Method and system for predicting ionospheric delay by combining a convolutional long short-term memory and a vision transformer

Abstract

This invention discloses a method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT. The method includes: acquiring single-moment TEC data of a target region provided by an empirical ionospheric model, as well as VTEC data and mask layers obtained from ground-based GNSS stations concurrently; inputting the acquired data into a trained fusion model, and outputting TEC prediction results for a preset future time period; wherein the fusion model combines a ConvLSTM model and a ViT model. This invention, by capturing local spatiotemporal details using ConvLSTM and combining it with ViT to capture global dependencies and multi-input fusion auxiliary information, can effectively predict ionospheric delay information even under conditions of ionospheric activity.

Description

Technical Field

[0001] This invention belongs to the field of ionospheric information forecasting technology, and in particular relates to a method and system for forecasting ionospheric delay using a fusion model combining ConvLSTM and ViT. Background Technology

[0002] Ionospheric activity (such as strong geomagnetic storms, ionospheric storms, and violent scintillation events) can cause violent disturbances in the TEC (Transient Electron Turbidity) system, which may surge several times in a short period of time, and the propagation delay may also change drastically, far exceeding the range predicted by ionospheric models.

[0003] Without accurate ionospheric delay estimation, GNSS positioning will produce large positioning errors. In addition, the PPP filtering convergence process will also be greatly affected. Rapid changes in the ionosphere will lengthen the convergence time, causing a gap in high-precision positioning and disrupting the continuity.

[0004] Currently, among the three mainstream methods for predicting ionospheric TEC, empirical ionospheric models, such as the International Reference Ionospheric Model (IRI), are difficult to predict instantaneous changes in TEC in actual predictions, and also lack accuracy in predicting local areas.

[0005] Traditional methods for predicting TEC (Transient Emissions and Dependencies) in time series include Autoregressive Moving Average (ARMA) and Autoregressive Integral Moving Average (ARIMA) models. However, these linear models are not effective at predicting the nonlinear changes in TEC. Meanwhile, the traditional ConvLSTM model is specifically designed to capture local spatiotemporal dependencies (such as video frame sequences and meteorological spatiotemporal data), but it also has certain limitations.

[0006] The terms used in this invention specification are:

[0007] TEC: Total Electron Content;

[0008] VTEC: Vertical TEC, total vertical electron content;

[0009] Vision Transformer (ViT): Vision Transformer, a model that applies the Transformer architecture to computer vision tasks;

[0010] ConvLSTM: Convolutional Long Short-Term Memory, a deep learning model that combines convolutional neural networks (CNN) and long short-term memory networks (LSTM);

[0011] Pixel Shuffle: Pixel rearrangement;

[0012] IRI2020 Model: International Reference Ionosphere 2020 Model. Summary of the Invention

[0013] To overcome the shortcomings of the prior art, this invention provides a method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT. By capturing local spatiotemporal details with ConvLSTM and combining it with ViT to capture global dependencies and multi-input fusion auxiliary information, the method can effectively predict ionospheric delay information even under active ionospheric conditions.

[0014] According to one aspect of the present invention, a method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT is provided, comprising:

[0015] Acquire single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period;

[0016] The acquired data is input into the trained fusion model, which outputs TEC forecast results for a future preset time period; wherein, the fusion model combines a ConvLSTM model and a ViT model, and its training includes:

[0017] A dataset is constructed, which includes single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period;

[0018] The TEC data from multiple consecutive times within the target area are used to construct a multi-time TEC sequence and input into the ConvLSTM model to output a spatiotemporal feature encoding map for a future preset duration.

[0019] The spatiotemporal feature encoding map output by the ConvLSTM model is used as a feature input containing temporal evolution information. It is combined with VTEC data obtained from ground GNSS stations at the same time and mask layers to form a multi-channel tensor. The multi-channel tensor is then segmented and embedded into non-overlapping small blocks to form an input sequence.

[0020] The input sequence is fed into the ViT model. Each Transformer block uses a multi-head self-attention mechanism to capture global dependencies in low-resolution ionospheric images. At the same time, the weights of VTEC data and mask layers are dynamically adjusted using an attention mechanism to output TEC prediction results for the target area in the future for a preset period.

[0021] Train the model on the constructed dataset until it converges, and save the model parameters.

[0022] As a further technical solution, the data acquisition process also includes:

[0023] VTEC data obtained from ground GNSS stations during the same period are used as the true values ​​of pixels, and a mask layer is generated based on whether the pixel includes the observation value.

[0024] As a further technical solution, the method also includes:

[0025] Cells containing observations are marked as valid, and those without observations are marked as invalid. The marking results of all cells are combined to form a binary mask layer.

[0026] As a further technical solution, the ConvLSTM model includes an encoder, a ConvLSTM network, and a decoder. The encoder contains three sets of 3D convolutional layers and three sets of 3D pooling layers. The PReLU function is used as the activation function and normalization is added in the ConvLSTM network. The decoder includes 3D convolutional layers and upsampling layers.

[0027] As a further technical solution, the dimensions of the multi-channel tensor formed by splicing are: batch size, height, width, and number of channels.

[0028] As a further technical solution, after splicing to obtain a multi-channel tensor, the following are included:

[0029] The multichannel tensor obtained by splicing is divided into several non-overlapping blocks, each of the same size; each non-overlapping block is flattened and embedded into a high-dimensional space to form the input sequence.

[0030] As a further technical solution, the weights of VTEC data and mask layers are dynamically adjusted using an attention mechanism, including:

[0031] For invalid pixels marked by the mask, their weight is automatically reduced; for pixels with high confidence in VTEC data, their weight is automatically increased.

[0032] According to one aspect of the present invention, a system for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT is provided, comprising:

[0033] The data acquisition module is used to acquire single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period;

[0034] The fusion forecast module is used to input the acquired data into the trained fusion model and output the TEC forecast results for a preset future time period; wherein, the fusion model combines a ConvLSTM model and a ViT model, and its training includes:

[0035] A dataset is constructed, which includes single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period;

[0036] The TEC data from multiple consecutive times within the target area are used to construct a multi-time TEC sequence and input into the ConvLSTM model to output a spatiotemporal feature encoding map for a future preset duration.

[0037] The spatiotemporal feature encoding map output by the ConvLSTM model is used as a feature input containing temporal evolution information. It is combined with VTEC data obtained from ground GNSS stations at the same time and mask layers to form a multi-channel tensor. The multi-channel tensor is then segmented and embedded into non-overlapping small blocks to form an input sequence.

[0038] The input sequence is fed into the ViT model. Each Transformer block uses a multi-head self-attention mechanism to capture global dependencies in low-resolution ionospheric images. At the same time, the weights of VTEC data and mask layers are dynamically adjusted using an attention mechanism to output TEC prediction results for the target area in the future for a preset period.

[0039] Train the model on the constructed dataset until it converges, and save the model parameters.

[0040] According to one aspect of the present invention, an apparatus for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT is provided, comprising a memory and a processor, the memory storing program instructions executable by the processor, the processor invoking the program instructions to perform the method for predicting ionospheric delay using the fusion model combining ConvLSTM and ViT.

[0041] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to perform the method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. This invention uses a ConvLSTM model to convert TEC data provided by an empirical model into multi-time-series evolution features, forming dynamic spatiotemporal features that combine spatial and temporal information. These dynamic spatiotemporal features are then combined with observations and mask layers to form multi-channel features, which are then input into a ViT model to output a TEC evolution sequence for a future period.

[0044] 2. The fusion model provided in this invention incorporates normalization, mixed sampling, and mask layer creation operations to improve the model's fitting performance. The fusion model captures local spatiotemporal details using ConvLSTM, combines this with ViT to capture global dependencies, and incorporates a multi-input fusion auxiliary information structure. This overcomes the shortcomings of traditional ConvLSTM (insufficient global features) and ViT (sensitive to local temporal variations), making it more suitable for predicting complex systems like the ionosphere where local dynamics are strongly coupled with global correlations.

[0045] 3. The fusion model provided by this invention can predict ionospheric delay information well under active ionospheric conditions, with high accuracy, reducing filtering convergence time and ensuring the continuity of high-precision positioning. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating a method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT, as provided in an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the data collection and ConvLSTM model processing flow provided in the embodiments of the present invention.

[0049] Figure 3 This is a schematic diagram of the data preprocessing and ViT model processing flow provided in an embodiment of the present invention.

[0050] Figure 4 This is a schematic diagram of the model training and application process provided in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0052] This invention provides a method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT, such as... Figure 1 As shown, firstly, the single-moment TEC data of the target area provided by the ionospheric empirical model is acquired, along with VTEC data and mask layers obtained from ground GNSS stations during the same period. Then, the acquired data is input into the trained fusion model, which outputs the TEC forecast results for the future preset time period.

[0053] The fusion model combines a ConvLSTM model and a ViT model, and its training includes:

[0054] Step 1, Data Preparation.

[0055] Long-term ionospheric TEC data is collected as input to the ConvLSTM model, and VTEC data obtained from ground GNSS stations during the same period are used as the true pixel values. A mask layer is generated based on whether a pixel includes an observation value. The input data is divided into training and validation sets according to a certain ratio.

[0056] Step 2, fusion model construction and training.

[0057] Step 2.1: Construct a ConvLSTM model and use ionospheric TEC data from multiple consecutive time points to form a TEC time series as input to the ConvLSTM model. The output is an ionospheric image with dynamic spatiotemporal characteristics. The ConvLSTM model consists of three parts: an encoder, a ConvLSTM network, and a decoder. The encoder contains three sets of 3D convolutional layers and three sets of 3D pooling layers. The ConvLSTM network uses the PReLU function as the activation function and incorporates normalization. The decoder consists of 3D convolutional layers and 3D upsampling layers. The input data is 1-hour TEC data for a certain region, and the output is a 1-hour data feature map. The input data shape is (number of samples, 24, 72, 72, 1), with dimensions interpreted as (batch_size, time step, latitude, longitude, number of channels). The output data shape is (number of samples, 24, 9, 9, 128), with dimensions interpreted as (batch_size, time step, latitude, longitude, number of channels). Where batch_size represents the batch size.

[0058] Step 2.2 involves preprocessing the ionospheric image with dynamic spatiotemporal characteristics output by the ConvLSTM model, as well as the VTEC data and mask layers obtained from ground GNSS stations during the same period.

[0059] Specifically, the ionospheric image with dynamic spatiotemporal characteristics output by the ConvLSTM model is used as the feature input containing temporal evolution information, and VTEC data obtained from ground GNSS stations during the same period is used as the true pixel values. Simultaneously, a mask layer is generated based on whether a pixel contains an observation value; pixels containing observation values ​​are marked as valid (represented by 1), and those without observation values ​​are marked as invalid (represented by 0). All pixel marking results are combined into a binary mask layer. These data are then concatenated and merged to form a multi-channel tensor, which is used as the input to the ViT model. Input data shape: (number of samples, 9, 9, 192), dimension interpretation: (batch_size, height, width, number of channels).

[0060] The concatenated multichannel tensor is segmented into several non-overlapping patches, each of size P×P. Each patch is flattened and embedded into a high-dimensional space to form the input sequence.

[0061] Step 2.3: Input the processed input sequence (i.e., the patch embedding sequence) into the ViT model. At the same time, use the multi-head self-attention mechanism to capture global information and introduce additional channel information (real observations and mask layers) into the input of each Transformer layer. The weights of these additional channel information are dynamically adjusted through the attention mechanism.

[0062] The attention weights for additional channel information change in real time with the distribution of input data / feature correlation, and are not fixed weights. For invalid pixels marked by the mask, the model automatically reduces their weights; for pixels with high confidence in VTEC, the model automatically increases their weights.

[0063] Specifically, the dynamic weights (dynamic associations across patches) in global dependency capture can be understood as follows: after an ionospheric image with dynamic spatiotemporal features is segmented into multiple patches, each patch corresponds to a token in the sequence. Multi-head self-attention calculates the association weights between any two patches. Patches in the active ionospheric region (high VTEC values) will be assigned higher attention weights by the model, while patches marked as "invalid" by the mask (without VTEC observations) will be assigned lower weights. For VTEC / mask data, in the step of concatenating VTEC (pixel ground truth), mask data, and ConvLSTM output features into a multi-channel tensor, and then segmenting it into patches and embedding it into a high-dimensional space, the mask data serves as a "validity marker." If the mask value of a patch is 0 (invalid pixel), the K / V vector of that patch will be weakened by the model, causing other patches to pay less attention to it (dynamically shielding invalid data). VTEC serves as "ground truth supervision," which will cause the attention weights to tilt towards "high confidence pixels." For example, if the VTEC value of a patch differs greatly from that of surrounding patches (large ionospheric gradient), the model will dynamically increase the attention weights of that patch and its neighboring patches.

[0064] Furthermore, an upsampling module (such as deconvolution or PixelShuffle) is added after the Transformer encoder to preserve local details. The reconstruction is further refined through methods such as convolution. The model output is the TEC data of the region to be determined for the next half hour. The data shape is: (number of samples, 1800, 144, 144, 1), and the dimension interpretation is: (batch_size, time step, latitude, longitude, number of channels).

[0065] Step 2.4: Input the training and validation set data into the model, train the model until it converges, and save the model parameters.

[0066] When applying the model, the ionospheric TEC data provided by the IRI2020 model at a predetermined time and the VTEC data and mask layers obtained by ground GNSS stations at the same time are obtained. The data are then input into the trained model, and the TEC forecast results for the future preset time period are output.

[0067] As a preferred embodiment, such as Figures 2 to 4 As shown, the method provided in this embodiment of the invention includes the following steps:

[0068] 1. Data preparation.

[0069] Download the ionospheric TEC data from the IRI2020 model for the region to be determined, spanning 11 years (one solar cycle) from 2010 to 2020. Collect VTEC data from ground-based GNSS stations during the same period as the true pixel values. Simultaneously, generate a mask layer based on whether a pixel contains an observation value; pixels containing observation values ​​are marked as valid (represented by 1), while those without are marked as invalid (represented by 0). All pixel markings are combined into a binary mask layer. The TEC data, true pixel values, and mask layer together constitute the model input layer, while the TEC forecast results for the region to be determined are used as the model output layer.

[0070] 2. Dataset construction and partitioning.

[0071] The data from 11 years was divided into training and validation sets according to a certain ratio, and the training and validation sets were then standardized.

[0072] 3. ConvLSTM model construction.

[0073] The TEC data for a certain region over one hour is used as input data and fed into the convolutional and pooling layers of the ConvLSTM model. Then, through the ConvLSTM network and decoder, feature extraction is performed, and the output data is the feature map of the TEC data for that region over one hour.

[0074] 4. Patch segmentation and embedding.

[0075] The ionospheric image output by ConvLSTM is used as an ionospheric image with dynamic spatiotemporal characteristics, and VTEC data obtained from ground GNSS stations during the same period are used as the true pixel values. At the same time, a mask layer is generated based on whether the pixel includes the observation value. These data are stitched together to form a multi-channel tensor as the model input.

[0076] Then, the concatenated multi-channel tensor is patch-segmented, dividing each input image into several fixed-size, non-overlapping patches, each of which is flattened into a one-dimensional vector. Each patch is then mapped to a high-dimensional space through a linear transformation (i.e., patch embedding) to form a new feature representation. Furthermore, positional encoding is added to each patch to preserve spatial location information within the image.

[0077] 5. Construct the Vision Transformer model.

[0078] The input patch embedding sequence is processed through ViT's Transformer layers. Each Transformer layer uses a multi-head self-attention mechanism to capture global dependencies in the low-resolution image. Additional channel information (ground observations and mask layers) is fused into the input of each layer.

[0079] 6. Upsampling and convolution.

[0080] The low-resolution features output by the Transformer network are input into the upsampling module (such as deconvolution and pixel shuffle) after the Transformer encoder. The output of the upsampling module is then subjected to a series of convolution operations for further refinement to improve the accuracy of the results. The model output is the TEC data for the next half hour of the region to be determined.

[0081] 7. Model training.

[0082] Input the training and validation set data into the model, train the model until it converges, and save the model parameters.

[0083] 8. Model application.

[0084] The model acquires ionospheric TEC data from the IRI2020 model for the region to be determined, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period. The data is then input into the trained model, which outputs TEC forecast results for the future preset time period.

[0085] The implementation of the various embodiments of this invention is based on programmed processing through a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of this invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of this invention provide a system for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT. This system is used to execute the method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT as described in the above method embodiments.

[0086] The system includes: a data acquisition module for acquiring single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period; and a fusion forecast module for inputting the acquired data into a trained fusion model and outputting TEC forecast results for a preset future time period. The fusion model combines a ConvLSTM model and a ViT model, and its training includes: constructing a dataset comprising single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period; and constructing a multi-moment TEC sequence from TEC data of multiple consecutive moments within the target area and inputting it into the ConvLSTM model. Output a spatiotemporal feature encoding map for a future preset duration; use the spatiotemporal feature encoding map output by the ConvLSTM model as a feature input containing temporal evolution information, combine it with VTEC data obtained from ground GNSS stations at the same time and mask layers to form a multi-channel tensor, and perform non-overlapping block segmentation and embedding on the multi-channel tensor to form an input sequence; input the input sequence into the ViT model, each Transformer block uses a multi-head self-attention mechanism to capture global dependencies in low-resolution ionospheric images, and simultaneously uses an attention mechanism to dynamically adjust the weights of VTEC data and mask layers, outputting the TEC forecast results for the target area for a future preset time period; train the model on the constructed dataset until convergence, and save the model parameters.

[0087] This invention provides a system for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT. Addressing the current situation where ionospheric delay information cannot be predicted well under active ionospheric conditions, this system employs several modules, using ConvLSTM to capture local spatiotemporal details and combining ViT to capture global dependencies and multi-input fusion auxiliary information. This allows for a good prediction of ionospheric delay information even under active ionospheric conditions.

[0088] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the modules in the above system embodiments by referring to the specific technical solutions in other method embodiments and combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, on the basis of the above system embodiments, and on the premise of ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.

[0089] Based on the same inventive concept as any of the foregoing embodiments, this embodiment of the invention also provides an apparatus for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT, including a memory and a processor. The memory stores program instructions that are executed by the processor, and the processor calls the program instructions to execute the method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT.

[0090] In embodiments of the present invention, the memory can be non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or it can be volatile memory, such as random-access memory (RAM). Memory is any other medium capable of carrying or storing desired program code having an instruction or data structure form and accessible by a computer, but is not limited thereto. The memory in embodiments of the present invention can also be a circuit or any other device capable of implementing a storage function for storing program instructions and / or data.

[0091] In this embodiment of the invention, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in this embodiment of the invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in this embodiment of the invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0092] Based on the same inventive concept as any of the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the method for predicting ionospheric delay using a fusion model combining ConvLSTM and ViT, the steps of which are as follows:

[0093] Acquire single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period;

[0094] The acquired data is input into the trained fusion model, which outputs TEC forecast results for a future preset time period; wherein, the fusion model combines a ConvLSTM model and a ViT model, and its training includes:

[0095] A dataset is constructed, which includes single-moment TEC data of the target area provided by the ionospheric empirical model, as well as VTEC data and mask layers obtained from ground GNSS stations during the same period;

[0096] The TEC data from multiple consecutive times within the target area are used to construct a multi-time TEC sequence and input into the ConvLSTM model to output a spatiotemporal feature encoding map for a future preset duration.

[0097] The spatiotemporal feature encoding map output by the ConvLSTM model is used as a feature input containing temporal evolution information. It is combined with VTEC data obtained from ground GNSS stations at the same time and mask layers to form a multi-channel tensor. The multi-channel tensor is then segmented and embedded into non-overlapping small blocks to form an input sequence.

[0098] The input sequence is fed into the ViT model. Each Transformer block uses a multi-head self-attention mechanism to capture global dependencies in low-resolution ionospheric images. At the same time, the weights of VTEC data and mask layers are dynamically adjusted using an attention mechanism to output TEC prediction results for the target area in the future for a preset period.

[0099] Train the model on the constructed dataset until it converges, and save the model parameters.

[0100] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0101] 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 or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting ionospheric delay by a fusion model combining ConvLSTM and ViT, characterized in that, The method comprises the following steps: Obtain single-time TEC data provided by an ionospheric empirical model, and VTEC data and a mask layer obtained by a ground GNSS station at the same time; Input the obtained data into a trained fusion model to output a TEC prediction result in a preset period in the future; wherein the fusion model combines a ConvLSTM model and a ViT model, and the training of the fusion model comprises the following steps: Construct a data set, wherein the data set comprises single-time TEC data provided by an ionospheric empirical model, and VTEC data and a mask layer obtained by a ground GNSS station at the same time; Form a multi-time TEC sequence by using TEC data at a plurality of continuous times in a target region, and input the multi-time TEC sequence into a ConvLSTM model to output a spatiotemporal feature encoding map in a preset period in the future; Input the spatiotemporal feature encoding map output by the ConvLSTM model as a feature containing time sequence evolution information, splice the VTEC data and the mask layer obtained by the ground GNSS station at the same time to form a multi-channel tensor, and perform non-overlapping small block segmentation and embedding on the multi-channel tensor to form an input sequence; Input the input sequence into a ViT model, use a multi-head self-attention mechanism in each Transformer block to capture global dependence in a low-resolution ionospheric image, dynamically adjust the weights of the VTEC data and the mask layer by using an attention mechanism, and output a TEC prediction result of the target region in a preset period in the future; Train the model on the constructed data set until convergence, and save the model parameters.

2. The method of claim 1, wherein the method comprises the following steps: 1) obtaining the ionospheric delay data; 2) preprocessing the ionospheric delay data; 3) constructing the training set and the test set; 4) training the model; 5) predicting the ionospheric delay. When obtaining data, the method further comprises the following steps: Obtain VTEC data obtained by a ground GNSS station at the same time as the pixel true value, and generate a mask layer according to whether the pixel contains an observation value.

3. The method of claim 2, wherein the method comprises the following steps: 1) obtaining the ionospheric delay data; 2) preprocessing the ionospheric delay data; 3) inputting the preprocessed ionospheric delay data into the ConvLSTM-ViT fusion model to obtain the ionospheric delay prediction results. The method further comprises the following steps: Mark the pixel containing an observation value as valid, mark the pixel not containing an observation value as invalid, and combine the marking results of all pixels to form a binary mask layer.

4. The method of claim 1, wherein the method comprises: The ConvLSTM model comprises an encoder, a ConvLSTM network and a decoder, wherein the encoder comprises three groups of 3D convolution layers and three groups of 3D pooling layers, a PReLU function is used as an activation function in the ConvLSTM network and normalization is added, and the decoder comprises a 3D convolution layer and an up-sampling layer.

5. The method of claim 1, wherein the method comprises: The dimensions of the spliced multi-channel tensor are batch size, height, width and channel number.

6. The method of claim 5, wherein the method comprises the following steps: After obtaining the multi-channel tensor by splicing, the method comprises the following steps: Segment the obtained multi-channel tensor to obtain a plurality of non-overlapping small blocks, and each non-overlapping small block has the same size; each non-overlapping small block is flattened and embedded into a high-dimensional space to form an input sequence.

7. The method of claim 3, wherein the method comprises the following steps: 1) obtaining the ionospheric delay data; 2) preprocessing the ionospheric delay data; 3) constructing the ionospheric delay data set; 4) training the model; 5) predicting the ionospheric delay. Dynamically adjusting the weights of the VTEC data and the mask layer by using an attention mechanism comprises the following steps: Automatically reduce the weight of the invalid pixel marked by the mask; and automatically increase the weight of the pixel with high confidence in the VTEC data.

8. A system for predicting ionospheric delay by a fusion model combining ConvLSTM and ViT, characterized in that, The method comprises the following steps: A data acquisition module is configured to obtain single-time TEC data provided by an ionospheric empirical model, and VTEC data and a mask layer obtained by a ground GNSS station at the same time; The fusion prediction module is configured to input the acquired data into the trained fusion model, and output a TEC prediction result of a future preset period; wherein the fusion model combines a ConvLSTM model and a ViT model, and the training of the fusion model comprises: constructing a data set, the data set including single-time TEC data of a target area provided by an ionospheric empirical model, and VTEC data and a mask layer obtained by a ground GNSS station at the same period; constructing a data set, the data set including single-time TEC data of a target area provided by an ionospheric empirical model, and VTEC data and a mask layer obtained by a ground GNSS station at the same period; constructing a data set, the data set including single-time TEC data of a target area provided by an ionospheric empirical model, and VTEC data and a mask layer obtained by a ground GNSS station at the same period; inputting the TEC data of the continuous multiple time points in the target area into the ConvLSTM model to form a multi-time TEC sequence, and outputting a spatio-temporal feature encoding map of a future preset time length; inputting the spatio-temporal feature encoding map output by the ConvLSTM model as a feature containing time sequence evolution information, splicing the VTEC data and the mask layer obtained by the ground GNSS station at the same period to form a multi-channel tensor, and performing non-overlapping small block segmentation and embedding on the multi-channel tensor to form an input sequence; 9. A device for predicting ionospheric delay by a fusion model combining ConvLSTM and ViT, characterized in that, inputting the input sequence into the ViT model, using a multi-head self-attention mechanism in each Transformer block to capture global dependency in a low-resolution ionospheric image, and using an attention mechanism to dynamically adjust the weights of the VTEC data and the mask layer, and outputting a TEC prediction result of a future preset period of the target area; 10. A non-transitory computer-readable storage medium, comprising: training the model on the constructed data set until convergence, and saving the model parameters. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the method of predicting ionospheric delay by combining ConvLSTM and ViT of the fusion model according to any one of claims 1 to 7. The non-transitory computer readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the method of predicting ionospheric delay by combining ConvLSTM and ViT of the fusion model according to any one of claims 1 to 7.

Images