Methods and systems for analyzing atmospheric composite anomalies applied to meteorological condition monitoring

Abstract

This invention provides a method and system for analyzing atmospheric complex anomalies applied to meteorological state monitoring. It collects atmospheric observation data from different time periods in a target monitoring area, performs spatiotemporal alignment on the data, generates a meteorological state tensor, reconstructs its anomaly-sensitive features, extracts multi-scale anomaly indication features, and inputs these features into a pre-trained coupled identification network. Through feature correlation modeling and cross-element dependency analysis, it generates a correlation pattern matrix. Based on this, it generates a complex anomaly evolution map containing the migration trajectory of the anomaly center through time series extrapolation and spatial diffusion simulation, calculates the spatiotemporal distribution parameters of the anomaly's influence range, and combines the element dependency strength in the correlation pattern matrix to generate an atmospheric complex anomaly assessment result including anomaly development trend prediction. This invention comprehensively considers the spatiotemporal distribution of the anomaly's influence range and element interactions, improving the reliability and accuracy of anomaly development trend prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method and system for analyzing atmospheric complex anomalies applied to meteorological condition monitoring. Background Technology

[0002] In the field of meteorological condition monitoring, atmospheric complex anomaly analysis is a key technology that identifies anomalous phenomena that may trigger extreme weather or environmental risks through systematic observation and data interpretation of atmospheric physical states. Its accuracy directly impacts the effectiveness of meteorological early warning and decision support. Current atmospheric anomaly analysis technologies typically do not fully consider the inherent correlations of atmospheric data across time, space, and elements, ignore the differences in the physical characteristics of the vertical atmospheric stratification, and lack in-depth modeling of the interaction mechanisms between meteorological elements. This results in limitations in capturing complex atmospheric complex anomalies and affects the reliability of predicting anomaly development trends. Therefore, developing anomaly analysis methods that can more accurately characterize the complexity of atmospheric systems has become an important research direction for improving meteorological condition monitoring capabilities. Summary of the Invention

[0003] This invention provides a method and system for analyzing atmospheric complex anomalies applied to meteorological condition monitoring.

[0004] In a first aspect, embodiments of the present invention provide a method for analyzing atmospheric complex anomalies applied to meteorological state monitoring. The method includes: collecting atmospheric observation data from different time periods in a target monitoring area; performing spatiotemporal alignment on the atmospheric observation data to generate a meteorological state tensor with a unified spatiotemporal resolution, wherein the dimensions of the meteorological state tensor include a time series axis, a spatial grid axis, and a meteorological element axis; reconstructing anomaly-sensitive features on the meteorological state tensor to extract multi-scale anomaly indication features reflecting anomaly fluctuations at different atmospheric levels, wherein the scale level of the multi-scale anomaly indication features corresponds to the spatial grid axis resolution of the meteorological state tensor; inputting the multi-scale anomaly indication features into a pre-trained coupled identification network, and generating an association pattern matrix describing the anomaly propagation path between atmospheric elements through feature correlation modeling and cross-element dependency analysis; based on the association pattern matrix, generating a complex anomaly evolution map containing the migration trajectory of the anomaly center through time series extrapolation and spatial diffusion simulation; calculating the spatiotemporal distribution parameters of the anomaly influence range according to the complex anomaly evolution map, and generating an atmospheric complex anomaly assessment result containing anomaly development trend prediction by combining the element dependency strength in the association pattern matrix.

[0005] Secondly, embodiments of the present invention provide a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the atmospheric composite anomaly analysis method for meteorological state monitoring as described above.

[0006] This invention provides an atmospheric complex anomaly analysis method for meteorological state monitoring. By collecting atmospheric observation data and generating a meteorological state tensor containing time series axes, spatial grid axes, and meteorological element axes, it transforms meteorological data from scattered observations to a structured high-dimensional representation, effectively preserving the intrinsic correlation of atmospheric states in multiple spatiotemporal dimensions and providing a more comprehensive data foundation for subsequent anomaly analysis. By reconstructing anomaly-sensitive features from the meteorological state tensor, multi-scale anomaly indication features reflecting anomalous fluctuations at different atmospheric levels are extracted. This allows for targeted capture of anomalous signals at different scales within the vertically layered structure, enhancing sensitivity to complex atmospheric anomalies. The multi-scale anomaly indication features are input into a coupled identification network to generate a correlation pattern matrix. Through feature correlation modeling and cross-element dependency analysis, the path and intensity relationships of anomaly propagation among atmospheric elements are revealed, overcoming the limitations of traditional isolated element analysis and achieving a deep characterization of the multi-element coupling characteristics of the atmospheric system. Based on the correlation pattern matrix, a complex anomaly evolution map containing the migration trajectory of the anomaly center is generated. By combining time series extrapolation with spatial diffusion simulation, the evolution process of anomalies in the spatiotemporal dimensions is dynamically tracked, overcoming the deficiency of static anomaly identification in failing to reflect development trends. Finally, by combining the element dependence intensity in the correlation model matrix, the atmospheric composite anomaly assessment results are generated. By comprehensively considering the spatiotemporal distribution of the anomaly's impact range and the interaction of elements, the reliability and accuracy of the anomaly development trend prediction are improved, providing a more systematic, dynamic, and accurate technical solution for composite anomaly analysis in meteorological condition monitoring. Attached Figure Description

[0007] Figure 1 This is a flowchart of an atmospheric composite anomaly analysis method for meteorological condition monitoring provided by an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 The flowchart below illustrates a method for analyzing atmospheric composite anomalies applied to meteorological state monitoring, as provided in an embodiment of the present invention. This method can be executed by a computer system and may include the following steps: Step S100: Collect atmospheric observation data of different time periods in the target monitoring area, perform spatiotemporal alignment on the atmospheric observation data, and generate a meteorological state tensor with uniform spatiotemporal resolution. The dimensions of the meteorological state tensor include time series axis, spatial grid axis and meteorological element axis.

[0010] Atmospheric observation data refers to various data about atmospheric conditions acquired at different times within a target monitoring area through various meteorological observation equipment and methods. These data include parameters such as temperature, air pressure, humidity, wind speed, and wind direction. Spatiotemporal alignment involves processing atmospheric observation data from different sources and at different temporal and spatial scales to achieve a unified resolution and alignment in both time and space, facilitating subsequent analysis and processing. The meteorological state tensor is a multidimensional data structure containing a time series axis, a spatial grid axis, and a meteorological element axis. The time series axis represents different points in time, the spatial grid axis represents different spatial grids within the target monitoring area, and the meteorological element axis represents different meteorological elements.

[0011] When collecting atmospheric observation data, various observation devices can be used, such as meteorological satellites, ground meteorological stations, and meteorological radars. For example, meteorological satellites can acquire atmospheric data over a wide area, while ground meteorological stations can provide detailed meteorological data for specific locations. For atmospheric observation data collected at different time periods and spatial locations, spatiotemporal alignment processing is required. Specifically, this can be achieved by first determining a unified time interval and spatial grid division, and then interpolating or resampling the raw data according to this unified standard. For example, assuming the time intervals of the raw data are inconsistent—some observations are taken hourly, and others every two hours—all data can be interpolated to an hourly time interval. For the spatial grid, the target monitoring area is divided into grids of a certain size, and the raw data is allocated and interpolated according to the grid, ensuring that each grid has corresponding meteorological element values. Through this spatiotemporal alignment processing, a meteorological state tensor with a unified spatiotemporal resolution is ultimately generated.

[0012] Step S200: Reconstruct the meteorological state tensor using anomaly-sensitive features, and extract multi-scale anomaly indication features that reflect anomalous fluctuations at different atmospheric levels. The scale level of the multi-scale anomaly indication features corresponds to the spatial grid axis resolution of the meteorological state tensor.

[0013] Anomaly-sensitive feature reconstruction processes the meteorological state tensor, highlighting features related to anomalous fluctuations while suppressing features of normal fluctuations, thus making anomalous features more obvious and easier to identify. Multi-scale anomaly indicator features reflect anomalous fluctuations at different atmospheric levels. These features have different scale levels, and the scale level corresponds to the spatial grid axis resolution of the meteorological state tensor. Different scale levels can correspond to different geographical coverage and the fineness of anomalous fluctuations; for example, high resolution corresponds to fine-grained anomalous features over a small area, while low resolution corresponds to overall anomalous features over a large area.

[0014] In one implementation, step S200 may include the following steps S210 to S260: Step S210: Decompose the meteorological state tensor into hierarchical sub-tensors corresponding to different pressure layers according to the atmospheric vertical stratification standard. The sub-tensors retain the time series axis and spatial grid axis information of the original meteorological state tensor. The pressure layer interval of each sub-tensor is dynamically divided according to the atmospheric vertical motion characteristics.

[0015] The atmospheric vertical stratification standard divides the atmosphere into different pressure layers based on the physical properties and vertical motion characteristics of the atmosphere. The stratified meteorological subtensor is a subtensor obtained by dividing the meteorological state tensor according to the atmospheric vertical stratification standard. Each stratified meteorological subtensor corresponds to a specific pressure layer and retains the time series axis and spatial grid axis information of the original meteorological state tensor. The pressure layer intervals are dynamically divided according to the characteristics of atmospheric vertical motion. This involves considering the velocity, direction, and other characteristics of atmospheric vertical motion when dividing the pressure layers, making the divided pressure layers more reflective of the actual atmospheric conditions.

[0016] When performing hierarchical decomposition, the boundaries of different pressure layers can be determined based on standard atmospheric models or actual observed vertical atmospheric motion data. For example, based on the variations of meteorological elements such as temperature and humidity with altitude, the atmosphere can be divided into different pressure layers such as the troposphere and stratosphere. Then, the meteorological state tensor is divided according to these pressure layers to obtain stratified meteorological sub-tensors corresponding to different pressure layers. The data in each stratified meteorological sub-tensor still includes information on the time series axis and the spatial grid axis, which ensures that temporal and spatial variations can be taken into account in subsequent analyses.

[0017] Step S220: Perform spatiotemporal domain joint filtering on each hierarchical meteorological subtensor, capture anomalous fluctuation signals at different scales by dynamically adjusting the spatiotemporal window, and generate an initial set of anomalous features containing local abrupt change features and regional gradual change features. During the dynamic adjustment process, the window size is automatically adapted according to the rate of change of meteorological elements. When the change is drastic, the window is shrunk to capture local details, and when the change is gradual, the window is expanded to capture regional trends.

[0018] Spatiotemporal joint filtering simultaneously filters hierarchical meteorological sub-tensors in both the time and spatial domains to remove noise and interference, highlighting anomalous fluctuation signals. Dynamically adjusting the spatiotemporal window automatically adjusts its size based on the rate of change of meteorological elements to better capture anomalous fluctuation signals at different scales. Local abrupt changes are characteristics of drastic changes in meteorological elements within a short period and a small area, while regional gradual changes are characteristics of continuous changes in meteorological elements over a longer period and a larger area. The initial set of anomalous features, obtained through spatiotemporal joint filtering and dynamic adjustment of the spatiotemporal window, comprises both local abrupt changes and regional gradual changes.

[0019] In one implementation, step S220 may include the following steps 221 to S226: Step S221: Divide the time series axis of the hierarchical meteorological sub-tensor into windows, adjust the window length according to the rate of change of meteorological elements, use a short window to capture fast-changing features when the rate of change exceeds a set threshold, and use a long window to capture slow-changing features when the rate of change is less than a set threshold, and generate a time segmentation result containing fast-changing feature windows and slow-changing feature windows. Each window contains a timestamp and statistical features of the elements within the window.

[0020] Time series axis windowing involves dividing the time series axis of a hierarchical meteorological sub-tensor into different windows according to certain rules. The rate of change of meteorological elements is the amount of change of a meteorological element per unit time. A threshold is a pre-set value used to determine the magnitude of the rate of change of meteorological elements. Rapid change characteristics are those of meteorological elements that change rapidly over a short period, while slow change characteristics are those that change slowly over a longer period. The time segmentation result is obtained after dividing the time series axis, including rapid change characteristic windows and slow change characteristic windows. Each window has a corresponding timestamp and statistical characteristics of the elements within the window, such as the average, maximum, and minimum values ​​of the meteorological elements within the window.

[0021] Step S222: Decompose the spatial grid axis into multiple scales within each time window to generate a spatial scale sequence from high resolution to low resolution. Each spatial scale corresponds to a different geographical coverage area. High resolution corresponds to small-scale fine features, while low resolution corresponds to large-scale overall features. The feature dimensions remain consistent across scales.

[0022] Multi-scale decomposition of spatial grid axes involves decomposing the hierarchical meteorological subtensor into spatial grid axes within each time window, generating spatial scale sequences at different resolutions. High resolution corresponds to fine features over a small area, such as capturing local meteorological anomalies; low resolution corresponds to overall features over a large area, such as reflecting meteorological trends over a larger region. Maintaining consistency of feature dimensions across scales ensures that spatial scale sequences at different resolutions remain identical in feature dimensions for subsequent analysis and processing.

[0023] When performing multi-scale decomposition of spatial grid axes, methods such as wavelet decomposition can be used. For example, for a two-dimensional spatial grid, a two-dimensional wavelet transform can be used to decompose it into sub-images of different resolutions. Specifically, the original spatial grid is taken as the highest resolution scale, and then the next lower resolution scale is obtained through wavelet transform, and so on, until the desired sequence of spatial scales with different resolutions is obtained.

[0024] Step S223: Based on the time segmentation results and spatial scale sequence, calculate the degree of deviation between the meteorological element values ​​in each spatiotemporal window unit and the historical average state. Calculate the historical benchmark value for the same period using the moving average method. The degree of deviation is obtained by the difference between the current value and the benchmark value. Generate an abnormal fluctuation intensity index, which is used to quantify the degree of deviation between the current state and the historical normal state.

[0025] The time segmentation result is the result obtained in step S221, which includes fast-changing feature windows and slow-changing feature windows. The spatial scale sequence is the spatial scale sequence from high resolution to low resolution obtained in step S222. The spatiotemporal window unit is a unit formed by combining the time window and the spatial scale. The historical average state is the average state of meteorological elements within the same historical time period. The abnormal fluctuation intensity index is obtained by calculating the degree of deviation between the current meteorological element value and the historical benchmark value, and is used to quantify the degree of deviation between the current state and the historical normal.

[0026] When calculating the deviation between meteorological element values ​​within each spatiotemporal window unit and the historical average for the same period, the historical baseline value is first calculated using the moving average method. For example, for a given spatiotemporal window unit, meteorological element data from the same time period over several past years are selected, and the average value of these data is calculated using the moving average method as the historical baseline value. Then, the degree of difference between the meteorological element values ​​within the current spatiotemporal window unit and the historical baseline value is calculated, for example, by representing the absolute value of the difference between the two. This degree of difference is used as an indicator of the intensity of abnormal fluctuations; the larger the indicator, the greater the deviation of the current state from the historical norm.

[0027] Step S224: Classify the spatiotemporal window units according to the abnormal fluctuation intensity index. Mark the window units whose abnormal fluctuation intensity index exceeds the deviation threshold determined by the statistical distribution characteristics of historical data of the same period as local mutation feature candidate areas. Mark the window units whose abnormal fluctuation intensity index does not exceed the deviation threshold but shows a continuous changing trend as regional gradual change feature candidate areas. The classification criteria are determined based on the intensity distribution characteristics of historical abnormal events.

[0028] Historical data distribution characteristics refer to the statistical features of meteorological element data within the same historical time period, such as mean and standard deviation. The deviation threshold is a value determined based on the historical data distribution characteristics and used to determine whether the intensity of abnormal fluctuations exceeds the normal range. Local abrupt change candidate areas are spatiotemporal window units where the intensity of abnormal fluctuations exceeds the deviation threshold; these areas may experience drastic changes in meteorological elements within a short period and on a small scale. Regional gradual change candidate areas are spatiotemporal window units where the intensity of abnormal fluctuations does not exceed the deviation threshold but shows a continuous changing trend; these areas may experience continuous changes in meteorological elements over a longer period and on a larger scale. The classification criteria are determined based on the intensity distribution characteristics of historical anomalous events. By analyzing the intensity distribution of past anomalous events, the classification thresholds for local abrupt change characteristics and regional gradual change characteristics are determined.

[0029] Step S225: Calculate the feature gradient for the two types of candidate regions respectively. Use the spatial gradient operator to calculate the spatial gradient of the local mutation feature candidate region to enhance the edge information. Use the smooth gradient to calculate the regional gradient feature candidate region to retain the transition information and enhance the distinguishability of different types of abnormal features.

[0030] Feature gradient is the rate of change of a meteorological element in space or time. Spatial gradient operators are mathematical operators used to calculate spatial gradients, such as the Sobel operator, which can highlight edge information in images or data. Gradient smoothing is a method of smoothing gradient calculations, preserving transitional information in the data and avoiding overly sharp gradient results. Enhancing the discriminability of different types of anomalous features is achieved by calculating feature gradients, making it easier to distinguish between local abrupt changes and regional gradual changes.

[0031] Step S226: Integrate the enhanced features of the two types of candidate regions, arrange them in chronological order and spatial location, and generate an initial set of abnormal features containing location coordinates, intensity change rate and morphological descriptor. Each feature element in the initial set of abnormal features contains a corresponding spatiotemporal coordinate marker to ensure that the source of the features can be traced in subsequent processing.

[0032] The integrated and enhanced candidate region features are obtained by merging and integrating the local abrupt change feature candidate region features and the regional gradual change feature candidate region features processed in step S225. Arranging by time and spatial location means arranging the integrated features in chronological and spatial order for subsequent analysis and processing. Location coordinates are the spatial coordinates of the feature's location, intensity change rate is the rate of change of the meteorological element within the feature region, and morphological descriptors are parameters used to describe the shape and structure of the feature region. The initial anomaly feature set is a feature set containing location coordinates, intensity change rate, and morphological descriptors. Each feature element has a corresponding spatiotemporal coordinate marker, ensuring that the source of the feature can be traced in subsequent processing.

[0033] When integrating the enhanced features of the two types of candidate regions, the features of the local abrupt change feature candidate region and the regional gradual change feature candidate region are merged. Then, these features are sorted according to time order and spatial location. For each feature element, its location coordinates, intensity change rate, and morphological descriptor are calculated. For example, the location coordinates can be determined by the index of the spatial grid, the intensity change rate can be obtained by calculating the ratio of the change in meteorological element values ​​within the feature region to the time interval, and the morphological descriptor can be determined by calculating parameters such as the perimeter and area of ​​the feature region. Finally, corresponding spatiotemporal coordinate markers are added to each feature element to generate an initial set of anomaly features.

[0034] Step S230: Calculate the characteristic gradient difference of adjacent layer meteorological subtensors to construct vertical correlation weights. The vertical correlation weights are obtained by calculating the rate of change of characteristic values ​​of corresponding spatial grid points in adjacent layers, reflecting the vertical propagation intensity and direction of anomalous characteristics at different atmospheric levels.

[0035] Adjacent stratified meteorological subtensors are two adjacent stratified meteorological subtensors within a vertical atmospheric hierarchy. The characteristic gradient difference is the difference in the rate of change of the eigenvalues ​​of adjacent stratified meteorological subtensors at corresponding spatial grid points. The vertical correlation weight is obtained by calculating the characteristic gradient difference between adjacent stratified meteorological subtensors and reflects the intensity and direction of vertical propagation of anomalous features across different atmospheric levels. For example, a high vertical correlation weight indicates a strong propagation intensity of anomalous features between adjacent atmospheric levels; a positive vertical correlation weight indicates that anomalous features propagate from lower to higher layers, and vice versa.

[0036] In one implementation, step S230 may include the following steps S231 to S236: Step S231: Extract the anomalous feature values ​​of adjacent layer meteorological sub-tensors at the same spatial grid position and the same timestamp to obtain the vertical feature sequence. The vertical feature sequence contains the feature values ​​of the corresponding positions of the upper and lower layers to ensure strict alignment of spatiotemporal positions.

[0037] Anomaly eigenvalues ​​are the eigenvalues ​​in the hierarchical meteorological subtensors that are associated with anomalous fluctuations. The vertical feature sequence is obtained by extracting anomalous eigenvalues ​​from adjacent hierarchical meteorological subtensors at the same spatial grid location and timestamp. It includes eigenvalues ​​from corresponding locations in the upper and lower layers, ensuring strict spatiotemporal alignment. This guarantees the accuracy of subsequent calculations, allowing the vertical correlation weights to accurately reflect the vertical propagation of anomalous features across different atmospheric levels.

[0038] Step S232: Calculate the feature gradient difference of the vertical feature sequence. Quantify the intensity and direction of changes in anomalous features at different atmospheric levels through the feature gradient difference. Positive gradient indicates feature enhancement, and negative gradient indicates feature weakening.

[0039] The feature gradient difference is the difference in the rate of change of feature values ​​between two layers in a vertical feature sequence. By calculating the feature gradient difference, the intensity and direction of changes in anomalous features at different atmospheric levels can be quantified. A positive gradient indicates that the anomalous features in the upper layer are stronger than those in the lower layer, meaning that the anomalous features are enhanced in the vertical direction; a negative gradient indicates that the anomalous features in the upper layer are weaker than those in the lower layer, meaning that the anomalous features are weakened in the vertical direction.

[0040] Step S233: Normalize the feature gradient difference and compress it to a uniform numerical range to generate the initial value of the vertical correlation weight.

[0041] Normalization compresses the numerical range of feature gradient differences into a uniform interval, such as [0, 1]. This makes feature gradient differences at different locations and times comparable, facilitating subsequent processing and analysis. The initial value of the vertical correlation weight is obtained from the normalized feature gradient difference; it represents the preliminary calculation result of the vertical correlation weight.

[0042] When performing normalization, a linear normalization method can be used. For example, for all values ​​of the feature gradient difference, first find its maximum and minimum values, then subtract the minimum value from each feature gradient difference value, and then divide by the difference between the maximum and minimum values ​​to obtain the normalized feature gradient difference. Use the normalized feature gradient difference as the initial value for the vertical correlation weights.

[0043] Step S234: Correct the initial weights according to the vertical motion law of the atmosphere. When there are rising airflow characteristics, increase the weight value of the upper layer to the lower layer. When there are descending airflow characteristics, increase the weight value of the lower layer to the upper layer, so that the weight distribution conforms to the physical motion characteristics of the atmosphere.

[0044] When correcting the initial weights based on the vertical motion patterns of the atmosphere, the first step is to determine whether there are updrafts or downdrafts in the atmosphere. This can be done by analyzing meteorological data such as temperature and air pressure to determine the direction of airflow. If updrafts are present, the weight of the upper layer over the lower layer is increased, because updrafts make it easier for anomalous features in the upper layer to affect the lower layer; if downdrafts are present, the weight of the lower layer over the upper layer is increased, because downdrafts make it easier for anomalous features in the lower layer to affect the upper layer. For example, a correction coefficient can be used to adjust the initial weights: when updrafts are present, the weight of the upper layer over the lower layer is multiplied by a correction coefficient greater than 1; when downdrafts are present, the weight of the lower layer over the upper layer is multiplied by a correction coefficient greater than 1.

[0045] Step S235: Spatial smoothing is performed on the corrected weight values, and isolated point noise is eliminated by spatial filtering to ensure the continuity and rationality of the spatial distribution of weights.

[0046] Spatial smoothing is the process of smoothing the corrected weight values ​​spatially to eliminate outlier noise. Outlier noise refers to individual anomalies in the weight value distribution that can affect the continuity and rationality of the spatial distribution of weights. Mean filtering can be used for spatial smoothing. For example, for the corrected weight value of each spatial grid point, the average of the weight values ​​within a certain range around it can be taken as the smoothed weight value for that point.

[0047] Step S236: Associate the processed weight values ​​with the corresponding spatial grid positions to generate a complete vertical correlation weight matrix. The matrix dimension is consistent with the spatial grid dimension of the hierarchical meteorological sub-tensor, which facilitates subsequent cross-layer feature fusion.

[0048] The processed weight values ​​are those corrected in step S234 and spatially smoothed in step S235. Associating the processed weight values ​​with their corresponding spatial grid locations involves mapping the processed weight value of each spatial grid point to its location information, forming a complete weight matrix. The vertical correlation weight matrix is ​​a matrix with the same spatial grid dimension as the stratified meteorological subtensor, containing vertical correlation information of anomalous features between different atmospheric levels. When generating the vertical correlation weight matrix, the spatial grid dimension of the stratified meteorological subtensor is first determined. Then, the processed weight values ​​are arranged according to their spatial grid locations to form a matrix.

[0049] Step S240: Perform cross-layer feature enhancement on the initial abnormal feature set through vertical correlation weights. Integrate the abnormal feature information of the upper layer into the feature expression of the lower layer through weighted summation, while retaining the abnormal fluctuation characteristics of each layer and enhancing the correlation expression of abnormal fluctuations in the vertical direction.

[0050] When performing cross-layer feature enhancement, for each spatial grid location, the anomalous features of the upper layer are weighted according to the weight values ​​in the vertical correlation weight matrix. For example, assuming the anomalous feature value of a certain spatial grid point in the upper layer is A and the vertical correlation weight is w, the weighted anomalous feature value is... Then, the weighted upper-layer anomaly feature values ​​are summed with the anomaly feature values ​​of the corresponding spatial grid points in the lower layer to obtain the cross-layer enhanced anomaly feature values. In this way, upper-layer anomaly feature information is integrated into the lower-layer feature representation, while preserving the anomaly fluctuation characteristics of each layer and enhancing the correlation expression of vertical anomaly fluctuations.

[0051] Step S250: Employ a multi-resolution feature fusion method to scale-align and integrate the anomalous features of different hierarchical meteorological subtensors, and achieve matching of features of different resolutions through Gaussian pyramid decomposition to generate a multi-scale anomalous indicator feature set with hierarchical correlation.

[0052] When using a multi-resolution feature fusion approach to scale-align and integrate anomalous features from different hierarchical meteorological subtenses, the anomalous features are first processed using Gaussian pyramid decomposition. For example, for each hierarchical meteorological subtense's anomalous features, it is used as the top layer of a Gaussian pyramid. Then, downsampling and Gaussian filtering are applied to obtain the next layer's lower-resolution features, and so on, until the desired feature sequences of different resolutions are obtained. Next, features of the same resolution from different hierarchical meteorological subtenses are matched and associated. For example, the similarity between features can be calculated, and features with high similarity can be associated and integrated. Through this process, a multi-scale anomalous indicator feature set with hierarchical association is generated.

[0053] Step S260: Standardize the feature dimensions of the multi-scale anomaly indicator feature set. By eliminating the dimensional differences of features at different atmospheric levels, ensure that the anomaly indicator features at different atmospheric levels have a unified feature space representation, and obtain the final multi-scale anomaly indicator features.

[0054] When standardizing the feature dimensions of a multi-scale anomaly indicator feature set, a standardization method such as Z-score standardization can be used. For each feature dimension in the multi-scale anomaly indicator feature set, its mean and standard deviation are calculated. Then, for each feature value, the mean is subtracted, and the value is divided by the standard deviation to obtain the standardized feature value. By performing this processing on all feature dimensions, the dimensional differences between features from different atmospheric levels are eliminated, ensuring that the anomaly indicator features from different atmospheric levels have a unified feature space representation, thus obtaining the final multi-scale anomaly indicator features.

[0055] Step S300: Input the multi-scale anomaly indicator features into the pre-trained coupled identification network, and generate an association pattern matrix describing the anomaly propagation path between atmospheric elements through feature correlation modeling and cross-element dependency analysis.

[0056] Multi-scale anomaly indicator features are those obtained in steps S200-S260, reflecting anomalous fluctuations at different atmospheric levels. The pre-trained coupled identification network is a pre-trained neural network used to analyze and process the input features. Feature correlation modeling establishes the correlation between multi-scale anomaly indicator features through the coupled identification network, analyzing the mutual influence between different features. Cross-element dependency analysis analyzes the dependencies between different meteorological elements, determining their interactions during anomaly propagation. The correlation pattern matrix is ​​a matrix used to describe the anomalous propagation paths between atmospheric elements; the elements of the matrix represent the anomalous propagation intensity and direction between different meteorological elements.

[0057] In one implementation, step S300 may include the following steps S310-S360: Step S310: Input the multi-scale anomaly indicator features into the feature input layer of the coupled identification network, perform feature dimension mapping, and convert it into a standard feature vector that meets the network input requirements. The dimension of the standard feature vector matches the number of neurons in the hidden layer of the network.

[0058] When mapping the feature dimensions of multi-scale anomaly indicator features into the feature input layer of a coupled identification network, the number of neurons in the hidden layer of the coupled identification network is first determined. Then, based on the number of hidden layer neurons, the multi-scale anomaly indicator features are dimensionality transformed. For example, if the dimension of the multi-scale anomaly indicator features does not match the number of hidden layer neurons, it can be converted into a standard feature vector through linear or nonlinear transformation. A fully connected layer can be used for linear transformation, multiplying the multi-scale anomaly indicator features by a weight matrix and adding a bias vector to obtain the standard feature vector.

[0059] Step S320: The first layer processing unit of the coupled identification network performs local correlation modeling on the standard feature vector, captures the correlation relationship of anomalous features at different spatial locations within the same atmospheric level, generates correlation weights by calculating the similarity of spatial neighborhood features, and generates a local correlation feature map.

[0060] The first processing unit of the coupled identification network is the first layer of the network structure used to process input features. Local correlation modeling analyzes the correlation between anomalous features at different spatial locations within the same atmospheric level. Spatial neighborhood features are anomalous features within a certain range around a given spatial location. Correlation weights are obtained by calculating the similarity of spatial neighborhood features and are used to represent the correlation strength between anomalous features at different spatial locations. The local correlation feature map is obtained through local correlation modeling and is used to describe the correlation between anomalous features at different spatial locations within the same atmospheric level.

[0061] In one implementation, step S320 may include the following steps S321 to S326: Step S321: Divide the standard feature vector according to atmospheric levels, and extract the level feature sub-vectors corresponding to each atmospheric level according to the pressure layer identifier. The level feature sub-vectors retain the time series information and spatial location information of the original standard feature vector to ensure that the level attributes of the features are not lost.

[0062] The standard feature vector is the feature vector obtained in step S310 that meets the input requirements of the coupled identification network. Dividing by atmospheric level involves dividing the standard feature vector into different subsets based on different atmospheric pressure layers. Pressure layer identifiers are used to distinguish different atmospheric levels and can be determined using pressure information from meteorological data. The level-specific feature sub-vectors are feature vectors extracted from the standard feature vectors for each atmospheric level, preserving the time-series and spatial location information of the original standard feature vectors.

[0063] When dividing the standard feature vectors by atmospheric level, the standard feature vectors are first classified according to the pressure layer identifier. For example, based on the range of air pressure values, the standard feature vectors can be divided into different subsets, each subset corresponding to an atmospheric level. Then, the corresponding level feature sub-vectors are extracted from each subset, while retaining both time-series and spatial location information. For example, elements corresponding to a specific atmospheric level in the standard feature vectors can be extracted using an index to form level feature sub-vectors.

[0064] Step S322: Spatial dimension expansion is performed on each level feature sub-vector. The one-dimensional vector is transformed into a two-dimensional grid structure of the hierarchical feature map through dimension transformation. Each node of the grid corresponds to an anomaly feature vector in a spatial location. The grid size is consistent with the spatial grid axis of the hierarchical meteorological sub-tensor.

[0065] The hierarchical feature vectors are the feature vectors corresponding to each atmospheric level obtained in step S321. Spatial dimension unrolling transforms the one-dimensional hierarchical feature vectors into a two-dimensional grid structure to better analyze the correlation between anomalous features at different spatial locations within the same atmospheric level. The hierarchical feature map is a two-dimensional grid structure obtained through spatial dimension unrolling. Each node of the grid corresponds to an anomalous feature vector at a spatial location. The grid size is consistent with the spatial grid axes of the hierarchical meteorological sub-tensor, ensuring spatial consistency between the hierarchical feature map and the original meteorological data.

[0066] Step S323: Calculate the feature similarity between each node in the hierarchical feature map and its surrounding neighboring nodes, and generate a similarity matrix using a preset similarity metric. The element values ​​of the similarity matrix represent the correlation strength between the corresponding spatial location anomaly features, and the higher the value, the stronger the correlation.

[0067] The hierarchical feature map is the feature map of the two-dimensional mesh structure obtained in step S322. Neighboring nodes are nodes within a certain range around a node in the hierarchical feature map. Feature similarity is the degree of similarity between anomalous feature vectors of two spatial locations. Preset similarity metrics are used to calculate feature similarity, such as cosine similarity and Euclidean distance. The similarity matrix is ​​a matrix whose element values ​​represent the correlation strength between anomalous features corresponding to two spatial locations; higher values ​​indicate a stronger correlation. When calculating the feature similarity between each node in the hierarchical feature map and its surrounding neighboring nodes, for each node, its surrounding neighboring nodes are selected. For example, the four neighboring nodes (up, down, left, and right) of the node can be selected. Then, the feature similarity between the node and its neighboring nodes is calculated using the preset similarity metrics.

[0068] Step S324: Construct association weights based on the similarity matrix, adjust the similarity values ​​of each row through normalization, and generate an attention weight distribution map, where the position with a high weight value represents the neighborhood features that have a greater impact on the current node.

[0069] Association weights are used to represent the strength of association between anomalous features at different spatial locations. Normalization adjusts the similarity values ​​in each row of the similarity matrix so that their sum is 1. The attention weight distribution map is the weight distribution map obtained after normalization, where positions with higher weight values ​​represent neighborhood features that have a greater impact on the current node.

[0070] Step S325: The attention weight distribution map and the hierarchical feature map are weighted and combined. The feature of each node is fused with the neighborhood weight through matrix multiplication to generate an enhanced feature representation of each spatial location that incorporates neighborhood information, thereby enhancing local spatial correlation.

[0071] The attention weight distribution map, obtained in step S324, represents the influence of neighborhood features on the current node. The hierarchical feature map is the feature map of the two-dimensional grid structure obtained in step S322. Weighted combination involves matrix multiplication of the attention weight distribution map and the hierarchical feature map, fusing the features of each node with their neighborhood weights. The enhanced feature representation is obtained through weighted combination, fusing neighborhood information at each spatial location and enhancing local spatial correlation.

[0072] Step S326: Arrange all enhanced feature representations according to their spatial location, and obtain a local correlation feature map describing the spatial correlation of anomalous features within the same atmospheric level through combination operations.

[0073] The enhanced feature representation is the feature representation of each spatial location fused with neighborhood information obtained in step S325. Arranging by spatial location means arranging all enhanced feature representations according to their spatial positions in the hierarchical feature map. The combination operation combines the arranged enhanced feature representations into a single feature map. The local correlation feature map is obtained through the combination operation and is used to describe the spatial correlation relationships of anomalous features at different spatial locations within the same atmospheric level.

[0074] Step S330: The second-layer processing unit of the coupled identification network performs cross-level correlation modeling on the local correlation feature map, establishes the vertical propagation path between anomaly features of different atmospheric levels, and generates a hierarchical correlation feature map through vertical feature fusion.

[0075] The second processing unit of the coupled identification network is the network structure used for further processing of local correlation feature maps. Cross-level correlation modeling analyzes the correlation between anomalous features at different atmospheric levels and establishes their vertical propagation paths. Vertical feature fusion merges the local correlation feature maps from different atmospheric levels in the vertical direction to generate a hierarchical correlation feature map. The hierarchical correlation feature map, obtained through cross-level correlation modeling and vertical feature fusion, is used to describe the vertical correlation between anomalous features at different atmospheric levels.

[0076] In one implementation, step S330 may include the following steps S331 to S335: Step S331: Arrange the local correlation feature maps of different atmospheric levels in order of vertical height, and stack them sequentially from low to high to form a three-dimensional feature cube. The height dimension corresponds to the atmospheric level, the horizontal dimension corresponds to the spatial location, and the channel dimension corresponds to the feature type, thus constructing a vertical feature space.

[0077] The local correlation feature maps of different atmospheric levels are the feature maps obtained in steps S320-S326 that describe the spatial correlation of anomalous features within the same atmospheric level. Arranging them in vertical order means arranging the local correlation feature maps of different atmospheric levels according to the vertical altitude of the atmosphere, from lower to higher layers. The three-dimensional feature cube is formed by stacking the local correlation feature maps of different atmospheric levels sequentially; its height dimension corresponds to the atmospheric level, its horizontal dimension corresponds to the spatial location, and its channel dimension corresponds to the feature type. The vertical feature space is obtained by constructing the three-dimensional feature cube and is used to analyze the vertical correlation between anomalous features of different atmospheric levels.

[0078] Step S332: Extract vertical features from the three-dimensional feature cube. Capture the propagation path signals of anomalous features between different atmospheric levels through vertical convolution operations, and extract the associated features in the vertical direction.

[0079] The three-dimensional feature cube is the cubic structure obtained in step S331, used to represent the vertical correlation of anomalous features across different atmospheric levels. Vertical feature extraction extracts feature information in the vertical direction from the three-dimensional feature cube to capture the propagation path signals of anomalous features across different atmospheric levels. Vertical convolution is an operation used to extract features in the vertical direction. By sliding the convolution kernel vertically, convolution calculations are performed on the three-dimensional feature cube to extract the correlation features in the vertical direction.

[0080] Step S333: Based on the extracted propagation path signal, calculate the vertical direction anomaly propagation probability, convert the propagation signal into a probability value within a preset range through probability transformation, and generate a bidirectional propagation probability matrix to describe the propagation probability of the anomaly feature from top to bottom and from bottom to top.

[0081] The propagation path signal is the propagation path information of anomalous features across different atmospheric levels extracted in step S332 through vertical convolution. The vertical anomalous propagation probability is the probability that an anomalous feature propagates vertically from top to bottom or from bottom to top. Probability transformation converts the propagation path signal into probability values ​​within a preset range, such as [0, 1]. The bidirectional propagation probability matrix is ​​a matrix used to describe the propagation probability of anomalous features from top to bottom and from bottom to top; the elements of the matrix represent the probability of anomalous feature propagation between different atmospheric levels.

[0082] When calculating the vertical anomaly propagation probability based on the extracted propagation path signal, the propagation path signal is first processed. A nonlinear function can be used to convert the propagation path signal into probability values. For example, the sigmoid function can be used to map the propagation path signal to the [0, 1] interval. For each atmospheric level pair, the propagation probabilities of the anomaly feature from top to bottom and from bottom to top are calculated, forming a bidirectional propagation probability matrix. For example, the upper triangular part of the matrix represents the propagation probability of the anomaly feature from top to bottom, and the lower triangular part represents the propagation probability of the anomaly feature from bottom to top.

[0083] Step S334: Fuse the bidirectional propagation probability matrix with the local correlation feature map. Combine the propagation probability with the original feature value through element-wise multiplication. Smooth the fusion result vertically to eliminate the discontinuity of feature transition between levels.

[0084] The bidirectional propagation probability matrix, obtained in step S333, describes the probability of anomalous features propagating in the vertical direction. The local correlation feature map, obtained in steps S320-S326, describes the spatial correlation of anomalous features within the same atmospheric level. Fusion involves element-wise multiplication of the bidirectional propagation probability matrix and the local correlation feature map, combining the propagation probability with the original eigenvalues. Vertical smoothing smooths the fusion result in the vertical direction to eliminate discontinuities in feature transitions between levels. When fusing the bidirectional propagation probability matrix and the local correlation feature map, for each spatial location and feature type, the corresponding propagation probability in the bidirectional propagation probability matrix is ​​multiplied by the corresponding eigenvalue in the local correlation feature map. Then, the fusion result is smoothed in the vertical direction. A vertical smoothing filter, such as a vertical mean filter, can be used to filter the fusion result and eliminate discontinuities in feature transitions between levels.

[0085] Step S335: Integrate the processed cross-level features, retain the salient features in the vertical direction through feature combination operations, and generate a hierarchical correlation feature map containing vertical propagation path information. The hierarchical correlation feature map retains the spatial structure information of each atmospheric level, while enhancing the feature correlation in the vertical direction.

[0086] The processed cross-level features are those obtained after fusion and vertical smoothing in step S334. The feature combination operation combines the processed cross-level features, retaining salient features in the vertical direction. The hierarchical correlation feature map is obtained through the feature combination operation; it contains vertical propagation path information, retains the spatial structure information of each atmospheric level, and enhances the vertical feature correlation.

[0087] Step S340: Perform cross-meteorological element dependency analysis based on hierarchical association feature map to identify the mutual influence relationship between different meteorological element anomalous features and quantify the interaction strength between elements.

[0088] When performing cross-meteorological element dependency analysis based on hierarchical association feature maps, the anomalous feature information of different meteorological elements is first extracted from the hierarchical association feature maps. Then, the correlation between the anomalous features of different meteorological elements is calculated to identify their mutual influence relationships. For example, the Pearson correlation coefficient can be used to calculate the correlation between the anomalous features of different meteorological elements. The larger the absolute value of the correlation coefficient, the stronger the mutual influence relationship between the anomalous features of the two meteorological elements. Based on the value of the correlation coefficient, the strength of the interaction between elements is quantified. For example, the absolute value of the correlation coefficient can be used as the quantified value of the interaction strength between elements.

[0089] Step S350: Quantify the intensity of the cross-element dependency relationship, convert the correlation intensity into a numerical representation, and generate a correlation intensity matrix describing the intensity of abnormal propagation among meteorological elements.

[0090] When quantifying the intensity of cross-element dependencies and generating a correlation strength matrix, the interaction strength between elements quantified in step S340 is used as an element of the correlation strength matrix. In this way, the intensity of cross-element dependencies is converted into a numerical representation, generating a correlation strength matrix describing the intensity of anomalous propagation among meteorological elements.

[0091] Step S360: Restructure the correlation strength matrix, rearrange the matrix rows and columns according to meteorological element types, and generate correlation pattern matrices with rows and columns corresponding to different meteorological element types. The element values ​​of the correlation pattern matrix represent the intensity of the abnormal propagation path of the corresponding meteorological element pair.

[0092] The correlation strength matrix, obtained in step S350, describes the intensity of anomalous propagation between meteorological elements. Structural restructuring rearranges the rows and columns of the correlation strength matrix so that each row and column corresponds to a different meteorological element type. The correlation pattern matrix is ​​obtained by restructuring the correlation strength matrix; its elements represent the intensity of the anomalous propagation path for the corresponding meteorological element pair.

[0093] Step S400: Based on the correlation pattern matrix, generate a composite anomaly evolution map containing the migration trajectory of the anomaly center through time series extrapolation and spatial diffusion simulation.

[0094] The correlation pattern matrix, obtained in steps S300-S360, describes the intensity of anomalous propagation paths among atmospheric elements. Time series extrapolation predicts future time points based on historical data variation patterns. Spatial diffusion simulation simulates the spatial diffusion process of anomalous features. The composite anomaly evolution map, obtained through time series extrapolation and spatial diffusion simulation, contains the migration trajectory of the anomaly center and describes the temporal and spatial evolution of atmospheric anomalies.

[0095] In one implementation, step S400 may include the following steps S410-S460: Step S410: Extract time series features from the correlation pattern matrix, calculate the rate of change of matrix element values ​​through a sliding time window, analyze the changing pattern of abnormal propagation paths between different meteorological elements over time, and generate a path evolution feature sequence. The length of the path evolution feature sequence is consistent with the length of the time series axis of the meteorological state tensor.

[0096] When extracting time-series features from the correlation pattern matrix, a sliding time window is used to slide along the time dimension of the correlation pattern matrix. The rate of change is obtained by calculating the difference between matrix element values ​​at adjacent time steps and dividing by the time interval. Then, the changing patterns of anomalous propagation paths among different meteorological elements over time are analyzed. For example, the sign and magnitude of the rate of change can be observed to determine whether the anomalous propagation path is strengthening or weakening. The rate of change and its pattern at each time step are used as an element of the path evolution feature sequence, generating a path evolution feature sequence whose length is consistent with the time-series axis length of the meteorological state tensor.

[0097] Step S420: Perform time series extrapolation based on the path evolution feature sequence, extrapolate the abnormal propagation path change trend in the future preset time period through the historical path evolution pattern, and generate path trend prediction results, including prediction information on propagation direction, intensity and range.

[0098] The path evolution characteristic sequence is the sequence obtained in step S410, used to describe the time-varying patterns of anomalous propagation paths among different meteorological elements. Time series extrapolation is based on historical path evolution patterns to predict the changing trends of anomalous propagation paths over a predetermined period. The path trend prediction results are obtained through time series extrapolation and include prediction information on propagation direction, intensity, and range, used to describe the changes in anomalous atmospheric propagation paths within the predetermined period.

[0099] In one implementation, step S420 may include the following steps S421 to S426: Step S421: Perform dynamic feature decomposition on the path evolution feature sequence, extract path evolution components at different time scales through multi-scale fluctuation separation, and generate a multi-scale feature sequence containing high-frequency fluctuation components and low-frequency trend components.

[0100] The path evolution feature sequence is the sequence obtained in step S410, used to describe the time-varying patterns of anomalous propagation paths among different meteorological elements. Dynamic feature decomposition decomposes the path evolution feature sequence into path evolution components at different time scales. Multi-scale fluctuation separation is a method used to decompose a time series into components at different time scales, such as wavelet decomposition. High-frequency fluctuation components are the rapidly changing components in the path evolution feature sequence over a short period, while low-frequency trend components are the slowly changing trend components over a long period. The multi-scale feature sequence is obtained through dynamic feature decomposition and includes both high-frequency fluctuation components and low-frequency trend components. When performing dynamic feature decomposition on the path evolution feature sequence, a multi-scale fluctuation separation method, such as wavelet decomposition, is used. The path evolution feature sequence is used as input for wavelet decomposition, and it is decomposed into components at different scales through wavelet transform.

[0101] Step S422: Divide the multi-scale feature sequence into a modeling subset and a validation subset in chronological order, so that the subsets maintain temporal continuity and do not overlap.

[0102] The multi-scale feature sequence is the feature sequence obtained in step S421, containing high-frequency fluctuation components and low-frequency trend components. The modeling subset is the dataset used to build a dynamic dependency model of path evolution, and the validation subset is the dataset used to verify the accuracy of the model. Dividing the multi-scale feature sequence into the modeling subset and the validation subset in chronological order ensures that the subsets maintain temporal continuity and do not overlap, thus ensuring the rationality of the model training and validation process.

[0103] Step S423: Construct a dynamic dependency model for path evolution. By capturing the interaction between different fluctuation components in a multi-scale feature sequence, establish a nonlinear mapping relationship between path parameters and time.

[0104] When constructing a dynamic dependency model of path evolution, neural network models, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), can be used. The model is trained by taking multi-scale feature sequences from the modeling subset as input and path parameters as output. During training, the neural network model automatically captures the interaction relationships between different fluctuation components in the multi-scale feature sequences, establishing a nonlinear mapping relationship between path parameters and time. For example, an LSTM model can use memory units to process time series data, capturing the dependencies between different time steps, thereby better establishing the mapping relationship between path parameters and time.

[0105] Step S424: Perform error analysis on the dynamic dependency model using a validation subset, and optimize the model's fitting accuracy to path evolution by adjusting the feature component weights and mapping relationship parameters.

[0106] When performing error analysis on a dynamic dependency model using a validation subset, the multi-scale feature sequences of the validation subset are input into the dynamic dependency model to obtain the predicted path parameters. Then, the error between the predicted and actual values ​​is calculated; for example, mean squared error (MSE) can be used to measure the magnitude of the error. The MSE is calculated by summing the squared differences between each predicted and actual value and then dividing by the number of samples. By calculating the MSE, the model's fit on the validation subset can be evaluated.

[0107] Based on the results of error analysis, the feature component weights and mapping parameters in the dynamic dependency model are adjusted. For example, in a neural network model, the gradient descent algorithm can be used to update the model parameters. The basic idea of ​​gradient descent is to calculate the gradient of the error function with respect to the parameters, and then update the parameters in the opposite direction of the gradient, so that the value of the error function gradually decreases. After each parameter update, error analysis is performed again using a validation subset until the error reaches a satisfactory level or a preset number of iterations is reached. By continuously adjusting the feature component weights and mapping parameters, the model's fitting accuracy to path evolution is optimized, enabling the model to more accurately predict future trends in anomaly propagation paths.

[0108] Step S425: Based on the optimized dynamic dependency model, recursively predict the latest multi-scale feature sequence to generate anomaly propagation path parameters for each future time step, including propagation direction, intensity, and range.

[0109] When recursively predicting the latest multi-scale feature sequence based on the optimized dynamic dependency model, the latest multi-scale feature sequence is input into the optimized dynamic dependency model. The model predicts the anomaly propagation path parameters for future time steps based on the nonlinear mapping relationship of previously learned path parameters over time. For example, for each future time step, the model outputs the anomaly propagation direction, intensity, and range for that time step. The propagation direction can be represented by an angle, the intensity can be quantified numerically, and the range can be described by area or geographical region. This recursive prediction method generates the anomaly propagation path parameters for future time steps.

[0110] Step S426: Perform spatiotemporal consistency verification on the path parameters at each time step, eliminate prediction jumps by using gradient constraints on parameters of adjacent time steps, and generate smooth path trend prediction results.

[0111] When performing spatiotemporal consistency verification on path parameters at each time step, the gradient of path parameters between adjacent time steps is first calculated. For example, for propagation direction, the angular difference of propagation direction between adjacent time steps is calculated; for intensity, the difference of intensity value between adjacent time steps is calculated; for range, the difference of range area between adjacent time steps is calculated. Then, the gradient of parameters between adjacent time steps is constrained according to a preset gradient threshold. If the gradient of a parameter exceeds the threshold, the parameter is considered to have experienced a prediction jump. For parameters with prediction jumps, interpolation or smoothing filtering methods can be used for correction. For example, for jumps in propagation direction, linear interpolation can be used to insert some intermediate values ​​between adjacent time steps to make the change in propagation direction smoother; for jumps in intensity and range, moving average filtering can be used to smooth the parameters. Through such spatiotemporal consistency verification and correction, a smooth path trend prediction result is generated.

[0112] Step S430: Perform spatial dimension analysis on the correlation pattern matrix, identify the spatial distribution pattern of abnormal features through matrix decomposition, and determine the initial anomaly center location and diffusion direction parameters. The initial center location is the spatial grid point with the highest anomaly intensity.

[0113] When performing spatial dimension analysis on the correlation pattern matrix, matrix decomposition is used. Taking singular value decomposition (SVD) as an example, the correlation pattern matrix is ​​decomposed into the product of three matrices. By analyzing the singular values ​​and their corresponding eigenvectors, the spatial distribution patterns of anomalous features can be identified. For example, the spatial distribution corresponding to eigenvectors with larger singular values ​​may represent the main distribution pattern of anomalous features. Then, the spatial grid point with the highest anomalous intensity is found in the correlation pattern matrix and used as the initial anomalous center location. The initial anomalous center location can be determined by traversing each element in the correlation pattern matrix and comparing the element values. For the diffusion direction parameter, it can be determined by analyzing the spatial distribution of anomalous features around the initial anomalous center. For example, the anomalous intensity gradient in different directions around the initial anomalous center can be calculated; the gradient direction represents the possible direction of anomalous diffusion.

[0114] Step S440: Combining the path trend prediction results and spatial diffusion direction parameters, simulate the spatial location change of the anomaly center in the future preset time period to generate the anomaly center migration trajectory.

[0115] In one implementation, step S440 may include the following steps S441 to S446: Step S441: Analyze the propagation direction angle parameter in the path trend prediction result, and construct a dynamic direction transfer probability matrix by combining it with the spatial diffusion direction parameter. The matrix elements represent the probability that the anomaly center will migrate from the current grid position to the neighboring position. The probability value is affected by both the path trend direction and spatial correlation.

[0116] The dynamic direction transition probability matrix is ​​a matrix whose elements represent the probability that an anomaly center will migrate from its current grid location to a neighboring location. The probability value is influenced by both the path trend direction and spatial correlation. This means that the probability of an anomaly center migrating to a certain neighboring location depends not only on the propagation direction of the path trend prediction but also on the spatial correlation between that neighboring location and the current location, such as factors like distance and terrain.

[0117] When analyzing the propagation direction angle parameter in the path trend prediction results and constructing a dynamic direction transition probability matrix, the main direction of anomaly propagation is first determined based on the propagation direction angle parameter. For example, the propagation direction angle is divided into different intervals, each interval corresponding to a general propagation direction. Then, combined with the spatial diffusion direction parameter, the spatial correlation between the current grid position and its neighboring positions is considered. For each neighboring position, its spatial distance and directional relationship with the current grid position are calculated. Euclidean distance can be used to measure spatial distance, and angle difference can be used to measure directional relationship. Based on the path trend direction and spatial correlation, a transition probability is assigned to each neighboring position. For example, if the direction of a neighboring position is consistent with the path trend direction and the spatial distance is relatively short, the transition probability of that neighboring position is high; otherwise, the transition probability is low. These transition probabilities are then combined to form a dynamic direction transition probability matrix.

[0118] Step S442: Perform spatial distribution analysis of anomaly intensity on the correlation pattern matrix. Determine the initial anomaly center location by calculating the weighted average of the anomaly intensity of the spatial grid points. The weighted average comprehensively considers the intensity of the grid point itself and the contribution of the intensity of its neighboring area.

[0119] When performing anomaly intensity spatial distribution analysis on the correlation pattern matrix and determining the initial anomaly center location, the process first iterates through each spatial grid point in the correlation pattern matrix. For each spatial grid point, its neighboring grid points are determined. The range of the neighborhood can be defined according to the actual situation. Then, a weight is assigned to each neighboring grid point. The magnitude of the weight is determined based on the distance between the neighboring grid point and the current grid point; the closer the distance, the greater the weight. For example, a Gaussian weighting function can be used to calculate the weight value based on the distance. Next, the weighted average anomaly intensity of each spatial grid point is calculated using the formula: Weighted average = Own anomaly intensity × Own weight + Σ(Neighboring anomaly intensity × Neighboring weight). Finally, the weighted average anomaly intensity of all spatial grid points is compared, and the spatial grid point with the highest weighted average is determined as the initial anomaly center location.

[0120] Step S443: Based on the dynamic direction transfer probability matrix and the initial anomaly center location, construct a spatial correlation migration model. Calculate the migration probability of the neighborhood grid points at the current location at each prediction time step. The probability value is positively correlated with the diffusion rate of the path trend prediction.

[0121] The dynamic direction transition probability matrix, constructed in step S441, represents the probability of anomaly centers migrating from their current grid location to neighboring locations. The initial anomaly center location is the spatial grid point with the highest anomaly intensity, determined in step S442. The spatial association migration model, built upon the dynamic direction transition probability matrix and the initial anomaly center location, simulates the spatial migration process of anomaly centers. The migration probability of neighboring grid points at each prediction time step is calculated; specifically, at each future time step, the migration probability of neighboring grid points is calculated based on the current anomaly center's location. The probability value is positively correlated with the diffusion rate of the path trend prediction; a higher diffusion rate results in a higher migration probability for neighboring grid points.

[0122] When constructing a spatial association migration model based on the dynamic direction transition probability matrix and the initial anomaly center location, the initial anomaly center location is used as the starting point. At each prediction time step, the transition probability between the current anomaly center location and its neighboring grid points is obtained from the dynamic direction transition probability matrix. Then, the transition probability is adjusted by combining the diffusion rate predicted by the path trend. For example, the transition probability can be multiplied by a scaling factor of the diffusion rate, so that the higher the diffusion rate, the higher the migration probability of the neighboring grid points. The adjusted transition probability is used as the migration probability of the neighboring grid points at the current location at that prediction time step. By continuously repeating this process, the migration probability of the neighboring grid points at the current location is calculated at each prediction time step, thus constructing the spatial association migration model.

[0123] Step S444: The spatial correlation migration model is sampled multiple times using the Monte Carlo simulation method to generate multiple possible anomaly center migration trajectory samples. Each trajectory contains a complete time step location sequence.

[0124] When performing multiple samplings on a spatially correlated migration model using the Monte Carlo simulation method, the first step is to determine the number of samplings, for example, 1000. For each sampling, starting from the initial anomaly center location, random sampling is performed based on the migration probabilities of neighboring grid points at the current location in each prediction time step of the spatially correlated migration model. For example, a random number generator can be used to generate a random number between 0 and 1, and then the cumulative distribution function of the migration probabilities can be used to determine which neighboring grid point the anomaly center will migrate to at that time step. This process is repeated until sampling is completed for all prediction time steps, resulting in one anomaly center migration trajectory sample. By repeatedly performing the above sampling process, multiple possible anomaly center migration trajectory samples are generated, each sample containing the spatial location of the anomaly center at each prediction time step.

[0125] Step S445: Perform statistical significance analysis on the migration trajectory samples, calculate the trajectory density of each spatial location at different time steps, and extract trajectories with densities exceeding a set threshold as candidate main trajectories.

[0126] When performing statistical significance analysis on migration trajectory samples and extracting candidate master trajectories, the first step is to iterate through all migration trajectory samples. For each time step, the trajectory density at each spatial location is calculated. This can be achieved by creating a two-dimensional array, where each element corresponds to a spatial grid point. At each time step, the number of migration trajectories passing through that grid point is recorded in the corresponding array element. Then, based on a set threshold, combinations of spatial locations and time steps with trajectory densities exceeding the threshold are selected. The migration trajectories corresponding to these combinations are then considered as candidate master trajectories.

[0127] Step S446: Optimize the candidate main trajectory using a trajectory smoothing algorithm to eliminate spatial position jumps and time step fluctuations, and generate an anomaly center migration trajectory with spatiotemporal continuity.

[0128] Trajectory smoothing algorithms include moving average filtering and spline interpolation. Candidate master trajectories are those extracted in step S445 whose trajectory density exceeds a set threshold.

[0129] When optimizing candidate master trajectories using trajectory smoothing algorithms, a suitable algorithm can be selected. Taking moving average filtering as an example, for each spatial location in the candidate master trajectory, its average value within a certain time window is calculated. For instance, a time window of length 3 can be defined. For each spatial location at a time step, the average value of that time step and the spatial locations one time step before and after it is calculated, and this average value is used as the smoothed spatial location for that time step. In this way, each time step in the candidate master trajectory is processed to eliminate abrupt changes in spatial location. For fluctuations in time steps, interpolation methods can be used for correction. For example, for sudden changes in spatial location at a time step, linear interpolation can be used to insert intermediate values ​​between adjacent time steps, making the changes at time steps more continuous. Through such trajectory smoothing algorithm optimization, an anomaly center migration trajectory with spatiotemporal continuity is generated.

[0130] Step S450: Integrate the migration trajectory of the anomaly center with the spatial distribution pattern of the anomaly at each time step, arrange the spatial distribution patterns at different times in chronological order, and generate a composite anomaly evolution map framework containing time axis and spatial axis.

[0131] When integrating the anomaly center migration trajectory with the anomaly spatial distribution patterns at each time step to generate a composite anomaly evolution map framework, the first step is to associate each time step in the anomaly center migration trajectory with its corresponding anomaly spatial distribution pattern. For example, for the spatial location of a certain time step in the anomaly center migration trajectory, the corresponding anomaly spatial distribution pattern for that time step is found, and the location of the anomaly center is marked in that spatial distribution pattern. Then, the anomaly spatial distribution patterns at different times are arranged in chronological order. A three-dimensional array can be created, where the first dimension represents the time step, and the second and third dimensions represent the spatial grid. The anomaly spatial distribution pattern for each time step is stored in the corresponding position in the array. Through this integration and arrangement, a composite anomaly evolution map framework containing both a time axis and a spatial axis is generated.

[0132] Step S460: Perform temporal alignment on the composite anomaly evolution map framework. By adjusting the spatial position of different time slices, ensure that the anomaly features of different time slices have spatiotemporal consistency, and obtain the final composite anomaly evolution map.

[0133] When performing time-series alignment on the composite anomaly evolution map framework, the migration trajectory of the anomaly center and the changing patterns of the anomaly spatial distribution pattern at each time step are first analyzed. Based on the migration of the anomaly center, translation and rotation operations are performed on the anomaly spatial distribution pattern in different time slices to ensure the continuity and rationality of the anomaly center's position across different time slices. For example, if the anomaly center moves a certain distance in a certain direction at a certain time step, the corresponding anomaly spatial distribution pattern at that time step is also moved the same distance in the same direction. The anomaly features within the anomaly spatial distribution pattern are also adjusted accordingly to ensure spatiotemporal consistency in their position and morphological changes across different time slices. Through this time-series alignment and adjustment, the final composite anomaly evolution map is obtained.

[0134] Step S500: Calculate the spatiotemporal distribution parameters of the anomaly's influence range based on the composite anomaly evolution map, and generate an atmospheric composite anomaly assessment result that includes anomaly development trend prediction by combining the element dependence strength in the correlation model matrix.

[0135] The composite anomaly evolution map, obtained in step S460, is a spatiotemporally consistent map used to illustrate the temporal and spatial evolution of atmospheric anomalies. The spatiotemporal distribution parameters of the anomaly's influence range are relevant parameters of the anomaly's influence range in different times and spaces, such as the area, shape, and uniformity of distribution of the anomaly's influence range. The element dependence strength in the correlation model matrix represents the strength of the element dependence relationship between different meteorological elements, as indicated by the anomaly propagation path strength in the correlation model matrix obtained in step S360. The atmospheric composite anomaly assessment result is obtained by combining the spatiotemporal distribution parameters of the anomaly's influence range and the element dependence strength in the correlation model matrix. It includes an assessment result predicting the anomaly's development trend and is used to comprehensively assess the development of atmospheric composite anomalies.

[0136] In one implementation, step S500 may include the following steps S510-S560: Step S510: Extract the abnormal region from each time slice of the composite anomaly evolution map, determine the boundary between the abnormal and normal regions by the region segmentation method, and use the region growing algorithm to expand outward from the anomaly center to determine the spatial range of the anomaly influence at different times. Each range includes spatial coordinates and boundary description.

[0137] When extracting anomalous regions and determining the spatial extent of anomalous influence in each time slice of a composite anomaly evolution map, a suitable region segmentation method is first selected. For example, a threshold segmentation method can be used, setting a threshold based on the intensity value of the anomalous features, and regions with intensity values ​​exceeding the threshold are considered candidate anomalous regions. Then, using the anomaly center as a seed point, a region growing algorithm is used to expand outward from the anomaly center. During the region growing process, the neighboring points of the current seed point are checked. If the anomalous feature intensity of a neighboring point is similar to that of the seed point (e.g., within a certain intensity range), the neighboring point is added to the anomalous region and used as a new seed point for further expansion. This process is repeated until no further expansion is possible. For each anomalous influence spatial extent, its spatial coordinates and boundary description are recorded. The spatial coordinates can be determined by the index of the spatial grid points contained in the anomalous region, and the boundary description can be achieved by recording the boundary grid points of the anomalous region.

[0138] Step S520: Calculate the geometric parameters of the anomaly influence range in each time slice, including area size, shape complexity and spatial distribution entropy value. Calculate the area using grid counting, calculate the shape complexity using boundary analysis, calculate the spatial distribution entropy using distribution statistics, and generate a sequence of spatial distribution characteristic parameters.

[0139] The geometric parameters of anomaly impact range describe its geometric characteristics, including area size, shape complexity, and spatial distribution entropy. Grid counting is used to calculate the area of ​​anomaly impact range by counting the number of spatial grid points within it and multiplying that number by the actual geographic area corresponding to each grid point. Boundary analysis is used to calculate shape complexity, such as by calculating the perimeter and curvature of the boundary. Spatial distribution entropy is an indicator of the uniformity of anomaly distribution in space, obtained through statistical analysis of the spatial distribution of anomaly features. The spatial distribution feature parameter sequence is obtained by calculating the geometric parameters of the anomaly impact range in each time slice, with each element corresponding to the spatial distribution feature parameters of a time slice.

[0140] In one implementation, step S520 may include the following steps S521 to S525: Step S521: Grid the spatial range of the anomaly impact, discretize the continuous area into regular grids, count the number of grid cells contained in the anomaly area, calculate the area size of the anomaly impact range based on the actual geographical area corresponding to the grid cells, and express the area value in an appropriate geographical unit.

[0141] When gridding the spatial extent of anomaly impacts and calculating its area, the grid size is first determined based on the actual situation. For example, the spatial extent of the anomaly impact can be divided into square grids of equal size based on the spatial resolution of the composite anomaly evolution map. Then, all grid cells are traversed to determine if they are within the anomaly region. This can be done by checking if the center coordinates of the grid cells are within the boundary of the anomaly region. Grid cells within the anomaly region are counted. Finally, the count is multiplied by the actual geographic area corresponding to each grid cell to obtain the area size of the anomaly impact range, expressed in appropriate geographic units.

[0142] Step S522: Extract the boundary contour lines of the spatial range affected by the anomaly, identify the region boundary through the edge detection algorithm, and calculate the ratio of the perimeter of the contour line to the area as the first indicator of shape complexity, reflecting the degree of irregularity of the region. The larger the ratio, the more complex the shape.

[0143] When extracting the boundary contour of the spatial range affected by anomalies and calculating the first indicator of shape complexity, a suitable edge detection algorithm is first selected. Taking the Canny edge detection algorithm as an example, the spatial range affected by anomalies is represented as a binary image, where the abnormal area is white and the normal area is black. Then, the Canny edge detection algorithm is used to process the binary image to obtain the boundary contour of the spatial range affected by anomalies. Next, the perimeter of the boundary contour is calculated. This can be done by traversing each point on the boundary contour, calculating the distance between adjacent points, and then adding all the distances together to obtain the perimeter. Finally, the perimeter is divided by the area of ​​the spatial range affected by anomalies to obtain the first indicator of shape complexity.

[0144] Step S523: Perform fractal feature analysis on the spatial range of the anomaly's influence, calculate the fractal dimension of the boundary contour as the second indicator of shape complexity, and quantify the complexity of the boundary. The higher the fractal dimension, the more complex the boundary.

[0145] Fractal feature analysis is a method used to analyze the fractal characteristics of objects. Fractals are geometric objects that possess self-similarity and fractional dimension. The fractal dimension of a boundary contour is an indicator used to describe the complexity of the boundary contour; the higher the fractal dimension, the more complex the structure of the boundary contour, with more details and irregularities. The second indicator of shape complexity, obtained by calculating the fractal dimension of the boundary contour, is used to further quantify the shape complexity of the area affected by anomalies.

[0146] Box counting can be used to perform fractal feature analysis and calculate the fractal dimension of the spatial extent of anomaly influence. The basic idea of ​​box counting is to cover the boundary contour line with boxes of different sizes and count the number of boxes required to cover the boundary contour line. Specifically, a series of box sizes of different sizes are selected, for example, starting with larger box sizes and gradually decreasing them. For each box size, the area containing the spatial extent of the anomaly influence is divided into a grid of boxes of the corresponding size. Then, the number of boxes covering the boundary contour line is counted. Based on the relationship between box size and the number of covering boxes, a double logarithmic coordinate graph is plotted, where the horizontal axis is the logarithm of the box size and the vertical axis is the logarithm of the number of covering boxes. By linearly fitting the data points in the double logarithmic coordinate graph, the slope of the fitted line is obtained, and the absolute value of this slope is the fractal dimension of the boundary contour line. The higher the fractal dimension, the more complex the boundary and the more irregular the shape of the anomaly influence range.

[0147] Step S524: Calculate the spatial distribution entropy value of the meteorological element anomaly intensity within the anomaly area, divide the anomaly intensity value into multiple intervals, count the proportion of grid cells in each interval, and calculate the distribution entropy using the information entropy formula to reflect the spatial uniformity of the anomaly intensity distribution. The higher the entropy value, the more uneven the distribution.

[0148] The spatial distribution entropy of meteorological element anomaly intensity within an anomalous region is an indicator used to measure the spatial uniformity of the anomalous intensity distribution within the region. Dividing the anomalous intensity values ​​into multiple intervals involves dividing the range of anomalous intensity values ​​into several non-overlapping intervals. The grid cell percentage of each interval represents the proportion of grid cells belonging to each interval out of the total number of grid cells within the anomalous region.

[0149] When calculating the spatial distribution entropy of meteorological element anomaly intensity within an anomalous region, the first step is to determine the intervals for dividing the anomaly intensity values. For example, the region can be divided into 10 equally wide intervals based on the minimum and maximum anomaly intensities. Next, the number of grid cells in each interval within the anomalous region is counted. Then, the percentage of grid cells in each interval is calculated, which is the number of grid cells in that interval divided by the total number of grid cells in the anomalous region. Finally, the distribution entropy is calculated using the information entropy formula.

[0150] Step S525: Integrate the area size, shape complexity index and spatial distribution entropy value into a feature vector. After normalizing each parameter, arrange them in order and form a feature vector sequence according to the time slice order to generate a spatial distribution feature parameter sequence. Each feature vector corresponds to the abnormal spatial distribution state at a specific time.

[0151] An eigenvector is a vector composed of multiple feature parameters, used to comprehensively represent the geometric features of the anomaly's influence range. Normalization maps the value of each parameter to a uniform range, such as [0, 1], to eliminate dimensional differences between different parameters. The spatial distribution feature parameter sequence is a sequence formed by arranging the feature vectors of each time slice in chronological order, with each feature vector corresponding to the anomaly's spatial distribution state at a specific moment.

[0152] When integrating area size, shape complexity index, and spatial distribution entropy value into a feature vector and generating a spatial distribution feature parameter sequence, the area size, the first shape complexity index, the second shape complexity index, and the spatial distribution entropy value are first normalized. For example, a linear normalization method can be used. For each parameter, its minimum and maximum values ​​are calculated, then the parameter value is subtracted from the minimum value, and then divided by the difference between the maximum and minimum values ​​to obtain the normalized parameter value. The normalized area size, the first shape complexity index, the second shape complexity index, and the spatial distribution entropy value are arranged in order to form a feature vector. For each time slice of the composite anomaly evolution map, the above process is repeated to obtain the feature vector corresponding to each time slice. Finally, these feature vectors are arranged in the order of the time slices to form a spatial distribution feature parameter sequence.

[0153] Step S530: Analyze the variation law of the spatial distribution characteristic parameter sequence over time, calculate the expansion rate, shape change rate and distribution uniformity change trend of the anomaly influence range through the change trend of the parameter sequence, and generate time dynamic characteristic parameters.

[0154] When analyzing the temporal variation of spatial distribution characteristic parameter sequences and generating temporal dynamic characteristic parameters, the changes in area size, shape complexity index, and spatial distribution entropy value over time are first extracted from the spatial distribution characteristic parameter sequences. For the area size variation sequence, the difference in area size between adjacent time steps is calculated and then divided by the time interval to obtain the expansion rate of the anomaly's influence range. For the shape complexity index variation sequence, the difference in shape complexity index between adjacent time steps is similarly calculated and then divided by the time interval to obtain the shape change rate. For the spatial distribution entropy value variation sequence, its trend over time is observed to determine whether it is increasing, decreasing, or remaining constant, and this trend is taken as the distribution uniformity variation trend. The expansion rate, shape change rate, and distribution uniformity variation trend are combined to form the temporal dynamic characteristic parameters.

[0155] Step S540: Extract the dependence strength values ​​between meteorological elements from the correlation pattern matrix, obtain the degree of mutual influence between elements through the matrix element values, construct the element correlation strength network, and identify the core meteorological elements that play a key role in the development of anomalies.

[0156] When extracting dependency strength values ​​between meteorological elements from the association pattern matrix and constructing an element association strength network, the process first iterates through each element in the association pattern matrix, using the values ​​of the matrix elements as the dependency strength values ​​between corresponding pairs of meteorological elements. Then, using meteorological elements as nodes and dependency strength values ​​as edge weights, the element association strength network is constructed. The element association strength network can be represented using an adjacency matrix from graph theory, where the element values ​​are the dependency strength values. Next, by analyzing the topology of the element association strength network, core meteorological elements that play a crucial role in the development of anomalies are identified. For example, indices such as degree centrality and betweenness centrality can be calculated for each node. Degree centrality represents the number of connections between a node and other nodes, while betweenness centrality represents the importance of a node as a bridge in the network. Meteorological elements corresponding to nodes with high degree centrality and betweenness centrality are likely to be core meteorological elements.

[0157] Step S550: Conduct a comprehensive analysis of the spatial distribution characteristic parameters, temporal dynamic characteristic parameters, and the dependence intensity values ​​of core meteorological elements, and establish an anomaly development assessment model through multi-dimensional information fusion.

[0158] In one implementation, step S550 may include the following steps S551-S556: Step S551: Construct an abnormal development assessment dimension system, mapping the area size in the spatial distribution characteristic parameters to the scale dimension, the shape complexity to the morphology dimension, the spatial distribution entropy to the uniformity dimension, and the time dynamic characteristic parameters to the velocity dimension, thus obtaining a multi-dimensional assessment framework.

[0159] The anomaly development assessment dimensional system is a framework for evaluating the development of anomalies, encompassing scale, morphology, uniformity, and velocity dimensions. The spatial distribution characteristic parameters—area size, shape complexity, and spatial distribution entropy—correspond to different physical characteristics, which are mapped to different assessment dimensions to measure different aspects of anomaly development. The temporal dynamic characteristic parameter reflects the rate of change of the anomaly's influence range over time, and is mapped to the velocity dimension. This multi-dimensional assessment framework is obtained by mapping spatial distribution characteristic parameters and temporal dynamic characteristic parameters to different assessment dimensions, and is used to comprehensively assess the development of anomalies.

[0160] When constructing an anomaly development assessment dimension system and obtaining a multi-dimensional assessment framework, the meaning and function of each assessment dimension are first clarified. The scale dimension measures the size of the anomaly's impact range, the morphology dimension measures the shape characteristics of the anomaly's impact range, the uniformity dimension measures the spatial distribution uniformity of the anomaly's intensity, and the velocity dimension measures the rate of change of the anomaly's impact range. Then, the area size among the spatial distribution characteristic parameters is used as an indicator of the scale dimension, shape complexity (including the first and second shape complexity indicators) as an indicator of the morphology dimension, spatial distribution entropy as an indicator of the uniformity dimension, and temporal dynamic characteristic parameters (including expansion rate, shape change rate, and distribution uniformity change trend) as indicators of the velocity dimension. These indicators are combined to form the multi-dimensional assessment framework.

[0161] Step S552: Based on the impact records in the historical abnormal event database, determine the weight coefficients of each evaluation dimension. The weights are dynamically adjusted according to historical data to reflect the importance of different dimensions to the development of the abnormality.

[0162] The historical anomalous event database contains information on past atmospheric anomalies, including records of the impact of these events. Weighting coefficients for each assessment dimension determine the importance of each dimension in the anomaly development assessment. Appropriate weighting allocation involves assigning suitable weighting coefficients to each assessment dimension based on its actual impact on anomaly development. Dynamic adjustment of weighting coefficients based on historical data involves updating and adjusting the weighting coefficients as new historical anomalous event data is added, ensuring that the weighting coefficients accurately reflect the importance of different dimensions in anomaly development.

[0163] When determining the weight coefficients of each evaluation dimension based on the impact records in a historical anomaly event database, the process begins by extracting information related to the anomaly's development from the database. This includes characteristics such as the scale, form, uniformity, and development speed of the anomaly, as well as its final impact. Then, machine learning algorithms, such as linear regression and decision trees, are used to establish a model relating the anomaly's features to its impact. During model building, the indicators of each evaluation dimension are used as input features, and the impact of the anomaly is used as the output label. By training the model, the contribution coefficients of each evaluation dimension to the impact of the anomaly are obtained, and these contribution coefficients are used as the weight coefficients for each evaluation dimension. As new historical anomaly event data accumulates, the model is periodically retrained, and the weight coefficients are updated to achieve dynamic adjustment of the weight coefficients.

[0164] Step S553: ​​Calculate the dependence strength value of the core meteorological elements and the correlation between each assessment dimension. Measure the correlation between the dependence strength of the elements and the dimension parameters through correlation analysis, and select elements with high correlation as key element combinations.

[0165] Correlation degree refers to the degree of correlation between the dependence strength value of core meteorological elements and the parameters of each assessment dimension. Correlation analysis is a method used to measure the degree of correlation between two variables, such as the Pearson correlation coefficient and the Spearman correlation coefficient. Key element combinations are combinations of core meteorological elements with high correlation. These elements may have a strong correlation with one or more assessment dimensions of anomaly development and have an important impact on the development of anomalies.

[0166] When calculating the correlation between the dependence strength values ​​of core meteorological elements and each assessment dimension, and selecting key element combinations, the dependence strength values ​​of the core meteorological elements and the parameters of each assessment dimension are first organized into a data matrix. Then, an appropriate correlation analysis method is selected, such as the Pearson correlation coefficient method. For each core meteorological element's dependence strength value and each assessment dimension's parameter, the Pearson correlation coefficient is calculated. Based on the calculated correlation coefficients, combinations of core meteorological elements and assessment dimensions with high correlation are selected. For example, a correlation threshold can be set, and combinations of core meteorological elements and assessment dimensions with correlation coefficients exceeding the threshold can be considered as key element combinations.

[0167] Step S554: Input the combination of key elements with high correlation and evaluation dimension parameters into the evaluation model framework, construct the abnormal development trend prediction function, and take the dimension parameters and element dependence strength as input and the degree of abnormal development as output.

[0168] When inputting highly correlated key element combinations and evaluation dimension parameters into the evaluation model framework and constructing an anomaly development trend prediction function, the first step is to select a suitable evaluation model framework, such as a neural network model or a support vector machine model. The core meteorological element dependence strength values ​​and evaluation dimension parameters from the highly correlated key element combinations are used as input features, and the degree of anomaly development (which can be quantified based on the impact records in a historical anomaly event database) is used as the output label. Then, the evaluation model framework is trained using training data. During training, the model learns the mapping relationship between input features and output labels, ultimately obtaining the anomaly development trend prediction function. For example, in a neural network model, by continuously adjusting the network weights and biases, the model's output is made as close as possible to the true degree of anomaly development, thus obtaining an accurate anomaly development trend prediction function.

[0169] Step S555: The current abnormal state parameters are extrapolated through the prediction function. By inputting the spatial distribution characteristic parameters, temporal dynamic characteristic parameters and element dependence strength values ​​at the current moment, the predicted value of the abnormal development speed for the future preset period is generated.

[0170] The prediction function is the function constructed in step S554 for predicting the development trend of anomalies. The current anomaly state parameters are the spatial distribution characteristic parameters, temporal dynamic characteristic parameters, and element dependence strength values ​​at the current moment. The future preset time period is a pre-defined time interval used to predict the development of anomalies. The anomaly development speed prediction value is obtained by inputting the current anomaly state parameters into the prediction function and is used to predict the speed of anomaly development within the future preset time period.

[0171] When using a prediction function to extrapolate current anomalous state parameters and generate predicted anomalous development rates for a future preset time period, the spatial distribution characteristic parameters (including area size, shape complexity, and spatial distribution entropy), temporal dynamic characteristic parameters (including expansion rate, shape change rate, and distribution uniformity trend), and element dependence strength values ​​(the dependence strength values ​​of core meteorological elements in highly correlated key element combinations) at the current moment are first organized into an input vector. This input vector is then fed into the prediction function. Based on the learned mapping relationship between the input features and the degree of anomalous development, the prediction function predicts the anomalous development rate for the future preset time period.

[0172] Step S556: Based on the predicted value of the abnormal development speed and the evolution law of the spatial distribution characteristic parameters, extrapolate the changes in the abnormal impact range within the future preset time period, integrate the identification results of key driving factors, and generate an atmospheric composite anomaly assessment result that includes an abnormal development trend curve, an impact range prediction map, and a list of key driving factors.

[0173] When extrapolating the changes in the anomaly's impact range within a predetermined future time period based on the evolution patterns of predicted anomaly development speed and spatial distribution characteristic parameters to generate an atmospheric complex anomaly assessment result, the process begins by using mathematical models or interpolation methods to predict changes in parameters such as the area size, shape complexity, and spatial distribution entropy of the anomaly's impact range within the predetermined future time period, based on the evolution patterns of the predicted anomaly development speed and spatial distribution characteristic parameters. For example, linear interpolation methods can be used to calculate the spatial distribution characteristic parameters at each time step within the predetermined future time period, based on the spatial distribution characteristic parameters at the current and past times, as well as the predicted anomaly development speed. Then, based on the predicted spatial distribution characteristic parameters, an anomaly development trend curve is plotted, with the horizontal axis representing time and the vertical axis representing relevant parameters of the anomaly's impact range (such as area size and shape complexity). Simultaneously, an impact range prediction map is generated based on the predicted spatial distribution of the anomaly's impact range, where different colors or markers can be used to represent areas of different intensities of the anomaly's impact range. Finally, the key driving factor identification results are compiled into a key driving factor list, and the anomaly development trend curve, impact range prediction map, and key driving factor list are combined to generate the atmospheric complex anomaly assessment result.

[0174] Step S560: Generate atmospheric complex anomaly assessment results based on the anomaly development assessment model, including anomaly development speed prediction, impact range prediction, and identification of key driving factors.

[0175] When generating atmospheric complex anomaly assessment results based on the anomaly development assessment model, the spatial distribution characteristic parameters, temporal dynamic characteristic parameters, and dependence intensity values ​​of core meteorological elements at the current moment are input into the anomaly development assessment model. Based on the learned mapping relationship between the input features and the degree of anomaly development, the anomaly development assessment model predicts the anomaly development speed for a predetermined future period, obtaining a predicted anomaly development speed value. Simultaneously, based on the evolution patterns of the predicted anomaly development speed and spatial distribution characteristic parameters, the changes in the anomaly's impact range within the predetermined future period are extrapolated, yielding an impact range prediction result. Furthermore, by analyzing the element correlation intensity network, core meteorological elements that play a crucial role in anomaly development are identified, yielding key driving element identification results. The anomaly development speed prediction, impact range prediction, and key driving element identification results are integrated to generate an atmospheric complex anomaly assessment result containing this information.

[0176] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0177] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0178] In one embodiment, the processor 101 executes the atmospheric composite anomaly analysis method for meteorological condition monitoring provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for analyzing atmospheric composite anomalies applied to meteorological condition monitoring, characterized in that, The method includes: Atmospheric observation data of the target monitoring area at different time periods are collected, and the atmospheric observation data is spatiotemporally aligned to generate a meteorological state tensor with a uniform spatiotemporal resolution. The dimensions of the meteorological state tensor include a time series axis, a spatial grid axis, and a meteorological element axis. Anomaly-sensitive feature reconstruction is performed on the meteorological state tensor to extract multi-scale anomaly indication features reflecting anomalous fluctuations at different atmospheric levels. The scale hierarchy of the multi-scale anomaly indication features corresponds to the spatial grid axis resolution of the meteorological state tensor. Specifically, this includes: hierarchically decomposing the meteorological state tensor into layered meteorological sub-tensors corresponding to different pressure layers according to the atmospheric vertical stratification standard. The layered meteorological sub-tensors retain the time series axis and spatial grid axis information of the original meteorological state tensor, and the pressure layer interval of each layered meteorological sub-tensor is dynamically divided according to the atmospheric vertical motion characteristics; spatiotemporal joint filtering is performed on each layered meteorological sub-tensor, and anomalous fluctuation signals at different scales are captured by dynamically adjusting the spatiotemporal window to generate an initial set of anomaly features containing local abrupt change features and regional gradual change features; and calculating the phase... Vertical correlation weights are constructed using the feature gradient differences of adjacent layer meteorological subtensors. These weights are calculated based on the rate of change of eigenvalues ​​of corresponding spatial grid points in adjacent layers, reflecting the vertical propagation intensity and direction of anomalous features at different atmospheric levels. The initial anomalous feature set is then enhanced across layers using these vertical correlation weights, integrating upper-layer anomalous feature information into lower-layer feature representations while preserving the anomalous fluctuation characteristics of each layer, thus strengthening the correlation expression of vertical anomalous fluctuations. Anomalous features from different layer meteorological subtensors are scale-aligned and integrated, and Gaussian pyramid decomposition is used to match features at different resolutions, generating a multi-scale anomalous indicator feature set with hierarchical correlation. Finally, the feature dimensions of this multi-scale anomalous indicator feature set are standardized to obtain the final multi-scale anomalous indicator features. The multi-scale anomaly indicator features are input into a pre-trained coupled identification network, and an association pattern matrix describing the anomaly propagation path between atmospheric elements is generated through feature correlation modeling and cross-element dependency analysis. Based on the aforementioned correlation pattern matrix, a composite anomaly evolution map containing the migration trajectory of the anomaly center is generated through time series extrapolation and spatial diffusion simulation. The spatiotemporal distribution parameters of the anomaly's impact range are calculated based on the composite anomaly evolution map, and atmospheric composite anomaly assessment results containing anomaly development trend predictions are generated by combining the element dependence strength in the correlation model matrix. The process of performing joint spatiotemporal filtering on each hierarchical meteorological sub-tensor, capturing anomalous fluctuation signals at different scales by dynamically adjusting the spatiotemporal window, and generating an initial set of anomalous features containing local abrupt change characteristics and regional gradual change characteristics includes: The time series axis of the hierarchical meteorological subtensor is divided into windows. The window length is adjusted according to the rate of change of meteorological elements to generate time segmentation results containing fast-changing characteristic windows and slow-changing characteristic windows. Each window contains a timestamp and statistical features of the elements within the window. Within each time window, the spatial grid axis is decomposed into multiple scales to generate a spatial scale sequence. Each spatial scale corresponds to a different geographical coverage area, and the feature dimension consistency is maintained between each scale. Based on the time segmentation results and spatial scale sequence, the deviation of meteorological element values ​​within each spatiotemporal window unit from the historical average state is calculated, and the historical benchmark value is calculated. The deviation is obtained by the difference between the current value and the benchmark value, and an abnormal fluctuation intensity index is generated. The abnormal fluctuation intensity index is used to quantify the deviation of the current state from the historical normal state. The spatiotemporal window units are classified according to the abnormal fluctuation intensity index. Window units whose abnormal fluctuation intensity index exceeds the deviation threshold determined by the statistical distribution characteristics of historical data of the same period are marked as local mutation feature candidate areas. Window units whose abnormal fluctuation intensity index does not exceed the deviation threshold but show a continuous changing trend are marked as regional gradual change feature candidate areas. The classification criteria are determined based on the intensity distribution characteristics of historical abnormal events. The feature gradients of the two types of candidate regions are calculated separately. The spatial gradient operator is used to calculate the spatial gradient of the local mutation feature candidate region to enhance the edge information, and the smooth gradient is used to calculate the regional gradual feature candidate region to retain the transition information, thereby enhancing the distinguishability of different types of abnormal features. The enhanced features of the two candidate regions are integrated and arranged in chronological order and spatial location to generate an initial set of abnormal features containing location coordinates, intensity change rate and morphological descriptor. Each feature element in the initial set of abnormal features contains a corresponding spatiotemporal coordinate marker to ensure that the source of the features can be traced in subsequent processing.

2. The method according to claim 1, characterized in that, The calculation of the characteristic gradient difference between adjacent hierarchical meteorological sub-tensors to construct vertical correlation weights includes: Extract the anomalous feature values ​​of adjacent hierarchical meteorological subtensors at the same spatial grid location and the same time stamp to obtain the vertical feature sequence; Calculate the feature gradient difference of the vertical feature sequence, and quantify the intensity and direction of changes in anomaly features at different atmospheric levels through the feature gradient difference; The feature gradient difference is normalized and compressed to a uniform numerical range to generate the initial value of the vertical correlation weight. The initial weights are corrected according to the vertical motion of the atmosphere. When there are updrafts, the weight of the upper layer over the lower layer is increased; when there are downdrafts, the weight of the lower layer over the upper layer is increased. Spatial smoothing is performed on the corrected weight values, and isolated point noise is eliminated by spatial filtering. The processed weight values ​​are associated with the corresponding spatial grid positions to generate a complete vertical correlation weight matrix.

3. The method according to claim 1, characterized in that, The step of inputting the multi-scale anomaly indicator features into a pre-trained coupled identification network, and generating an association pattern matrix describing the propagation path of anomalies among atmospheric elements through feature correlation modeling and cross-element dependency analysis includes: The multi-scale anomaly indicator features are input into the feature input layer of the coupled identification network, and feature dimension mapping is performed to convert them into standard feature vectors. The dimension of the standard feature vectors matches the number of neurons in the hidden layer of the network. The first layer of the coupled identification network is used to model the local correlation of standard feature vectors, capture the correlation of anomalous features at different spatial locations within the same atmospheric level, generate correlation weights by calculating the similarity of spatial neighborhood features, and generate a local correlation feature map. By using the second-layer processing unit of the coupled identification network to model the cross-level correlation of the local correlation feature map, the vertical propagation path between anomaly features of different atmospheric levels is established, and the hierarchical correlation feature map is generated by vertical feature fusion. Based on the hierarchical correlation feature map, cross-meteorological element dependency analysis is performed to identify the mutual influence relationship between different meteorological element anomalous characteristics and quantify the interaction strength between elements. The intensity of the inter-element dependency is quantified, the correlation intensity is converted into a numerical representation, and a correlation intensity matrix describing the intensity of anomaly propagation among meteorological elements is generated. The correlation strength matrix is ​​restructured, and the matrix rows and columns are rearranged according to the meteorological element type to generate a correlation pattern matrix with rows and columns corresponding to different meteorological element types. The element values ​​of the correlation pattern matrix represent the intensity of the abnormal propagation path of the corresponding meteorological element pair.

4. The method according to claim 3, characterized in that, The first-layer processing unit of the coupled identification network performs local correlation modeling on the standard feature vectors to capture the correlation relationships of anomalous features at different spatial locations within the same atmospheric level, generating a local correlation feature map, including: The standard feature vector is divided according to atmospheric levels, and the level feature sub-vector corresponding to each atmospheric level is extracted according to the pressure layer identifier. The level feature sub-vector retains the time series information and spatial location information of the original standard feature vector. Each level feature vector is expanded in space dimension, and the one-dimensional vector is transformed into a hierarchical feature map with a two-dimensional grid structure through dimension transformation. Each node of the grid corresponds to an anomaly feature vector in a spatial location, and the grid size is consistent with the spatial grid axis of the hierarchical meteorological sub-tensor. Calculate the feature similarity between each node in the hierarchical feature map and its surrounding neighboring nodes to generate a similarity matrix. The element values ​​of the similarity matrix represent the correlation strength between two spatial location anomaly features, with higher values ​​indicating a stronger correlation. Based on the similarity matrix, association weights are constructed, and the similarity values ​​of each row are adjusted through normalization to generate an attention weight distribution map. The attention weight distribution map and the hierarchical feature map are weighted and combined, and the features of each node and the neighborhood weights are fused through matrix multiplication to generate an enhanced feature representation of each spatial location that incorporates neighborhood information. All enhanced feature representations are arranged according to their spatial location, and a combination operation is performed to obtain a local correlation feature map describing the spatial correlation of anomalous features within the same atmospheric level.

5. The method according to claim 4, characterized in that, The second-layer processing unit of the coupled identification network performs cross-level correlation modeling on the local correlation feature map, establishes vertical propagation paths between anomaly features of different atmospheric levels, and generates a hierarchical correlation feature map through vertical feature fusion, including: The local correlation feature maps of different atmospheric levels are arranged in vertical order and stacked from low to high to form a three-dimensional feature cube. The height dimension corresponds to the atmospheric level, the horizontal dimension corresponds to the spatial location, and the channel dimension corresponds to the feature type, thus constructing a vertical feature space. Vertical features are extracted from the three-dimensional feature cube. The propagation path signals of anomalous features between different atmospheric levels are captured by vertical convolution operations, and the correlation features in the vertical direction are extracted. Based on the extracted propagation path signal, the abnormal propagation probability in the vertical direction is calculated, and the propagation signal is converted into a probability value within a preset range through probability transformation to generate a bidirectional propagation probability matrix. The bidirectional propagation probability matrix is ​​fused with the local correlation feature map. The propagation probability is combined with the original feature value through element-wise multiplication. The fusion result is then smoothed in the vertical direction to eliminate the discontinuity of feature transition between levels. The integrated cross-level features are combined and the salient features in the vertical direction are retained through feature combination operations to generate a hierarchical association feature map containing information on the vertical propagation path.

6. The method according to claim 1, characterized in that, The process of generating a composite anomaly evolution map containing the migration trajectory of the anomaly center based on the correlation pattern matrix through time series extrapolation and spatial diffusion simulation includes: Time series features are extracted from the correlation pattern matrix to generate a path evolution feature sequence. The length of the path evolution feature sequence is consistent with the time series axis length of the meteorological state tensor. Based on the path evolution feature sequence, time series extrapolation is performed, and the abnormal propagation path change trend in the future preset period is extrapolated through the historical path evolution pattern to generate path trend prediction results. Spatial dimension analysis is performed on the correlation pattern matrix. The spatial distribution pattern of the abnormal features is identified by matrix decomposition. The initial anomaly center location and diffusion direction parameters are determined. The initial center location is the spatial grid point with the highest anomaly intensity. By combining path trend prediction results and spatial diffusion direction parameters, the spatial location change of the anomaly center in a future preset time period is simulated to generate the migration trajectory of the anomaly center. The migration trajectory of the anomaly center is integrated with the spatial distribution pattern of the anomaly at each time step, and the spatial distribution patterns at different times are arranged in chronological order to generate a composite anomaly evolution map framework that includes a time axis and a spatial axis. The composite anomaly evolution map framework is time-aligned by adjusting the spatial position of different time slices to ensure that the anomaly features of different time slices have spatiotemporal consistency, thus obtaining the final composite anomaly evolution map.

7. The method according to claim 1, characterized in that, The step of calculating the spatiotemporal distribution parameters of the anomalous impact range based on the composite anomaly evolution map, and generating an atmospheric composite anomaly assessment result including anomaly development trend prediction by combining the element dependence strength in the correlation model matrix, includes: Abnormal regions are extracted from each time slice of the composite anomaly evolution map. The boundary between the abnormal and normal regions is determined by the region segmentation method. The region growing algorithm is used to expand outward from the anomaly center to determine the spatial range of the anomaly influence at different times. Each range includes spatial coordinates and boundary description. Calculate the geometric parameters of the anomaly influence range in each time slice, including area size, shape complexity, and spatial distribution entropy. Calculate the area using grid counting, the shape complexity using boundary analysis, and the spatial distribution entropy using distribution statistics to generate a sequence of spatial distribution characteristic parameters. Analyze the variation law of spatial distribution characteristic parameter sequence over time, calculate the expansion rate, shape change rate and distribution uniformity change trend of the anomaly influence range through the change trend of parameter sequence, and generate time dynamic characteristic parameters; The dependency strength values ​​between meteorological elements are extracted from the correlation pattern matrix. The degree of mutual influence between elements is obtained through the matrix element values. An element correlation strength network is constructed to identify the core meteorological elements that play a key role in the development of anomalies. By comprehensively analyzing the spatial distribution characteristic parameters, temporal dynamic characteristic parameters, and the dependence intensity values ​​of core meteorological elements, an anomaly development assessment model is established through multi-dimensional information fusion. Based on the aforementioned abnormal development assessment model, atmospheric complex anomaly assessment results are generated, including predictions of abnormal development speed, predictions of impact range, and identification of key driving factors.

8. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is used to load the computer program to implement the atmospheric complex anomaly analysis method for meteorological condition monitoring as described in any one of claims 1-7.

Images