatial correlation feature parameters.

[0076] Step 2: Perform correlation analysis on the solar and geomagnetic activity characteristic parameters and time characteristic parameters in the above period of time with the TEC historical data of the same period of time and the TEC historical data after one day, and select characteristic parameters with strong correlation. The above selected strong correlation characteristic parameters, the current point and the TEC historical data of the last four points are used as feature vectors, and the TEC value of the current point in the next day to be predicted is used as sample data; wherein, the characteristic parameters in each sample data are the historical data of the previous n days, and the value to be predicted is the TEC value of the n+1th day.

[0077] The specific steps include:

[0078] (2.1) Assume that the feature vector formed by the filtered feature parameters is as follows. In units of hours, the feature parameters of the i-th hour in the previous n days are constructed as the feature vector f i :

[0079] f i ={SSN i ,F10.7 i ,Kp i ,ap i ,…,TEC 0,i ,TEC 1,i ,TEC 2,i TEC 3,i ,TEC 4,i}

[0080] Wherein, i=1,2,...,24*n;

[0081] (2.2) All the characteristic parameters of the previous n days are constructed into a characteristic vector F = {f1,f2,...,f n};

[0082] (2.3) The predicted TEC value for the 24 hours of the previous day corresponding to the feature vector of the previous day is constructed as the predicted label vector Y:

[0083] Y={TEC0,TEC1,TEC2,...,TEC 23}

[0084] Step 3: In units of hours, use the Transformer model self-attention encoding mechanism to encode the feature vectors of T = 24*n time points in the previous n days, calculate the query vector, key vector and value vector of each feature parameter in each group of feature vectors, and calculate the self-attention weight vector corresponding to the position of each feature vector based on the query vector and the key vector. Finally, for each feature vector, after the weighted summation between the weight vector and the value vector, the self-attention encoding process is completed through forward feedback and layer normalization to obtain the self-attention score matrix. The output dimension is N×T×S, where S is the number of feature elements and N is the number of samples.

[0085] The specific steps include:

[0086] (3.1) Substitute the eigenvectors f1, f2, ...f in the eigenvector F i ...f N (where i = 1, 2, ..., N, N = 24*n) parallel input, using the self-attention mechanism to calculate the query vector q corresponding to the input vector i , key vector k i Sum value vector v i :

[0087]

[0088] (3.2) Calculate the input vector f for each position i The weight of the query vector in . Taking the feature vector f1 as an example, calculate the attention weight value of position j (j = 1, 2, ..., N) on f1:

[0089]

[0090] Where, d k is the key matrix dimension;

[0091] (3.3) Calculate the feature information f extracted by the self-attention mechanism for the feature vector f1 1,Attention :

[0092]

[0093] (3.4) Similarly, for the input vectors f2,f3,...,f N , extract the corresponding feature information according to steps (2.2) and (2.3), and combine them to get the single-head self-attention score matrix F Attention :

[0094] F Attention ={f 1,Attention ,f 2,Attention ,...,f N,Attention}

[0095] (3.5) Based on the calculation of the single-head self-attention mechanism in step (3.4), a multi-head self-attention mechanism is used to convert all feature vectors in step (3.1) into the query matrix Q = {q1, q2, ..., q 120}, key matrix K = {k1, k2, ..., k 120} and value matrix V = {v1,v2,...,v 120} is divided into h equal parts, and h groups of query matrices Q are obtained m , key matrix K m Sum value matrix Vm (m=1,2,..,h);

[0096] (3.6) Using the initialized linear change matrix and Map the above matrix to different spaces and execute steps (2.1) to (2.4) to obtain h groups of self-attention score matrices:

[0097]

[0098] Among them, head m The dimension is T×(S / h);

[0099] (3.7) Merge h groups of attention score matrices and initialize the linear transformation matrix W O Get the multi-head self-attention score matrix F with dimension T×S Attention :

[0100] Head=[head1,head2,...,head h ]

[0101] F Attention =Head·W O

[0102] (3.8) The above self-attention score matrix F Attention The self-attention encoding process is then completed through forward feedback and layer normalization to obtain the self-attention encoding output matrix Z with a dimension of N×T×S, where S is the number of feature elements and N is the number of samples, that is, the model predicts the sample size batch_size for each batch.

[0103] Step 4: Use the output of the self-attention encoding layer as the input information of the decoding layer, and use linear transformation to convert the input data into 24 TEC prediction values ​​to be predicted, and obtain a vector containing 24 elements.

[0104] The specific steps include:

[0105] (4.1) The output matrix Z of each batch of encoding part in step (3.8) is used as the input matrix of the decoding part, and it is converted into the prediction result TEC using linear transformation pred :

[0106]

[0107] Among them, W L and b L Preset initial value for linear transformation. For a single sample, Z is a 1×(T*S) vector, W L is a linear transformation matrix of (T*S)×24, b Lis a 1×24 vector. The TEC obtained from the above formula is pred That is, the TEC predicted value for 24 hours in the n+1th day obtained by the prediction model proposed in the present invention based on the historical data of the previous n days.

[0108] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. The TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism is characterized by: The steps include: (1) Extract solar and geomagnetic activity parameters, and use the TEC historical data of the current grid point and the last four grid points as spatial characteristic parameters. According to the characteristics of the solar activity cycle, construct indexes with year, season, month, and hour as time scales as temporal characteristic parameters. (2) Perform correlation analysis on the solar and geomagnetic activity characteristic parameters and time characteristic parameters in the above period of time with the TEC historical data of the same period of time and the TEC historical data after one day, and select characteristic parameters with strong correlation. The above selected strongly correlated characteristic parameters, the TEC historical data of the current point and the last four points are used as feature vectors, and the TEC value of the current point in the next day to be predicted is used as sample data; wherein the characteristic parameters in each sample data are the historical data of the previous n days, and the value to be predicted is the TEC value of the n+1th day; (3) In units of hours, the Transformer model self-attention encoding mechanism is used to encode the feature vectors of T = 24*n time points in the previous n days, and the query vector, key vector and value vector of each feature parameter in each set of feature vectors are calculated. The self-attention weight vector of the position corresponding to each feature vector is calculated based on the query vector and the key vector. Finally, for each feature vector, the weighted sum between the weight vector and the value vector is calculated, and then the self-attention encoding process is completed through forward feedback and layer normalization to obtain the self-attention score matrix. The output dimension is N×T×S, where S is the number of feature elements and N is the number of samples. (4) The output of the self-attention encoding layer is used as the input information of the decoding layer. The input data is converted into 24 TEC prediction values ​​to be predicted by linear transformation to obtain a vector containing 24 elements.

2. The TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism according to claim 1, characterized in that: Step (1) specifically includes: (1.1) Extract the solar activity index SSN and F10.7, the geomagnetic activity intensity index Kp, ap, AE and Dst, and the TEC historical data of the current grid point and the last four points as characteristic parameters; (1.2) Taking 11 years as the cycle, the years of the 23rd and 24th solar activity cycles are marked as FY = {1, 2, ..., 11, 1, 2, ..., 11}, and the rest of the years are deduced in the same way. Then, the following cycle conversion formula is used to convert FY into annual characteristic parameters FYS and FYC: (1.3) Taking the total number of days in a year T as the period, where T = 365 in a normal year and T = 366 in a leap year, the annual cumulative day DOY is constructed into characteristic parameters FDS and FDC according to the sine and cosine formulas: Where, DOY=1,2,...,T; (1.4) With a 24-hour period, the hourly HOD of a day is constructed into characteristic parameters FHS and FHC according to the sine and cosine formulas: Where, HOD=0,1,2,...,23; (1.5) Take the historical TEC data of a grid point in a low-latitude area during the period of strong solar and geomagnetic activity, and calculate the correlation with the solar and geomagnetic characteristic parameters SSN, F10.7, Kp, ap, AE and Dst, and the time scale characteristic parameters FYS, FYC, FDS, FDC, FHS and FHC during the same period; at the same time, take the historical TEC data of the same time length after one day, and calculate the correlation with the above characteristic parameters, and select the characteristic parameters with the highest correlation; (1.6) Extract the current grid point TEC0 and the nearest four grid points TEC1, TEC2, TEC3 and TEC4 as spatial correlation feature parameters.

3. The TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism according to claim 1, characterized in that: Step (2) specifically includes: (2.1) Assume that the feature vector formed by the filtered feature parameters is as follows. In units of hours, the feature parameters of the i-th hour in the previous n days are constructed as the feature vector f i : f i ={SSN i ,F10.7 i ,Kp i ,ap i ,…,TEC 0,i ,TEC 1,i ,TEC 2,i TEC 3,i ,TEC 4,i } Wherein, i=1,2,...,24*n; (2.2) All the characteristic parameters of the previous n days are constructed into a characteristic vector F = {f1,f2,...,f n }; (2.3) The predicted TEC value for the 24 hours of the previous day corresponding to the feature vector of the previous day is constructed as the predicted label vector Y: Y={TEC0,TEC1,TEC2,...,TEC 23 }。 4. The TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism according to claim 1, characterized in that: Step (3) specifically includes: (3.1) Substitute the eigenvectors f1, f2, ...f in the eigenvector F i ...f N , where i = 1, 2, ..., N, N = 24*n parallel input, and the self-attention mechanism is used to calculate the query vector q corresponding to the input vector i , key vector k i Sum value vector v i : (3.2) Calculate the input vector f for each position i The weight of the query vector in . Taking the feature vector f1 as an example, calculate the attention weight value of position j, j=1,2,...,N for f1: Where, d k is the key matrix dimension; (3.3) Calculate the feature information f extracted by the self-attention mechanism for the feature vector f1 1,Attention : (3.4) Similarly, for the input vectors f2,f3,...,f N , extract the corresponding feature information according to steps (2.2) and (2.3), and combine them to get the single-head self-attention score matrix F Attention : F Attention ={f 1,Attention ,f 2,Attention ,...,f N,Attention } (3.5) Based on the calculation of the single-head self-attention mechanism in step (3.4), a multi-head self-attention mechanism is used to convert all feature vectors in step (3.1) into the query matrix Q = {q1, q2, ..., q 120 }, key matrix K = {k1, k2, ..., k 120 } and value matrix V = {v1,v2,...,v 120 } is divided into h equal parts, and h groups of query matrices Q are obtained m , key matrix K m Sum value matrix V m (m=1,2,..,h); (3.6) Using the initialized linear change matrix and Map the above matrix to different spaces and execute steps (2.1) to (2.4) to obtain h groups of self-attention score matrices: Among them, head m The dimension is T×(S / h); (3.7) Merge h groups of attention score matrices and initialize the linear transformation matrix W O Get the multi-head self-attention score matrix F with dimension T×S Attention : Head=[head1,head2,...,head h ] F Attention =Head·W O (3.8) The above self-attention score matrix F Attention The self-attention encoding process is then completed through forward feedback and layer normalization to obtain the self-attention encoding output matrix Z with a dimension of N×T×S, where S is the number of feature elements and N is the number of samples, that is, the model predicts the sample size batch_size for each batch.

5. The TEC spatiotemporal prediction method integrating feature analysis and self-attention mechanism according to claim 1, characterized in that: Step (4) specifically includes: (4.1) The output matrix Z of each batch of encoding part obtained by step (3.8) is used as the input matrix of the decoding part, and it is converted into the prediction result TEC using linear transformation pred : Among them, W L and b L Preset initial values ​​for linear transformation. For a single sample, Z is a vector of 1×(T*S), W L is a linear transformation matrix of (T*S)×24, b L is a 1×24 vector, and the TEC obtained from the above formula is pred That is, the TEC predicted value for 24 hours in the n+1th day obtained by the prediction model proposed in the present invention based on the historical data of the previous n days.

Images