TEC spatio-temporal prediction method combining feature analysis and self-attention mechanism

Abstract

The TEC spatio-temporal prediction method based on feature analysis and self-attention mechanism is to screen strong and effective features by analyzing the correlation between TEC historical data and each feature parameter, to calculate the attention weight between data points on each time series by using the self-attention encoding mechanism, and to obtain the attention output by weighted summation. In the ionospheric total electron content prediction task, firstly, the correlation between each parameter and the TEC sequence at the same period and the TEC sequence after one day is evaluated according to the solar and geomagnetic activity feature parameters, the strong correlation features are screened, and the adjacent grid point TEC sequence is taken as the feature quantity to establish a multi-temporal and spatial feature matrix of different time scales in the global range. Then, the input feature information is encoded according to the self-attention encoding mechanism to obtain the attention output result, and then the future TEC prediction value is obtained by linear model decoding in the decoding layer. The method can efficiently extract spatio-temporal information and realize higher precision prediction of TEC through the self-attention mechanism.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically the TEC spatiotemporal prediction method that integrates feature analysis and self-attention mechanism. Background Technology

[0002] As the requirements for navigation and positioning performance of global satellite navigation systems continue to increase, the impact of space weather disturbances on satellite signal transmission has gradually become a significant factor affecting signal quality. Satellite signals are affected to varying degrees when passing through the magnetosphere, ionosphere, and atmosphere, with the ionospheric effect in particular interfering with the transmission path and energy of navigation signals to varying degrees. Specifically, the charged particles filling the ionosphere cause varying degrees of signal refraction in satellite signals of different frequencies. This phenomenon results in a delay in signal transmission time, altering the time it takes for the signal to reach the ground receiver. The resulting time error reduces navigation and positioning accuracy, sometimes reaching meter-level errors. This ionospheric delay phenomenon is closely related to the ionospheric TEC (Transient Electron Delay) within a columnar region per unit cross-sectional area along the satellite's path through the ionosphere, and has a significant impact on navigation and positioning performance.

[0003] To address the interference from the complex ionospheric environment on satellite navigation systems in practical applications and minimize pseudorange errors caused by ionospheric delay, a dual-frequency differential method can be used to eliminate ionospheric delay errors, achieving a correction effect of over 90%. However, 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 ionospheric delay using traditional empirical models, including the International Reference Ionosphere (IRI), Klobuchar, NeQuick, and related improved models. These methods are simple and efficient but have limited correction accuracy and struggle to reflect long-term TEC series and spatial characteristics. Furthermore, 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, primarily improve the model's TEC prediction performance by better capturing TEC time-series information. However, the aforementioned methods, due to their inherent model structure, struggle to acquire longer-term TEC information or their sequential operation structure limits their efficiency. Therefore, fully utilizing TEC spatiotemporal information and designing reasonable prediction models based on data characteristics is beneficial for further improving TEC prediction accuracy and efficiency, which is of great significance for maintaining stable operation of satellite navigation and positioning.

[0004] The technical differences compared to existing technologies are as follows:

[0005] Technical Comparison with Patent CN111814398B "A Graph-Based Fusion Spatiotemporal Attention Method for Predicting Earth's Surface Solar Irradiance"

[0006] Patent CN111814398B constructs a measured station SWR subnet based on the geographical location of the surface solar radiance to be predicted. The extracted feature data is the measured solar radiance of each subnet point, which is directly used as the feature parameters for the first P time steps to predict the solar radiance of the subnet points and the point to be measured in the next Q time steps. In our TEC prediction task, a set of feature vectors is constructed for each point to be predicted globally. The parameters include not only the historical TEC data of the current point, but also its related influencing parameters such as the solar activity index and the geomagnetic activity index. The historical TEC data around the point to be measured are extracted to capture its spatial correlation, and the temporal periodicity of TEC is transformed into multi-timescale feature parameters. Before model encoding, feature analysis and screening are also performed. The above features are correlated with the historical TEC data of the current time period and 24 hours later to screen effective features. This achieves the goal of enriching feature parameters from multiple perspectives, fully extracting spatiotemporal information, optimizing feature parameters, improving data quality, and thus improving 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 heatmap TEC map, mapping latitude and longitude to image height and width, and four features to the number of image channels. The sample is global TEC data from 2020-2023. Our approach, however, constructs feature vectors for all grid points globally. These vectors include historical TEC data for the current point and surrounding points, more solar and geomagnetic activity-related parameters, and multi-timescale feature parameters. During data preprocessing, we use correlation analysis between these features and the current TEC sequence and the predicted TEC sequence one day later to filter high-quality feature information, thereby optimizing feature parameters and improving data quality. Furthermore, our sample consists of historical global TEC data from 1998-2018, as well as solar and geomagnetic activity data, covering the 23rd and 24th solar cycles, which is more conducive to extracting richer TEC feature information and constructing a more generalized prediction model.

[0009] Patent CN114492978B employs an attention mechanism to dynamically adjust the output weights of an LSTM network. This method calculates the attention between the input sequence and the decoder, requiring interaction between the decoder's state and the input sequence to generate the output, a step-by-step calculation. In contrast, we use a self-attention mechanism. This method eliminates the need for an external decoder, emphasizing self-attention within the input sequence. Self-attention can be achieved between each position by calculating a self-attention score, handling long-distance dependencies in TEC sequences during modeling and offering stronger parallel computing capabilities. Summary of the Invention

[0010] To address the above problems, this invention proposes a spatiotemporal prediction method for TEC (Thermal Activity Domain) that integrates feature analysis and a self-attention mechanism. This method combines solar and geomagnetic activity indices with historical TEC data to construct spatiotemporal feature parameters for TEC. First, the correlation between historical TEC data and each feature parameter is analyzed, and features with high correlation are selected. Then, a Transformer model self-attention encoding mechanism is used to encode the input feature vector, thereby extracting the spatiotemporal feature information of TEC. Finally, the predicted TEC value for future times is obtained through decoding using a linear model at the decoding layer. This method can efficiently extract spatiotemporal information and achieve higher-precision TEC prediction through the self-attention mechanism.

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

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

[0013] (1) Extract solar and geomagnetic activity parameters, and use the current grid point and the TEC historical data of the last 4 grid points as spatial feature parameters. Based on the solar activity cycle characteristics, construct indices with year, season, month and hour as time scales as time feature parameters.

[0014] (2) Correlation analysis was performed on the solar and geomagnetic activity characteristic parameters and time characteristic parameters within the above period with the historical TEC data of the same period and the historical TEC data one day later, and the characteristic parameters with strong correlation were selected. The selected strongly correlated characteristic parameters, the historical TEC data of the current point and the last 4 points were used as feature vectors, and the TEC value of the current point to be predicted for the next day was used as sample data; wherein, the characteristic parameter in each sample data is the historical data of the previous n days, and the value to be predicted is the TEC value of the (n+1)th day;

[0015] (3) Using the Transformer model self-attention encoding mechanism as the unit of hours, the feature vectors of the previous n days, a total of T = 24*n time points, are encoded. The query vector, key vector and value vector of each feature parameter in each group of feature vectors are calculated. The self-attention weight vector of the corresponding position of each feature vector is calculated according to the query vector and key vector. Finally, for each feature vector, the self-attention encoding process is completed by weighted summation between the weight vector and the value vector, and then by 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 coding layer is used as the input information of the decoding layer. A linear transformation is used to convert the input data into 24 TEC prediction values ​​to be predicted, resulting in a vector containing 24 elements.

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

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

[0019] (1.2) Using an 11-year cycle, the years of the 23rd and 24th solar activity cycles are labeled as FY = {1, 2, ..., 11, 1, 2, ..., 11}, and so on for the remaining years. Then, the following cycle conversion formula is used to convert FY into the annual characteristic parameters FYS and FYC:

[0020]

[0021] (1.3) Using the total number of days in a year, T, as the period, where T = 365 in a common year and T = 366 in a leap year, the accumulated days of the year (DOY) are constructed as characteristic parameters FDS and FDC according to the sine and cosine formulas:

[0022]

[0023] In the formula, DOY = 1, 2, ..., T;

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

[0025]

[0026] In the formula, HOD = 0, 1, 2, ..., 23;

[0027] (1.5) Take the historical TEC data of a grid point in a low-latitude region during periods of strong solar and geomagnetic activity, and perform correlation calculations with the solar and geomagnetic characteristic parameters SSN, F10.7, Kp, ap, AE and Dst, as well as the time-scale characteristic parameters FYS, FYC, FDS, FDC, FHS and FHC. At the same time, take the historical TEC data of the same time length after a one-day interval and perform correlation calculations 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 four most recent grid points TEC1, TEC2, TEC3 and TEC4 as spatial correlation feature parameters.

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

[0030] (2.1) Assume the feature vector formed by the filtered feature parameters is as follows, with the feature parameters of the i-th hour of the previous n days as the unit, construct 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] Where i = 1, 2, ..., 24*n;

[0033] (2.2) Construct an eigenvector F = {f1, f2, ..., fn} from all the feature parameters of the previous n days. n};

[0034] (2.3) The predicted TEC value of the 24-hour period corresponding to the feature vector of the day before yesterday 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 includes:

[0036] (3.1) The eigenvectors f1, f2, ... f in the eigenvector F i ...f N Where i = 1, 2, ..., N, N = 24*n parallel inputs, and a self-attention mechanism is used to compute the query vector q corresponding to the input vector.i Key vector k i Sum vector v i :

[0037]

[0038] (3.2) Calculate the input vector f at each position i The weights of the query vector are calculated. Taking feature vector f1 as an example, the attention weights of positions j, j = 1, 2, ..., N, on f1 are calculated as follows:

[0039]

[0040] In the formula, d k The dimension of the key matrix;

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

[0042]

[0043] (3.4) Similarly, for the input vectors f2, f3, ..., f N Based on steps (2.2) and (2.3), the corresponding feature information is extracted and combined to obtain 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 single-head self-attention mechanism calculation in step (3.4), a multi-head self-attention mechanism is adopted to convert all feature vectors obtained in step (3.1) into the query matrix Q = {q1,q2,...,q...}. 120}, the bond matrix K = {k1,k2,...,k 120 The sum matrix V = {v1, v2, ..., v} 120 The query matrix Q is divided into h equal parts, resulting in h query matrices. m Key matrix K m Sum matrix V m (m = 1, 2, ..., h);

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

[0047]

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

[0049] (3.7) Merge the attention score matrices of h groups and initialize the linear transformation matrix W. O We obtain 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, resulting in the self-attention encoding output matrix Z, with dimensions N×T×S, where S is the number of feature elements and N is the number of samples, i.e., the batch size predicted by the model for each batch.

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

[0054] (4.1) The output matrix Z of each batch of encoded parts obtained in step (3.8) is used as the input matrix of the decoding part, and a linear transformation is used to convert it into the prediction result TEC. pred :

[0055]

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

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] This invention proposes a spatiotemporal prediction method for TEC (Thermal Activity Chronicity) that integrates feature analysis and a self-attention mechanism. Besides extracting solar and geomagnetic activity indices closely related to TEC changes as feature parameters, the method evaluates the correlation between each parameter and TEC sequences within the same time period and TEC sequences one day later, filters strongly correlated features, and constructs a TEC sequence of inverse-distance adjacent grid points considering spatial correlation as feature quantities. Combined with historical TEC data, a multi-spatial-temporal feature matrix at different time scales globally is established to improve the richness of the sample data. Next, in the Transformer model encoding stage, a self-attention mechanism is used to extract features from the spatiotemporal information of TEC, fully capturing the temporal periodicity and spatial correlation of long-term TEC sequences. Finally, a linear transformation is applied in the decoding layer, using the self-attention output matrix as the input to obtain the predicted TEC values ​​for the next 24 hours. This method employs feature selection and a parallel computing structure for the model, significantly improving the model training and prediction speed, and has unique advantages in processing solar activity-related data with an 11-year cycle. Meanwhile, the model's multi-head self-attention mechanism can map feature information to different dimensional spaces, enabling full extraction of TEC feature data and helping to improve the model's learning effect. Therefore, compared with existing prediction models, it can effectively improve the TEC prediction accuracy in different regions, which is of great significance for reducing the impact of ionospheric disturbances on satellite navigation system positioning errors. Attached Figure Description

[0059] Figure 1 This is a flowchart of the TEC spatiotemporal prediction process that integrates feature analysis and self-attention mechanism according to the present invention.

[0060] Figure 2 This is a flowchart of the Transformer self-attention encoding TEC prediction process based on spatiotemporal information fusion of the present invention. Detailed Implementation

[0061] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

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

[0063] Step 1: Extract solar and geomagnetic activity parameters, and use the historical TEC data of the current grid point and the four most recent grid points as spatial feature parameters. Based on the characteristics of solar activity cycles, construct indices with time scales of year, season, month, and hour as time feature parameters.

[0064] Specifically, the following steps are included:

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

[0066] (1.2) Using an 11-year cycle, the years of the 23rd and 24th solar activity cycles (1997 to 2018) are labeled as FY = {1, 2, ..., 11, 1, 2, ..., 11}, and so on for the remaining years. Then, the following cycle conversion formulas are used to convert FY into the annual characteristic parameters FYS and FYC:

[0067]

[0068] (1.3) Using the total number of days in a year, T, as the period, where T = 365 in a common year and T = 366 in a leap year, the accumulated days of the year (DOY) are constructed as characteristic parameters FDS and FDC according to the sine and cosine formulas:

[0069]

[0070] In the formula, DOY = 1, 2, ..., T;

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

[0072]

[0073] In the formula, HOD = 0, 1, 2, ..., 23;

[0074] (1.5) Take the historical TEC data of a grid point in a low-latitude region during periods of strong solar and geomagnetic activity, and perform correlation calculations with the solar and geomagnetic characteristic parameters SSN, F10.7, Kp, ap, AE and Dst, as well as the time-scale characteristic parameters FYS, FYC, FDS, FDC, FHS and FHC. At the same time, take the historical TEC data of the same time length after a one-day interval and perform correlation calculations 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 four most recent 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 within the aforementioned period with historical TEC data for the same period and historical TEC data one day later, and select the characteristic parameters with strong correlations. Use the selected strongly correlated characteristic parameters, the current point, and the historical TEC data of the last four points as feature vectors, and the TEC value of the current point to be predicted for the next day as sample data; wherein, the characteristic parameter in each sample data is the historical data of the previous n days, and the value to be predicted is the TEC value of the (n+1)th day.

[0077] Specifically, the following steps are included:

[0078] (2.1) Assume the feature vector formed by the filtered feature parameters is as follows, with the feature parameters of the i-th hour of the previous n days as the unit, construct 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] Where i = 1, 2, ..., 24*n;

[0081] (2.2) Construct an eigenvector F = {f1, f2, ..., fn} from all the feature parameters of the previous n days. n};

[0082] (2.3) The predicted TEC value of the 24-hour period corresponding to the feature vector of the day before yesterday is constructed as the predicted label vector Y:

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

[0084] Step 3: Using the Transformer model's self-attention encoding mechanism, the feature vectors of the previous n days (T = 24*n) are encoded hourly. The query vector, key vector, and value vector for each feature parameter in each feature vector group are calculated. Based on the query vector and key vector, the self-attention weight vector corresponding to each feature vector is calculated. Finally, for each feature vector, the weighted sum of the weight vector and value vector is performed, followed by forward feedback and layer normalization to complete the self-attention encoding process and 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] Specifically, the following steps are included:

[0086] (3.1) The eigenvectors f1, f2, ... f in the eigenvector F i ...f N (Where i = 1, 2, ..., N, N = 24*n) Parallel input, using a self-attention mechanism to compute the query vector q corresponding to the input vector. i Key vector k i Sum vector v i :

[0087]

[0088] (3.2) Calculate the input vector f at each position i The weights of the query vector are calculated. Taking feature vector f1 as an example, the attention weights of position j (j=1,2,...,N) with respect to f1 are calculated as follows:

[0089]

[0090] In the formula, d k The dimension of the key matrix;

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

[0092]

[0093] (3.4) Similarly, for the input vectors f2, f3, ..., f N Based on steps (2.2) and (2.3), the corresponding feature information is extracted and combined to obtain 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 single-head self-attention mechanism calculation in step (3.4), a multi-head self-attention mechanism is adopted to convert all feature vectors obtained in step (3.1) into the query matrix Q = {q1,q2,...,q...}. 120}, the bond matrix K = {k1,k2,...,k 120 The sum matrix V = {v1, v2, ..., v} 120 The query matrix Q is divided into h equal parts, resulting in h query matrices. m Key matrix K m Sum matrix Vm (m = 1, 2, ..., h);

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

[0097]

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

[0099] (3.7) Merge the attention score matrices of h groups and initialize the linear transformation matrix W. O We obtain 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, resulting in the self-attention encoding output matrix Z, with dimensions N×T×S, where S is the number of feature elements and N is the number of samples, i.e., the batch size predicted by the model for each batch.

[0103] Step 4: Use the output of the self-attention coding 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, resulting in a vector containing 24 elements.

[0104] Specifically, the following steps are included:

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

[0106]

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

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A spatiotemporal prediction method for TEC that integrates feature analysis and self-attention mechanism, characterized by: It comprises the following steps: (1) Extract the solar and geomagnetic activity parameters, and the current grid point and the TEC history data of the nearest 4 grid points as spatial feature parameters, and construct the index with year, season, month and hour as time scale as time feature parameters according to the characteristics of solar activity cycle; (2) The solar and geomagnetic activity characteristic parameters and time characteristic parameters in the solar activity cycle time are respectively correlated with the TEC history data in the same time period and the TEC history data after one day interval, and the strongly correlated feature parameters are screened out, and the above screened strong correlation feature parameters, the current point and the TEC history data of the nearest 4 points are taken as the feature vector, and the future one day TEC value of the current point to be predicted is taken as the sample data; wherein the feature parameters in each sample data are the history data of the previous n days, and the predicted value is the TEC value of the n+1 day; (3) Taking hour as the unit, the feature vectors of T=24*n time points in the previous n days are encoded by using the self-attention encoding mechanism of the Transformer model, the query vector, the key vector and the value vector of each feature parameter in each feature vector are calculated, and the self-attention weight vector of the corresponding position of each feature vector is calculated according to 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, and the output dimension is N×T×S, wherein S is the number of feature elements, and N is the sample number; (4) The output result of the self-attention encoding layer is taken as the input information of the decoding layer, the input data is converted into the 24 TEC prediction values to be predicted by using linear transformation, and a vector containing 24 elements is obtained.

2. The TEC spatiotemporal prediction method of claim 1, wherein: Step (1) specifically comprises: (1.1) Extract the solar activity index SSN and F10.7, the geomagnetic activity intensity index Kp, ap, AE and Dst, and the TEC history data of the current grid point and the nearest 4 grid points as feature parameters; (1.2) The years of the 23rd and 24th solar cycles are marked as and so on, and then the following cycle conversion formula is used to convert into the annual characteristic parameters and : ； (1.3) with the total number of days T in a year as a period, wherein T = 365 for a common year and T = 366 for a leap year, the annual accumulated day is calculated according to the sine and cosine formula configured as a characteristic parameter and : ； In the formulae, ; (1.4) The hour of the day HOD is constructed as a characteristic parameter according to the sine and cosine formula for a 24-hour cycle and : ； In the formulae, ; (1.5) Take the TEC history data of a grid point in low latitude area in the period of strong solar and geomagnetic activity, respectively, 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; at the same time, take the TEC history data of the same time length after one day interval, respectively, and calculate the correlation with the solar and geomagnetic characteristic parameters and the time scale characteristic parameters, and screen out the feature parameters with higher correlation; (1.6) Extract the current grid point TEC0 and the nearest 4 grid points TEC1, TEC2, TEC3 and TEC4 as spatial correlation feature parameters.

3. The TEC spatiotemporal prediction method of claim 1, wherein: Step (2) specifically comprises: (2.1) Assuming that the feature vector composed of the feature parameters after the screening is as follows, the feature parameters of the first hour of the day are constructed as a feature vector :​​ ； wherein ; (2.2) the previous all feature parameters of the day are constructed as a feature vector ; (2.3) The 24-hour TEC prediction value of the day corresponding to the previous day's eigenvector is constructed as a prediction label vector : 。 4. The TEC spatiotemporal prediction method of claim 1, wherein: Step (3) specifically comprises: (3.1) The characteristic sub-vectors in the query vector , , , ... ... , , are input in parallel, and the query vector , key vector and value vector corresponding to the input vector are calculated by using a self-attention mechanism. ； (3.2) Calculate each position input vector weight of the query vector in the feature vector For example, calculate the position , Attention weight value of ​ ； In the formula, is the dimension of the bond matrix; (3.3) Computing self-attention mechanism on feature vectors Extracted feature information : ； (3.4) Similarly, for the input vector , the corresponding feature information is extracted according to steps (2.2) and (2.3), and the single-head self-attention score matrix is obtained by combination: ； (3.5) On the basis of the single-head self-attention mechanism calculation in step (3.4), a multi-head self-attention mechanism is used, and the query matrix , the key matrix and the value matrix obtained in step (3.1) are divided into equal parts to obtain h groups of query matrix , key matrix and value matrix , ; (3.6) Using initialized linearly varying matrices , and Map the query matrix , the key matrix and the value matrix to different spaces and perform steps (2.1) to (2.4) to obtain h sets of self-attention score matrices: ； wherein, the dimension is T x (S / h); (3.7) merge the h group attention score matrices and pass through an initialized linear transformation matrix obtain a multi-head self-attention score matrix of dimension T x S : ； ； (3.8) The self-attention score matrix described above The self-attention encoding process is completed again through forward feedback and layer normalization to obtain a self-attention encoding output matrix Z with a dimension of N x T x S, where S is the number of characteristic elements, and N is the number of samples, i.e., the model predicts the sample size of each batch .

5. The TEC spatiotemporal prediction method fused with feature analysis and self-attention mechanism according to claim 1, characterized in that: Step (4) specifically comprises: (4.1) The batch-wise coded portion output matrix obtained in step (3.8) is converted into a prediction result by using a linear transformation as a decoding portion input matrix As a decoding portion input matrix, it is converted into a prediction result by using a linear transformation : ； wherein, and is a linear transformation preset initial value for a single sample, is a 1 x (T*S) vector, is a (T*S) x 24 linear transformation matrix, is a 1 x 24 vector, and the above equation results in is the 24-hour TEC prediction value of the first day predicted by the proposed prediction model according to the previous day historical data. day historical data.

Images