Atmospheric electric field prediction method under transformer hybrid model

Abstract

This invention relates to the field of lightning detection technology, and more particularly to an atmospheric electric field prediction method based on a Transformer hybrid model. This invention solves the problem of temporal misalignment between radar image features and atmospheric electric field signals due to their asynchronous nature. The technical solution of this invention is as follows: Step S1: Determine the target area; extract radar echo intensity (REI) data corresponding to all radar image pixels within the target area; Step S2: Feature extraction and fusion: Based on the target area of ​​the radar image, design multiple artificial features, and perform deep feature extraction and multimodal feature fusion based on a convolutional neural network (CNN); Step S3: Design a cross-modal attention mechanism to achieve dynamic temporal alignment. The beneficial effects of this invention are: achieving spatial consistency between the radar image and the electric field meter's detection range, avoiding noise introduced by the mismatch in detection range, and improving the relevance and usability of the data.

Description

[0001] This invention relates to the field of lightning detection technology, and in particular to a method for predicting atmospheric electric fields using a Transformer hybrid model. Background Technology

[0002] Atmospheric electric field signal prediction is crucial for thunderstorm activity monitoring and early warning, agricultural production, and aviation safety. References: Li J, Zhao P, Yuan H, et al. Differential structure to improve performance of DC electric field sensors[J]. IEEE Sensors Journal, 2023, 23(9): 9909-9916. and Yang X, Xing H, Su X, et al. Entropy-based thunderstorm imaging system with real-time prediction and early warning[J]. IEEE Transactions on Instrumentation and Measurement, 2022, 73: 1-12. Traditional atmospheric electric field signal prediction methods based on machine learning and long short-term memory networks mainly rely on single atmospheric electric field data or single-architecture statistical models, as referenced in the papers: Callaghan KMS, Martinez-Hernandez U. Low-Cost, Multi-Sensor Non-Destructive Banana Ripeness Estimation Using Machine Learning[J]. IEEE Sensors Journal, 2025. and Hanyu E, Cui Y, Pedrycz W, et al. Strategies for designing machine learning models in renewable energy with insufficient data[J]. Energy, 2024, 311: 133475. and Yang X, Xing H, Xu W, et al. 3DAEFA-based thunderstorm prediction system with higher performance[J]. IEEE Sensors Journal, 2022, 22(23): 22865-22884., which make it difficult to capture the changing characteristics of complex atmospheric electric field signals.Furthermore, considering that radar images carry a wealth of thunderstorm activity information, rainfall intensity, thunderstorm area range, and future rainfall intensity characteristics can be obtained based on the echo intensity and region of the radar image. (See the papers: Feng K, Yu X, Wang M, et al. Multi-Function Radar Signal Sequence Prediction Via Trend Filtering and Similarity Comparison[J]. IEEE Transactions on Aerospace and Electronic Systems, 2025. and Feng X, Chen P, Weng Y, et al. CMDN: Continuous Human Activity Recognition Based on Multi-domain Radar Data Fusion[J]. IEEE SensorsJournal, 2025.). Therefore, the features carried by radar images can serve as an important data source for predicting atmospheric electric field signals.

[0003] However, using radar image features to predict atmospheric electric field signals faces three major problems:

[0004] First, the radar image area does not match the detection range of the atmospheric electric field meter. The radar image coverage area corresponding to the radar detection range is generally much larger than the electric field meter detection range (for example, the S-band radar detection range reaches 460 km, while the electric field meter detection range is usually 15-25 km). (See the paper: Qiu Z, Song Y, Chen W, et al. Dependence Mappingof the Static Electric Field Distribution Characteristics to Dielectric Strength of Long Air Gap Based on Random Forest[J]. IEEE Transactions on Dielectrics and Electrical Insulation, 2025). Therefore, it is necessary to ensure that the corresponding ranges of the two are consistent in spatial dimension.

[0005] Second, it is uncertain which radar image features to design and extract. Designing quantifiable and physically meaningful radar image features can enhance physical interpretability.

[0006] Third, the asynchronous nature of radar image features and atmospheric electric field signals leads to a misalignment in their timing, as illustrated in the paper: SuN, Chen X, Guan J, et al. Radar maritime target detection via spatial–temporal feature attention graph convolutional network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 1-15. This is because radar images are typically output every 6 minutes, while the sampling frequency of the atmospheric electric field is 1 Hz.

[0007] In summary, the key to improving the accuracy of atmospheric electric field prediction using radar image data lies in how to select radar image regions, extract radar image features, and process heterogeneous radar image features and electric field data.

[0008] How to solve the above problems is the problem that this invention aims to solve. Summary of the Invention

[0009] The purpose of this invention is to provide an atmospheric electric field prediction method under the Transformer hybrid model, which solves the problems of inconsistency between radar image area and atmospheric electric field instrument detection range, uncertainty in designing and extracting radar image features, and time misalignment between radar image features and atmospheric electric field signals due to their asynchronous nature. This method has the beneficial effect of improving the accuracy of atmospheric electric field prediction.

[0010] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0011] A method for predicting atmospheric electric fields using a Transformer mixture model is provided, including the following steps:

[0012] Step S1: Determine the target area; extract the radar echo intensity (REI) data corresponding to all radar image pixels within the target area.

[0013] Step S2: Feature Extraction and Fusion: Based on the target area of ​​the radar image, five artificial features with clear physical meanings are designed, and deep feature extraction and multimodal feature fusion are performed based on Convolutional Neural Network (CNN).

[0014] Step S3: Design a cross-modal attention mechanism to achieve dynamic temporal alignment.

[0015] In step S1, the longitude covered by the radar image measured by any radar station and latitude The range is shown in equation (1):

[0016] (1)

[0017] In equation (1), and They represent longitude respectively. The minimum and maximum values; and Representing latitude The minimum and maximum values;

[0018] Based on the radar image, longitude Divide into P parts, and divide the latitude Divided into Q parts, forming a uniform grid, with a resolution in the longitude direction. Resolution in the latitudinal direction for:

[0019] (2)

[0020] The pixel coordinates corresponding to the selected target region location are: At latitude At this location, the actual ground distance corresponding to each pixel in longitude and latitude is... and ;

[0021] For any radar image pixel (m,n), its distance to the station to be measured is... The Euclidean distance d(m,n) is:

[0022] (3)

[0023] The selection criterion for the target region is that the distance does not exceed D, thus obtaining the set. :

[0024] (4)

[0025] REI dataset extracted from the target region for:

[0026] (5)

[0027] Let m be the pixel at time t. t ,n t The REI value of ), where T is the time series set;

[0028] Where: m t n tLet be the x and y coordinates of the pixel (m, n) at time t.

[0029] The first artificial feature in step S2, the regional average intensity AV(t) (AV, Average Value), reflects the overall energy level of the target region.

[0030] (6)

[0031] The second artificial feature: Maximum Value (MV) characterizes the intensity of the strongest echo core, indicating strong local convection.

[0032] (7)

[0033] The third artificial feature: the standard deviation of echo intensity SD(t) (SD, Standard Deviation) describes the spatial dispersion of echo intensity.

[0034] (8)

[0035] Fourth artificial feature: Strong echo ratio (StrongRatio(t)) quantifies the range of strong convection activity: Intensity ≥ 5 dBZ.

[0036] (9)

[0037] Fifth artificial feature: Correlation coefficient (Corr, Correlation) between adjacent frames captures the dynamic evolution characteristics of thunderstorm systems.

[0038] (10)

[0039] The five features constitute the artificial feature vector. CNN is used to extract spatial features from the target area of ​​the radar image;

[0040] The input is the cropped spatial map. H and W represent the height and width of the spatial map, respectively, in pixels. This spatial map is processed by a three-layer convolutional network, each layer containing convolution, activation functions, and pooling operations.

[0041] The output of the global average pooling layer is a 3D deep feature vector :

[0042] (11)

[0043] Where CNN(·) represents the forward propagation process of a convolutional neural network, The feature dimension of the final output of the network;

[0044] Artificial feature vectors With deep feature vectors By splicing and merging, a fused feature vector is formed. :

[0045] (12)

[0046] Where [·; ·] denotes vector concatenation operation; This represents the total dimension of the fused feature vectors.

[0047] This fusion strategy combines the physical interpretability of artificial features with the spatial representation capabilities of CNN features to form complementary multimodal feature representations.

[0048] The cross-modal attention mechanism includes a feature encoding layer:

[0049] Atmospheric electric field and radar image features are uniformly encoded; radar image features are transformed into a hidden space through linear projection. :

[0050] (13)

[0051] in, and These are the weight matrix and bias vector for the linear projection, respectively. This represents the projected radar characteristics. The hidden layer dimension of the Transformer model;

[0052] By analyzing high-frequency atmospheric electric field signal sequences Inject location information and encode it as temporal features:

[0053] (14)

[0054] Where PE(·) represents the position encoding function, used to add position information to each time step in the sequence. The atmospheric electric field characteristics are represented by T, which incorporates location information, and T is the sequence length.

[0055] The cross-modal attention mechanism includes cross-modal attention:

[0056] Based on the encoded features, an asymmetric attention mechanism is constructed.

[0057] Using high-frequency atmospheric electric field characteristics as the query (Q, Query), and low-frequency radar characteristics as the key (K, Key) and value (V, Value), the electric field signal can dynamically query relevant information from radar characteristics.

[0058] (15)

[0059] in, For querying the matrix, To query the projection weights; For querying the matrix, To query the projection weights; For querying the matrix, To query the projection weights; and These are the dimensions of key K / query Q and value V, respectively; This is the output of the cross-modal attention mechanism.

[0060] The cross-modal attention mechanism includes output normalization:

[0061] Features aligned through residual connections and layer normalization outputs:

[0062] (16)

[0063] in, Presentation layer normalization operation, The fused features are after cross-modal alignment;

[0064] This operation preserves the temporal characteristics of the original atmospheric electric field signal while incorporating spatial information from the radar image.

[0065] Multimodal features after time alignment This process is used to capture the multi-scale temporal dependencies within the data.

[0066] Multimodal features after time alignment An improved temporal convolutional network, TCN (Temporal Convolutional Networks), is used to capture its internal multi-scale temporal dependencies.

[0067] Improved temporal convolutional networks: including dilated convolutional layers, feature fusion, and prediction output;

[0068] The dilated convolutional layers: TCN contains 8 dilated convolutional layers, the first... The calculation process for the layer is as follows:

[0069] (17)

[0070] in, For layer index ( = 1, 2, ..., L). For the first The input features of the layer For the first The convolutional kernel weights of the layer, This represents a causal convolution operation, ensuring that the output at time t depends only on the input at time t and before. It is a non-linear activation function. For the first Output characteristics of the layer.

[0071] Feature fusion: The outputs of each layer are fused through residual connection and splicing.

[0072] (18)

[0073] (19)

[0074] in, It is the first The output of the layer residual connection, [·] represents the splicing operation along the feature channel dimension. It is the final feature that integrates multi-scale temporal information;

[0075] The prediction result is output as follows: The atmospheric electric field signal prediction result is achieved through a fully connected layer:

[0076] (20)

[0077] in, and These are the weights and biases of the output layer, respectively. This represents the atmospheric electric field value predicted by the model for future moments.

[0078] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0079] (1) Multimodal feature fusion enhances feature representation ability

[0080] Traditional methods often rely on a single data source, making it difficult to comprehensively capture the complex characteristics of thunderstorm activity. This invention innovatively designs five artificial features with clear physical meaning (such as regional average intensity, maximum intensity, and standard deviation), and combines them with deep spatial features extracted by CNNs, forming a complementary multimodal feature representation through splicing and fusion. This strategy not only improves the physical interpretability of the features but also enhances the model's comprehensive perception of spatial structure and intensity distribution in radar images, providing richer and more discriminative input for subsequent predictions.

[0081] 2. Cross-modal attention mechanism for dynamic temporal alignment

[0082] To address the temporal misalignment between radar images (low frequency) and atmospheric electric field signals (high frequency), this invention proposes an asymmetric cross-modal attention mechanism. This mechanism uses the high-frequency electric field signal as a query to dynamically retrieve relevant information from low-frequency radar features. This adaptively matches the temporal scale differences between the two, effectively solving the information fusion difficulties caused by inconsistent sampling frequencies in traditional methods, and significantly improving the accuracy and robustness of multi-source data fusion.

[0083] 3. A hybrid Transformer+TCN architecture captures multi-scale temporal dependencies.

[0084] In terms of temporal modeling, this invention integrates the global attention mechanism of the Transformer with an improved Temporal Convolutional Network (TCN). The Transformer is responsible for capturing long-range dependencies, while the TCN captures multi-scale local temporal patterns through dilated convolutions and residual connections. This hybrid architecture combines global perception with local detail modeling capabilities, enabling a more accurate characterization of the dynamic evolution of atmospheric electric field signals, thereby improving the timeliness and accuracy of predictions.

[0085] 4. Spatial range adaptation improves data consistency

[0086] By setting a target area centered on the electric field meter and filtering effective radar image pixels based on Euclidean distance, this invention achieves spatial consistency between the radar image and the detection range of the electric field meter, avoiding noise introduced by mismatch in detection range and improving the relevance and usability of the data.

[0087] 5. Significant practical application value

[0088] This invention has broad application prospects in thunderstorm activity monitoring, agricultural disaster prevention, and aviation safety. By improving the accuracy and timeliness of atmospheric electric field prediction, it can provide more reliable technical support for lightning early warning, and has significant social and economic value.

[0089] In summary, this patent solves key challenges in atmospheric electric field prediction through innovations in multimodal feature fusion, cross-modal attention alignment, and hybrid temporal modeling, demonstrating strong technological advancement and practical application value. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail, taking into account the thunderstorm activity characteristics within a circular area with a radius of 15 km centered on the Nanjing University of Information Science and Technology (NUIST) site. Of course, the specific embodiments described herein are merely illustrative of the invention and are not intended to limit it.

[0091] Example 1

[0092] The atmospheric electric field prediction method under the Transformer mixture model includes the following steps:

[0093] Step S1: Determine the target area; extract the radar echo intensity (REI) data corresponding to all radar image pixels within the target area.

[0094] Step S2: Feature Extraction and Fusion: Based on the target area of ​​the radar image, five artificial features with clear physical meanings are designed, and deep feature extraction and multimodal feature fusion are performed based on Convolutional Neural Network (CNN).

[0095] Step S3: Design a cross-modal attention mechanism to achieve dynamic temporal alignment.

[0096] In this embodiment, the radar echo intensity (REI) data corresponding to all radar image pixels in the target area are extracted from the thunderstorm activity characteristics within a circular area with a radius of 15km (the detection range of the atmospheric electric field meter) centered on the Nanjing University of Information Science and Technology (NUIST) site.

[0097] In step S1, the longitude covered by the radar image measured by any radar station and latitude The range is shown in equation (1):

[0098] (1)

[0099] In equation (1), and They represent longitude respectively. The minimum and maximum values; and Representing latitude The minimum and maximum values;

[0100] In this embodiment, the latitude and longitude range covered by the radar image measured by the Nanjing radar station is selected.

[0101]

[0102] Based on the radar image, longitude Divide into P parts, and divide the latitude Divided into Q parts, forming a uniform grid, with a resolution in the longitude direction. Resolution in the latitudinal direction for:

[0103] (2)

[0104] In this embodiment, longitude is... Divided into 560 parts, latitude Divide into 480 parts, and substitute the data from formula (1) into...

[0105] The pixel coordinates corresponding to the selected target region location are: At latitude At this location, the actual ground distance corresponding to each pixel in both longitude (x-direction) and latitude (y-direction) is: and ;

[0106] For any radar image pixel (m,n), its distance to the station to be measured is... The Euclidean distance d(m,n) is:

[0107] (3)

[0108] In this embodiment, the Nanjing University of Information Science and Technology site location is selected. Corresponding pixel coordinates =(153,240); at latitude At 32.2066°, the actual distances corresponding to pixels in each direction are:

[0109]

[0110] any pixel The Euclidean distance d to pixel (153, 240) is:

[0111]

[0112] The selection criterion for the target region is that the distance does not exceed D, thus obtaining the set. :

[0113] (4)

[0114] In this embodiment, (5)

[0115] The search scope is as follows:

[0116]

[0117] REI dataset extracted from the target region for:

[0118] (5)

[0119] Let m be the pixel at time t. t ,n t The REI value of ), where T is the time series set;

[0120] Where: m t n t Let be the x and y coordinates of the pixel (m, n) at time t.

[0121] In step S2:

[0122] First artificial characteristic: Regional average intensity AV(t) (AV, Average Value) reflects the overall energy level of the target area.

[0123] (6)

[0124] The second artificial feature: Maximum Value (MV) characterizes the intensity of the strongest echo core, indicating strong local convection.

[0125] (7)

[0126] The third artificial feature: the standard deviation of echo intensity SD(t) (SD, Standard Deviation) describes the spatial dispersion of echo intensity.

[0127] (8)

[0128] Fourth artificial feature: Strong echo ratio (StrongRatio(t)) quantifies the range of strong convection activity: Intensity ≥ 5 dBZ.

[0129] (9)

[0130] Fifth artificial feature: Correlation coefficient (Corr, Correlation) between adjacent frames captures the dynamic evolution characteristics of thunderstorm systems.

[0131] (10)

[0132] The five features constitute the artificial feature vector. CNN is used to extract spatial features from the target area of ​​the radar image;

[0133] The input is the cropped spatial map. H and W represent the height and width of the spatial map, respectively, in pixels. This spatial map is processed by a three-layer convolutional network, each layer containing convolution, activation functions, and pooling operations.

[0134] The output of the global average pooling layer is a 3D deep feature vector :

[0135] (11)

[0136] Where CNN(·) represents the forward propagation process of a convolutional neural network, The feature dimension of the final output of the network;

[0137] Artificial feature vectors With deep feature vectors By splicing and merging, a fused feature vector is formed. :

[0138] (12)

[0139] Where [·; ·] denotes vector concatenation operation; This represents the total dimension of the fused feature vectors.

[0140] This fusion strategy combines the physical interpretability of artificial features with the spatial representation capabilities of CNN features to form complementary multimodal feature representations.

[0141] The cross-modal attention mechanism includes a feature encoding layer:

[0142] Atmospheric electric field and radar image features are uniformly encoded; radar image features are transformed into a hidden space through linear projection. :

[0143] (13)

[0144] in, and These are the weight matrix and bias vector for the linear projection, respectively. This represents the projected radar characteristics. The hidden layer dimension of the Transformer model;

[0145] By analyzing high-frequency atmospheric electric field signal sequences Inject location information and encode it as temporal features:

[0146] (14)

[0147] Where PE(·) represents the position encoding function, used to add position information to each time step in the sequence. The atmospheric electric field characteristics are represented by T, which incorporates location information, and T is the sequence length.

[0148] The cross-modal attention mechanism includes cross-modal attention:

[0149] Based on the encoded features, an asymmetric attention mechanism is constructed.

[0150] Using high-frequency atmospheric electric field characteristics as the query (Q, Query), and low-frequency radar characteristics as the key (K, Key) and value (V, Value), the electric field signal can dynamically query relevant information from radar characteristics.

[0151] (15)

[0152] in, For querying the matrix, To query the projection weights; For querying the matrix, To query the projection weights; For querying the matrix, To query the projection weights; and These are the dimensions of key K / query Q and value V, respectively; This is the output of the cross-modal attention mechanism.

[0153] The cross-modal attention mechanism includes output normalization:

[0154] Features aligned through residual connections and layer normalization outputs:

[0155] (16)

[0156] in, Presentation layer normalization operation, The fused features are after cross-modal alignment;

[0157] This operation preserves the temporal characteristics of the original atmospheric electric field signal while incorporating spatial information from the radar image.

[0158] Multimodal features after time alignment This process is used to capture the multi-scale temporal dependencies within the data.

[0159] Multimodal features after time alignment An improved temporal convolutional network, TCN (Temporal Convolutional Networks), is used to capture its internal multi-scale temporal dependencies.

[0160] Improved temporal convolutional networks: including dilated convolutional layers, feature fusion, and prediction output;

[0161] The dilated convolutional layers: TCN contains 8 dilated convolutional layers, the first... The calculation process for the layer is as follows:

[0162] (17)

[0163] in, For layer index ( = 1, 2, ..., L). For the first The input features of the layer For the first The convolutional kernel weights of the layer, This represents a causal convolution operation, ensuring that the output at time t depends only on the input at time t and before. It is a non-linear activation function. For the first Output characteristics of the layer.

[0164] Feature fusion: The outputs of each layer are fused through residual connection and splicing.

[0165] (18)

[0166] (19)

[0167] in, It is the first The output of the layer residual connection, [·] represents the splicing operation along the feature channel dimension. It is the final feature that integrates multi-scale temporal information;

[0168] The prediction result is output as follows: The atmospheric electric field signal prediction result is achieved through a fully connected layer:

[0169] (20)

[0170] in, and These are the weights and biases of the output layer, respectively. This represents the atmospheric electric field value predicted by the model for future moments.

[0171] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting atmospheric electric fields using a Transformer hybrid model, characterized in that, Includes the following steps: Step S1: Determine the target area; extract the radar echo intensity (REI) data corresponding to all radar image pixels within the target area; Step S2: Feature Extraction and Fusion: Based on the target region of the radar image, design multiple artificial features and perform deep feature extraction and multimodal feature fusion based on a convolutional neural network (CNN). Step S3: Design a cross-modal attention mechanism to achieve dynamic temporal alignment; The pixel coordinates corresponding to the selected target region location are: At latitude At this location, the actual ground distance corresponding to each pixel in longitude and latitude is... and ; For any radar image pixel (m,n), its distance to the station to be measured is... The Euclidean distance d(m,n) is: （3） The selection criterion for the target region is that the distance does not exceed D, thus obtaining the set. : （4） REI dataset extracted from the target region for: （5） Let m be the pixel at time t. t ,n t The REI value of ), where T is the time series set; Where: m t n t Let (m, n) be the x and y coordinates of the pixel at time t. In step S2 First artificial characteristic: Regional average intensity AV(t) reflects the overall energy level of the target area. （6） The second artificial feature: the maximum intensity MA(t) characterizes the intensity of the strongest echo core, indicating strong local convection. （7） The third artificial feature: the standard deviation SD(t) of the echo intensity describes the spatial dispersion of the echo intensity. （8） Fourth artificial characteristic: Strong echo ratio (t) quantifies the range of strong convection activity: Intensity ≥ 5 dBZ: （9） Fifth artificial feature: Correlation coefficient Corr(t) between adjacent frames captures the dynamic evolution characteristics of thunderstorm systems. （10） The cross-modal attention mechanism includes a feature encoding layer: Atmospheric electric field and radar image features are uniformly encoded; radar image features are transformed into a hidden space through linear projection. : （13） in, and These are the weight matrix and bias vector for the linear projection, respectively. This represents the projected radar characteristics. The hidden layer dimension of the Transformer model; By analyzing high-frequency atmospheric electric field signal sequences Inject location information and encode it as temporal features: （14） Where PE(·) represents the position encoding function, used to add position information to each time step in the sequence. The atmospheric electric field feature representation incorporates location information, where T is the sequence length; The cross-modal attention mechanism includes cross-modal attention: Based on the encoded features, an asymmetric attention mechanism is constructed. Using high-frequency atmospheric electric field characteristics as the query: Q, and low-frequency radar characteristics as the key: K and value: V, the electric field signal can dynamically query relevant information in the radar characteristics; （15） in, For querying the matrix, To query the projected weights; For querying the matrix, To query the projected weights; For querying the matrix, To query the projected weights; and These are the dimensions of key K / query Q and value V, respectively; This is the output of the cross-modal attention mechanism; The cross-modal attention mechanism includes output normalization: Features aligned through residual connections and layer normalization outputs: （16） in, Presentation layer normalization operation, The fused features are after cross-modal alignment; Multimodal features after time alignment This process is used to capture the multi-scale temporal dependencies within the data.

2. The atmospheric electric field prediction method under the Transformer hybrid model according to claim 1, characterized in that, In step S1, the longitude covered by the radar image measured by any radar station and latitude The range is shown in equation (1): （1） In equation (1), and They represent longitude respectively. The minimum and maximum values; and Representing latitude The minimum and maximum values; Based on the radar image, longitude Divide into P parts, and divide the latitude Divided into Q parts, forming a uniform grid, with a resolution in the longitude direction. Resolution in the latitudinal direction for: （2）。 3. The atmospheric electric field prediction method under the Transformer hybrid model according to claim 1, characterized in that, The five artificial features constitute the artificial feature vector. CNN is used to extract spatial features from the target area of ​​the radar image; The input is the cropped spatial map. , where H and W are the height and width of the spatial map, respectively, in pixels; The spatial graph is processed by a three-layer convolutional network, each layer containing convolution, activation functions, and pooling operations. The output of the global average pooling layer is a 3D deep feature vector : （11） Where CNN(·) represents the forward propagation process of a convolutional neural network, The feature dimension of the final output of the network; Artificial feature vectors With deep feature vectors By splicing and merging, a fused feature vector is formed. : （12） Where [·; ·] denotes vector concatenation operation; This represents the total dimension of the fused feature vectors.

4. The atmospheric electric field prediction method under the Transformer hybrid model according to claim 1, characterized in that, Multimodal features after time alignment An improved temporal convolutional network, TCN, is employed to capture its internal multi-scale temporal dependencies.

5. The atmospheric electric field prediction method under the Transformer hybrid model according to claim 4, characterized in that, Improved temporal convolutional networks: This includes dilated convolutional layers, feature fusion, and the output of prediction results; The dilated convolutional layers: TCN contains 8 dilated convolutional layers, the first... The calculation process for the layer is as follows: （17） in, For layer index ( = 1, 2, ..., L). For the first The input features of the layer For the first The convolutional kernel weights of the layer, This represents a causal convolution operation, ensuring that the output at time t depends only on the input at time t and before. It is a non-linear activation function. For the first The output features of the layer; Feature fusion: The outputs of each layer are fused through residual connection and splicing. （18） （19） in: It is the first The output of the layer residual connection, [·] represents the splicing operation along the feature channel dimension. It is the final feature that integrates multi-scale temporal information; The prediction result is output as follows: The atmospheric electric field signal prediction result is achieved through a fully connected layer: （20） in, and These represent the weights and biases of the output layer, respectively. This represents the atmospheric electric field value predicted by the model for future moments.

Images