Systemic risk prediction method and system

Abstract

The present application discloses a method and system for predicting systemic risks, which relates to the field of deep learning technology, including: converting meteorological sequence data recorded by a meteorological system into meteorological frequency domain data, and fusing the sequence data and frequency domain data to obtain frequency-time-space fusion data; splitting the frequency-time-space fusion data based on the channel dimension, and calculating the spatial correlation of the first data group obtained by splitting and the time-frequency correlation of the second data group obtained by splitting; splicing the spatial correlation and the time-frequency correlation based on the channel dimension to obtain frequency-time-space cascade data; linearly mapping the frequency-time-space cascade data in the time dimension, channel dimension and spatial dimension in turn to obtain meteorological risk prediction results. The present application takes into account the low-frequency and high-frequency change characteristics through frequency domain processing, overcomes high-frequency noise interference, and uses spatial and time-frequency correlation calculations to mine feature interactions, thereby achieving deep fusion of multi-dimensional features, and ultimately improving the accuracy and robustness of systemic meteorological risk prediction.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and in particular to a method and system for predicting systemic risks. Background Art

[0002] Early warning of systemic meteorological disaster risks has always been a research hotspot in academia. Traditional meteorological risk prediction methods are mainly based on linear models. However, since linear models cannot capture the nonlinear relationship between meteorological data variables, nonlinear models have gradually replaced linear models in meteorological risk prediction and warning in recent years, and have effectively improved the performance of risk prediction. However, existing nonlinear models still have certain defects.

[0003] Traditional recurrent neural network models (such as LSTM and GRU) can only process features in a single main dimension (time) and have difficulty capturing cross-dimensional correlations (such as the coupling relationship between periodic risk signals in high-frequency fluctuations in meteorological data and cross-domain contagion effects). In addition, data in the meteorological field will have certain high-frequency noise and non-stationary characteristics. Existing methods are easily misled by this short-term disturbance. At the same time, the presence of noise will also interfere with some regular small changes in the sequence, resulting in the final accuracy of the systemic meteorological risk forecast results being difficult to effectively guarantee.

[0004] Therefore, how to improve the accuracy and robustness of forecasting systemic meteorological risks is an urgent problem that needs to be solved. Summary of the Invention

[0005] The main purpose of this application is to provide a systemic risk prediction method and system, aiming to solve the technical problem of how to improve the accuracy and robustness of forecasting systemic meteorological risks.

[0006] To achieve the above objectives, this application proposes a systemic risk prediction method, which includes:

[0007] Converting meteorological sequence data recorded by a meteorological system into meteorological frequency domain data, and fusing the meteorological sequence data with the meteorological frequency domain data to obtain frequency-time-space fusion data, wherein the dimensions of the meteorological sequence data include a time dimension and a spatial dimension determined based on the number of feature points, and the feature points are meteorological indicators that have changed in the meteorological system;

[0008] Splitting the frequency-space-time fusion data based on a preset model channel dimension to obtain a first data group and a second data group, and respectively calculating the spatial correlation of the first data group and the time-frequency correlation of the second data group;

[0009] splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data;

[0010] The frequency-time-space cascade data is linearly mapped in the time dimension, the model channel dimension and the space dimension in sequence to obtain a meteorological risk prediction result of the meteorological system.

[0011] In one embodiment, the meteorological sequence data at any time point in the meteorological sequence data includes multiple feature points, and the step of converting the meteorological sequence data recorded by the meteorological system into meteorological frequency domain data includes:

[0012] For any feature point, extract a feature trajectory of the feature point changing with time in the meteorological sequence data;

[0013] Performing discrete cosine transform on the feature trajectory to obtain discrete cosine transform coefficients of the feature point at each time point;

[0014] After traversing each characteristic point, the corresponding characteristic trajectory is encoded according to each discrete cosine transform coefficient to obtain meteorological frequency domain data corresponding to the meteorological sequence data.

[0015] In one embodiment, the step of fusing the meteorological sequence data and the meteorological frequency domain data to obtain frequency-time-space fusion data includes:

[0016] splicing the meteorological sequence data and the meteorological frequency domain data to obtain candidate frequency spatiotemporal fusion data;

[0017] Linear mapping is performed on the target dimension in the candidate frequency-time-space fusion data to obtain frequency-time-space fusion data including the preset model channel dimension.

[0018] In one embodiment, the steps of splitting the frequency-space-time fusion data based on a preset model channel dimension to obtain a first data group and a second data group, and respectively calculating the spatial correlation of the first data group and the time-frequency correlation of the second data group include:

[0019] Performing layer normalization on the frequency-space-time fusion data to obtain a target vector matrix;

[0020] Evenly splitting the target vector matrix based on a preset model channel dimension to obtain a first data group and a second data group;

[0021] Calculating the correlation between the feature points in the first data group multiple times in parallel to obtain the spatial correlation of the first data group;

[0022] The attention affinity between the feature points in the second data group is calculated multiple times in parallel to obtain the time-frequency correlation of the second data group.

[0023] In one embodiment, the first data group includes a spatial query vector matrix, a spatial key vector matrix, and a spatial value vector matrix, and the step of calculating the correlation between the feature points in the first data group in parallel multiple times to obtain the spatial correlation of the first data group includes:

[0024] In any calculation process, calculating the first inter-vector similarity according to the spatial query vector matrix and the spatial key vector matrix, and normalizing the first inter-vector similarity to obtain a first attention weight matrix;

[0025] Calculate a first result vector matrix according to the first attention weight matrix and the spatial value vector matrix;

[0026] The first result vector matrices obtained from multiple calculation processes are connected to obtain the spatial correlation of the first data group.

[0027] In one embodiment, the first data group includes at least a spatial value vector matrix, the second data group includes a time-frequency query vector matrix, a time-frequency key vector matrix, and a time-frequency value vector matrix, and the step of calculating the attention affinity between feature points in the second data group in parallel multiple times to obtain the time-frequency correlation of the second data group includes:

[0028] In any calculation process, calculating the second inter-vector similarity according to the time-frequency query vector matrix and the time-frequency key vector matrix, and normalizing the second inter-vector similarity to obtain a second attention weight matrix;

[0029] Calculating a candidate result vector matrix according to the second attention weight matrix and the time-frequency value vector matrix, and adjusting the dimension order of the candidate result vector matrix to be consistent with the spatial value vector matrix to obtain a second result vector matrix;

[0030] The second result vector matrices obtained from multiple calculation processes are connected to obtain the time-frequency correlation of the second data group.

[0031] In one embodiment, the step of splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data includes:

[0032] splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain candidate frequency-space-time cascade data;

[0033] Inputting the candidate frequency-space-time cascade data into a preset first multilayer perceptron and a preset second multilayer perceptron for processing, respectively, to obtain time-domain weight adjustment data output by the first multilayer perceptron and frequency-domain weight adjustment data output by the second multilayer perceptron, wherein the architecture of the second multilayer perceptron sequentially includes an inverse discrete cosine transform part, the architecture of the first multilayer perceptron, and a discrete cosine transform part;

[0034] The time domain weight adjustment data and the frequency domain weight adjustment data are spliced ​​based on the time dimension to obtain frequency-time-space cascade data.

[0035] In one embodiment, after the step of splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data, the step further includes:

[0036] Using the frequency-space-time cascade data as the frequency-space-time fusion data, and returning to perform the step of splitting the frequency-space-time fusion data based on the preset model channel dimension;

[0037] When the number of return executions reaches the preset number of stacking times, based on the frequency-time-space cascade data obtained after returning to execute the stacking times, the step of linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension and the space dimension in sequence is executed.

[0038] In one embodiment, the step of linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension, and the space dimension in sequence to obtain the meteorological risk prediction result of the meteorological system includes:

[0039] Performing a first linear mapping on the frequency-space-time cascade data in the time dimension to obtain a first candidate risk prediction result of the meteorological system at each target time point;

[0040] Performing a second linear mapping on the first candidate risk prediction result in the channel dimension to obtain a second candidate risk prediction result of the meteorological system under the channel dimension fusion;

[0041] Perform a third linear mapping on the second candidate risk prediction result in the spatial dimension to obtain a meteorological risk prediction result of the meteorological system under the spatial dimension fusion.

[0042] In addition, to achieve the above objectives, the present application also proposes a systemic risk prediction system, which includes:

[0043] a conversion and fusion module for converting meteorological sequence data recorded by the meteorological system into meteorological frequency domain data, and fusing the meteorological sequence data with the meteorological frequency domain data to obtain frequency-time-space fusion data, wherein the dimensions of the meteorological sequence data include a time dimension and a spatial dimension determined based on the number of feature points, and the feature points are meteorological indicators that have changed in the meteorological system;

[0044] a spatiotemporal calculation module, configured to split the frequency-space-time fusion data based on a preset model channel dimension to obtain a first data group and a second data group, and respectively calculate the spatial correlation of the first data group and the time-frequency correlation of the second data group;

[0045] A dimension splicing module, configured to splice the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data;

[0046] A mapping output module is used to linearly map the frequency-time-space cascade data in the time dimension, the model channel dimension and the space dimension in sequence to obtain the meteorological risk prediction result of the meteorological system.

[0047] In addition, to achieve the above-mentioned purpose, the present application also proposes an electronic device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program is configured to implement the steps of the systemic risk prediction method as described above.

[0048] One or more technical solutions proposed in this application have at least the following technical effects:

[0049] This application first converts the meteorological series data recorded by the meteorological system into meteorological frequency domain data, and fuses the meteorological series data with the meteorological frequency domain data to obtain frequency-time-space fusion data, thereby realizing the conversion of the meteorological series data in the time-space domain to the frequency domain, which not only retains the time and space dimension information of the original time series, but also captures the periodic characteristics and the change laws of different frequency components in the data through the frequency domain data, providing a basis for subsequent cross-dimensional feature extraction and fusion; then, the frequency-time-space fusion data is split based on the preset model channel dimension to obtain the first data group and the second data group, and the spatial correlation of the first data group and the time-frequency correlation of the second data group are calculated respectively, so as to focus on the feature relationship of the spatial dimension and the time-frequency dimension respectively, wherein the spatial correlation calculation can mine the spatial correlation of different feature points. pattern, and the time-frequency correlation calculation captures the mutual influence of feature points in the time-frequency domain, so as to more comprehensively reflect the feature interaction relationship of different dimensions in the data; then, based on the model channel dimension, the spatial correlation and the time-frequency correlation are spliced ​​to obtain frequency-time-space cascade data, realizing the preliminary fusion of spatial and time-frequency features; finally, the frequency-time-space cascade data are linearly mapped in the time dimension, the model channel dimension and the space dimension in turn to obtain the meteorological risk prediction result of the meteorological system, so as to gradually focus on the key features of different dimensions through layer-by-layer linear mapping. The final risk prediction result of the meteorological system comprehensively considers the interactive influence of time series, frequency domain characteristics and spatial characteristics, can more accurately reflect the dynamic change trend of systemic meteorological risks, and improves the accuracy and robustness of the prediction results.

[0050] In summary, this application converts meteorological series data in the spatiotemporal domain into data in the frequency domain through discrete cosine transform and fuses the original sequence data to form frequency-time-space fusion data, then calculates the spatial correlation and time-frequency correlation of the data respectively, and splices them to form frequency-time-space cascade data, and finally performs linear mapping in the time dimension, channel dimension and spatial dimension in turn to obtain meteorological risk prediction results. This avoids the problems that traditional recurrent neural networks can only process meteorological characteristics in a single time dimension, it is difficult to capture cross-dimensional correlations of data, and the prediction process is easily affected by high-frequency noise, resulting in difficulty in ensuring prediction accuracy. It achieves the goal of taking into account both low-frequency and high-frequency change characteristics through frequency domain processing, overcoming high-frequency noise interference, and using spatial and time-frequency correlation calculations to mine feature interactions. It also realizes multi-dimensional deep fusion of meteorological features based on the Transformer architecture, ultimately improving the accuracy and robustness of systematic meteorological risk prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 A flowchart of the first embodiment of the systemic risk prediction method of this application is provided;

[0054] Figure 2 A flow chart illustrating the second embodiment of the systemic risk prediction method of this application;

[0055] Figure 3 A schematic diagram of a simplified process of the systemic risk prediction method provided in Example 2 of this application;

[0056] Figure 4 This is a schematic diagram of the module structure of the systemic risk prediction system according to an embodiment of the present application;

[0057] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the systemic risk prediction method in the embodiment of the present application.

[0058] The purpose, features and advantages of this application will be further explained with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0059] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0060] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0061] The main solution of the embodiment of the present application is: converting the meteorological sequence data recorded by the meteorological system into meteorological frequency domain data, and fusing the meteorological sequence data and the meteorological frequency domain data to obtain frequency-time-space fusion data, wherein the dimensions of the meteorological sequence data include a time dimension and a spatial dimension determined based on the number of feature points, and the feature points are meteorological indicators that have changed in the meteorological system; splitting the frequency-time-space fusion data based on a preset model channel dimension to obtain a first data group and a second data group, and calculating the spatial correlation of the first data group and the time-frequency correlation of the second data group respectively; splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data; linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension and the spatial dimension in turn to obtain a meteorological risk prediction result of the meteorological system.

[0062] Traditional recurrent neural network models (such as LSTM and GRU) can only process features along a single primary dimension (time), making it difficult to capture cross-dimensional correlations (such as the coupling between cyclical risk signals and cross-domain contagion effects in high-frequency fluctuations in meteorological data). Furthermore, meteorological data often exhibits high-frequency noise and non-stationary characteristics, making existing methods easily misled by these short-term disturbances. Furthermore, the presence of noise can also interfere with subtle, regular variations in the sequence, making it difficult to effectively guarantee the accuracy of the final systemic meteorological risk forecast. Therefore, improving the accuracy and robustness of systemic meteorological risk forecasts is a pressing issue.

[0063] The present application provides a solution, which converts meteorological sequence data in the spatiotemporal domain into data in the frequency domain through discrete cosine transform and fuses the original sequence data to form frequency-time-space fusion data, then calculates the spatial correlation and time-frequency correlation of the data respectively, and splices them to form frequency-time-space cascade data, and finally performs linear mapping in the time dimension, channel dimension and spatial dimension in turn to obtain meteorological risk prediction results. It avoids the problems that traditional recurrent neural networks can only process meteorological characteristics in a single time dimension, it is difficult to capture cross-dimensional correlations of data, and the prediction process is easily disturbed by high-frequency noise, resulting in difficulty in ensuring prediction accuracy. It achieves the goal of taking into account both low-frequency and high-frequency change characteristics through frequency domain processing, overcoming high-frequency noise interference, using spatial and time-frequency correlation calculations to mine feature interactions, and realizing multi-dimensional deep fusion of meteorological features based on the Transformer architecture, ultimately improving the accuracy and robustness of systematic meteorological risk prediction.

[0064] It should be noted that the execution subject of this embodiment may be a computing service device with data processing, network communication, and program execution capabilities, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of implementing the aforementioned functions, such as a systemic risk prediction system. This embodiment and the following embodiments will be described below using a systemic risk prediction system as an example.

[0065] Based on this, the present invention provides a method for predicting systemic risks. Figure 1 , Figure 1 This is a flow chart of the first embodiment of the systemic risk prediction method of this application.

[0066] In this embodiment, the systemic risk prediction method includes steps S10 to S40:

[0067] Step S10: converting meteorological sequence data recorded by the meteorological system into meteorological frequency domain data, and fusing the meteorological sequence data with the meteorological frequency domain data to obtain frequency-time-space fusion data, wherein the dimensions of the meteorological sequence data include a time dimension and a spatial dimension determined based on the number of feature points, and the feature points are meteorological indicators that have changed in the meteorological system;

[0068] It should be noted that meteorological series data is a set of random meteorological variables arranged in chronological order, reflecting the state or characteristic changes of the meteorological system at different time points; meteorological frequency domain data refers to the data obtained by converting meteorological series data through methods such as discrete cosine transform, reflecting the distribution and change law of meteorological data in different frequency components; frequency-time-space fusion data refers to the data formed by splicing and fusing meteorological series data and meteorological frequency domain data in the time dimension, which contains information in three dimensions of time, space and channel; the time dimension is a dimension determined based on the time series of meteorological data, which indicates the continuity and sequentiality of meteorological data in time; the number of feature points refers to the number of data points in the time series used to describe the characteristics of the meteorological system; the spatial dimension is a dimension determined based on the number of feature points, reflecting the spatial distribution characteristics of the meteorological system at different feature points.

[0069] It is understandable that since traditional methods only process meteorological features in a single time dimension, it is difficult to capture cross-dimensional correlations and is easily interfered by high-frequency noise, making it difficult to ensure prediction accuracy. Therefore, performing step S10 can avoid the problem of difficulty in capturing cross-dimensional correlations and susceptibility to high-frequency noise interference caused by processing data only in the time dimension. By converting time series data into frequency domain data and fusing it, it can take into account both low-frequency and high-frequency change characteristics, overcome the disturbance of high-frequency noise on the regularity of meteorological data to a certain extent, and provide a more comprehensive data foundation for subsequent feature extraction and fusion.

[0070] For example, for a meteorological system training dataset with an input format of A*N, a sliding window method can be used to collect time series data, for example, with a window length hyperparameter of L = 12. The window length can be modified as needed. Theoretically, a longer window length contains more time-related information, resulting in better model predictions. By further dividing the training dataset into batches using the sliding window, the following data can be obtained (only the time of the data is shown):

[0071] The first batch of data: 2000.1, 2000.2, …, 2000.12;

[0072] The second batch of data: 2000.2, 2000.3, …, 2001.1;

[0073] …

[0074] Therefore, the dimensions of each batch of data are: L*N. L represents the window length, or the length of the sequence after partitioning; N represents the feature dimension. In deep learning, multiple data sets are also integrated to obtain a batch of data with the dimensions: B*L*C. B represents the batch size, also known as batchsize. During model training, the input is data of type B*L*N, and the output is data of type B*L_new*1. L_new*1 represents the systemic meteorological risk situation at the future time to be predicted.

[0075] For the convenience of the example, we set batchsize=1, and the data dimension of the collected meteorological sequence data is 1*L*N. , where L and N represent the number of sequence elements selected by the sliding window (i.e., the time dimension determined based on the time series) and the number of features contained in each time element (i.e., the spatial dimension determined based on the number of feature points). The data is subjected to discrete cosine transform (DCT) to obtain meteorological frequency domain data The meteorological sequence data and meteorological frequency domain data are fused to obtain frequency-time-space fusion data. .

[0076] In a feasible implementation manner, the meteorological sequence data at any time point in the meteorological sequence data includes multiple feature points. The step of converting the meteorological sequence data recorded by the meteorological system into meteorological frequency domain data in step S10 may include steps S11 to S13:

[0077] Step S11, for any characteristic point, extracting a characteristic trajectory of the characteristic point changing with time in the meteorological sequence data;

[0078] It should be noted that characteristic points refer to specific indicators or data points in meteorological series data that are used to describe the characteristics of the meteorological system, such as atmospheric pressure and surface temperature in meteorological data; characteristic trajectories refer to the data sequence formed by the changes of characteristic points over time, reflecting the changes in the values ​​of the characteristic points at different time points.

[0079] It is understandable that when processing meteorological sequence data, the entire sequence is often analyzed directly, ignoring the dynamic characteristics of a single meteorological feature point changing over time, resulting in insufficient in-depth exploration of the potential patterns in the data. Therefore, step S11 is performed to extract the trajectory of the meteorological feature points changing over time, thereby avoiding the problem of traditional methods processing the meteorological sequence data as a whole while ignoring the dynamic changes of individual feature points. This can more accurately capture the changing trend of each meteorological feature point in the time dimension, provide a more accurate data basis for subsequent frequency domain conversion and feature coding, and help to deeply explore the potential patterns in meteorological sequence data.

[0080] For example, in a 1*L*N input time series data, taking the jth feature in N dimensions will result in the L j features being combined as a whole. This whole reflects the change in feature j over a period of length L, known as the "feature trajectory" of feature j. In meteorology, suppose we need to analyze multiple years of meteorological data for a region to predict the risk of extreme weather events (such as heavy rain or drought). Each feature point can represent a specific meteorological element (such as temperature, humidity, or air pressure) recorded by a weather monitoring station. Specifically, the sequence of temperature changes (feature points) recorded at a particular weather monitoring station over time is extracted from the time series data. This sequence is the feature trajectory of that feature point. For example, if we collect the average daily temperature data for a particular weather station over the past 10 years, the chronological sequence of these data is the temperature feature trajectory of that weather station.

[0081] Step S12, performing discrete cosine transform on the feature trajectory to obtain discrete cosine transform coefficients of the feature point at each time point;

[0082] It is understandable that, since it is usually difficult to effectively separate and analyze the changing patterns of different frequency components when processing meteorological sequence data, the distinction between high-frequency noise and low-frequency trends is not clear enough. Therefore, step S12 is performed. Through discrete cosine transform, the characteristic trajectory can be converted from the time domain to the frequency domain, and the transformation coefficients under different frequency components are obtained. This avoids the shortcomings of traditional methods in the frequency domain analysis of meteorological data, and can effectively decompose the characteristic trajectory into different frequency components. At the same time, low-frequency changing trends and high-frequency fluctuation components can be distinguished. This not only helps to extract the periodic laws of the characteristic trajectory, but also can suppress the influence of high-frequency noise to a certain extent, and enhance the regular characteristics of meteorological data.

[0083] For example, DCT is applied to each track (N tracks in total), and for track , the i-th DCT coefficient is:

[0084]

[0085] Among them, when i=1 =1, otherwise = 0. Each time step in the trajectory produces a DCT coefficient, i∈1,2,…,L.

[0086] Step S13: After traversing each feature point, the corresponding feature trajectory is encoded according to each discrete cosine transform coefficient to obtain meteorological frequency domain data corresponding to the meteorological sequence data.

[0087] It is understandable that when meteorological sequence data is directly converted into frequency domain data, there is often a lack of systematic integration of the frequency domain characteristics of different feature points, resulting in the frequency domain data being unable to fully reflect the comprehensive characteristics of the meteorological system in the frequency domain. Therefore, step S13 is performed to encode and integrate the corresponding feature trajectories according to the discrete cosine transform coefficients of each feature point, so that data that fully reflects the frequency domain characteristics of the meteorological system can be obtained, avoiding the shortcomings of traditional methods in frequency domain data integration, and being able to systematically integrate the frequency domain characteristics of all feature points to form a complete frequency domain data representation, providing richer frequency domain information for subsequent frequency-time-space fusion and risk prediction, and enhancing the model's ability to grasp the laws of meteorological data.

[0088] For example, the DCT coefficients encode the input meteorological sequence data to obtain the frequency domain data corresponding to the meteorological sequence data. , where the low-frequency DCT coefficients encode the rough outline of the meteorological sequence data, and the high-frequency DCT coefficients encode the details of the meteorological sequence data. It can be understood that with a characteristic trajectory (take j = 1, that is, [ ], that is, an L*1 column vector data), if the first half of the value of the column vector is padded with zeros in the second half to the original length, and then the Inverse Discrete Cosine Transform (IDCT) is used to convert it back to the original time and space domain, it will be found that the converted data is almost completely fitted with the original data, which shows that the first half of the DCT coefficient column vector (that is, the "low-frequency coefficient") can characterize / encode the trajectory of the original meteorological sequence data; by the same token, if the second half of the value of the DCT coefficient column vector is padded with zeros in the first half to the original length, and then the IDCT is used to convert it back to the original time and space domain, it will be found that the converted data are all jitter values ​​with very small amplitudes, so they represent a kind of trajectory details.

[0089] In this embodiment, the characteristic trajectory of any meteorological feature point in the time series is first extracted, and then the characteristic trajectory is discrete cosine transformed to obtain the transformation coefficient of each time point. Finally, after traversing all meteorological feature points, the characteristic trajectory is encoded to obtain frequency domain data. This avoids the problem that traditional methods only process meteorological data in the time domain, which makes it difficult to capture periodic laws and is easily interfered by high-frequency noise. At the same time, it overcomes the limitation of traditional frequency domain conversion methods that cannot retain complete time-frequency information. Through discrete cosine transform, meteorological data can be converted to the frequency domain to extract periodic features, and a certain amount of time-frequency localization information can be retained, thereby enhancing the extraction effect of high-frequency features. Ultimately, the technical effect of more comprehensive and accurate mining of potential laws in time series data is achieved, providing a richer and more accurate feature representation for subsequent systemic risk prediction, which helps to improve the performance and reliability of the prediction model.

[0090] In a feasible implementation manner, the step of fusing the meteorological sequence data and the meteorological frequency domain data in step S10 to obtain frequency-time-space fusion data may include steps S14 to S15:

[0091] Step S14, splicing the meteorological sequence data and the meteorological frequency domain data to obtain candidate frequency spatiotemporal fusion data;

[0092] It should be noted that the candidate frequency spatiotemporal fusion data is the data obtained by directly splicing the original meteorological series data and the meteorological frequency domain data. It contains information in both the time domain and the frequency domain, but has not undergone further dimensional adjustment.

[0093] For example, the original input meteorological series data Meteorological frequency domain data after DCT processing Splicing to obtain candidate frequency spatiotemporal fusion data .

[0094] Step S15 , linearly mapping the target dimension in the candidate frequency-time-space fusion data to obtain frequency-time-space fusion data including preset model channel dimensions.

[0095] It is understandable that, since traditional methods lack effective organization and feature mapping means for fused data after data fusion, it is impossible to adjust the fused data to a dimension and structure suitable for subsequent processing, resulting in limited feature expression capabilities of the fused data, making it difficult to directly use it as the input of complex models. Therefore, step S15 is performed, and the linear mapping operation is performed to avoid the problems of dimensional mismatch and insufficient feature expression of the fused data. Not only is the dimensional structure of the data adjusted and the channel dimension increased, but the features are also optimized and mapped, so that the frequency-temporal-spatial fusion data has stronger feature expression capabilities and adaptability, and can better support subsequent deep feature mining and risk prediction tasks.

[0096] For example, the candidate frequency-time-space fusion data X is input into the linear mapping layer (essentially a MLP), and the last dimension of X, i.e. the target dimension, is linearly mapped to obtain the frequency-time-space fusion data including the preset model channel dimension C. .

[0097] In this implementation, candidate frequency spatiotemporal fusion data is formed by splicing meteorological sequence data and meteorological frequency domain data, and linear mapping is performed on them to introduce channel dimensions, which effectively avoids the problems of incomplete information, mismatched fusion data dimensions, and insufficient feature expression caused by traditional methods that only use a single data type. At the same time, it overcomes the defect of not being able to fully explore the potential correlations in time series and frequency domain data, and achieves the effect of integrating time domain and frequency domain information and improving feature expression capabilities, providing more comprehensive and well-structured data support for subsequent systemic risk prediction, and enhancing the model's ability to capture risk characteristics.

[0098] Step S20, splitting the frequency-time-space fusion data based on a preset model channel dimension to obtain a first data group and a second data group, and calculating the spatial correlation of the first data group and the time-frequency correlation of the second data group respectively;

[0099] It should be noted that the preset model channel dimension refers to a dimension in the data, which is used to distinguish different feature channels or feature groups; spatial correlation refers to the correlation between the feature points in the first data group, reflecting the mutual influence and correlation degree of the feature points in space; time-frequency correlation refers to the attention affinity between the feature points in the second data group, reflecting the interaction and correlation degree of the feature points in the time-frequency domain.

[0100] It is understandable that since traditional methods are difficult to capture cross-dimensional correlations and cannot effectively mine the spatial and time-frequency relationships between feature points, performing step S20 can avoid the problem of difficulty in capturing cross-dimensional correlations. By splitting the frequency-space fusion data and calculating the spatial correlation and time-frequency correlation respectively, the correlation pattern of the feature points in the space and time-frequency domains is deeply mined, and the complex relationship between the features can be more comprehensively mined, providing richer information for the subsequent meteorological feature fusion.

[0101] In a feasible implementation, step S20 may include steps S21 to S24:

[0102] Step S21, performing layer normalization on the frequency-space-time fusion data to obtain a target vector matrix;

[0103] It should be noted that the target vector matrix is ​​the frequency-time-space fusion data after layer normalization processing, with a standardized numerical range, which is suitable for subsequent deep learning model processing.

[0104] It is understandable that when processing multi-dimensional meteorological data, the differences in the numerical ranges of different features may lead to unstable model training, slow convergence, and limited prediction performance. Therefore, step S21 is performed to standardize the frequency-space fusion data through layer normalization operations, thereby avoiding the problems of unstable model training and slow convergence caused by differences in the numerical ranges of features, thereby improving the training efficiency and prediction performance of the model and enhancing the model's adaptability to different meteorological characteristics.

[0105] For example, the aforementioned frequency-time-space fusion data is received After that, the data is layer normalized without changing its data dimension, and the Z map is generated into a target vector matrix, that is, three sets of matrices: query Q, key K and value V , , where the three W are all learnable weight matrices.

[0106] Step S22, evenly splitting the target vector matrix based on a preset model channel dimension to obtain a first data group and a second data group;

[0107] It is understandable that, since it is usually difficult to effectively separate and process features of different channels when processing high-dimensional data, this leads to mutual interference between features, affecting the model's ability to extract key features. Therefore, step S22 is performed to obtain a first data group and a second data group by evenly splitting the target vector matrix along the preset model channel dimension, thereby avoiding mutual interference between features. Different feature extraction operations can be performed on these two data groups respectively, thereby more effectively mining the spatial correlation and time-frequency correlation in the data.

[0108] For example, the three matrices in the target vector matrix are evenly divided into two groups along the channel dimension C. For clarity, the divided feature matrix is ​​represented as the time-frequency group Q T , K T 、V T and space group Q S , K S 、V S Next, the temporal-frequency and spatial correlations are computed in two separate self-attention modules.

[0109] Step S23, calculating the correlation between the feature points in the first data group multiple times in parallel to obtain the spatial correlation of the first data group;

[0110] It is understandable that since serial processing is usually used when calculating the correlation between feature points, the calculation efficiency is low and it is difficult to process large-scale data. Therefore, step S23 is performed. Through multiple parallel calculations, the correlation between the feature points in the first data group can be efficiently mined to obtain a spatial correlation matrix, avoiding the problem of low calculation efficiency caused by serial processing and providing richer spatial feature information for subsequent feature fusion.

[0111] For example, the first data set is used as input. After linear transformation to obtain Q, K, and V, the dot product between Q and K is calculated to obtain a similarity matrix between feature points. The similarity matrix is ​​normalized (e.g., using softmax) to obtain an attention weight matrix. Finally, the attention weight matrix is ​​multiplied by V to obtain a spatial correlation matrix.

[0112] Step S24 , calculating the attention affinity between the feature points in the second data group multiple times in parallel to obtain the time-frequency correlation of the second data group.

[0113] It is understandable that when calculating the time-frequency correlation, traditional methods find it difficult to effectively capture the complex relationships between feature points, especially long-distance dependencies, resulting in insufficient feature extraction. Therefore, step S24 is performed. By calculating the attention affinity multiple times in parallel, the time-frequency correlation between the feature points in the second data group can be effectively captured, and the time-frequency correlation matrix is ​​obtained. This avoids the problem that traditional methods cannot effectively capture long-distance dependencies, and can more comprehensively explore the complex relationships between feature points, thereby enhancing the model's ability to express time-frequency features.

[0114] For example, a linear transformation is performed on the second data set to obtain Q, K, and V. The dot product between Q and K is calculated to obtain a similarity matrix, which is then normalized to obtain an attention weight matrix, and finally multiplied by V to obtain a time-frequency correlation matrix. It should be noted that time-frequency correlation may involve a complex relationship between time steps and feature points. Therefore, the data may need to be appropriately arranged or deformed during calculation to ensure that the attention mechanism can capture the required time-frequency relationship.

[0115] In this implementation, the defects of insufficient feature extraction, low computational efficiency and inability to effectively capture complex relationships when processing high-dimensional data in traditional methods are overcome, and efficient and comprehensive feature mining is achieved, providing more accurate data support for subsequent systemic risk prediction and enhancing the predictive performance of the model.

[0116] Step S30, splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data;

[0117] It should be noted that frequency-space cascade data is data formed by splicing spatial correlation and time-frequency correlation along the channel dimension. It integrates spatial and time-frequency characteristics and more comprehensively reflects the characteristic change laws of the meteorological system.

[0118] It is understandable that since traditional methods cannot achieve deep fusion of multi-dimensional features, resulting in insufficient feature utilization, step S30 can avoid the problem of insufficient fusion of multi-dimensional features. By splicing spatial correlation and time-frequency correlation along the channel dimension, deep fusion of multi-dimensional meteorological features is achieved, the expressive ability and discrimination of the features are enhanced, and a more comprehensive and accurate feature representation is provided for subsequent meteorological risk prediction.

[0119] In a feasible implementation, step S30 may include steps S31 to S33:

[0120] Step S31, splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain candidate frequency-space temporal cascade data;

[0121] It should be noted that the candidate frequency spatiotemporal cascade data is the data obtained by splicing the spatial correlation and time-frequency correlation along the channel dimension, which integrates the spatial and time-frequency features, but has not yet undergone further weight adjustment and optimization.

[0122] Additionally, it's important to note that after completing the cascade, positional embeddings are also required. One of the key elements of the Transformer is positional embedding, which indicates the absolute or relative position of each token. For this positional embedding in this model, the classic sine-cosine positional encoding can be used.

[0123] For example, the outputs of the two attention layers are concatenated along the channel dimension (two C / 2 concatenations become C) to obtain the candidate frequency-space cascade data :

[0124]

[0125] Step S32: inputting the candidate frequency-space-time cascade data into a preset first multilayer perceptron and a preset second multilayer perceptron for processing, respectively, to obtain time-domain weight adjustment data output by the first multilayer perceptron and frequency-domain weight adjustment data output by the second multilayer perceptron, wherein the architecture of the second multilayer perceptron sequentially includes an inverse discrete cosine transform part, the architecture of the first multilayer perceptron, and a discrete cosine transform part;

[0126] It should be noted that the time domain weight adjustment data refers to the data output by the first multi-layer perceptron, which reflects the importance weight of the time domain features in the fused feature representation; the frequency domain weight adjustment data refers to the data output by the second multi-layer perceptron, which reflects the importance weight of the frequency domain features in the fused feature representation; the inverse discrete cosine transform part is the part in the second multi-layer perceptron used to convert the frequency domain data back to the time domain data for comparison and fusion with the output of the first multi-layer perceptron; the discrete cosine transform part is the part in the second multi-layer perceptron used to convert the processed time domain data back to the frequency domain data to maintain the consistency and integrity of the data.

[0127] It is understandable that since traditional methods lack an effective weight adjustment mechanism for time domain and frequency domain features when processing fused feature data, the importance of features cannot be dynamically reflected, thereby affecting the prediction performance of the model. Therefore, step S32 is performed, and the time domain weight adjustment data and frequency domain weight adjustment data are obtained by using the first multi-layer perceptron and the second multi-layer perceptron to process the candidate frequency-space cascade data respectively. This can avoid the problem of fixed feature weights and dynamically adjust the weights of time domain and frequency domain features according to data characteristics, thereby enhancing the model's attention to key features and improving the efficiency of feature utilization.

[0128] For example, in order to further narrow the gap between time domain data and frequency domain data, in the feedforward network of this part, the candidate frequency-space cascade data are respectively Time domain data in With frequency domain data Apply the first multi-layer perceptron MLP and the second multi-layer perceptron FreqMLP. In MLP, there can be two layers of MLP. The first layer of MLP become , the second layer MLP then Change back In FreqMLP, the IDCT and DCT components are designed before and after the Vanilla MLP. Specifically, an IDCT transform is added before the two MLP layers, and a DCT transform is added after the two MLP layers. This allows FreqMLP to be viewed as a trainable time-domain filter, allowing for adaptive adjustment of the weights of various components in the frequency-domain feature embedding, further simplifying the utilization of frequency-domain information.

[0129] Step S33: splicing the time-domain weight adjustment data and the frequency-domain weight adjustment data based on the time dimension to obtain frequency-time-space cascade data.

[0130] For example, after the respective applications are completed, the time domain weight adjustment data and the frequency domain weight adjustment data are cascaded into a frequency-time-space cascade data. .

[0131] In this implementation, spatial correlation and time-frequency correlation are first spliced ​​along the channel dimension to form candidate frequency-time-space cascade data. Then, the candidate data are respectively input into a preset first multi-layer perceptron and a second multi-layer perceptron including an inverse discrete cosine transform, a first multi-layer perceptron architecture and a discrete cosine transform for processing to dynamically adjust the weights of time domain and frequency domain features. Finally, the adjusted time domain and frequency domain weight data are spliced ​​along the time dimension to obtain the final frequency-time-space cascade data. This avoids the problems of insufficient fusion of meteorological features, fixed weights and insufficient integration of time-frequency features in traditional methods, and achieves the effect of efficient fusion of spatial and temporal features, dynamic weight adjustment and capturing dynamic changes in the time dimension, thereby providing more accurate and comprehensive data support for systemic risk prediction and improving the accuracy and robustness of the prediction.

[0132] Step S40 , linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension, and the space dimension in sequence to obtain a meteorological risk prediction result of the meteorological system.

[0133] It should be noted that the meteorological risk forecast results are obtained by performing multi-dimensional linear mapping on frequency-time-space cascade data, reflecting the probability or degree of risks that may occur in the meteorological system in the future.

[0134] It is understandable that since traditional methods cannot fully mine the key information of multi-dimensional meteorological characteristics, resulting in insufficient accuracy and robustness of the prediction results, step S40 can avoid the problem of being unable to fully mine the key information of multi-dimensional meteorological characteristics. By performing linear mapping in the time dimension, channel dimension and space dimension in turn, the key features of each dimension are gradually extracted, and finally the meteorological risk prediction results that comprehensively consider multi-dimensional information are obtained, which effectively improves the accuracy and robustness of the prediction.

[0135] Furthermore, it's understandable that meteorological series data reflect the temporal evolution of meteorological systems and the changes in meteorological indicators at different spatial locations. For example, the temporal variations in indicators such as temperature, air pressure, and humidity at different geographic locations (feature points) within meteorological series data follow natural meteorological patterns, such as the diurnal and seasonal variations in temperature within a region, and the correlation between air pressure and weather systems. Meteorological risk forecasts, in essence, explore the meteorological patterns inherent in meteorological series data and quantify risks. By analyzing the temporal evolution of various meteorological indicators within meteorological series data, it's possible to predict potential anomalies or risk states within the meteorological system. For example, when meteorological series data show a sustained, abnormally high temperature in a particular region, accompanied by changes in humidity, this may indicate the occurrence of meteorological risk events such as heat waves.

[0136] For example, a linear regression model is constructed to estimate the systemic meteorological risk at the predicted future time. The regression head can be composed of three simple linear layers (multi-layer perceptrons, MLPs). The first linear layer fuses information at the time level (turning L+L into L_new), the second linear layer fuses information at the channel level (turning C into 1), and the last linear layer fuses information at the feature index level (turning N into 1).

[0137] In a feasible implementation, step S40 may include steps S41 to S43:

[0138] Step S41, performing a first linear mapping on the frequency-time-space cascade data in the time dimension to obtain a first candidate risk prediction result of the meteorological system at each target time point;

[0139] It should be noted that the first candidate risk prediction result refers to the preliminary meteorological risk prediction result obtained after linear mapping of frequency-time-space cascade data in the time dimension, which reflects the risk change trend of the meteorological system at different time points.

[0140] It is understandable that, since it is usually difficult to effectively integrate information in the time dimension when processing multi-dimensional data, the ability to predict the dynamic change patterns in meteorological series data is insufficient, thereby affecting the accuracy of risk prediction. Therefore, step S41 is performed to linearly map the frequency-space cascade data in the time dimension to obtain the key features of each future time point, avoiding the problem that traditional methods cannot effectively integrate time dimension information, and realizing the effective prediction of the dynamic change patterns in future time series data, providing a basis for the subsequent further integration of other dimensional information.

[0141] Step S42, performing a second linear mapping on the first candidate risk prediction result in the channel dimension to obtain a second candidate risk prediction result of the meteorological system under the channel dimension fusion;

[0142] It should be noted that the second candidate risk prediction result refers to the risk prediction result obtained after linear mapping the first candidate risk prediction result in the channel dimension, which integrates the key features of different channels and further optimizes the accuracy of risk prediction.

[0143] It is understandable that, due to the lack of an effective weight adjustment mechanism when fusing features from different channels, the importance of features from each channel cannot be reflected dynamically, which also affects the accuracy of risk prediction. Therefore, step S42 is performed to further fuse information from different channels by linearly mapping the first candidate risk prediction result in the channel dimension, thus avoiding the problem of fixed weights of features from each channel in traditional methods, and realizing dynamic weight adjustment of features from different channels, thereby enhancing the model's attention to key features and improving the efficiency of feature utilization.

[0144] Step S43: performing a third linear mapping on the second candidate risk prediction result in the spatial dimension to obtain a meteorological risk prediction result of the meteorological system under the spatial dimension fusion.

[0145] It is understandable that, since it is often difficult to effectively integrate the spatial relationship between different feature points when processing spatial dimension data, the spatial features are not fully utilized, which affects the comprehensiveness and accuracy of meteorological risk prediction. Therefore, step S43 is performed. By linearly mapping the second candidate risk prediction result in the spatial dimension, the spatial information of different feature points can be further integrated, avoiding the problem that traditional methods cannot effectively integrate spatial dimension information, and realizing deep mining and integration of spatial dimension features, so that the prediction results can more comprehensively reflect the risk status of the meteorological system.

[0146] In this implementation, the frequency-time-space cascade data is first linearly mapped in the time dimension to extract key features at the time point, obtaining the first candidate risk prediction result. A second linear mapping is then performed in the channel dimension to fuse information from different channels and dynamically adjust weights to obtain the second candidate risk prediction result. Finally, a third linear mapping is performed in the spatial dimension to integrate spatial information from different feature points to obtain the final meteorological risk prediction result. This layer-by-layer mapping and fusion approach avoids the shortcomings of traditional methods in integrating multi-dimensional meteorological information, overcomes the limitations of fixed feature weights and single-dimensional analysis, and achieves comprehensive capture of the dynamic laws of meteorological sequence data and deep integration of multi-dimensional features, significantly improving the accuracy and robustness of meteorological risk prediction and providing a more accurate decision-making basis for systematic meteorological risk prediction.

[0147] This embodiment provides a method for predicting systemic risks. By using discrete cosine transform, meteorological series data in the spatiotemporal domain are converted into data in the frequency domain and the original series data are fused to form frequency-time-space fusion data. Then, the spatial correlation and time-frequency correlation of the data are calculated respectively and spliced ​​to form frequency-time-space cascade data. Finally, linear mapping is performed in the time dimension, channel dimension and spatial dimension in sequence to obtain meteorological risk prediction results. This avoids the problems that traditional recurrent neural networks can only process meteorological features in a single time dimension, have difficulty in capturing cross-dimensional correlations of data, and the prediction process is easily disturbed by high-frequency noise, resulting in difficulty in ensuring prediction accuracy. It achieves the goal of taking into account both low-frequency and high-frequency change characteristics through frequency domain processing, overcoming high-frequency noise interference, using spatial and time-frequency correlation calculations to mine feature interactions, and realizing multi-dimensional deep fusion of meteorological features based on the Transformer architecture, ultimately improving the accuracy and robustness of systemic meteorological risk prediction.

[0148] In a feasible implementation, the first data group includes a spatial query vector matrix, a spatial key vector matrix, and a spatial value vector matrix. Step S23 may include steps S231 to S233:

[0149] Step S231: in any calculation process, calculating the first inter-vector similarity according to the spatial query vector matrix and the spatial key vector matrix, and normalizing the first inter-vector similarity to obtain a first attention weight matrix;

[0150] It should be noted that the spatial query vector matrix is ​​the vector matrix used to query feature point information in the attention mechanism, usually represented by Q; the spatial key vector matrix is ​​the key vector matrix used to match the query vector in the attention mechanism, usually represented by K; the first vector similarity refers to the similarity between the query vector and the key vector, usually calculated by dot product or other similarity measurement methods; the first attention weight matrix is ​​the normalized similarity matrix, which represents the attention weights between feature points and is used for subsequent information fusion.

[0151] It is understandable that since simple linear or nonlinear transformations are usually used when calculating the correlation between meteorological feature points, it is difficult to effectively capture the complex relationships and dynamic changes between feature points, resulting in insufficient feature extraction. Therefore, step S231 is performed to calculate the similarity between the spatial query vector matrix and the spatial key vector matrix, and perform normalization processing, thereby avoiding the problem that traditional methods cannot effectively capture the complex relationships between feature points. The similarity between feature points can be dynamically calculated, and the stability and comparability of the calculation results are enhanced through normalization processing, thereby realizing dynamic capture of the complex relationships between feature points and providing more accurate weight information for subsequent spatial correlation calculations.

[0152] For example, first, the spatial query vector matrix and the spatial key vector matrix The matching degree between vectors is measured by matrix multiplication, and the similarity between the first vectors is calculated. . Perform softmax operation and normalization on the obtained similarity (divided by ), the results obtained determine the matching relationship between the element at the current encoding position and the elements at all positions (that is, the first attention weight matrix). It can be understood that, assuming that all elements of Q and K are independent random variables and satisfy zero mean and unit variance, in order to ensure that the variance of the dot product is still 1 regardless of the length of the vector, the dot product is divided by , after scaling, using softmax can make the final weight distribution more evenly between 0 and 1.

[0153] Step S232, calculating a first result vector matrix according to the first attention weight matrix and the spatial value vector matrix;

[0154] It should be noted that the spatial value vector matrix is ​​the value vector matrix containing the specific information of the feature points in the attention mechanism, usually expressed as V; the first result vector matrix refers to the result matrix obtained by multiplying the attention weight matrix and the value vector matrix, which represents the fused feature information.

[0155] It is understandable that due to the lack of an effective weight adjustment mechanism when fusing feature point information, the feature fusion is insufficient and the importance of key feature points cannot be highlighted. Therefore, step S232 is performed to avoid the problem of fixed weights during feature fusion by multiplying the first attention weight matrix with the spatial value vector matrix. The contribution of feature points can be dynamically adjusted according to the attention weight, so that the model pays more attention to key feature points, enhances the richness and discrimination of feature expression, and improves the effect of feature fusion.

[0156] For example, the first attention weight matrix and the spatial value vector matrix V matrix ( ) are multiplied together to obtain the final matrix (i.e. the first result vector matrix, with a dimension of (L+L)*N*(C / 2))). Each result vector is the collective embodiment of all previous V vectors corresponding to different weights (different attention levels).

[0157] Step S233 , connecting the first result vector matrices obtained from multiple calculation processes to obtain the spatial correlation of the first data group.

[0158] It is understandable that, since when processing the multi-head attention mechanism, the outputs of multiple heads are usually simply averaged or spliced, it is difficult to fully tap the complementary information between different heads, resulting in incomplete feature expression. Therefore, step S233 is performed. By connecting the first result vector matrix obtained by multiple calculation processes, the problem of feature information loss caused by simple averaging or splicing can be avoided, and the output of the multi-head attention mechanism is fully integrated, thereby enhancing the feature expression ability and the discrimination ability of the model.

[0159] For example, after linearly projecting the query, key, and value multiple times, multiple self-attentions are applied in parallel, and finally the results are concatenated and output to obtain the spatial correlation of the first data set. This approach is called "multi-head self-attention":

[0160]

[0161]

[0162] Can be written as:

[0163] in, , each W is a learnable weight matrix.

[0164] In this embodiment, during the calculation process, the similarity between the first vectors is first calculated based on the time-frequency query vector matrix and the time-frequency key vector matrix, and then normalized to obtain the first attention weight matrix. This effectively avoids the problem that traditional methods have difficulty in dynamically capturing the complex relationships between meteorological feature points. The first attention weight matrix and the spatial value vector matrix are then used to calculate the first result vector matrix, avoiding the defect of fixed weights during the feature fusion process. Finally, the first result vector matrices calculated multiple times are concatenated to form spatial correlation, overcoming the problem of feature information loss caused by simple averaging or splicing. Overall, through the multi-head attention mechanism and dynamic weight adjustment, the complex relationships between meteorological feature points are fully explored and integrated, significantly improving the richness and accuracy of feature expression, thereby enhancing the model's predictive performance for systemic risk.

[0165] In a feasible implementation, the first data group includes at least a spatial value vector matrix, the second data group includes a time-frequency query vector matrix, a time-frequency key vector matrix, and a time-frequency value vector matrix, and step S24 may include steps S241 to S243:

[0166] Step S241: in any calculation process, calculating the second inter-vector similarity according to the time-frequency query vector matrix and the time-frequency key vector matrix, and normalizing the second inter-vector similarity to obtain a second attention weight matrix;

[0167] It should be noted that the time-frequency query vector matrix is ​​the vector matrix used to query feature point information in the attention mechanism, usually expressed as Q; the time-frequency key vector matrix is ​​the key vector matrix used to match the query vector in the attention mechanism, usually expressed as K; the second inter-vector similarity refers to the similarity between the query vector and the key vector, usually calculated by dot product or other similarity measurement methods; the second attention weight matrix is ​​the normalized similarity matrix, which represents the attention weights between feature points and is used for subsequent information fusion.

[0168] It is understandable that since simple linear or nonlinear transformations are usually used when calculating the correlation between meteorological feature points, it is difficult to effectively capture the complex relationships and dynamic changes between feature points, resulting in insufficient feature extraction. Therefore, step S241 is performed to calculate the similarity between the time-frequency query vector matrix and the time-frequency key vector matrix, and perform normalization processing, thereby avoiding the problem that traditional methods cannot effectively capture the complex relationships between feature points. The similarity between feature points can be dynamically calculated, and the stability and comparability of the calculation results are enhanced through normalization processing, thereby realizing dynamic capture of the complex relationships between feature points and providing more accurate weight information for subsequent time-frequency correlation calculations.

[0169] For example, first, the time-frequency query vector matrix and the time-frequency key vector matrix The matching degree between vectors is measured by matrix multiplication, and the similarity between the second vectors is calculated. The obtained similarity is subjected to softmax operation and normalization, and the obtained results determine the matching relationship between the element at the current encoding position and the elements at all positions (that is, the second attention weight matrix).

[0170] Step S242: Calculate a candidate result vector matrix based on the second attention weight matrix and the time-frequency value vector matrix, and adjust the dimension order of the candidate result vector matrix to be consistent with the spatial value vector matrix to obtain a second result vector matrix;

[0171] It should be noted that the time-frequency value vector matrix is ​​the value vector matrix containing the specific information of the feature points in the attention mechanism, usually expressed as V; the candidate result vector matrix is ​​the intermediate result matrix obtained by multiplying the attention weight matrix and the value vector matrix; the second result vector matrix refers to the candidate result vector matrix after adjusting the dimension order to ensure that its dimension is consistent with the spatial value vector matrix for subsequent processing.

[0172] It is understandable that due to the lack of an effective weight adjustment mechanism when fusing feature point information, the feature fusion is insufficient, and the importance of key feature points cannot be highlighted. Therefore, step S242 is performed to obtain a candidate result vector matrix by multiplying the second attention weight matrix with the time-frequency value vector matrix, and then adjust the dimensional order of the candidate result vector matrix to make it consistent with the spatial value vector matrix, avoiding the problem of fixed weights during feature fusion, and being able to dynamically adjust the contribution of feature points according to the attention weight, highlighting the importance of key feature points, and enhancing the richness and discrimination of feature expression.

[0173] For example, the second attention weight matrix and the V matrix ( ) to obtain the candidate result vector matrix, and then adjust the dimension order of the obtained candidate result data to make it consistent with Keeping it consistent, we get the final matrix (i.e., the second result vector matrix, with a dimension of N* (L+L)* (C / 2)). Each result vector is the collective embodiment of all previous V vectors corresponding to different weights (different attention levels).

[0174] Step S243 , connecting the second result vector matrices obtained from multiple calculation processes to obtain the time-frequency correlation of the second data group.

[0175] It is understandable that, since when processing the multi-head attention mechanism, it is usually difficult to fully tap the complementary information between different heads by simply averaging or splicing the outputs of multiple heads, resulting in incomplete feature expression, step S243 is performed. By connecting the second result vector matrix obtained by multiple calculation processes, the problem of feature information loss caused by simple averaging or splicing is avoided, and the output of the multi-head attention mechanism is fully integrated, thereby enhancing the feature expression ability and the discrimination ability of the model.

[0176] For example, after linearly projecting the query, key, and value multiple times, multiple self-attentions are used in parallel, and the results are concatenated and output. This approach is called "multi-head self-attention":

[0177]

[0178]

[0179] Can be written as:

[0180] in, , each W is a learnable weight matrix.

[0181] In this embodiment, during the calculation process, the similarity between the second vectors is first calculated based on the time-frequency query vector matrix and the time-frequency key vector matrix, and normalized to obtain the second attention weight matrix. This process effectively avoids the problem that traditional methods are difficult to dynamically capture the complex relationship between feature points; then, the candidate result vector matrix is ​​calculated using the second attention weight matrix and the time-frequency value vector matrix, and its dimension order is adjusted to be consistent with the spatial value vector matrix to obtain the second result vector matrix, avoiding the defect of fixed weights in the process of meteorological feature fusion; finally, the second result vector matrices calculated multiple times are connected to form time-frequency correlation, overcoming the problem of feature information loss caused by simple averaging or splicing. Overall, through the multi-head attention mechanism and dynamic weight adjustment, the complex relationship between meteorological feature points is fully explored and integrated, significantly improving the richness and accuracy of feature expression, thereby enhancing the model's predictive performance for systemic risks.

[0182] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above embodiment 1 can be referred to the above introduction and will not be described in detail later. Figure 2 After step S30, the systemic risk prediction method further includes steps S301-S302:

[0183] Step S301: using the frequency-space-time cascade data as the frequency-space-time fusion data, and returning to the step of splitting the frequency-space-time fusion data based on the preset model channel dimension;

[0184] It is understandable that, since only a single processing is usually performed during the feature extraction and fusion process, it is difficult to fully explore the deep features and complex relationships in the data, resulting in insufficient comprehensive and in-depth feature expression, which in turn affects the performance of the prediction model. Therefore, step S301 is performed to gradually build a deeper feature representation by feeding back the frequency-time-space cascade data to the previous step and performing multiple iterative processing, thereby avoiding the problem of insufficient feature extraction caused by a single processing. By stacking multiple iterative processes, the feature extraction and fusion can be gradually deepened, thereby enhancing the model's ability to capture complex meteorological feature relationships.

[0185] Exemplarily, the data dimension of the frequency-time-space cascade data is 2L*N*C, which is consistent with the data dimension of the frequency-time-space fusion data, thereby forming a closed loop, that is, the newly processed data can be used as data input again, thereby repeating the processing multiple times.

[0186] Step S302, when the number of returned executions reaches the preset number of stacking times, based on the frequency-time-space cascade data obtained after returning to execute the said number of stacking times, execute the step of linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension and the space dimension in sequence.

[0187] It should be noted that the number of stacking times refers to the number of times the frequency-space-time cascade data is returned as frequency-space-time fusion data to perform the splitting step, that is, the number of iterative processing times.

[0188] It is understandable that due to the lack of deep integration and optimization mechanism of multi-dimensional features when processing complex feature data, the information of features in different dimensions cannot be fully integrated, affecting the accuracy and robustness of the prediction results. Therefore, step S302 is performed. After reaching the preset number of stacking times, the frequency-time-space cascade data is linearly mapped in the time dimension, channel dimension and space dimension in sequence, thereby avoiding the problem of insufficient multi-dimensional feature integration in traditional methods. More comprehensive and in-depth feature integration can be performed through the frequency-time-space cascade data after multiple stacking, ensuring that the information of features in different dimensions is fully integrated.

[0189] In this embodiment, the frequency-time-space cascade data is fed back to the previous step for multiple iterative processing. Each iteration further deepens the extraction and fusion of features until the preset number of stacking times is reached. Then, the frequency-time-space cascade data after multiple stacking are linearly mapped in the time dimension, channel dimension and spatial dimension in turn. This avoids the problems of insufficient meteorological feature extraction and insufficient multi-dimensional feature integration caused by single processing in traditional methods, overcomes the defect of being unable to fully explore the deep-level features and complex relationships of the data, realizes the progressive refinement of features and deep multi-dimensional integration, and significantly improves the accuracy and robustness of the model for systemic meteorological risk prediction.

[0190] For example, to help understand the implementation process of the systemic risk prediction method obtained by combining this embodiment with the above embodiment 1, please refer to Figure 3 , Figure 3 A brief flowchart of a systematic risk prediction method is provided, specifically:

[0191] The input meteorological sequence data has a dimensionality of 1*L*N. The xy plane of the input example in the figure represents a two-dimensional space spanned by L (time dimension) and N (number of spatial dimensions), with the z-axis representing the amplitude of each element. The discrete cosine transform (DCT) first transforms the input data sequence along the time dimension to the frequency domain, resulting in a data dimension of 1*L*N. The original sequence data is then concatenated with the frequency domain result, resulting in a data dimension of 1*2L*N. After concatenation, the 1*2L*N span is still a two-dimensional plane (ignoring the initial batchsize=1), with one axis representing the number of features N. This two-dimensional plane is then linearly mapped (implemented using a multi-layer perceptron (MLP)) into a high-dimensional feature space of 1*2L*N*C. The frequency-space-time cross-attention module (marked by the dashed line) aggregates information from the frequency, time, and space domains to update the representation of each feature. Finally, based on the learned features, the regression head (consisting of three linear layers) estimates the meteorological risk level for the next L_new time points.

[0192] Furthermore, the entire model architecture is minimized and truth value The mean square error (MSE) between them is optimized, and the loss function L is shown in the formula:

[0193]

[0194] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the systemic risk prediction method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0195] This application also provides a systemic risk prediction system, please refer to Figure 4 , the systemic risk prediction system includes:

[0196] The transformation and fusion module 10 is used to transform the meteorological sequence data recorded by the meteorological system into meteorological frequency domain data, and fuse the meteorological sequence data with the meteorological frequency domain data to obtain frequency-time-space fusion data, wherein the dimensions of the meteorological sequence data include a time dimension and a spatial dimension determined based on the number of feature points, and the feature points are meteorological indicators that have changed in the meteorological system;

[0197] a spatiotemporal calculation module 20 for splitting the frequency-space-time fusion data based on a preset model channel dimension to obtain a first data group and a second data group, and respectively calculating the spatial correlation of the first data group and the time-frequency correlation of the second data group;

[0198] A dimension splicing module 30 is used to splice the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data;

[0199] The mapping output module 40 is used to linearly map the frequency-time-space cascade data in the time dimension, the model channel dimension and the space dimension in sequence to obtain the meteorological risk prediction result of the meteorological system.

[0200] Optionally, the meteorological sequence data at any time point in the meteorological sequence data includes multiple feature points, and the conversion and fusion module 10 is further configured to:

[0201] For any feature point, extract a feature trajectory of the feature point changing with time in the meteorological sequence data;

[0202] Performing discrete cosine transform on the feature trajectory to obtain discrete cosine transform coefficients of the feature point at each time point;

[0203] After traversing each characteristic point, the corresponding characteristic trajectory is encoded according to each discrete cosine transform coefficient to obtain meteorological frequency domain data corresponding to the meteorological sequence data.

[0204] Optionally, the transformation and fusion module 10 is further configured to:

[0205] splicing the meteorological sequence data and the meteorological frequency domain data to obtain candidate frequency spatiotemporal fusion data;

[0206] Linear mapping is performed on the target dimension in the candidate frequency-time-space fusion data to obtain frequency-time-space fusion data including the preset model channel dimension.

[0207] Optionally, the spatiotemporal calculation module 20 is further configured to:

[0208] Performing layer normalization on the frequency-space-time fusion data to obtain a target vector matrix;

[0209] Evenly splitting the target vector matrix based on a preset model channel dimension to obtain a first data group and a second data group;

[0210] Calculating the correlation between the feature points in the first data group multiple times in parallel to obtain the spatial correlation of the first data group;

[0211] The attention affinity between the feature points in the second data group is calculated multiple times in parallel to obtain the time-frequency correlation of the second data group.

[0212] Optionally, the first data group includes a spatial query vector matrix, a spatial key vector matrix, and a spatial value vector matrix, and the spatiotemporal calculation module 20 is further configured to:

[0213] In any calculation process, calculating the first inter-vector similarity according to the spatial query vector matrix and the spatial key vector matrix, and normalizing the first inter-vector similarity to obtain a first attention weight matrix;

[0214] Calculate a first result vector matrix according to the first attention weight matrix and the spatial value vector matrix;

[0215] The first result vector matrices obtained from multiple calculation processes are connected to obtain the spatial correlation of the first data group.

[0216] Optionally, the first data group includes at least a spatial value vector matrix, the second data group includes a time-frequency query vector matrix, a time-frequency key vector matrix, and a time-frequency value vector matrix, and the spatiotemporal calculation module 20 is further configured to:

[0217] In any calculation process, calculating the second inter-vector similarity according to the time-frequency query vector matrix and the time-frequency key vector matrix, and normalizing the second inter-vector similarity to obtain a second attention weight matrix;

[0218] Calculating a candidate result vector matrix according to the second attention weight matrix and the time-frequency value vector matrix, and adjusting the dimension order of the candidate result vector matrix to be consistent with the spatial value vector matrix to obtain a second result vector matrix;

[0219] The second result vector matrices obtained from multiple calculation processes are connected to obtain the time-frequency correlation of the second data group.

[0220] Optionally, the dimension splicing module 30 is further configured to:

[0221] splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain candidate frequency-space-time cascade data;

[0222] Inputting the candidate frequency-space-time cascade data into a preset first multilayer perceptron and a preset second multilayer perceptron for processing, respectively, to obtain time-domain weight adjustment data output by the first multilayer perceptron and frequency-domain weight adjustment data output by the second multilayer perceptron, wherein the architecture of the second multilayer perceptron sequentially includes an inverse discrete cosine transform part, the architecture of the first multilayer perceptron, and a discrete cosine transform part;

[0223] The time domain weight adjustment data and the frequency domain weight adjustment data are spliced ​​based on the time dimension to obtain frequency-time-space cascade data.

[0224] Optionally, the dimension splicing module 30 is further configured to:

[0225] Using the frequency-space-time cascade data as the frequency-space-time fusion data, and returning to perform the step of splitting the frequency-space-time fusion data based on the preset model channel dimension;

[0226] When the number of return executions reaches the preset number of stacking times, based on the frequency-time-space cascade data obtained after returning to execute the stacking times, the step of linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension and the space dimension in sequence is executed.

[0227] Optionally, the mapping output module 40 is further configured to:

[0228] Performing a first linear mapping on the frequency-space-time cascade data in the time dimension to obtain a first candidate risk prediction result of the meteorological system at each target time point;

[0229] Performing a second linear mapping on the first candidate risk prediction result in the channel dimension to obtain a second candidate risk prediction result of the meteorological system under the channel dimension fusion;

[0230] Perform a third linear mapping on the second candidate risk prediction result in the spatial dimension to obtain a meteorological risk prediction result of the meteorological system under the spatial dimension fusion.

[0231] The systemic risk prediction system provided in this application, employing the systemic risk prediction method described in the aforementioned embodiments, can address the technical problem of improving the accuracy and robustness of systemic meteorological risk predictions. Compared to the prior art, the systemic risk prediction system provided in this application has the same beneficial effects as the systemic risk prediction method described in the aforementioned embodiments. Other technical features of the systemic risk prediction system are the same as those disclosed in the aforementioned embodiments and are not further elaborated upon here.

[0232] The present application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the systemic risk prediction method in the above-mentioned embodiment one.

[0233] Reference below Figure 5, which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic devices in the embodiments of the present application may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (such as in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The electronic device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0234] like Figure 5 As shown, the electronic device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in a read-only memory 1002 or programs loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the electronic device. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems may be connected to the input / output interface 1006: an input device 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; a storage device 1003 including, for example, a magnetic tape or hard disk; and a communication device 1009. The communication device 1009 may allow the electronic device to communicate with other devices wirelessly or wired to exchange data. Although the figures show electronic devices with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have instead.

[0235] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device 1003, or installed from a read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are performed.

[0236] The electronic device provided in this application, employing the systemic risk prediction method described in the aforementioned embodiment, can address the technical problem of improving the accuracy and robustness of systemic meteorological risk predictions. Compared to the prior art, the electronic device provided in this application achieves the same beneficial effects as the systemic risk prediction method described in the aforementioned embodiment. Other technical features of this electronic device are the same as those disclosed in the aforementioned embodiment and are not further elaborated upon here.

[0237] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0238] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A systemic risk prediction method, characterized in that: The systemic risk prediction method includes: For any feature point in the meteorological sequence data recorded by the meteorological system at any time point, extract a feature trajectory of the feature point changing with time, wherein the dimensions of the meteorological sequence data include a time dimension and a spatial dimension determined based on the number of feature points, and the feature point is a meteorological indicator that has changed in the meteorological system; Performing discrete cosine transform on the feature trajectory to obtain discrete cosine transform coefficients of the feature point at each time point; After traversing each feature point, encoding the corresponding feature trajectory according to each discrete cosine transform coefficient to obtain meteorological frequency domain data corresponding to the meteorological sequence data, and fusing the meteorological sequence data with the meteorological frequency domain data to obtain frequency-time-space fusion data; Performing layer normalization on the frequency-space-time fusion data to obtain a target vector matrix; Evenly splitting the target vector matrix based on a preset model channel dimension to obtain a first data group and a second data group; Calculating the correlation between the feature points in the first data group multiple times in parallel to obtain the spatial correlation of the first data group; Calculating the attention affinity between the feature points in the second data group multiple times in parallel to obtain the time-frequency correlation of the second data group; splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data; linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension, and the space dimension in sequence to obtain a meteorological risk prediction result of the meteorological system; The first data group includes a spatial query vector matrix, a spatial key vector matrix, and a spatial value vector matrix. The step of calculating the correlation between the feature points in the first data group in parallel multiple times to obtain the spatial correlation of the first data group includes: In any calculation process, calculating the first inter-vector similarity according to the spatial query vector matrix and the spatial key vector matrix, and normalizing the first inter-vector similarity to obtain a first attention weight matrix; Calculate a first result vector matrix according to the first attention weight matrix and the spatial value vector matrix; The first result vector matrices obtained from multiple calculation processes are connected to obtain the spatial correlation of the first data group.

2. The systemic risk prediction method according to claim 1, wherein: The step of fusing the meteorological sequence data and the meteorological frequency domain data to obtain frequency-time-space fusion data comprises: splicing the meteorological sequence data and the meteorological frequency domain data to obtain candidate frequency spatiotemporal fusion data; Linear mapping is performed on the target dimension in the candidate frequency-time-space fusion data to obtain frequency-time-space fusion data including the preset model channel dimension.

3. The systemic risk prediction method according to claim 1, wherein: The first data group includes at least a spatial value vector matrix, the second data group includes a time-frequency query vector matrix, a time-frequency key vector matrix, and a time-frequency value vector matrix, and the step of calculating the attention affinity between each feature point in the second data group in parallel multiple times to obtain the time-frequency correlation of the second data group includes: In any calculation process, calculating the second inter-vector similarity according to the time-frequency query vector matrix and the time-frequency key vector matrix, and normalizing the second inter-vector similarity to obtain a second attention weight matrix; Calculating a candidate result vector matrix according to the second attention weight matrix and the time-frequency value vector matrix, and adjusting the dimension order of the candidate result vector matrix to be consistent with the spatial value vector matrix to obtain a second result vector matrix; The second result vector matrices obtained from multiple calculation processes are connected to obtain the time-frequency correlation of the second data group.

4. The systemic risk prediction method according to claim 1, wherein: The step of splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data includes: splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain candidate frequency-space-time cascade data; Inputting the candidate frequency-space-time cascade data into a preset first multilayer perceptron and a preset second multilayer perceptron for processing, respectively, to obtain time-domain weight adjustment data output by the first multilayer perceptron and frequency-domain weight adjustment data output by the second multilayer perceptron, wherein the architecture of the second multilayer perceptron sequentially includes an inverse discrete cosine transform part, the architecture of the first multilayer perceptron, and a discrete cosine transform part; The time domain weight adjustment data and the frequency domain weight adjustment data are spliced ​​based on the time dimension to obtain frequency-time-space cascade data.

5. The systemic risk prediction method according to claim 1, wherein: After the step of splicing the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data, the following further comprises: Using the frequency-space-time cascade data as the frequency-space-time fusion data, and returning to perform the step of splitting the frequency-space-time fusion data based on the preset model channel dimension; When the number of return executions reaches the preset number of stacking times, based on the frequency-time-space cascade data obtained after returning to execute the stacking times, the step of linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension and the space dimension in sequence is executed.

6. The systemic risk prediction method according to claim 1, wherein: The step of linearly mapping the frequency-time-space cascade data in the time dimension, the model channel dimension, and the space dimension in sequence to obtain the meteorological risk prediction result of the meteorological system includes: Performing a first linear mapping on the frequency-space-time cascade data in the time dimension to obtain a first candidate risk prediction result of the meteorological system at each target time point; Performing a second linear mapping on the first candidate risk prediction result in the channel dimension to obtain a second candidate risk prediction result of the meteorological system under the channel dimension fusion; Perform a third linear mapping on the second candidate risk prediction result in the spatial dimension to obtain a meteorological risk prediction result of the meteorological system under the spatial dimension fusion.

7. A systemic risk prediction system, characterized in that: The method for implementing the systemic risk prediction method according to any one of claims 1 to 6, wherein the systemic risk prediction system comprises: A transformation and fusion module is used to extract, for any feature point of the meteorological sequence data at any time point in the meteorological sequence data recorded by the meteorological system, a feature trajectory of the feature point changing with time, wherein the dimension of the meteorological sequence data includes a time dimension and a spatial dimension determined based on the number of feature points, and the feature point is a meteorological indicator that has changed in the meteorological system; perform discrete cosine transform on the feature trajectory to obtain discrete cosine transform coefficients of the feature point at each time point; after traversing each feature point, encode the corresponding feature trajectory according to each discrete cosine transform coefficient to obtain meteorological frequency domain data corresponding to the meteorological sequence data, and fuse the meteorological sequence data with the meteorological frequency domain data to obtain frequency-time-space fusion data; a spatiotemporal calculation module, configured to perform layer normalization on the frequency-space-time fusion data to obtain a target vector matrix; evenly split the target vector matrix based on a preset model channel dimension to obtain a first data group and a second data group; calculate the correlation between the feature points in the first data group multiple times in parallel to obtain the spatial correlation of the first data group; and calculate the attention affinity between the feature points in the second data group multiple times in parallel to obtain the time-frequency correlation of the second data group; A dimension splicing module, configured to splice the spatial correlation and the time-frequency correlation based on the model channel dimension to obtain frequency-time-space cascade data; a mapping output module, configured to linearly map the frequency-time-space cascade data in the time dimension, the model channel dimension, and the space dimension in sequence to obtain a meteorological risk prediction result of the meteorological system; The first data group includes a spatial query vector matrix, a spatial key vector matrix and a spatial value vector matrix. The spatiotemporal calculation module is also used to calculate the similarity between the first vectors based on the spatial query vector matrix and the spatial key vector matrix in any calculation process, and normalize the similarity between the first vectors to obtain a first attention weight matrix; calculate the first result vector matrix based on the first attention weight matrix and the spatial value vector matrix; and connect the first result vector matrices obtained from multiple calculation processes to obtain the spatial correlation of the first data group.

Images