Ionospheric TEC spatio-temporal prediction method and device based on CBAM enhanced ED-ConvLSTM

Abstract

The application provides an ionospheric TEC space-time prediction method and equipment based on CBAM enhanced EDConvLSTM, belongs to the technical field of ionosphere, and comprises the following steps: integrating ionospheric TEC grid data, geomagnetic index and solar activity index, and preprocessing; constructing a multi-channel space-time grid data set; identifying a magnetic storm event based on the geomagnetic index, extracting data before a first preset time length before the magnetic storm event, during the magnetic storm and after a second preset time length, and constructing a magnetic storm exclusive data set focusing on magnetic storm disturbance; constructing a prediction model with an encoder-decoder structure ConvLSTM as the main part and embedding a CBAM module at a key position of the encoder; training and parameter optimizing the model by using the magnetic storm exclusive data set; inputting the preprocessed multi-channel data into the trained prediction model, and outputting ionospheric TEC space-time prediction results. The application can accurately capture the complex nonlinear space-time evolution characteristics of the ionosphere, and significantly improve the prediction accuracy and robustness under the extreme space weather of the magnetic storm.

Description

Technical Field

[0001] This invention belongs to the field of ionospheric technology, specifically relating to a spatiotemporal prediction method and system for total electron content (TEC) of the ionosphere based on an encoder-decoder convolutional long short-term memory network (ED-ConvLSTM) enhanced by convolutional block attention modules (CBAM), which is particularly suitable for accurate prediction of ionospheric disturbances during geomagnetic storms. Background Technology

[0002] Total electron content (TEC) in the ionosphere is a key parameter describing the state of the ionosphere, and its drastic changes (especially during geomagnetic storms) pose a serious threat to satellite navigation and communication systems. Modeling and predicting ionospheric errors is one of the important error sources in high-precision GNSS data processing. Existing TEC prediction methods, such as empirical models, mathematical statistics methods, and basic deep learning models (e.g., LSTM, CNN), are insufficient in capturing the complex nonlinear, multi-scale spatiotemporal evolution characteristics of ionospheric TEC, particularly in focusing on and extracting geomagnetic storm disturbance characteristics. This results in prediction accuracy and robustness that are insufficient to meet the requirements of high-precision applications.

[0003] Specific problems include: 1) Traditional models do not adequately model spatiotemporal correlations. Traditional models usually process spatial and temporal features separately, making it difficult to effectively capture the coupling evolution of ionospheric TEC in the spatiotemporal dimension; 2) They fail to effectively distinguish and reinforce the features of extreme space weather events such as geomagnetic storms. Models trained on general datasets mainly use data from calm periods, resulting in insufficient learning of the disturbance features of extreme space weather events such as geomagnetic storms, leading to a significant decrease in prediction performance; 3) They lack attention-focusing mechanisms for key spatiotemporal features. Existing models lack adaptive recognition capabilities for the importance of different channels (such as solar activity index and geomagnetic index) and different spatial regions, and cannot dynamically focus on the key features most related to geomagnetic storm disturbances, limiting the model's prediction accuracy and physical interpretability, resulting in low efficiency in utilizing important information. Summary of the Invention

[0004] The present invention mainly addresses the problem that existing methods are insufficient in capturing the complex spatiotemporal evolution patterns of the ionospheric TEC, especially the non-stationary and nonlinear characteristics under geomagnetic storm disturbances.

[0005] Another objective is to overcome the poor prediction performance of models trained on general datasets for extreme events such as geomagnetic storms, and to achieve targeted reinforcement learning. This will improve the model's ability to select and focus features on key channels (such as the solar activity index) and key spatial regions (such as areas of strong disturbance), thereby enhancing prediction accuracy and physical interpretability.

[0006] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions: A CBAM-enhanced ED The ConvLSTM spatiotemporal prediction method for ionospheric TEC is characterized by comprising: Integrate ionospheric TEC grid data, geomagnetic index, and solar activity index, and perform preprocessing to construct a multi-channel spatiotemporal grid dataset; Based on geomagnetic index identification of geomagnetic storm events, data on the first preset duration before the occurrence of a geomagnetic storm, during the geomagnetic storm, and the second preset duration after the end of the geomagnetic storm are extracted from a multi-channel spatiotemporal grid dataset to construct a geomagnetic storm-specific dataset focusing on geomagnetic storm disturbances; Build with encoder The decoder structure uses ConvLSTM as the backbone and embeds CBAM modules at key locations in the encoder as a prediction model. The model was trained and its parameters optimized using the aforementioned geostorm-specific dataset; The preprocessed multi-channel data is input into the trained prediction model, and the spatiotemporal prediction results of the ionosphere TEC are output.

[0007] Furthermore, the preprocessing includes outlier handling, spatiotemporal resolution unification, data normalization, and multi-channel data stitching, which expands the one-dimensional exponential data into two-dimensional data aligned with the TEC grid space.

[0008] Moreover, the construction of the geomagnetic storm-specific dataset involves collecting data from two days before the geomagnetic storm, during the geomagnetic storm, and one day after the geomagnetic storm for each geomagnetic storm event identified and extracted based on the Dst index within a complete solar activity cycle. This forms a complete sample event to ensure the capture of the full-process characteristics of the geomagnetic storm precursors, main phase, and recovery phase.

[0009] Moreover, the encoder of the prediction model adopts a cascaded structure of multi-level convolutional layers and convolutional long short-term memory layers, and the decoder adopts a transposed convolutional layer and convolutional long short-term memory layer structure that is symmetrical to the encoder. The CBAM module sequentially refines the features through channel attention and spatial attention.

[0010] Moreover, the CBAM module is embedded after each convolutional layer of the encoder and after the last convolutional long short-term memory layer to focus on key features related to geomagnetic storms.

[0011] Moreover, the model training uses an adaptive optimizer, which combines dropout regularization, dynamic learning rate adjustment and early stopping mechanism, and optimizes the model parameters with mean squared error as the loss function.

[0012] Moreover, the model outputs a multi-step TEC grid prediction sequence after cross-channel feature fusion, achieving end-to-end prediction.

[0013] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method as described above.

[0014] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method as described above.

[0015] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method as described above.

[0016] Compared with the prior art, the present invention has the following advantages: 1. Significantly Improved Prediction Accuracy: By introducing a dual mechanism of channel attention and spatial attention into the encoder-decoder structure, the model can dynamically focus on the input features and spatial locations most relevant to geomagnetic storm disturbances, effectively modeling spatiotemporal evolution. The combination of these two mechanisms greatly improves prediction accuracy in complex scenarios such as geomagnetic storms. The method of this invention outperforms the comparative models in terms of root mean square error (RMSE) and mean absolute error (MAE) during high, medium, and low solar activity years. Especially during high solar activity years with frequent geomagnetic storms, it significantly improves prediction accuracy compared to existing techniques such as ConvLSTM and COPG.

[0017] 2. Highly targeted and robust: By constructing a dedicated dataset for geomagnetic storm disturbance periods, including the precursor period (short-term background 48 hours before the geomagnetic storm), the main phase, and the recovery period, the model can learn the complete evolution process from calm to disturbance and then to recovery. At the same time, it eliminates the interference of long-term calm period data, significantly improving the adaptability and prediction stability to extreme space weather events.

[0018] 3. Highly efficient end-to-end forecasting: This method enables end-to-end forecasting from multi-source data input to TEC grid sequence output, with a high degree of automation and easy integration into operational forecasting systems.

[0019] 4. Systematic and representative data construction: This invention identifies and extracts geomagnetic storm event samples within a complete solar activity cycle, ensuring that the model training data covers geomagnetic storm disturbance characteristics of different intensities and types, avoiding overfitting or underfitting problems caused by solar activity phase deviations. Simultaneously, by retaining small geomagnetic storm samples, its disturbance characteristics are prevented from being smoothed, improving the model's generalization ability and predictive stability for various geomagnetic storm events.

[0020] 5. The present invention is simple and convenient to implement, highly practical, and solves the problems of low practicality and inconvenience in actual application of related technologies. It can improve user experience and has significant market value. The present invention can accurately capture the complex nonlinear spatiotemporal evolution characteristics of the ionosphere, significantly improve the prediction accuracy and robustness under extreme space weather conditions such as geomagnetic storms, and is suitable for operational forecasting of ionospheric disturbances in scenarios such as satellite navigation and space communication. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the data processing flow in an embodiment of the present invention.

[0022] Figure 2 This is a structural diagram of a convolutional neural network-long short-term memory neural network-attention mechanism neural network model according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the structure of the Convolutional Block Attention (CBAM) module according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram illustrating the fusion of the CBAM module and ED-ConvLSTM in an embodiment of the present invention. Detailed Implementation

[0025] The following will further explain the concept, specific structure and technical effects of the present invention in conjunction with the accompanying drawings and embodiments, so as to fully understand the purpose, features and effects of the present invention.

[0026] Example 1: This invention provides a spatiotemporal prediction method for ionospheric TEC based on CBAM-enhanced ED-ConvLSTM, comprising the following steps: 1. Data preparation and preprocessing: Integrating TEC grid data and various geomagnetic and solar activity indices (such as the magnetic storm ring current index Dst and the global interplanetary index Kp, and the 10.7 cm solar radio flux F10.7), a unified format multi-channel spatiotemporal grid dataset is constructed through outlier processing, spatiotemporal resolution unification, normalization and other methods.

[0027] The global TEC grid data (2-hour temporal resolution, 5°×2.5° spatial resolution) provided by IGS can be used, which is the vertical total electron content (VTEC).

[0028] 2. Construct a dedicated geomagnetic storm dataset: Automatically identify and classify geomagnetic storm levels based on the Dst index (preferably classified into three levels: small, medium, and large). Extract data from the multi-channel spatiotemporal grid dataset for the first preset duration before the geomagnetic storm event, during the geomagnetic storm, and the second preset duration after the geomagnetic storm event, to obtain complete time-period data including precursors, main phase, and recovery phase, and construct a dataset focused on perturbation feature learning.

[0029] In one possible embodiment, the construction of the geomagnetic storm-specific dataset involves selecting a complete solar activity cycle and collecting data for each geomagnetic storm event, including the two days before the geomagnetic storm, during the geomagnetic storm, and the one day after the geomagnetic storm, to form a complete sample event, thereby ensuring the capture of the full-process characteristics of the geomagnetic storm precursors, main phase, and recovery phase.

[0030] For each identified geomagnetic storm event (including minor, moderate, and major storms), data from the 48 hours before its occurrence (precursor period), during the storm (main phase), and the 24 hours after its end (recovery period) are extracted to form a complete geomagnetic storm sample. The precursor period uses 48 hours of calm time data immediately preceding the storm's occurrence to provide the ionospheric background state before the storm, enhancing the model's ability to perceive the abrupt transition from calm to disturbance.

[0031] All geomagnetic storm samples (including small, medium, and large geomagnetic storms) are combined into a dataset, which can be divided into training, validation, and test sets according to the proportions in the specific implementation.

[0032] It is worth noting that the extraction of small geomagnetic storm samples is also of significant technical importance. During model training, if only medium and large geomagnetic storm samples are used, the perturbation features of small geomagnetic storms will be smoothed out by the model due to insufficient sample size or dominance by strong geomagnetic storm samples, resulting in the model's inability to effectively identify and predict weak geomagnetic storm events. By including small geomagnetic storm samples in the dataset, the model can learn the perturbation features of geomagnetic storms of different intensities, improving the model's sensitivity and generalization ability to various geomagnetic storm events.

[0033] 3. Constructing a CBAM-enhanced ED-ConvLSTM model: Using an encoder-decoder structured ConvLSTM (ED-ConvLSTM) as the backbone network to capture spatiotemporal dependencies. Convolutional Block Attention (CBAM) modules are embedded at key locations in the encoder, enabling it to first perform channel attention to filter important feature channels, and then perform spatial attention to focus on key spatial regions, thereby adaptively enhancing the ability to extract features of geomagnetic storm disturbances. Through this design, the weighting mechanism for multi-channel data evolves from traditional empirical weighting to an adaptive process based on the dynamic evolution of input features. Specifically, a nonlinear recalibration of the feature map is achieved through a channel attention mechanism (CAM), establishing a flexible weight allocation logic between physical driving parameters and ionospheric response. That is, through shared multilayer perceptron (MLP) computation, the model generates a weight factor between 0 and 1 for each channel (Kp, Dst, F10.7), considered as a "channel influence index." This factor acts as a dynamic scale, weighting the original feature map channel by channel, thereby achieving real-time identification and adjustment of the importance of different physical driving sources.

[0034] In one possible embodiment, the encoder-decoder core structure of ED-ConvLSTM comprises four main modules: an input layer, an encoder, a decoder, and an output layer. The encoder, through a three-level cascaded structure of "Conv2D + ConvLSTM," progressively extracts local features and compresses spatial dimensions to capture temporal dynamics at different spatial granularities. The decoder, through a mirror structure of "ConvLSTM + DeConv2D," reconstructs temporal features and restores the original spatial resolution. Finally, cross-channel feature fusion is achieved through the Conv2D layer, outputting a single-channel TEC grid prediction result. The core is to enhance the extraction and reconstruction capabilities of multi-scale spatiotemporal features.

[0035] See Figure 2 The embodiment is an improvement on the overall architecture of the ED-ConvLSTM model, which includes four main modules: input layer, encoder, decoder, and output layer. The preferred implementation method is as follows: The input layer receives dimensions as follows: The multichannel tensor, in which To input the time step, and These are the space height and width, respectively. Number of input channels (in this embodiment) (Corresponding to the four channels: TEC, Dst, Kp, and F10.7).

[0036] The encoder employs a three-stage cascaded structure, with each stage consisting of a Conv2D layer and a ConvLSTM layer connected sequentially. The Conv2D layers progressively extract local spatial features and compress the spatial dimension, followed by the ConvLSTM layers capturing temporal dynamic information, thus obtaining spatiotemporal features at different spatial granularities. In each stage of the encoder, a Convolutional Block Attention (CBAM) module is embedded after the Conv2D layer for adaptive refinement of the features in terms of channel and spatial dimension. The feature map sizes output by each stage of the encoder are as follows: Output of the first ConvLSTM1 layer:

[0037] Output of the second-stage ConvLSTM2 layer:

[0038] Output of the third-level ConvLSTM3 layer:

[0039] in, , , The number of channels at each level, The spatial dimensions are defined for each level. As the coding level deepens, the spatial dimensions gradually decrease while the number of channels gradually increases, enabling the extraction of multi-scale spatiotemporal features.

[0040] The decoder is a symmetrical mirror image of the encoder, consisting of three cascaded units, each consisting of a DeConv2D layer and a ConvLSTM layer connected sequentially. Spatial resolution is gradually restored through the DeConv2D layer, and temporal features are reconstructed by the ConvLSTM layer, achieving step-by-step upsampling and reconstruction of the deep features extracted by the encoder.

[0041] The output layer achieves cross-channel feature fusion through a 1×1 Conv2D layer, with an output dimension of [missing value]. The future TEC grid prediction sequence, in which To predict the time step.

[0042] In one possible embodiment, the ED-ConvLSTM model structure includes: 1) Input layer: receiving dimension is × × × × A tensor, where B is the batch size. The input time step is H, and the input space height and width are H and W, respectively. =4 input channels; 2) Encoder: It consists of multi-level cascaded units, each of which contains a convolutional layer Conv2D and a convolutional long short-term memory layer ConvLSTM, used to extract and compress multi-scale spatiotemporal features step by step; Attention mechanism: A CBAM module is embedded after each Conv2D layer and after the last ConvLSTM layer in the encoder. The CBAM module is composed of a channel attention submodule (CAM) and a spatial attention submodule (SAM) in sequence. 3) Decoder: As a symmetrical structure of the encoder, it consists of multiple ConvLSTM layers and transposed convolutional layers DeConv2D, used to progressively reconstruct temporal dynamics and restore spatial resolution; 4) Output layer: Cross-channel feature fusion is achieved by a Conv2D layer with a 1×1 convolutional kernel, and the output dimension is... × × × ×1 future TEC prediction sequence.

[0043] The CBAM module achieves dual refinement of the input feature map through the sequential cascading of channel attention and spatial attention submodules: the channel attention submodule adaptively adjusts the feature weights of different channels to highlight key channels relevant to the prediction task; the spatial attention submodule adaptively focuses on key regions in the feature map that change significantly with dynamics. The final refined feature map output simultaneously enhances the feature responses of important channels and key spatial regions.

[0044] In the CBAM module, the channel attention submodule first adaptively adjusts the channel feature weights based on global max pooling and global average pooling to highlight key channel information. Then, the spatial attention submodule focuses on key regions with significant dynamic changes in spatiotemporal features. After optimization by the dual attention mechanism, irrelevant information is suppressed and core features are amplified, providing more targeted feature representations for subsequent predictions.

[0045] The CBAM module is embedded after each Conv2D layer and the last ConvLSTM layer of the encoder. Specifically, CBAM is embedded after the first and second Conv2D layers of the encoder to refine shallow local spatial features; CBAM is embedded after the third Conv2D layer and subsequent ConvLSTM layers of the encoder to further enhance key information in deep features. By embedding CBAM at these specific locations in the encoder, the model can dynamically focus on key channels and key spatial regions related to geomagnetic storm disturbances at various stages of feature extraction.

[0046] 4. Model Training and Optimization: The model is trained using a geomagnetic storm-specific dataset. A strategy combining Dropout, dynamic learning rate adjustment, and early stopping is employed to optimize model parameters and prevent overfitting.

[0047] In one possible implementation, the model training employs the following strategy: 1) Optimizer: The Adam optimizer is used, with an initial learning rate set to 0.0002 to 0.0003; 2) Regularization: Insert a Dropout layer after each ConvLSTM layer, with a dropout rate set to 0.2; 3) Learning rate scheduling: The ReduceLROnPlateau strategy is adopted. When the validation loss does not decrease for 5 consecutive epochs, the learning rate is reduced to 0.5 to 0.6 times the original value. 4) Early stopping mechanism: The EarlyStopping strategy is adopted. When the validation loss has not improved for 10 to 12 consecutive epochs, the training is terminated early and the optimal model weights are saved.

[0048] 5. TEC Prediction: Input the preprocessed real-time or historical multi-channel data into the trained model, which can directly output global or regional TEC grid prediction sequences for multiple future time steps. In practice, the data preprocessing method in this step is the same as that in step 1.

[0049] Example 2: An example of an ED-ConvLSTM ionospheric TEC spatiotemporal prediction method based on CBAM enhancement is provided, comprising the following steps: Step S1: Data Preparation and Preprocessing 1. Data source: Global TEC grid data (2-hour time resolution, 5°×2.5° spatial resolution) provided by IGS was used, and Dst, Kp and F10.7 indices were collected simultaneously for the same time period.

[0050] 2. Outlier Handling: Outliers / missing values ​​marked in the TEC and F10.7 indices are imputed using linear interpolation along the time dimension. In practice, outlier markers from different data source files may be "999" or "9999". 3. Spatiotemporal Resolution Standardization: All data is standardized to a 2-hour time resolution. The one-dimensional Dst, Kp, and F10.7 indices are spatially expanded into a two-dimensional grid with the same size as the TEC grid (e.g., 73×72), forming spatially aligned multi-channel data.

[0051] 4. Data Normalization: Perform Min-Max normalization for each channel (TEC, Dst, Kp, F10.7) to eliminate the influence of dimensions. The formula is as follows: Xnorm = (X - Xmin) / (Xmax - Xmin) in, X represents the raw data, i.e. the channel data to be normalized (such as the measured values ​​of TEC, Dst, Kp, or F10.7). Xmin is the minimum value of this channel in the dataset; Xmax is the maximum value of this channel in the dataset; Xnorm represents the normalized data, with a value range of [value range missing]. .

[0052] 5. Construct a five-dimensional dataset: Concatenate the processed data from the four channels along the channel dimensions, construct samples according to a sliding time window (e.g., the past 48 hours as input), and finally form a dataset with dimensions (number of samples, input time step, height, width, number of channels), for example (number of samples, input time step T_in, H, W, 4).

[0053] Step S2: Construct a dedicated dataset for geomagnetic storms First, a complete solar activity cycle is selected. Within this timeframe, geomagnetic storm events are automatically identified based on the Dst index threshold, and geomagnetic storm levels are classified according to the following criteria: Minor geomagnetic storm: -50 nT <Dst ≤ -30 nT Moderate geomagnetic storm: -100 nT <Dst ≤ -50 nT Major geomagnetic storm: -200 nT <Dst ≤ -100 nT For each identified geomagnetic storm event, data from the 48 hours before its occurrence (precursor period), during the storm (main phase), and the 24 hours after its end (recovery period) are extracted to form a complete geomagnetic storm sample. The precursor period uses 48 hours of calm time data immediately preceding the storm to provide the ionospheric background state before the storm, enhancing the model's ability to perceive the abrupt transition from calm to disturbance. All geomagnetic storm samples are combined into a dataset, which can be proportionally divided into training, validation, and test sets in practice.

[0054] The main purpose of this step is to construct geomagnetic storm samples. Since geomagnetic storms are infrequent, to avoid excessively long precursor periods (such as data exceeding two days) affecting the construction of these samples, while ensuring the data captures the geomagnetic storm phenomenon, a 48-hour precursor period is preferred. The dataset of this invention focuses on the geomagnetic storm disturbance period (including necessary short-term precursor background, main phase, and recovery phase), excluding longer calm periods. This avoids diluting the learning of disturbance features with a large amount of invalid calm data, ensuring the model focuses on learning key features in the geomagnetic storm triggering and evolution process.

[0055] In one possible implementation, the sample generation logic based on multi-channel TEC grid data constructs the model's input-output sample pairs using a sliding window method. To capture the weight evolution patterns before and after a geomagnetic storm, three representative observation windows were selected: the Quiet Phase, the Main Phase, and the Recovery Phase. By averaging the weight vectors of all samples within each time window to eliminate random noise interference, the "average focus of attention" of the model at different ionospheric response stages is obtained. For example, during the ionospheric quiet phase, the weight of the geomagnetic index channel is automatically reduced, and the model mainly relies on historical TEC sequences to extract seasonal and diurnal variation patterns. Once a geomagnetic storm occurs, it automatically identifies the channel based on the fluctuation patterns of the Dst or Kp indices. If these indices reach a maximum value, the CAM module can keenly capture the jumps in these channels' global max-pooling characteristics, thereby instantaneously and significantly increasing the weight of the geomagnetic index channel. See also Figure 1The embodiment provides a process for converting raw multi-channel time-series data into input-output sample pairs required for model training using the sliding window method. "Multi-channel TEC grid data" refers to a multi-channel TEC grid data sequence, with data arranged chronologically and time points labeled 1, 2, ..., T-3, T-2, T-1, T, etc., where T is the current time and Tn is the nth past time. The sliding window diagram illustrates a fixed-length window sliding along the time axis, each time capturing a continuous segment of historical data as an input sample and using the data from the next consecutive time point as the corresponding output sample. The example in the figure shows the generation of several samples: Sample 1 and Sample 2…Sample T-3. Each sample consists of an input part and an output part. For example, the input of Sample 1 is the data at times 1, 2, and 3, and the output is the data at time 4; the input of Sample 2 is the data at times 2, 3, and 4, and the output is the data at time 5; and so on; the input of Sample T-3 is the data at times T-3, T-2, and T-1, and the output is the data at time T. In this way, the original grid time series is transformed into a five-dimensional tensor dataset (number of samples, input time step, height, width, number of channels) that can be used for supervised learning, laying the foundation for the training of subsequent deep learning models.

[0056] Step S3: Construct a CBAM-enhanced ED-ConvLSTM model See Figure 3 The CBAM module in the embodiment includes two sequentially connected sub-modules: Channel Attention and Spatial Attention.

[0057] For the input feature map First, it passes through the channel attention module: feature map The channels are then processed using global max pooling (MaxPool) and global average pooling (AvgPool) to generate two distinct channel descriptors. These two descriptors are fed into a multilayer perceptron (MLP) with shared parameters (consisting of two fully connected layers), yielding two output vectors. These two output vectors are then summed element-wise and passed through a sigmoid activation function to generate channel attention weights. .Will Compared with the original feature map Element-wise multiplication yields the refined feature map of the channel. (i.e., Channel refinedInput).

[0058] Subsequently, The input spatial attention module performs global max pooling and global average pooling along the channel dimension to generate two spatial feature descriptors. These two descriptors are concatenated along the channel dimension. The resulting feature map is then passed through a standard convolutional layer to reduce the number of channels to 1, and finally through a sigmoid activation function to generate spatial attention weights. .Will and Element-wise multiplication yields the final refined output feature map. .

[0059] The core of the model is a 4-layer encoder-decoder structure (including a three-level cascaded structure at the encoding and decoding ends and a bottleneck layer). An example of a preferred configuration is shown in the table below: Table 1: ED-ConvLSTM Model Parameter Configuration Table

[0060] The above configuration enables the model to achieve a balance between spatial feature extraction and computational efficiency, effectively capturing the multi-scale spatiotemporal evolution of ionospheric TEC.

[0061] In this invention, the CBAM module is embedded in two key locations of the encoder: Encoding layer CBAM: Located after each Conv2D layer, it adaptively calibrates the feature importance of different channels through channel attention, and then highlights key regions with significant dynamic changes through spatial attention. The refined feature map is captured by ConvLSTM to capture temporal dynamic features. The final encoded CBAM is located after the last ConvLSTM layer of the encoder. It applies global attention weighting to the high-level semantic features extracted by the encoder to further enhance the global spatiotemporal features related to geomagnetic storm disturbances.

[0062] CBAM uses the aforementioned dual attention mechanism to enable the network to adaptively focus on important channels and key spatial regions while suppressing irrelevant information, thereby enhancing its feature representation capabilities.

[0063] See Figure 4 This demonstrates the specific embedding location of the CBAM module in the ED-ConvLSTM model encoder: The input layer receives dimensions as follows: × × × The tensor, in which The input time step is H, and the height and width of the feature map are H and W, respectively. Number of input channels (in this embodiment) =4, corresponding to the four channels TEC, Dst, Kp, and F10.7.

[0064] The encoder employs a three-stage cascaded structure, with each stage consisting of a Conv2D layer and a ConvLSTM layer: the first stage uses a 3×3 Conv2D convolutional kernel with a stride of 1 and 64 output channels, followed by a CBAM module and then a ConvLSTM1 (64 output channels); the second stage uses a 3×3 Conv2D convolutional kernel with a stride of 2 and 128 output channels, followed by a CBAM module and then a ConvLSTM2 (128 output channels); the third stage uses a 3×3 Conv2D convolutional kernel with a stride of 2 and 256 output channels, followed by a CBAM module and then a ConvLSTM3 (256 output channels). Through this structure, the encoder progressively extracts local spatial features and compresses spatial dimensions, capturing temporal dynamic information at different spatial granularities. A CBAM module is embedded again after the last ConvLSTM layer of the encoder to further refine the features; this layer also serves as an intermediate layer (i.e., a bottleneck layer) connecting the encoder and decoder.

[0065] The decoder is a three-stage mirror structure symmetrical to the encoder: the first stage consists of DeConv2D (transposed convolution, 3×3 kernel, stride 2, 128 output channels) and ConvLSTM (128 output channels); the second stage consists of DeConv2D (3×3 kernel, stride 2, 64 output channels) and ConvLSTM (64 output channels); the third stage consists of DeConv2D (3×3 kernel, stride 1, 32 output channels) and ConvLSTM (32 output channels). The decoder progressively reconstructs the temporal features and restores the original spatial resolution.

[0066] Finally, the output layer achieves cross-channel feature fusion through a 1×1 Conv2D layer, with an output dimension of... × × A future TEC grid prediction sequence of ×1.

[0067] Step S4: Model training, using the following preferred scheme 1. Loss function: The mean squared error (MSE) is used as the loss function.

[0068] 2. Optimizer: Use the Adam optimizer with an initial learning rate of 0.00025.

[0069] 3. Regularization and scheduling: The dropout rate after the ConvLSTM layer is 0.2; the ReduceLROnPlateau scheduler is used, and the learning rate is multiplied by 0.55 (preferred empirical value) when the validation loss does not decrease after 5 epochs; EarlyStopping is used, and training is terminated early when the validation loss does not improve after 12 epochs.

[0070] In practice, the geomagnetic storm-specific dataset (containing samples of small, medium, and large geomagnetic storms) constructed in step S2 can be divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1 or 8:1:1). The training set is used for backpropagation of model parameters and weight updates; the validation set is used to monitor overfitting during training and trigger learning rate decay and early stopping mechanisms; the test set is used to finally evaluate the model's prediction accuracy and generalization ability, and does not participate in any parameter optimization process. Through this strategy, overfitting can be effectively avoided while ensuring the model's fitting ability, thus improving the model's robustness in predicting unknown geomagnetic storm events.

[0071] Step S5: TEC Prediction The preprocessed multi-channel data (dimension 1, T_in, H, W, 4) corresponding to the time period to be predicted is input into the trained model. The model directly outputs the TEC prediction grid (dimension 1, T_out, H, W, 1) for the next T_out time steps. After inverse normalization of the results, the predicted TEC value for physical units can be obtained.

[0072] To facilitate understanding of the technical effects of this invention, the TEC (Transient Electron Occurrence) of three typical ionospheric activity years—low, middle, and high solar activity years—was predicted using the method of this invention. The prediction results were compared with those of the traditional physical model COPG and the existing convolutional neural network model ConvLSTM, as shown in Table 1. The results show that under the conditions of the three typical solar activity years, RMSE and MAE are minimized, and R... 2 The results show that the method of the present invention is optimal in all indicators.

[0073] Table 1 Comparison of TEC Prediction Effects

[0074] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0075] The following describes the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction electronic device provided by the present invention. The CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction electronic device described below can be referred to in correspondence with the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method described above.

[0076] The electronic device may include a processor, a communications interface, memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can call logical instructions from the memory to execute the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method, which mainly includes the software processing portion described above.

[0077] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the software processing part of the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method provided by the above methods.

[0079] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the software processing portion of the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method provided by the above methods.

[0080] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0081] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spatiotemporal prediction method for ionospheric TEC based on CBAM-enhanced ED-ConvLSTM, characterized in that, include: Integrate ionospheric TEC grid data, geomagnetic index, and solar activity index, and perform preprocessing to construct a multi-channel spatiotemporal grid dataset; Based on geomagnetic index identification of geomagnetic storm events, data on the first preset duration before the occurrence of a geomagnetic storm, during the geomagnetic storm, and the second preset duration after the end of the geomagnetic storm are extracted from a multi-channel spatiotemporal grid dataset to construct a geomagnetic storm-specific dataset focusing on geomagnetic storm disturbances; A prediction model is constructed with a ConvLSTM encoder-decoder structure as the backbone and CBAM modules embedded at key locations in the encoder. The implementation includes: the encoder of the prediction model adopts a cascaded structure of multi-level convolutional layers and convolutional long short-term memory layers; the decoder adopts a transposed convolutional layer and convolutional long short-term memory layer structure symmetrical to the encoder; the CBAM module sequentially refines features through channel attention and spatial attention; the CBAM module is embedded after each convolutional layer in the cascaded structure of multi-level convolutional layers and convolutional long short-term memory layers in the encoder, and after the last convolutional long short-term memory layer, to focus on key features related to geomagnetic storms. The model was trained and its parameters optimized using the aforementioned geostorm-specific dataset; The preprocessed multi-channel data is input into the trained prediction model, and the spatiotemporal prediction results of the ionosphere TEC are output.

2. The spatiotemporal prediction method for ionospheric TEC based on CBAM enhancement according to claim 1, characterized in that: The preprocessing includes outlier handling, spatiotemporal resolution unification, data normalization, and multi-channel data stitching, which expands the one-dimensional exponential data into two-dimensional data aligned with the TEC grid space.

3. The spatiotemporal prediction method for ionospheric TEC based on CBAM enhancement according to claim 1, characterized in that: The construction of the geomagnetic storm-specific dataset involves collecting data from two days before the geomagnetic storm, during the geomagnetic storm, and one day after the geomagnetic storm for each geomagnetic storm event identified and extracted based on the Dst index within a complete solar activity cycle. This data forms a complete sample event to ensure the capture of the full-process characteristics of the geomagnetic storm precursors, main phase, and recovery phase.

4. The spatiotemporal prediction method for ionospheric TEC based on CBAM enhancement according to claim 1, characterized in that: The model training employs an adaptive optimizer, combining dropout regularization, dynamic learning rate adjustment, and early stopping mechanism, with mean squared error as the loss function to optimize the model parameters.

5. The spatiotemporal prediction method for ionospheric TEC based on CBAM enhancement according to claim 1, characterized in that: After cross-channel feature fusion, the model outputs a multi-step TEC grid prediction sequence to achieve end-to-end prediction.

6. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the CBAM-enhanced ED-ConvLSTM ionospheric TEC spatiotemporal prediction method as described in any one of claims 1 to 5.

8. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the spatiotemporal prediction method for ED-ConvLSTM ionospheric TEC based on CBAM enhancement as described in any one of claims 1 to 5.

Images