Weather early warning data pushing method and device based on large model decision

Abstract

The invention provides a weather early warning data pushing method and device based on large model decision, and the method comprises the steps: obtaining the geographic label data and historical early warning response records of a target region user, building an original data set, carrying out the deep semantic modeling of the original data set based on a large model, and building a personalized parameter set. Key early warning element screening is carried out on real-time meteorological data transmitted by a real-time meteorological monitoring network according to the personalized parameter set, the screened early warning elements are converted into an early warning description vector set, and the early warning description vector set and meteorological situation evolution process data output by a meteorological situation prediction system are input into a large-model space-time reasoning network; the method comprises the steps of obtaining an early warning priority evaluation sequence, fusing equipment operation state data and network environment parameters uploaded by a user terminal, performing push strategy optimization calculation on the early warning priority evaluation sequence, and outputting push strategy parameters for pushing. According to the invention, the accuracy and applicability of weather early warning data pushing can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method and device for pushing weather warning data based on large model decision-making. Background Technology

[0002] With the development of meteorological monitoring and information technology, it has become possible to deliver meteorological warnings to target users to provide meteorological risk alerts. Currently, warnings are typically generated based on meteorological observation data and forecast model outputs, and then pushed to user terminals through preset push rules or common user group segmentation methods. However, this method of generating and pushing warnings fails to fully consider the differences in users' perceptions of meteorological risks and their information needs, easily leading to a low degree of matching between the pushed warnings and users' actual needs, thus affecting users' effective reception and response to the warnings. Therefore, how to improve the effectiveness of weather warning data push has become a current research hotspot. Summary of the Invention

[0003] This invention provides a method and device for pushing weather early warning data based on large model decision-making.

[0004] In a first aspect, embodiments of the present invention provide a weather warning data push method based on large-scale model decision-making. The method includes: acquiring geographic tag data and historical warning response records of users in a target area; establishing a raw dataset of user behavior and warning feedback through data association mapping; the raw dataset containing records of user opening times for different warning types, content interaction frequency records, and cross-platform forwarding behavior records; performing deep semantic modeling on the raw dataset based on a large-scale model to analyze the evolution of user behavior patterns under different meteorological risk scenarios, and constructing a personalized parameter set including risk perception sensitivity, information receiving channel preferences, and content presentation format preferences; and filtering key warning elements from real-time meteorological data transmitted via a real-time meteorological monitoring network based on the personalized parameter set, and then using semantic analysis... The encoding process transforms the selected early warning elements into a set of early warning description vectors. These vectors contain semantic descriptions of meteorological element intensity, spatial descriptions of impact range, and temporal descriptions of urgency. The early warning description vectors, along with meteorological situation evolution data output from the meteorological situation prediction system, are input into the spatiotemporal inference network of a large-scale model. This model models the dynamic correlation between changes in meteorological elements and user risk perception, resulting in an early warning priority assessment sequence ranked according to user risk exposure levels. Furthermore, by integrating device operation status data and network environment parameters uploaded from user terminals, the large-scale model's multi-objective decision layer optimizes the push strategy for the early warning priority assessment sequence, outputting push strategy parameters. These parameters are then transformed into multimodal early warning information that meets user interaction habits for subsequent push notifications.

[0005] Secondly, embodiments of the present invention provide a computer device, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the weather warning data push method based on large model decision as described above.

[0006] This invention acquires geographic tag data and historical early warning response records of users in a target area and establishes a raw dataset of user behavior and early warning feedback. This clarifies the semantic correspondence between user behavior and early warning feedback. Based on a large model, deep semantic modeling is performed on the raw dataset, and the evolution of user behavior patterns under different meteorological risk scenarios is analyzed. A personalized parameter set is constructed, including risk perception sensitivity, information receiving channel preferences, and content presentation format preferences, which can more accurately reflect the dynamic differences in users' needs for meteorological early warnings. Based on the personalized parameter set, key early warning elements are selected from real-time meteorological data and converted into an early warning description vector set through semantic feature encoding. This ensures that the selected early warning elements are aligned with users' key concerns. Point matching is used to preserve the semantic connotation of the warning information, achieving semantic alignment between the warning information and the user's risk perception. The warning description vector set and meteorological situation evolution data are input into a large-scale model's spatiotemporal inference network to model the dynamic correlation between changes in meteorological elements and user risk perception. The resulting warning priority evaluation sequence, ranked by user risk exposure level, more accurately reflects the user's level of attention to different warning information. By integrating device operating status data uploaded by user terminals and network environment parameters, and optimizing the push strategy parameters through the multi-objective decision layer of the large-scale model, multimodal warning information that meets user interaction habits is generated. This allows the push strategy to adapt to the real-time status of the terminal, improving user acceptance of the warning information. Through the synergistic effect of the above steps, this method can effectively improve the accuracy and applicability of weather warning data push. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the application environment provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the weather warning data push method based on large model decision-making provided in an embodiment of this application; Figure 3 This is a structural block diagram of the computer device provided in the embodiments of this application. Detailed Implementation

[0008] In some embodiments, the weather warning data push method based on large model decision-making provided in this application is applied to, for example... Figure 1 The application environment shown. For example, as... Figure 1 As shown, the application environment includes a data provider 10 and a computer device 20.

[0009] The data provider 10 can be an electronic device such as a global positioning system, a user terminal, or a PC (Personal Computer), and this embodiment of the application is not limited to this. The computer device 20 can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center. Data transmission between the data provider 10 and the aforementioned computer device 20 occurs via a network.

[0010] Please refer to Figure 2 The diagram illustrates a flowchart of a weather warning data push method based on large model decision-making, according to an embodiment of this application. The steps in this method can be derived from the above... Figure 1 The method is executed by computer device 20. The method may include the following steps:

[0011] Step S100: Obtain geographic tag data and historical early warning response records of users in the target area, and establish a raw dataset of user behavior and early warning feedback through data association mapping. The raw dataset includes records of the timeliness of users opening different types of early warnings, records of content interaction frequency, and records of cross-platform forwarding behavior.

[0012] Geographic tag data reflects the geographical location of users in a target area and can be obtained through technologies such as the Global Positioning System (GPS) and cell tower triangulation. Its purpose is to accurately determine the specific area where a user is located, enabling the creation of highly relevant early warning data based on local weather conditions. Historical early warning response records cover the user's specific behaviors when faced with various weather warnings in the past, including whether the warning was opened, the specific time it was opened, the number of times the user interacted with the warning content, and whether the warning was forwarded to other platforms. Data association mapping is the process of integrating and matching geographic tag data with historical early warning response records. In practice, each user is assigned a unique identifier, which serves as an index to associate geographic tag data with corresponding historical early warning response records, resulting in a unified dataset. In this dataset, each record contains the user's geographical location information and their response behavior to the warning information.

[0013] The opening time record in the original dataset represents the time interval between when a user receives the warning information and when they actually open it. Analyzing this data reveals how promptly users pay attention to different types of warnings; for example, some users may open a rainstorm warning more quickly than a fog warning. The content interaction frequency record reflects how often users interact with the content after opening the warning information, such as the number of times they click on details, view images, or read related links. This record reflects the user's level of interest in the warning content. The cross-platform forwarding behavior record tracks users' sharing of warning information to different online platforms, such as social media groups and personal profiles. Analyzing these records allows us to assess the impact of the warning information's dissemination and its level of attention across different social networks.

[0014] Step S200: Based on the large model, perform deep semantic modeling on the original dataset, analyze the evolution of user behavior patterns under different meteorological risk scenarios, and construct a personalized parameter set that includes risk perception sensitivity, information receiving channel preference, and content presentation format preference.

[0015] Large-scale models, such as those employing deep learning architectures like the Transformers architecture, consist of multiple encoder and decoder layers, each containing multiple self-attention mechanisms and feedforward neural network sublayers. The encoder is responsible for feature extraction and representation learning from the input data, while the decoder makes predictions or generates data based on the encoder's output. Large-scale models are pre-trained on massive amounts of data, learning common patterns and features across various data types, including language, images, and sequences, thus enabling deep semantic analysis of the original datasets.

[0016] Meteorological risk scenarios are different situations categorized based on varying meteorological conditions and potential risks, such as heavy rain scenarios, strong wind scenarios, and high-temperature scenarios. User behavior patterns may change under different meteorological risk scenarios. By analyzing user behavior records under various meteorological risk scenarios, such as opening time, interaction frequency, and forwarding behavior, we can observe trends that change with the scenario, thereby understanding the evolution of user behavior patterns.

[0017] Risk perception sensitivity reflects a user's level of sensitivity to different weather risks. Different users may have different levels of attention and reaction speed to different types of weather risks. For example, some users are more sensitive to lightning risks and will quickly take corresponding preventive measures once a lightning warning is issued; while other users may be more concerned about frost risks.

[0018] In one implementation, step S200 may specifically include the following steps S210-S260: Step S210: Divide the original dataset into meteorological risk scenario levels. Based on the warning type field, decompose the original dataset into multiple scenario sub-datasets. Simultaneously extract the temporal correlation features of user behavior records in each scenario sub-dataset. The temporal correlation features are woven together in chronological order by the continuous timestamps of open time records, content interaction frequency records, and cross-platform forwarding behavior records.

[0019] Meteorological risk scenario stratification is the process of classifying and stratifying different meteorological risks based on meteorological conditions and risk levels. The warning type field is an identifier in the original dataset used to clearly define the type of warning information, such as "rainstorm warning," "strong wind warning," and "high temperature warning." By filtering and classifying the original dataset based on this field, it is decomposed into multiple subsets corresponding to different meteorological risk scenarios. For example, all user behavior records related to rainstorm warnings are extracted to form a rainstorm scenario subset; similarly, records related to strong wind warnings are grouped into a strong wind scenario subset, and so on.

[0020] In each scenario subset, user behavior records are arranged chronologically, including records of open timestamps, content interaction frequency, and cross-platform forwarding behavior. The temporal correlation feature is the result of integrating and associating these different types of behavior records along a timeline.

[0021] Step S220: Perform behavioral trajectory spectrum analysis on the temporal correlation features of each scenario subset, extract behavioral segments of different durations through a sliding window, calculate the dynamic time regularization distance between segments, and aggregate behavioral segments with a distance less than a threshold into behavioral trajectory clusters. Each behavioral trajectory cluster contains a set of user behavioral segments with similar response patterns.

[0022] Behavioral trajectory spectrum analysis is a method used to mine user behavior patterns and regularities. In this embodiment of the invention, a sliding window technique is used to process the temporal correlation features of each scenario subset. The sliding window is a fixed-length time window that slides along the time axis on the temporal correlation features, capturing a fixed-length behavioral segment each time. By adjusting the length of the sliding window and the sliding step size, behavioral segments of different durations and positions can be captured. Dynamic time warping distance is a method to measure the similarity between two time series, handling the problem of temporal asynchrony between two series. When calculating the dynamic time warping distance between behavioral segments, the changing patterns of features such as opening timeliness, content interaction frequency, and cross-platform forwarding behavior in each behavioral segment are compared to find the best matching path between them, and the distance along that path is calculated. This distance reflects the degree of similarity between the two behavioral segments in terms of behavioral patterns. A threshold is a pre-set distance standard used to determine whether two behavioral segments have similar response patterns. The calculated dynamic time warping distance between behavioral segments is compared with this threshold; behavioral segments with a distance less than the threshold are grouped into the same category and aggregated into a behavioral trajectory cluster. The behavioral segments within each behavioral trajectory cluster share similar response patterns. For example, behavioral segments within a certain behavioral trajectory cluster may all exhibit behaviors such as opening the warning information shortly after the warning is issued, and engaging in multiple content interactions and cross-platform forwarding.

[0023] Step S230: Input the behavior trajectory cluster into the large model for context-aware processing, perform weighted modeling on the timestamp sequence in the behavior trajectory cluster, and generate context-enhanced behavior vectors. In the context-enhanced behavior vectors, the feature weight of behavior segments with newer timestamps is higher than that of behavior segments with older timestamps.

[0024] Context-aware processing is the process by which the large model understands the specific context and background information represented by the behavioral trajectory clusters. After receiving the behavioral trajectory clusters, the large model analyzes and processes the timestamp sequence, which records the chronological order of behavioral segments and reflects changes in user behavior over time. Weighted modeling assigns different weights to each behavioral segment in the timestamp sequence to reflect the importance of behavior at different points in time to the current context. Since newer behavioral segments better reflect the user's current behavioral patterns and needs, they are given higher weights in weighted modeling. By comprehensively considering the features of the timestamp sequence and behavioral segments, and combining this weighted information, the large model models and extracts features from the behavioral trajectory clusters, generating a context-enhanced behavior vector. This vector contains the contextual information of the behavioral trajectory clusters and the features of behavioral segments at different points in time, highlighting the impact of newer behavioral segments, enabling the model to more accurately capture the user's current behavioral patterns and needs.

[0025] In one implementation, step S230 may specifically include the following steps S231-S236: Step S231: Normalize the user behavior segments in the behavior trajectory cluster by time granularity, and interpolate the behavior segments of different durations into time series of equal length. Each time series contains a fixed number of timestamp nodes, and each node corresponds to a user behavior record of one time unit.

[0026] The purpose of time granularity normalization is to address the issue of inconsistent durations among different behavioral segments within a behavioral trajectory cluster, enabling them to be compared and analyzed on the same time scale. Within a behavioral trajectory cluster, the duration of behavioral segments may vary due to differences in user activity and time distribution. Interpolation is used to convert these behavioral segments of varying lengths into time series of equal length. The interpolation process estimates the missing time points based on known behavioral segment data, thus expanding the behavioral segments into time series containing a fixed number of timestamp nodes. Each timestamp node corresponds to a user behavior record within a time unit, such as per second, per minute, or per hour, with the specific time unit determined based on actual needs and data characteristics. Time granularity normalization ensures that each time series has the same temporal resolution and length. The normalization method can employ general normalization techniques, without specific limitations.

[0027] Step S232: Extract the behavioral feature triplet for each timestamp node. The behavioral feature triplet consists of the opening time deviation value, the cumulative interaction frequency value, and the forwarding platform entropy value. The forwarding platform entropy value represents the uniformity of the distribution of user forwarding behavior across different platforms.

[0028] After time granularity normalization, behavioral feature triples are extracted for each timestamp node. The opening timeliness deviation value represents the deviation between the user's opening time of the warning information and the average opening time within the time corresponding to that timestamp node. By calculating the opening timeliness deviation value, we can understand whether the user's timely opening of the warning information at that time point deviates from the normal level. The cumulative interaction frequency value is the total number of times the user interacts with the warning content from the start of the behavior to that timestamp node, such as the cumulative number of operations such as clicking details, viewing images, and reading text descriptions. It reflects the user's attention to and participation in the warning information before that time point. The forwarding platform entropy value is an indicator used to measure the uniformity of user forwarding behavior across different network platforms. If the user's forwarding behavior is evenly distributed across multiple platforms, the forwarding platform entropy value is high; if the user mainly forwards the warning information to a few platforms, the forwarding platform entropy value is low. By analyzing the forwarding platform entropy value, we can understand the user's forwarding behavior patterns and social dissemination tendencies.

[0029] Step S233: Input the behavioral feature triples into the query generation layer in timestamp order to generate a query vector for each timestamp node, and use the preset context feature template as the key vector and value vector to calculate the similarity score between the query vector and the key vector.

[0030] The query generation layer is a component of the larger model that receives and processes behavioral feature triples ordered by timestamp. For example, this layer includes a linear transformation sublayer and an activation function sublayer. The linear transformation sublayer consists of a set of learnable weight matrices and bias vectors. When behavioral feature triples are input to this sublayer in timestamp order, the linear transformation sublayer performs a linear transformation on the input triples. Specifically, it multiplies the input behavioral feature triples by the weight matrix and adds the bias vector to obtain an intermediate vector. The activation function sublayer, following the linear transformation sublayer, performs a non-linear transformation on the intermediate vector. Activation functions such as ReLU and Sigmoid are used to introduce non-linearity and enhance the model's expressive power. After processing by the activation function, the intermediate vector is converted into query vectors, each query vector corresponding to a timestamp node and containing important information about the behavioral feature triples at that timestamp node.

[0031] The preset context feature templates are a predefined set of feature vectors representing different meteorological risk scenarios and user behavior contexts. These templates are designed based on historical data and domain knowledge, and include various possible context features. In this embodiment of the invention, the preset context feature templates are used as key vectors and value vectors, respectively. The key vectors are used to calculate similarity with the query vector, while the value vectors are used in the subsequent weighted summation process.

[0032] The similarity score between the query vector and the key vector is calculated using similarity calculation methods, such as the cosine similarity algorithm. The similarity score reflects the degree of matching between the behavioral characteristics of the current timestamp node and the preset contextual feature template. The higher the score, the more similar the two are, indicating that the current behavior may correspond to a specific context.

[0033] Step S234: Based on the similarity score, perform a weighted summation of the value vectors to generate a timestamp-level context vector. The dimension of the timestamp-level context vector is the same as that of the value vector, and it contains the interaction information between the context feature template and the behavioral feature triplet.

[0034] After obtaining the similarity scores between the query vector and the key vector, the value vectors are weighted and summed based on these scores. Specifically, each value vector is multiplied by its corresponding similarity score, and then all weighted value vectors are summed to obtain the timestamp-level context vector. Since the similarity score reflects the degree of matching between the query vector and the key vector, the information of the context feature template that best matches the current behavioral feature is fused into the timestamp-level context vector through weighted summation. The timestamp-level context vector has the same dimension as the value vector, which ensures that the timestamp-level context vector can be effectively computed with other vectors of the same dimension in subsequent calculations. At the same time, the timestamp-level context vector contains the interaction information between the context feature template and the behavioral feature triplet, combining user behavioral features and preset context features, which can more accurately reflect the context information of each timestamp node.

[0035] Step S235: Perform gated recurrent unit processing on the timestamp-level context vectors of all timestamp nodes to capture the dependencies on the time series and generate trajectory-level context vectors. The dimension of the trajectory-level context vectors is consistent with the dimension of the timestamp-level context vectors.

[0036] The gated loop unit includes update and reset gates. The update gate controls how much hidden state information from the previous time step needs to be passed to the current time step, while the reset gate determines how much hidden state information from the previous time step needs to be reset. Through these two gates, the gated loop unit can dynamically update the hidden state at the current time step based on the context vector of the current timestamp node and the hidden state from the previous time step. After processing the context vectors of all timestamp nodes, the final hidden state output by the gated loop unit is the trajectory-level context vector. The trajectory-level context vector has the same dimension as the timestamp-level context vector, containing context information for the entire behavioral trajectory. By capturing dependencies in the time series, it can better reflect the evolution and correlation of user behavior over time.

[0037] Step S236: Map the trajectory-level context vector to a preset dimension space through a fully connected layer to generate context-enhancing behavior vectors. The dimension of the context-enhancing behavior vectors is the same as the feature processing dimension of the large model.

[0038] The fully connected layer processes the trajectory-level context vector through a series of linear transformations and non-linear activation functions. The linear transformation multiplies the input vector by a weight matrix and adds a bias vector to obtain an intermediate vector. Then, the intermediate vector is processed by non-linear activation functions, introducing non-linearity and increasing the model's expressive power. Finally, the vector output by the fully connected layer is the context-enhancing behavior vector, which retains the contextual information of the trajectory-level context vector while being adjusted to a dimension suitable for large-scale model processing.

[0039] Step S240: Calculate the similarity between the context-enhanced behavior vector and the preset meteorological risk scenario prototype vector to generate a scenario fit vector. Modulate the context-enhanced behavior vector based on the scenario fit vector to obtain a scenario-specific behavior vector.

[0040] The preset meteorological risk scenario prototype vectors are a predefined set of vectors, each representing a typical characteristic of a specific meteorological risk scenario. These prototype vectors are obtained through analysis and summarization of a large amount of historical meteorological data and user behavior data, containing key features and patterns under each meteorological risk scenario. Similarity scores are calculated between the context-enhancing behavior vector and each preset meteorological risk scenario prototype vector using similarity calculation methods such as cosine similarity or Euclidean distance. These similarity scores are then combined into a vector, which is the scenario fit vector. Each element in the scenario fit vector represents the degree of matching between the context-enhancing behavior vector and the corresponding meteorological risk scenario prototype vector; a higher score indicates a greater similarity between the current context and the meteorological risk scenario.

[0041] Scene modulation of the context enhancement behavior vector based on the scene fit vector is a process of adjusting the context enhancement behavior vector according to the value of the scene fit vector. Specifically, for meteorological risk scenarios with high scores in the scene fit vector, the features related to that scenario in the context enhancement behavior vector are enhanced accordingly; for scenarios with low scores, the relevant features are weakened or ignored.

[0042] Step S250: Perform feature deentanglement processing on the scene-specific behavior vector. By minimizing mutual information, the behavior vector is decomposed into three statistically independent sub-vectors, which correspond to the risk perception sensitivity sub-vector, the information receiving channel preference sub-vector, and the content presentation form preference sub-vector, respectively.

[0043] The purpose of feature deentanglement processing is to separate and independent different features in the scene-specific behavior vector, so that each sub-vector contains only information about a specific aspect. Mutual information is an indicator that measures the correlation between two random variables. In this embodiment of the invention, feature deentanglement is achieved by minimizing the mutual information between different sub-vectors.

[0044] In one implementation, step S250 may specifically include the following steps S251-S256: Step S251: Initialize three learnable feature mapping matrices, corresponding to the risk perception sensitivity mapping matrix, the information receiving channel preference mapping matrix, and the content presentation format preference mapping matrix, respectively. The number of rows in each mapping matrix is ​​equal to the dimension of the scene-specific behavior vector, and the number of columns is equal to the dimension of the target sub-vector.

[0045] Learnable feature mapping matrices are matrices whose parameters can be adjusted during model training. These matrices map scene-specific behavior vectors to different sub-vector spaces. Risk perception sensitivity mapping matrices, information receiving channel preference mapping matrices, and content presentation format preference mapping matrices are initialized separately. The number of rows in each matrix is ​​set to be the same as the dimension of the scene-specific behavior vectors to ensure matrix multiplication can proceed correctly; the number of columns is set to be the same as the dimension of the target sub-vectors, so that the scene-specific behavior vectors can be projected into the dimension space of the target sub-vectors. During initialization, the parameters of the feature mapping matrices can be randomly initialized, for example, by randomly generating each element of the matrix using a Gaussian distribution. These parameters will be continuously adjusted during subsequent optimization to ensure that the sub-vectors obtained through matrix multiplication meet the requirement of minimizing mutual information.

[0046] Step S252: Perform matrix multiplication operations on the scene-specific behavior vector with the three feature mapping matrices respectively to generate three initial sub-vectors, namely the initial risk perception sensitivity sub-vector, the initial information receiving channel preference sub-vector, and the initial content presentation format preference sub-vector.

[0047] In this embodiment of the invention, the scenario-specific behavior vector is multiplied by the risk perception sensitivity mapping matrix, the information receiving channel preference mapping matrix, and the content presentation format preference mapping matrix, respectively. Specifically, the scenario-specific behavior vector is treated as a row vector and multiplied by each feature mapping matrix to obtain the corresponding initial sub-vector. Through matrix multiplication, the feature information in the scenario-specific behavior vector is projected into different sub-vector spaces to obtain the initial risk perception sensitivity sub-vector, the initial information receiving channel preference sub-vector, and the initial content presentation format preference sub-vector.

[0048] Step S253: Calculate the mutual information value between any two initial sub-vectors, measure the statistical correlation between the sub-vectors by the divergence between the joint probability distribution and the marginal probability distribution, and generate a mutual information matrix, where the matrix elements represent the mutual information magnitude of the corresponding two sub-vectors.

[0049] Mutual information is an indicator used to measure the correlation between two random variables. In this embodiment of the invention, it is used to evaluate the statistical correlation between any two initial sub-vectors. The process of calculating the mutual information value is based on the divergence between the joint probability distribution and the marginal probability distribution. The joint probability distribution describes the probability that two sub-vectors occur simultaneously, while the marginal probability distribution describes the probability that a single sub-vector occurs. The mutual information value can be obtained by calculating the difference (divergence) between the joint probability distribution and the marginal probability distribution.

[0050] The mutual information values ​​between all pairwise initial subvectors are combined into a matrix, known as the mutual information matrix. The mutual information matrix is ​​a symmetric matrix with zero diagonal elements (because the mutual information between a subvector and itself is zero). Each off-diagonal element represents the magnitude of the mutual information between the corresponding two subvectors. A larger mutual information value indicates a stronger correlation between the two subvectors; a smaller mutual information value indicates a weaker correlation.

[0051] Step S254: Construct an optimization function with the trace norm of the mutual information matrix as the objective, and adjust the parameters of the three feature mapping matrices to minimize the trace norm of the mutual information matrix until the mutual information values ​​of any two sub-vectors are lower than the preset threshold.

[0052] The optimization function is constructed with the trace norm of the mutual information matrix as the objective. The trace norm is the sum of the absolute values ​​of all eigenvalues ​​of a matrix and can be used to measure the "size" or "complexity" of the matrix. In this embodiment of the invention, the trace norm of the mutual information matrix is ​​minimized by adjusting the parameters of the three eigenmap matrices.

[0053] Optimization algorithms, such as gradient descent, are used to update the parameters of the feature map matrix. Gradient descent calculates the gradient of the optimization function with respect to the parameters, updates the parameters in the opposite direction of the gradient, and gradually decreases the value of the optimization function. After each parameter update, the initial sub-vectors and mutual information matrix are recalculated until the trace norm of the mutual information matrix reaches its minimum, and the mutual information value of any two sub-vectors is lower than a preset threshold. The preset threshold is a value pre-set based on actual needs and data characteristics. When the mutual information value is lower than this threshold, the correlation between two sub-vectors is considered sufficiently low, and they can be considered statistically independent.

[0054] Step S255: After the optimization function converges, the scenario-specific behavior vector is multiplied by the three optimized feature mapping matrices respectively to generate statistically independent risk perception sensitivity sub-vectors, information receiving channel preference sub-vectors, and content presentation form preference sub-vectors.

[0055] When the optimization algorithm causes the optimization function to converge, it indicates that the parameters of the feature mapping matrix have been adjusted to the optimal state. At this point, the trace norm of the mutual information matrix reaches its minimum, and the mutual information value between any two sub-vectors is lower than a preset threshold. Performing matrix multiplication operations between the scene-specific behavior vector and the optimized risk perception sensitivity mapping matrix, information receiving channel preference mapping matrix, and content presentation format preference mapping matrix respectively yields statistically independent risk perception sensitivity sub-vectors, information receiving channel preference sub-vectors, and content presentation format preference sub-vectors.

[0056] Step S256: Perform L2 normalization on the risk perception sensitivity sub-vector, information receiving channel preference sub-vector, and content presentation form preference sub-vector to make the magnitude of each sub-vector equal to 1, and generate unentangled sub-vectors that satisfy the unit vector constraint. The features of each dimension of the unentangled sub-vectors are mutually orthogonal.

[0057] L2 normalization divides a vector by its L2 norm (the square root of the sum of the squares of its elements), making the vector's magnitude equal to 1. In this embodiment, L2 normalization is applied to the risk perception sensitivity sub-vector, the information receiving channel preference sub-vector, and the content presentation format preference sub-vector. Through L2 normalization, the magnitude of each sub-vector is adjusted to 1, satisfying the constraint condition of a unit vector. Simultaneously, since the correlation between sub-vectors has been minimized in the preceding feature de-entanglement process, after L2 normalization, the dimensional features of the de-entangled sub-vectors are orthogonal to each other, meaning there is no linear correlation between them. This ensures that the dimensional features in each sub-vector independently represent specific aspects of information, avoiding information redundancy and interference, and improving the accuracy and efficiency of subsequent analysis and application.

[0058] Step S260: Standardize and concatenate the risk perception sensitivity sub-vector, information receiving channel preference sub-vector, and content presentation format preference sub-vector to generate a personalized parameter set with unified dimensions. The magnitude of each sub-vector in the personalized parameter set is equal to 1 and they are mutually orthogonal.

[0059] The three sub-vectors are concatenated in a pre-defined order to obtain a vector with a unified dimension, namely the personalized parameter set. During the concatenation process, the properties of each sub-vector being of magnitude 1 and being mutually orthogonal are maintained. This ensures that the personalized parameter set contains accurate information about the user's risk perception sensitivity, information receiving channel preferences, and content presentation format preferences, while also guaranteeing the independence and standardization of each part of the information.

[0060] Step S300: Based on the personalized parameter set, key early warning elements are screened from the real-time meteorological data transmitted by the real-time meteorological monitoring network. The screened early warning elements are converted into an early warning description vector set through semantic feature encoding. The early warning description vector set includes semantic description information of meteorological element intensity, spatial description information of impact range, and time description information of time urgency.

[0061] A real-time weather monitoring network is a network system composed of multiple weather monitoring devices distributed in different geographical locations. These devices include weather stations, satellites, radars, etc., and can monitor and collect data on various meteorological elements in real time, such as temperature, humidity, wind speed, and rainfall.

[0062] In one implementation, step S300 may specifically include the following steps S310-S360: Step S310: Analyze the risk perception sensitivity sub-vector in the personalized parameter set, calculate the variance value of each dimension feature, sort the variance values ​​from largest to smallest to generate an element sensitivity ranking table. The meteorological elements corresponding to the high variance dimension in the element sensitivity ranking table have higher user attention potential.

[0063] Analyzing the risk perception sensitivity subvector involves extracting and analyzing the features of each dimension within it. The risk perception sensitivity subvector contains information about users' sensitivity to different meteorological elements, with each dimension corresponding to a specific meteorological element. By calculating the variance of each dimension's features, the volatility of that feature can be measured. A larger variance indicates greater instability in the feature's changes, suggesting potentially greater fluctuations in user attention to that meteorological element, and thus higher user attention potential. The calculated variance values ​​of each dimension's features are then sorted from largest to smallest to generate an element sensitivity ranking table. In this ranking table, meteorological elements corresponding to dimensions with high variance are listed first, indicating that these elements are more likely to attract user attention.

[0064] Step S320: Based on the element sensitivity ranking table, perform hierarchical screening of meteorological elements in real-time meteorological data, retain the elements ranked at the top of the ranking table as the high-sensitivity element set, extract the elements that are strongly correlated with the high-sensitivity element set from the remaining elements as the associated element set, and merge them into a candidate early warning element set.

[0065] Based on the element sensitivity ranking table, the top-ranked meteorological elements are retained to obtain the high-sensitivity element set. These elements are selected according to users' risk perception sensitivity and have high importance and potential for user attention. Then, correlation analysis is performed on the remaining meteorological elements to identify those strongly correlated with the high-sensitivity element set. Correlation analysis can be performed by calculating correlation coefficients; a higher correlation coefficient indicates a stronger correlation between the two elements. Elements strongly correlated with the high-sensitivity element set are then grouped into the associated element set. Although these elements may not be ranked highly in users' risk perception sensitivity ranking, their close association with the high-sensitivity element set may still make them important to users. Finally, the high-sensitivity element set and the associated element set are merged to obtain the candidate warning element set.

[0066] Step S330: Perform multicollinearity diagnosis on the candidate early warning element set, calculate the degree of multicollinearity among elements through variance inflation factor, remove elements whose variance inflation factor exceeds the threshold, and retain elements whose variance inflation factor is below the threshold as key early warning elements.

[0067] Multicollinearity refers to a high degree of linear correlation among certain elements in a candidate early warning element set. This correlation may affect the accuracy of subsequent data analysis and model building. The variance inflation factor (VIF) is an indicator that measures the degree of multicollinearity. It assesses the variance inflation of an element by calculating the correlation between each element and other elements.

[0068] When calculating the variance inflation factor, a regression model is constructed with each candidate early warning element as the dependent variable and the other candidate early warning elements as independent variables. The coefficient of determination for each regression model is then calculated. The coefficient of determination represents the goodness of fit of the regression model to the data, ranging from 0 to 1; a value closer to 1 indicates a better fit. Based on the coefficient of determination, the variance inflation factor is calculated. A larger variance inflation factor indicates stronger collinearity between that element and other elements.

[0069] In one implementation, step S330 may specifically include the following steps S331-S336: Step S331: Construct the element correlation matrix of the candidate early warning element set. The rows and columns of the matrix correspond to the candidate early warning elements, and the matrix elements are the correlation coefficients of the corresponding two elements. The value range of the correlation coefficients is within a preset range.

[0070] The correlation matrix is ​​a square matrix whose number of rows and columns equals the number of elements in the candidate early warning element set. Each row and column of the matrix corresponds to a candidate early warning element, and each element in the matrix represents the correlation coefficient between two corresponding elements. The correlation coefficient is an indicator that measures the linear correlation between two variables, and its value typically ranges from -1 to 1, where -1 indicates a perfect negative correlation, 1 indicates a perfect positive correlation, and 0 indicates no correlation.

[0071] When constructing the element correlation matrix, the correlation coefficient between each pair of candidate early warning elements is calculated, and these coefficients are then filled into the corresponding positions in the matrix. For example, if the candidate early warning element set contains three elements: rainfall, wind speed, and humidity, then the element correlation matrix is ​​a 3×3 matrix, where the elements represent the correlation coefficients between rainfall and wind speed, rainfall and humidity, and wind speed and humidity, respectively. By constructing the element correlation matrix, the correlation between the elements in the candidate early warning element set can be intuitively understood.

[0072] Step S332: Construct a regression model with each candidate early warning element as the dependent variable and other candidate early warning elements as independent variables, calculate the determination coefficient of each regression model, and calculate the variance inflation factor based on the determination coefficient. The variance inflation factor is expressed by the reciprocal of the complementary value of the determination coefficient.

[0073] For each element in the candidate early warning element set, a regression model is constructed using it as the dependent variable and the other elements as independent variables. The coefficient of determination (COD) is calculated for each regression model. The COD reflects the goodness of fit of the regression model to the data, i.e., the degree to which the independent variables can explain the changes in the dependent variable. The closer the COD is to 1, the better the regression model fits the data, and the stronger the linear relationship between the dependent and independent variables. The variance inflation factor (VOP) is calculated based on the COD, represented by the reciprocal of the complementary value of the COD. The larger the VOP, the stronger the multicollinearity between that element and other elements, meaning the variance of that element is more significantly affected by other elements. By calculating the VOP for each candidate early warning element, the degree of multicollinearity among the elements can be assessed.

[0074] Step S333: Compare the variance inflation factor of each candidate early warning element with the preset threshold, mark the elements whose variance inflation factor exceeds the threshold as highly collinear elements, and record the position of highly collinear elements in the element sensitivity ranking table.

[0075] The calculated variance inflation factor of each candidate early warning element is compared with a preset threshold. This preset threshold is a standard value set based on actual conditions and experience, used to determine whether elements exhibit severe collinearity. If the variance inflation factor of an element exceeds the threshold, that element is marked as a highly collinear element. The position of highly collinear elements in the element sensitivity ranking table is also recorded. This is because subsequent processing requires a comprehensive consideration of both element sensitivity and collinearity to determine whether to retain the element. By recording the position of highly collinear elements in the ranking table, it is possible to remove collinearity while preserving elements important to the user as much as possible.

[0076] Step S334: Use a stepwise elimination strategy to process highly collinear elements. Prioritize eliminating highly collinear elements that are ranked lower in the element sensitivity ranking table. After each elimination, recalculate the variance inflation factor of the remaining elements until the variance inflation factor of all elements is lower than the threshold.

[0077] The stepwise elimination strategy is a method to gradually remove highly collinear elements. It prioritizes eliminating highly collinear elements ranked lower in the element sensitivity list, as these elements are relatively less important to users. After eliminating one highly collinear element, the candidate warning element set is reconstructed, and the variance inflation factor of the remaining elements is calculated. It is then checked whether the newly calculated variance inflation factors are all below a threshold. If there are still elements with variance inflation factors exceeding the threshold, the process of eliminating lower-ranking highly collinear elements continues, and the variance inflation factors are recalculated again. This process is repeated until the variance inflation factors of all elements are below the threshold, at which point the collinearity problem in the candidate warning element set is considered resolved.

[0078] Step S335: Check whether the removed element set contains the core elements of meteorological early warning. If there are missing elements, select the element with the highest sensitivity ranking from the highly collinear elements to supplement it. After supplementation, re-verify the collinearity.

[0079] Core elements of weather warnings are essential components, such as rainfall and wind speed. After gradually eliminating highly collinear elements, the remaining element set is checked to see if it contains these core elements. If some core elements are missing from the element set, the elements with the highest sensitivity ranking from the previously marked highly collinear elements need to be selected to supplement them.

[0080] After adding elements, perform a new collinearity diagnosis on the element set, calculate the variance inflation factor of each element, and check if collinearity issues still exist. If there are still elements with variance inflation factors exceeding the threshold, it may be necessary to adjust the removal and addition strategy again until the element set contains both core elements and no serious collinearity exists.

[0081] Step S336: The set of elements that have undergone collinearity verification and core element supplementation is determined as key early warning elements. The number of key early warning elements is a preset ratio range of the number of candidate early warning elements.

[0082] After collinearity verification and supplementation of core elements, the final set of elements determined is the key early warning element set. The number of key early warning elements is usually set as a preset proportion range of the number of candidate early warning elements. This proportion range is determined based on actual conditions and experience to ensure that the key early warning elements contain important meteorological information without being overly redundant. For example, the preset proportion range can be set within a certain interval, so that the key early warning elements can accurately reflect the meteorological conditions and the user's focus, while avoiding interference from too many irrelevant elements, thereby improving the quality and effectiveness of the early warning information.

[0083] Step S340: Extract the semantic description text, numerical change range, and duration of impact of key early warning elements, and combine the three into structured early warning element units according to semantic association rules. Each structured early warning element unit contains a text description segment, a numerical range segment, and a time segment marker segment.

[0084] The semantic description text of key warning elements is the content describing the element using natural language. The numerical variation range is the fluctuation of the key warning element's value within a certain range, and the duration of impact indicates the time range during which the key warning element's impact will last. Semantic association rules are rules that reasonably combine the semantic description text, numerical variation range, and duration of impact, combining them into structured warning element units. Each structured warning element unit contains three parts: a text description segment, i.e., the semantic description text, used to concisely express the type of meteorological element; a numerical variation range segment, containing information about the numerical variation range, providing users with more specific intensity information; and a time period marker segment, corresponding to the duration of impact, informing users of the time range of the meteorological element's impact.

[0085] Step S350: Input the structured early warning element unit into the dynamic semantic encoder, perform contextual semantic encoding on the text description segment to generate a text semantic vector, perform interval boundary encoding on the numerical interval segment to generate a numerical boundary vector, and perform time interval encoding on the duration of influence to generate a time interval vector.

[0086] In one implementation, step S350 may specifically include the following steps S351-S356: Step S351: Perform word segmentation on the text description segment of the structured early warning element unit, divide the continuous text into independent word units, remove stop words and retain the valid word units, and the valid word units correspond to meteorological early warning professional terms.

[0087] Word segmentation is the process of dividing continuous text in a text description segment into independent word units according to certain rules. For the text description segment of structured early warning element units, word segmentation algorithms are used to divide it into individual words.

[0088] Step S352: Input the effective word units into the pre-trained word embedding model, and convert each word unit into a fixed-dimensional word vector by word lookup. Perform average pooling on all word vectors to generate preliminary text vectors. The dimension of the preliminary text vectors is consistent with the dimension of the word vectors.

[0089] Pre-trained word embedding models are models pre-trained on large-scale text data. They map words into a low-dimensional vector space, ensuring that semantically similar words are close together in the vector space. Valid word units are input into the pre-trained word embedding model, and the model converts them into fixed-dimensional word vectors by looking up the vector representation corresponding to each word unit in the vocabulary. Each word vector contains the semantic information of that word.

[0090] Average pooling is an operation that sums all word vectors and takes their average. It adds up all word vectors corresponding to valid word units and then divides by the number of word vectors to obtain a preliminary text vector. The preliminary text vector has the same dimension as the word vectors and is a comprehensive representation of the semantic information of all valid word units in the text description segment. Average pooling can reduce local noise interference from individual words, resulting in a preliminary vector representing the overall semantics of the text.

[0091] Step S353: Input the initial text vector into the bidirectional recurrent network. The forward recurrent layer calculates the hidden state from the beginning of the word sequence backward, and the backward recurrent layer calculates the hidden state from the end of the word sequence forward. At each time step, the forward and backward hidden states are concatenated into a fusion vector.

[0092] A bidirectional recurrent neural network (BRNN) consists of a forward recurrent layer and a backward recurrent layer. Taking the initial text vector as input, the forward recurrent layer starts from the beginning of the word sequence and processes the vector representation of each word sequentially, calculating the hidden state at each time step. The hidden state is an intermediate variable in the recurrent network used to store and transmit sequence information, containing information about previously processed words. The backward recurrent layer starts from the end of the word sequence and processes the vector representation of each word in reverse order, similarly calculating the hidden state at each time step. At each time step, the hidden states calculated by the forward and backward recurrent layers are concatenated to obtain a fused vector.

[0093] Step S354: Perform attention weighting on the fusion vectors at all time steps, calculate the contribution weight of each fusion vector to the semantics of the text, generate an attention weight vector, and sum the fusion vectors using the weight vectors to obtain the context enhancement vector.

[0094] Attention-weighted processing can assign different weights to each fusion vector based on its contribution to the semantics of the text. In this embodiment of the invention, the contribution weight of each fusion vector to the semantics of the text is first calculated. This weight reflects the importance of the fusion vector in representing the overall semantics of the text.

[0095] An attention weight vector is generated, where each element corresponds to a weight in a fusion vector. Then, the attention weight vector and the fusion vector are weighted and summed; that is, each fusion vector is multiplied by its corresponding weight, and all weighted fusion vectors are summed to obtain the context enhancement vector. The context enhancement vector is obtained after considering the importance of the fusion vectors at each time step, highlighting the information of the fusion vectors that contribute more to the semantics of the text, and thus more accurately representing the contextual semantic information of the text.

[0096] Step S355: Input the context enhancement vector into the convolutional layer, extract features from the context enhancement vector through multiple convolutional kernels, and generate multi-scale text feature maps. The number of multi-scale text feature maps is the same as the number of convolutional kernels.

[0097] Each convolutional kernel performs a sliding convolution operation on the context enhancement vector, extracting features at different scales. Different sized kernels can capture text features of varying lengths; for example, smaller kernels extract local, detailed features, while larger kernels extract more global features. Each kernel generates a corresponding text feature map during the convolution process. This text feature map is the output of the convolution operation and contains the extracted feature information. The number of multi-scale text feature maps is the same as the number of convolutional kernels, and these feature maps reflect the text's feature information from different scales and perspectives.

[0098] Step S356: Perform global max pooling on the multi-scale text feature maps to extract the salient features of each feature map and generate text semantic vectors. The dimension of the text semantic vectors is positively correlated with the number of multi-scale text feature maps.

[0099] Global max pooling reduces dimensionality by decreasing the dimension of the feature maps while preserving the most salient information in each map. After performing global max pooling on all multi-scale text feature maps, the salient features extracted from each map are combined to generate a text semantic vector. Since each multi-scale text feature map corresponds to one salient feature, the dimensionality of the text semantic vector is positively correlated with the number of multi-scale text feature maps; that is, the more feature maps there are, the higher the dimensionality of the text semantic vector.

[0100] Step S360: Input the text semantic vector, numerical boundary vector and time interval vector into the feature fusion processor, dynamically adjust the fusion weight of each vector through an adaptive gating mechanism, and generate a warning description vector containing multi-dimensional semantic information. Multiple warning description vectors are grouped according to feature type to form a warning description vector set.

[0101] A feature fusion unit is a module used to fuse vectors of different types. It can comprehensively consider the information of different vectors and generate a new vector containing multi-dimensional semantic information. In this embodiment of the invention, the text semantic vector, the numerical boundary vector, and the time interval vector are input into the feature fusion unit.

[0102] The adaptive gating mechanism is the core mechanism in the feature fusion machine. It can dynamically adjust the weights of each vector in the fusion process based on the features and contextual information of the input vectors. Through this mechanism, the importance of each vector can be flexibly allocated according to different situations, so that the fused vector can better reflect the comprehensive information of the early warning elements.

[0103] Under the control of an adaptive gating mechanism, the feature fusion unit performs a weighted summation of the text semantic vector, numerical boundary vector, and time interval vector to generate a warning description vector. This warning description vector contains multi-dimensional semantic information about meteorological elements, including text semantics, numerical ranges, and time intervals, enabling a more comprehensive description of the characteristics of the warning elements.

[0104] Multiple warning description vectors are grouped according to element type. For example, all warning description vectors related to rainfall are grouped into one group, and warning description vectors related to wind speed are grouped into another group. The grouped warning description vectors are then combined to obtain a warning description vector set, which provides a complete representation of warning information for subsequent warning priority assessment and delivery.

[0105] Step S400: Input the warning description vector set and the meteorological situation evolution process data output by the meteorological situation prediction system into the spatiotemporal inference network of the large model, model the dynamic correlation between changes in meteorological elements and user risk perception, and obtain the warning priority assessment sequence ranked according to the user's risk exposure level.

[0106] In one implementation, step S400 may specifically include the following steps S410-S460: Step S410: Analyze the meteorological situation evolution data, extract the predicted values ​​and trend of meteorological elements at different time points, and combine them into a meteorological situation time series in chronological order. The length of the time series is equal to the number of time intervals in the prediction period.

[0107] Analyzing meteorological trend evolution data involves detailed analysis and extraction of key information. This data contains forecasts and trends of meteorological elements over a future period. Predicted values ​​for meteorological elements at different time points are extracted; these values ​​represent the expected values ​​of meteorological elements (such as temperature, wind speed, and rainfall) at specific points in time.

[0108] Simultaneously, the information on the changing trends of meteorological elements is extracted, reflecting the direction and extent of changes in these elements across different time points. These predicted values ​​and changing trends of meteorological elements at different time points are arranged chronologically to form a meteorological situation time series. The length of the time series is equal to the number of time intervals in the forecast period. For example, if the forecast period is the next 24 hours and the time interval is 1 hour, then the length of the time series is 24, with each element corresponding to one hour's predicted value and changing trend information of the meteorological element.

[0109] Step S420: Match the warning description vector set with the meteorological situation time series using timestamps, establish the association between the warning elements and the meteorological situation data at the corresponding time nodes, and generate a spatiotemporal correlation feature matrix, where the rows of the matrix correspond to the time nodes and the columns correspond to the warning elements.

[0110] Matching the warning description vector set with the meteorological situation time series using timestamps is to determine the association between each warning element and the corresponding time point in the meteorological situation data. Each warning description vector in the warning description vector set represents a warning element, containing information such as the intensity, impact range, and timeliness of the meteorological element. The meteorological situation time series contains the predicted values ​​and trends of meteorological elements at different time points.

[0111] By matching timestamps, warning elements are associated with meteorological situation data at specific time points. For example, if a warning description vector indicates "heavy rainfall is expected within the next 3 hours," then the warning element is associated with the rainfall forecast values ​​and trends at the corresponding 3 time points in the meteorological situation time series.

[0112] Based on this correlation, a spatiotemporal correlation feature matrix is ​​generated. The rows of the matrix correspond to time nodes, and the columns correspond to warning elements. Each element in the matrix represents the correlation characteristics of the warning element at a specific time node, such as the meteorological element forecast value of the warning element at that time node, its correlation with other elements, etc.

[0113] Step S430: Enhance the spatial dimension features of the spatiotemporal correlation feature matrix, calculate the contribution weight of each early warning element in the user's risk perception, generate a spatial attention weight matrix, and weight the correlation feature matrix through the weight matrix to obtain the spatial enhancement feature matrix.

[0114] In one implementation, step S430 may specifically include the following steps S431-S436: Step S431: Perform feature compression processing on the spatiotemporal correlation feature matrix. Compress the time dimension of the matrix by global average pooling, retain the spatial dimension features, and generate a spatial feature vector. The dimension of the spatial feature vector is consistent with the column dimension of the correlation feature matrix.

[0115] Global average pooling is used to compress the time dimension of the spatiotemporal correlation feature matrix by averaging the values ​​of each column (corresponding to a warning element) over time. This involves summing the elements of each column at different time points and dividing by the number of time points to obtain an average value. Combining the average values ​​of all columns generates the spatial feature vector. Because only the time dimension is compressed, the spatial dimension (i.e., the warning element dimension) is preserved, so the dimension of the spatial feature vector is the same as the column dimension of the correlation feature matrix. This method simplifies the complex temporal information in the spatiotemporal correlation feature matrix, highlighting the spatial characteristics of each warning element.

[0116] Step S432: Input the spatial feature vector into a two-layer fully connected network. The first fully connected network performs dimensionality reduction on the spatial feature vector to generate an intermediate vector. The second fully connected network performs dimensionality increase on the intermediate vector to generate a spatial weight vector with the same dimension as the spatial feature vector.

[0117] A two-layer fully connected network consists of two fully connected layers. Fully connected layers are a common layer structure in neural networks, connecting each element of the input vector to each neuron in the output layer. The spatial feature vector is input into the first fully connected layer, which processes it through a series of linear transformations and non-linear activation functions to achieve dimensionality reduction. Dimensionality reduction reduces data complexity, extracts more critical feature information, and generates an intermediate vector. The intermediate vector has a lower dimension than the spatial feature vector. Next, the intermediate vector is input into the second fully connected layer, which performs dimensionality upscaling. Through further linear transformations and non-linear activation functions, the dimensionality of the intermediate vector is increased to the same level as the spatial feature vector, generating a spatial weight vector. This process is similar to a feature filtering and reconstruction of the spatial feature vector, enabling the generated spatial weight vector to better reflect the potential importance of each early warning element in the user's risk perception.

[0118] Step S433: Perform element-wise multiplication of the spatial weight vector with the risk perception sensitivity sub-vector in the personalized parameter set, adjust the weight values ​​of each early warning element, enhance the element weights corresponding to high-value dimensions in the risk perception sensitivity sub-vector, and generate the adjusted spatial weight vector.

[0119] The spatial weight vector is multiplied element-wise with the risk perception sensitivity sub-vector in the personalized parameter set, that is, the elements at corresponding positions of the two vectors are multiplied. The risk perception sensitivity sub-vector reflects the user's sensitivity to different warning elements, where the warning elements corresponding to higher-value dimensions indicate that the user is more concerned and sensitive to those elements. Through the element-wise multiplication operation, the weight value of the elements corresponding to the higher-value dimensions in the risk perception sensitivity sub-vector is enhanced in the spatial weight vector; while the weight value of the elements corresponding to the lower-value dimensions is reduced accordingly. In this way, the spatial weight vector is adjusted according to the user's risk perception sensitivity, generating an adjusted spatial weight vector that better reflects the user's actual risk perception.

[0120] Step S434: Normalize the adjusted spatial weight vector to ensure that the values ​​of each element in the weight vector are within a preset range, so as to avoid the weight values ​​being too large or too small and affecting subsequent processing.

[0121] Normalization involves scaling the elements of a vector to ensure their values ​​fall within a preset range. Common normalization methods include max-min normalization and Z-score normalization. In this embodiment, the adjusted spatial weight vector is normalized to ensure that the value of each element is neither too large nor too small. If the weight value is too large, the impact of the warning element may be excessively amplified in subsequent weighted processing; if the weight value is too small, the warning element may be ignored in subsequent analysis.

[0122] Step S435: Copy and expand the normalized spatial weight vector to generate a spatial attention weight matrix with the same size as the spatiotemporal correlation feature matrix. The weight row vector corresponding to each time node in the matrix is ​​the same as the normalized spatial weight vector.

[0123] The copying and expansion process involves replicating the normalized spatial weight vectors to the same number as the number of time nodes in the spatiotemporal correlation feature matrix. These replicated vectors are then arranged sequentially to form a matrix. Since each row of the spatiotemporal correlation feature matrix corresponds to a time node, and each column corresponds to a warning element, the spatial attention weight matrix generated through this copying and expansion method has the same size as the spatiotemporal correlation feature matrix. Furthermore, the weight row vector corresponding to each time node in the matrix is ​​the same as the normalized spatial weight vector. This ensures that the warning elements at each time node are assigned the same weight, allowing for unified weight adjustments for warning elements at different time nodes in subsequent weighted processing.

[0124] Step S436: Perform element-wise multiplication on the spatiotemporal correlation feature matrix using the spatial attention weight matrix to enhance the feature values ​​of warning elements with weight values ​​greater than the average weight value, and weaken the feature values ​​of warning elements with weight values ​​less than the average weight value, thereby generating a spatial enhancement feature matrix.

[0125] The spatial attention weight matrix and the spatiotemporal correlation feature matrix are multiplied element-wise, meaning corresponding elements within the matrices are multiplied. For warning elements in the spatial attention weight matrix with weights greater than the average weight, their eigenvalues ​​in the spatiotemporal correlation feature matrix are enhanced after multiplication. This means these warning elements will be given more attention in subsequent analysis because they contribute more to user risk perception. Conversely, for warning elements with weights less than the average weight, their eigenvalues ​​are weakened after multiplication, reducing their impact on the overall analysis.

[0126] Step S440: Perform dynamic modeling of the spatial augmented feature matrix in the time dimension. Capture the dependence of meteorological elements on time through multi-layer convolution processing to generate a time evolution feature vector. The dimension of the evolution feature vector is consistent with the column dimension of the spatial augmented feature matrix.

[0127] Multi-layer convolutional processing utilizes multiple convolutional layers to process the spatial augmented feature matrix, capturing the dynamic changes and dependencies of meteorological elements over time. Convolutional layers perform convolution operations on the input data using convolutional kernels, extracting features at different scales and levels. In this embodiment of the invention, multiple convolutional layers process the spatial augmented feature matrix sequentially, with each layer learning temporal features at different levels of abstraction.

[0128] In one implementation, step S440 may specifically include the following steps S441-S446: Step S441: Perform temporal dimension expansion processing on the spatial enhancement feature matrix by adding a preset number of zero-padding time steps before and after the temporal dimension of the matrix, so that the time length of the expanded matrix meets the input requirements of the convolutional layer.

[0129] Convolutional layers have certain requirements regarding the size of the input data when performing convolution operations. To ensure that the spatial augmentation feature matrix can be successfully input into the convolutional layer for processing, its temporal dimension needs to be expanded. Specifically, a predetermined number of zero-padding time steps are added before and after the temporal dimension of the spatial augmentation feature matrix. Zero-padding time steps add rows with zero values ​​to the time dimension of the matrix. These zero values ​​do not change the actual characteristics of the meteorological elements in the original matrix; they are only used to adjust the time length of the matrix to meet the input requirements of the convolutional layer. This expansion ensures that the convolutional layer can perform effective convolution operations on the entire time series, thereby better capturing the changes of meteorological elements over time.

[0130] Step S442: Input the expanded spatial enhancement feature matrix into the first convolutional layer, and perform parallel convolution operations on the matrix through multiple convolution kernels of different sizes to extract temporal feature maps at different time scales. The number of feature maps is consistent with the number of convolution kernels.

[0131] The first convolutional layer contains multiple convolutional kernels of different sizes, which simultaneously perform convolution operations on the expanded spatial augmentation feature matrix. Different sized kernels can capture meteorological feature characteristics at different time scales. For example, smaller kernels can extract local variations of meteorological elements over a short period, while larger kernels can capture overall trends over a longer time span. Each kernel slides across the matrix to perform the convolution operation, generating a corresponding temporal feature map. The temporal feature map is the output of the convolution operation, reflecting the meteorological feature characteristics extracted by that kernel at a specific time scale. Because multiple kernels perform convolution operations simultaneously, the number of temporal feature maps equal to the number of kernels is generated. These feature maps reflect the changes in meteorological elements from different time scales and perspectives.

[0132] Step S443: Perform batch normalization on the time feature map, adjust the numerical distribution of elements in the feature map, enhance the nonlinear expressive power of the feature map through a nonlinear activation function, and generate an activated feature map.

[0133] Batch normalization can accelerate the training process of neural networks and improve model stability. In this embodiment of the invention, batch normalization is performed on the time feature map, adjusting the value of each element in the feature map to make its distribution more stable and uniform. Specifically, the mean and variance of the elements in the feature map are calculated, and then the elements are standardized so that the mean is 0 and the variance is 1. After batch normalization, the feature map is input into a nonlinear activation function. The nonlinear activation function can introduce nonlinear factors and enhance the expressive power of the feature map. Common nonlinear activation functions include ReLU and Sigmoid. By processing the batch-normalized feature map with a nonlinear activation function, the feature map can learn more complex patterns and relationships, generating an activated feature map. The activated feature map has a stronger nonlinear expressive power and can better reflect the dynamic changes of meteorological elements.

[0134] Step S444: Perform pooling on the activation feature map. Reduce the temporal resolution of the feature map by max pooling, retain key temporal features, and generate a dimensionality-reduced feature map. The temporal length of the dimensionality-reduced feature map is a preset ratio of the original feature map.

[0135] Max pooling reduces the temporal resolution of the feature map, meaning it shortens the feature map's length in the time dimension. The time length of the generated dimensionality-reduced feature map is a preset proportion of the original feature map, which is pre-set based on actual needs and model requirements. While reducing the temporal resolution, max pooling preserves the maximum value within each window, which is the key temporal feature. This allows the dimensionality-reduced feature map to retain important meteorological element change information while reducing the amount of data.

[0136] Step S445: Input the dimensionality-reduced feature map into the subsequent convolutional layer, and repeat the convolution, batch normalization, activation and pooling processes until the time length of the feature map reaches the preset threshold to generate a multi-scale temporal feature map set.

[0137] The reduced-dimensionality feature map is then fed into subsequent convolutional layers for further convolution, batch normalization, activation, and pooling. Subsequent convolutional layers extract higher-level temporal features, deepening the understanding of the temporal dependencies of meteorological elements. Each convolutional layer extracts features at different scales and levels through convolutional kernels; batch normalization adjusts the numerical distribution of the feature map; activation functions enhance the non-linear expressiveness of the feature map; and pooling reduces the dimensionality of the feature map while preserving key features.

[0138] These processing steps are repeated until the time length of the feature map reaches a preset threshold. The preset threshold is a time length standard set according to the model design and actual needs. When the time length of the feature map reaches this threshold, processing stops, and a multi-scale time feature map set is generated. This set contains time feature maps at different levels and scales, reflecting the changes of meteorological elements over time from multiple perspectives.

[0139] Step S446: Perform feature fusion processing on the multi-scale time feature map set, and generate a time evolution feature vector containing multi-time scale information by feature splicing. The dimension of the time evolution feature vector is consistent with the total number of channels of the multi-scale time feature map.

[0140] Feature fusion is the process of integrating various feature maps from a multi-scale temporal feature map set. In this embodiment of the invention, a feature stitching method is used for fusion. Feature stitching connects multiple feature maps along the channel dimension, combining their feature information together. All feature maps in the multi-scale temporal feature map set are stitched together along the channel dimension to generate a vector containing multi-timescale information, namely the temporal evolution feature vector. The dimension of this vector is consistent with the total number of channels in the multi-scale temporal feature map, integrating temporal features at different levels and scales, and can comprehensively reflect the changes of meteorological elements over time.

[0141] Step S450: Connect the time evolution feature vector with the risk perception sensitivity sub-vector in the personalized parameter set, calculate the matching degree between each early warning element and the user's risk perception, and generate a risk perception matching degree vector.

[0142] Association modeling involves comprehensively analyzing the temporal evolution feature vector and the risk perception sensitivity sub-vector from the personalized parameter set to calculate the matching degree between each warning element and the user's risk perception. The temporal evolution feature vector contains information about the changes of meteorological elements over time, while the risk perception sensitivity sub-vector reflects the user's sensitivity to different warning elements. In association modeling, the feature values ​​of each warning element in the temporal evolution feature vector and the sensitivity values ​​of the corresponding elements in the risk perception sensitivity sub-vector are considered. The information from these two vectors is fused to calculate the matching degree between each warning element and the user's risk perception. A higher matching degree indicates a greater impact of the warning element on the user's risk and a better match with the user's risk perception characteristics. The matching degree values ​​of all warning elements are combined to generate a risk perception matching degree vector, where each element corresponds to the matching degree of a warning element.

[0143] Step S460: Sort the warning description vectors in the warning description vector set based on the element values ​​in the risk perception matching degree vector, and generate a warning priority evaluation sequence in descending order of element values. The order of vectors in the sequence reflects the user's attention to the warning information.

[0144] Based on the element values ​​in the risk perception matching vector, the warning description vectors in the warning description vector set are sorted. The sorting rule is to arrange them in descending order of element value. The larger the element value, the higher the matching degree between the warning information represented by the corresponding warning description vector and the user's risk perception, and the higher the user's attention to the warning information is likely to be.

[0145] This sorting method generates a warning priority assessment sequence. In this sequence, warning information corresponding to warning description vectors that appear earlier in the sequence has higher priority and should be pushed to users first; while warning information that appears later in the sequence has lower priority. The warning priority assessment sequence provides a clear order for subsequent warning information delivery, enabling targeted delivery of warning information based on users' risk perception characteristics, thereby improving the effectiveness of warning information and user satisfaction.

[0146] Step S500: Integrate device operation status data and network environment parameters uploaded by user terminals, optimize the push strategy calculation of the early warning priority evaluation sequence through the multi-objective decision layer of the large model, output the push strategy parameters, and transform the push strategy parameters into multimodal early warning information that meets user interaction habits for push.

[0147] The device operation status data uploaded by the user terminal includes various operational information about the user device, such as network access type, background process usage, and screen display mode. This information reflects the current operating status and resource usage of the device. Network environment parameters include network latency, packet loss rate, and signal coverage, which describe the quality of the network environment in which the user terminal is located.

[0148] As one implementation method, in step S500, the device operation status data uploaded by the user terminal and network environment parameters are integrated, and the push strategy optimization calculation of the early warning priority evaluation sequence is performed through the multi-objective decision layer of the large model, and the push strategy parameters are output. Specifically, this may include the following steps S510-S560: Step S510: Parse the device operation status data, extract device network access type information, background process occupancy information and screen display mode information. The device operation status data is actively uploaded to the data processing center by the user terminal at preset time intervals.

[0149] Device operation status data is data actively uploaded by user terminals and contains multiple operational status information of the device. Parsing device operation status data involves analyzing and extracting this data. First, the device network access type information is extracted from the data. Device network access types can be divided into different types such as Wi-Fi and mobile data networks. This information reflects the device's current network connection method. Next, background process utilization information is extracted. Background process utilization is the proportion of system resources used by processes running in the background, reflecting the device's resource usage. If the background process utilization is too high, it may affect device performance and the reception of warning information. Finally, screen display mode information is extracted. Screen display modes can be divided into different modes such as screen on, screen off, and screen locked. This information has an important impact on determining the push method and presentation effect of warning information.

[0150] Step S520: Collect the current network environment parameters of the user terminal, including network latency, packet loss rate and signal coverage. The network environment parameters are obtained in real time through network monitoring tools and are timestamped with the device operation status data.

[0151] In this embodiment of the invention, network monitoring tools are used to collect network environment parameters such as network latency, packet loss rate, and signal coverage. Timestamp alignment ensures that network environment parameters and device operating status data correspond in time, enabling subsequent fusion processing to accurately consider the real-time status of devices and the network.

[0152] Step S530: The device operation status data and network environment parameter input data fusion module combine the device network access type information, background process occupancy information, screen display mode information, network latency duration, data packet loss rate and signal coverage range into a terminal environment feature vector through feature splicing. The dimension of the terminal environment feature vector is the sum of the dimensions of each information.

[0153] In one implementation, step S530 may specifically include the following steps S531-S536: Step S531: Standardize the network access type information of the device, divide the network access type into preset access categories, and convert the access category into an access type feature vector through one-hot encoding. The dimension of the access type feature vector is consistent with the number of access categories.

[0154] Standardization is the process of unifying and standardizing device network access type information. First, network access types are categorized into preset access classes, such as Wi-Fi, mobile data network, and Bluetooth network. Then, one-hot encoding is used to transform the access classes into access type feature vectors. One-hot encoding is a method of converting categorical variables into vector representations; for each access class, the corresponding position in the vector is 1, and the remaining positions are 0. The dimension of the access type feature vector is the same as the number of access classes. Through this encoding method, the device network access type information is transformed into a vector form suitable for feature concatenation.

[0155] Step S532: Extract the current occupancy percentage from the background process occupancy information, map the current occupancy percentage to multiple occupancy intervals, each occupancy interval corresponds to an occupancy level, and convert the occupancy level into a process feature vector through numerical encoding. The process feature vector is a one-dimensional vector and its value is positively correlated with the occupancy level.

[0156] The current resource usage percentage is extracted from background process usage information. This percentage represents the proportion of system resources currently being used by background processes. The current usage percentage is then mapped to multiple usage ranges, for example, dividing the usage percentage into low, medium, and high ranges, each corresponding to a usage level. Next, the usage level is converted into a process feature vector through numerical encoding. Numerical encoding represents the usage level with a numerical value. The process feature vector is a one-dimensional vector, and its value is positively correlated with the usage level; that is, the higher the usage level, the larger the value of the process feature vector.

[0157] Step S533: Perform mode classification processing on the screen display mode information, divide the screen display mode into preset display categories, and convert the display category into a display mode feature vector through category encoding. The display mode feature vector is a multi-dimensional binary vector.

[0158] The screen display mode information is classified into preset display categories, such as screen on, screen off, and screen locked. Category encoding is then used to convert these display categories into display mode feature vectors. Category encoding represents each display category using a multi-dimensional binary vector, where a 1 in one position indicates belonging to that category, and 0 in the others. This multi-dimensional binary encoding method transforms the screen display mode information into a vector form suitable for feature concatenation.

[0159] Step S534: Divide the network latency duration into multiple latency levels according to a preset threshold, and convert the latency levels into latency feature vectors through sequential encoding. The dimension of the latency feature vectors is the same as the number of latency levels.

[0160] Network latency is divided into multiple latency levels based on a preset threshold, such as low latency, medium latency, and high latency. These latency levels are then converted into latency feature vectors using sequential encoding. Sequential encoding represents each latency level as a vector with the same dimension as the number of latency levels. Each level corresponds to a position in the vector; a 1 at that position indicates belonging to that latency level, while other positions are 0.

[0161] Step S535: Classify the packet loss rate into levels. Divide the packet loss rate into multiple loss levels according to the percentage of loss rate. Convert the loss level into a loss rate feature vector through numerical mapping. The loss rate feature vector is a one-dimensional vector and the value increases as the loss rate increases.

[0162] Packet loss rates are categorized into different levels based on the percentage of loss, such as low loss rate, medium loss rate, and high loss rate. A numerical mapping is then used to convert these loss levels into a loss rate feature vector. This numerical mapping represents the loss level with a numerical value; the loss rate feature vector is a one-dimensional vector whose value increases as the loss rate increases.

[0163] Step S536: Divide the signal coverage area into multiple coverage levels, convert the coverage levels into coverage feature vectors through encoding, and concatenate the access type feature vector, process feature vector, display mode feature vector, delay feature vector, loss rate feature vector and coverage feature vector in sequence to generate terminal environment feature vector.

[0164] The signal coverage area is divided into multiple coverage levels, such as strong coverage, medium coverage, and weak coverage. The coverage levels are then converted into coverage feature vectors through encoding, which can be similar to the previous category encoding or sequential encoding. Next, the access type feature vector, process feature vector, display mode feature vector, latency feature vector, loss rate feature vector, and coverage feature vector are concatenated sequentially to generate the terminal environment feature vector.

[0165] Step S540: Input the early warning priority assessment sequence and the terminal environment feature vector into the multi-objective decision layer of the large model, and make multi-objective trade-off adjustments to the push time, push channel and information presentation format of the early warning information, and generate a push strategy optimization scheme, which includes multiple push parameter combinations.

[0166] The multi-objective decision layer can consist of an input layer, a fully connected layer, a multi-objective optimization module, and an output layer. The input layer is responsible for receiving the warning priority evaluation sequence and the terminal environment feature vector. The warning priority evaluation sequence reflects the ranking of the importance of different warning information, while the terminal environment feature vector contains comprehensive information such as device operating status and network environment.

[0167] The fully connected layer connects the input layer and the multi-objective optimization module. It contains multiple neurons, each of which is connected to all elements of the input layer. The fully connected layer performs linear transformations and non-linear activations on the input data, mapping the input data to a higher-dimensional feature space to extract more complex feature information.

[0168] The multi-objective optimization module is responsible for balancing and adjusting the push time, push channels, and information presentation format of early warning information. This module uses preset optimization algorithms, such as multi-objective genetic algorithms and Pareto optimization algorithms, to comprehensively consider multiple objective factors and attempt to find an optimal push strategy through continuous iteration and search. During the optimization process, different combinations of push parameters are evaluated and compared based on the input data and the preset objective function.

[0169] The output layer receives the output from the multi-objective optimization module and outputs the optimized push strategy as a combination of push parameters, forming a push strategy optimization scheme. The push strategy optimization scheme contains multiple possible combinations of push parameters, each representing a push strategy.

[0170] Step S550: Based on the information receiving channel preference sub-vector and content presentation format preference sub-vector in the personalized parameter set, filter the push parameter combinations in the push strategy optimization scheme, and retain the push parameter combination with the highest matching degree with user preferences as the optimal push strategy parameters.

[0171] The information reception channel preference subvector in the personalized parameter set reflects the user's preference for different information reception channels, while the content presentation format preference subvector reflects the user's preference for the presentation format of warning information. Based on these two subvectors, the push parameter combinations in the push strategy optimization scheme are filtered.

[0172] For each combination of push parameters, evaluate the degree of match between its push channel and information presentation format and user preferences. The matching degree can be evaluated using a calculation method that comprehensively considers information from both the information reception channel preference sub-vector and the content presentation format preference sub-vector. The push parameter combination with the highest matching degree to user preferences is retained as the optimal push strategy parameters, thus ensuring that the pushed alerts better meet user needs and preferences.

[0173] Step S560: Convert the optimal push strategy parameters to generate a parameter format that conforms to the push protocol requirements. The parameter format includes the push timestamp, channel identifier, and presentation format encoding.

[0174] The optimal push strategy parameters are the most suitable push strategy for users after screening, but they may need to be formatted to comply with the push protocol requirements. The push protocol specifies the standard format and specifications for warning information pushes, including push timestamps, channel identifiers, and presentation format encodings. The optimal push strategy parameters are formatted by converting information such as push time, push channel, and information presentation format into a format compliant with the push protocol. The push timestamp indicates the time the warning information is pushed, the channel identifier specifies the push channel used, and the presentation format encoding describes how the warning information is presented. Through format conversion, a parameter format compliant with the push protocol requirements is generated, preparing for subsequent warning information pushes.

[0175] As one implementation method, in step S500, the push strategy parameters are converted into multimodal warning information that meets user interaction habits and then pushed out. Specifically, this may include the following steps S570-S5120: Step S570: Analyze the target presentation format in the optimal push strategy parameters, determine the components of the multimodal warning information and the order of each component. The components include text description content, image illustration content and audio prompt content.

[0176] The target presentation format in the optimal push strategy parameters specifies the presentation method and structure of multimodal warning information. Analyzing the target presentation format clarifies the components of the multimodal warning information, namely, text descriptions, image illustrations, and audio prompts. Simultaneously, it determines the order in which these components are arranged, for example, whether to display the text descriptions first, then the image illustrations, and finally the audio prompts, or to use another arrangement.

[0177] Step S580: Generate text description content based on the warning description vector ranked first in the warning priority assessment sequence, extract the meteorological element intensity semantic description information, impact range spatial description information, and time urgency description information from the warning description vector, and organize them into text paragraphs according to natural language logic.

[0178] The warning description vector ranked first in the warning priority assessment sequence represents the most important warning information at present. From this warning description vector, we extract semantic descriptions of meteorological element intensity (e.g., "heavy rainfall," "strong winds"), spatial descriptions of impact range (e.g., "a certain city," "a certain region"), and temporal descriptions of urgency (e.g., "about to happen," "lasting for a period of time"). This information is then organized according to the logic of natural language to obtain a complete text paragraph. The text paragraph should clearly and accurately convey the key content of the warning information, allowing users to quickly understand the weather conditions and potential risks.

[0179] Step S590: Based on the spatial description information of the impact range in the text description, call the map service interface to obtain the map image of the target area, mark the spatial area of ​​the impact range of the weather warning on the map image, add a color gradient effect to represent the difference in impact intensity, and generate the image illustration content.

[0180] Based on the spatial description of the impact range in the text description, the map service interface is invoked. The map service interface is an interface that can acquire map data and images; this interface is used to obtain a map image of the target area. On the acquired map image, the corresponding spatial areas are marked according to the impact range of the weather warning. To more intuitively represent the differences in impact intensity, a color gradient effect is added; for example, using different shades of color to indicate the intensity of the impact.

[0181] Step S5100: Select the corresponding audio prompt template based on the timeliness and urgency information in the text description, convert the key warning information in the text description into speech segments using speech synthesis technology, mix the speech segments with the audio prompt template, and generate audio prompt content.

[0182] Based on the timeliness and urgency information in the text description, an appropriate audio prompt template is selected. Different levels of urgency may correspond to different styles and rhythms of audio prompt templates. For example, a faster-paced, more urgent audio prompt template might be chosen for an imminent emergency warning, while a relatively calmer template might be selected for a warning lasting for a period of time. Key warning information in the text description is then converted into speech segments using speech synthesis technology. Speech synthesis technology can convert text information into speech signals. The speech segments are then mixed with the audio prompt template, and the two are merged together to generate the audio prompt content.

[0183] Step S5110: Combine the text description content, image illustration content, and audio prompt content into a multimodal early warning information data packet according to the arrangement order specified by the target presentation format. The multimodal early warning information data packet contains the format identifier and data length information of each content part.

[0184] Following the order specified by the target presentation format, text descriptions, images, and audio prompts are combined to obtain a multimodal warning information data packet. During the combination process, a format identifier and data length information are added to each content component. The format identifier indicates the format type of the content component, such as text format, image format, or audio format; the data length information indicates the data size of the content component. The multimodal warning information data packet contains complete warning information, and the format identifier and data length information facilitate parsing and processing of the data packet by the receiving end.

[0185] Step S5120: Select the corresponding push interface according to the target push channel in the optimal push strategy parameters, and send the multimodal warning information data packet to the user terminal through the push interface when the target push time arrives.

[0186] Based on the target push channel in the optimal push strategy parameters, the corresponding push interface is selected. Different push channels may have different push interfaces, such as SMS push interfaces, APP push interfaces, and email push interfaces. When the target push time arrives, the multimodal warning information data packet is sent to the user terminal through the selected push interface. The push interface is responsible for transmitting the data packet according to the corresponding protocol and format to ensure that the warning information can be delivered to the user accurately and in a timely manner, thereby realizing a personalized weather warning information push service that meets the user's interaction habits.

[0187] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can be obtained from relevant content in the prior art, and will not be elaborated on in the embodiments of the present invention. In addition, when implementing the solution of the present invention, those skilled in the art can supplement the details based on common knowledge in the art. For example, based on common knowledge in the art, normalization or standardization can be used to eliminate the dimensional conflicts before feature fusion (for example, standardizing the opening time record, content interaction frequency record and cross-platform forwarding behavior record under continuous timestamps respectively, eliminating the dimensional differences, and then organizing them in chronological order), interpolation can be used to eliminate dimensional differences, and thresholds can be reasonably set in combination with historical data, experience or business scenario requirements, the model can be trained based on a general model training method, and the number of layers in the model structure and the selection of activation functions can be set based on actual needs, etc. The present invention will not provide redundant descriptions of the implementation process in excessive detail.

[0188] Please refer to Figure 3 This diagram illustrates the structural block diagram of a computer device 20 provided in one embodiment of this application. This computer device can be used to implement the functions of the aforementioned weather warning data push method based on large model decision-making. Specifically:

[0189] Computer device 20 includes a central processing unit (CPU) 21, a system memory 24 including random access memory (RAM) 22 and read-only memory (ROM) 23, and a system bus 25 connecting the system memory 24 and the CPU 21. Computer device 20 also includes a basic input / output system (I / O system) 26 that facilitates information transfer between various devices within the computer, and a mass storage device 27 for storing the operating system 271.

[0190] The input / output system 26 may include a display for showing information and input devices such as a mouse and keyboard for user input. Both the display and the input devices are connected to the central processing unit 21 via an input / output controller connected to the system bus 25.

[0191] Mass storage device 27 is connected to central processing unit 21 via a mass storage controller (not shown) connected to system bus 25. Mass storage device 27 and its associated computer-readable media provide non-volatile storage for computer device 20. That is, mass storage device 27 may include computer-readable media (not shown) such as hard disk or CD-ROM (Compact Disc Read-Only Memory) drive.

[0192] The memory also includes a computer program stored in the memory and configured to be executed by one or more processors to implement the above-described weather warning data push method based on large model decision-making.

Claims

1. A weather warning data pushing method based on large model decision, characterized in that, The method includes: Obtain geographic tag data and historical early warning response records of users in the target area, and establish a raw dataset of user behavior and early warning feedback through data association mapping. The raw dataset includes records of the opening time of users for different types of early warnings, records of content interaction frequency, and records of cross-platform forwarding behavior. Based on the large model, deep semantic modeling is performed on the original dataset to analyze the evolution of user behavior patterns under different meteorological risk scenarios and construct a personalized parameter set that includes risk perception sensitivity, information receiving channel preference and content presentation format preference. Based on the personalized parameter set, key early warning elements are screened from the real-time meteorological data transmitted by the real-time meteorological monitoring network. The screened early warning elements are then converted into an early warning description vector set through semantic feature encoding. The early warning description vector set includes semantic description information of meteorological element intensity, spatial description information of impact range, and time description information of time urgency. The warning description vector set and the meteorological situation evolution process data output by the meteorological situation prediction system are input into the spatiotemporal inference network of the large model to model the dynamic correlation between changes in meteorological elements and user risk perception, and obtain the warning priority evaluation sequence ranked according to the user's risk exposure level. By integrating device operation status data and network environment parameters uploaded by user terminals, the push strategy is optimized and calculated for the warning priority evaluation sequence through the multi-objective decision layer of the large model, and the push strategy parameters are output. The push strategy parameters are then transformed into multimodal warning information that meets user interaction habits for push.

2. The method of claim 1, wherein, The method involves performing deep semantic modeling on the original dataset based on a large model, analyzing the evolution of user behavior patterns under different meteorological risk scenarios, and constructing a personalized parameter set that includes risk perception sensitivity, information reception channel preferences, and content presentation format preferences, including: The original dataset is divided into meteorological risk scenario levels. Based on the warning type field, the original dataset is decomposed into multiple scenario sub-datasets. The temporal correlation features of user behavior records in each scenario sub-dataset are extracted simultaneously. The temporal correlation features are woven together by continuous timestamp opening time records, content interaction frequency records, and cross-platform forwarding behavior records in chronological order. Behavioral trajectory spectrum analysis is performed on the temporal correlation features of each scenario subset. Behavioral segments of different durations are extracted by sliding window, and the dynamic time regularization distance between segments is calculated. Behavioral segments with a distance less than a threshold are aggregated into behavioral trajectory clusters. Each behavioral trajectory cluster contains a set of user behavioral segments with similar response patterns. The behavior trajectory cluster is input into a large model for context-aware processing. The timestamp sequence in the behavior trajectory cluster is weighted and modeled to generate context-enhanced behavior vectors. In the context-enhanced behavior vectors, the feature weights of behavior segments with newer timestamps are higher than those of behavior segments with older timestamps. The similarity between the context-enhanced behavior vector and the preset meteorological risk scenario prototype vector is calculated to generate a scenario fit vector. The context-enhanced behavior vector is then modulated based on the scenario fit vector to obtain a scenario-specific behavior vector. The scenario-specific behavior vector is subjected to feature deentanglement processing. By minimizing mutual information, the behavior vector is decomposed into three statistically independent sub-vectors, which correspond to the risk perception sensitivity sub-vector, the information receiving channel preference sub-vector, and the content presentation form preference sub-vector, respectively. The risk perception sensitivity sub-vector, information receiving channel preference sub-vector, and content presentation form preference sub-vector are standardized and concatenated to generate a personalized parameter set with unified dimensions. The magnitude of each sub-vector in the personalized parameter set is equal to 1 and they are mutually orthogonal.

3. The method of claim 2, wherein, The step of inputting the behavior trajectory cluster into a large model for context-aware processing, performing weighted modeling on the timestamp sequences in the behavior trajectory cluster, and generating context-enhanced behavior vectors includes: The user behavior segments in the behavior trajectory cluster are normalized in terms of time granularity. Behavior segments of different durations are interpolated into time series of equal length. Each time series contains a fixed number of timestamp nodes, and each node corresponds to a user behavior record of one time unit. Extract behavioral feature triples for each timestamp node. The behavioral feature triples consist of the opening timeliness deviation value, the cumulative interaction frequency value, and the forwarding platform entropy value. The forwarding platform entropy value represents the uniformity of the distribution of user forwarding behavior across different platforms. The behavioral feature triples are input into the query generation layer in timestamp order to generate a query vector for each timestamp node. The preset context feature template is used as the key vector and value vector to calculate the similarity score between the query vector and the key vector. The value vectors are weighted and summed based on the similarity scores to generate a timestamp-level context vector. The dimensions of the timestamp-level context vector are the same as those of the value vectors, and it contains the interaction information between the context feature template and the behavioral feature triplet. Gated recurrent units are used to process the timestamp-level context vectors of all timestamp nodes to capture the dependencies in the time series and generate trajectory-level context vectors. The dimensions of the trajectory-level context vectors are the same as those of the timestamp-level context vectors. The trajectory-level context vector is mapped to a preset dimension space through a fully connected layer to generate a context-enhancing behavior vector. The dimension of the context-enhancing behavior vector is the same as the feature processing dimension of the large model.

4. The method of claim 3, wherein, The feature deentanglement processing of the scene-specific behavior vector involves decomposing the behavior vector into three statistically independent sub-vectors by minimizing mutual information, including: Initialize three learnable feature mapping matrices, corresponding to the risk perception sensitivity mapping matrix, the information receiving channel preference mapping matrix, and the content presentation format preference mapping matrix, respectively. The number of rows in each mapping matrix is ​​equal to the dimension of the scenario-specific behavior vector, and the number of columns is equal to the dimension of the target sub-vector. The scenario-specific behavior vector is multiplied by the three feature mapping matrices to generate three initial sub-vectors, namely the initial risk perception sensitivity sub-vector, the initial information receiving channel preference sub-vector, and the initial content presentation format preference sub-vector. Calculate the mutual information value between any two initial sub-vectors, measure the statistical correlation between the sub-vectors by the divergence between the joint probability distribution and the marginal probability distribution, and generate a mutual information matrix, where the matrix elements represent the magnitude of the mutual information between the corresponding two sub-vectors; Construct an optimization function with the trace norm of the mutual information matrix as the objective, and adjust the parameters of the three feature mapping matrices to minimize the trace norm of the mutual information matrix until the mutual information values ​​of any two sub-vectors are lower than a preset threshold. Once the optimization function converges, the scenario-specific behavior vector is multiplied by the three optimized feature mapping matrices to generate statistically independent risk perception sensitivity sub-vectors, information receiving channel preference sub-vectors, and content presentation format preference sub-vectors. The risk perception sensitivity subvector, information receiving channel preference subvector, and content presentation form preference subvector are subjected to L2 normalization to make the magnitude of each subvector equal to 1, thereby generating a de-entangled subvector that satisfies the unit vector constraint. The features of each dimension of the de-entangled subvector are mutually orthogonal.

5. The method according to claim 1, characterized in that, The process involves filtering key early warning elements from real-time meteorological data transmitted through the real-time meteorological monitoring network based on the personalized parameter set, and then converting the filtered early warning elements into an early warning description vector set through semantic feature encoding, including: The risk perception sensitivity sub-vector in the personalized parameter set is analyzed, the variance value of each dimension feature is calculated, and the variance values ​​are sorted from largest to smallest to generate an element sensitivity ranking table. The meteorological elements corresponding to the high variance dimension in the element sensitivity ranking table have higher user attention potential. Based on the aforementioned element sensitivity ranking table, meteorological elements in real-time meteorological data are hierarchically filtered. Elements ranked at the top of the ranking table are retained as a set of highly sensitive elements. Elements that are strongly correlated with the set of highly sensitive elements are extracted from the remaining elements as a set of associated elements and merged into a set of candidate early warning elements. Multicollinearity diagnosis is performed on the candidate early warning element set. The degree of multicollinearity among the elements is calculated by the variance inflation factor. Elements with variance inflation factors exceeding the threshold are removed, and elements with variance inflation factors below the threshold are retained as key early warning elements. Extract the semantic description text, numerical change range, and duration of impact of the key early warning elements, and combine the three into a structured early warning element unit according to semantic association rules. Each structured early warning element unit contains a text description segment, a numerical range segment, and a time segment marker segment. The structured early warning element unit is input into the dynamic semantic encoder, which performs contextual semantic encoding on the text description segment to generate a text semantic vector, performs interval boundary encoding on the numerical interval segment to generate a numerical boundary vector, and performs time interval encoding on the duration of the impact to generate a time interval vector. The text semantic vector, numerical boundary vector, and time interval vector are input into the feature fusion processor. The fusion weights of each vector are dynamically adjusted through an adaptive gating mechanism to generate a warning description vector containing multi-dimensional semantic information. Multiple warning description vectors are grouped according to feature type to form a warning description vector set.

6. The method according to claim 5, characterized in that, The step of performing multicollinearity diagnosis on the candidate early warning element set, calculating the degree of multicollinearity among elements using the variance inflation factor, and removing elements whose variance inflation factor exceeds a threshold, includes: Construct an element correlation matrix for the candidate early warning element set, where the rows and columns of the matrix correspond to the candidate early warning elements, and the matrix elements are the correlation coefficients between two corresponding elements. The values ​​of the correlation coefficients are within a preset range. A regression model is constructed with each candidate early warning element as the dependent variable and other candidate early warning elements as independent variables. The coefficient of determination of each regression model is calculated, and the variance inflation factor is calculated based on the coefficient of determination. The variance inflation factor is represented by the reciprocal of the complementary value of the coefficient of determination. Compare the variance inflation factor of each candidate early warning element with a preset threshold, mark elements whose variance inflation factor exceeds the threshold as highly collinear elements, and record the position of highly collinear elements in the element sensitivity ranking table. A stepwise elimination strategy is adopted to process highly collinear features. After each elimination, the variance inflation factor of the remaining features is recalculated until the variance inflation factor of all features is lower than the threshold. Check whether the removed element set contains the core elements of meteorological early warning. If there are missing elements, select the element with the highest sensitivity ranking from the highly collinear elements to supplement it. After supplementation, re-verify the collinearity. The set of elements that have undergone collinearity verification and core element supplementation is determined as key early warning elements, and the number of key early warning elements is a preset ratio range of the number of candidate early warning elements.

7. The method according to claim 5, characterized in that, The step of inputting the structured early warning element unit into a dynamic semantic encoder to perform contextual semantic encoding on the text description segment to generate a text semantic vector includes: The text description segment of the structured early warning element unit is processed by word segmentation, and the continuous text is divided into independent word units. After removing stop words, the valid word units are retained. The valid word units correspond to meteorological early warning professional terms. The effective word units are input into the pre-trained word embedding model. Each word unit is converted into a fixed-dimensional word vector by word lookup. All word vectors are subjected to average pooling to generate preliminary text vectors. The dimensions of the preliminary text vectors are consistent with the dimensions of the word vectors. The initial text vector is input into a bidirectional recurrent network. The forward recurrent layer calculates the hidden state from the beginning of the word sequence backward, and the backward recurrent layer calculates the hidden state from the end of the word sequence forward. At each time step, the forward and backward hidden states are concatenated into a fusion vector. Attention weighting is applied to the fusion vectors at all time steps, the contribution weight of each fusion vector to the semantics of the text is calculated, and an attention weight vector is generated. The fusion vectors are then weighted and summed using the weight vectors to obtain the context enhancement vector. The context enhancement vector is input into a convolutional layer, and features are extracted from the context enhancement vector through multiple convolutional kernels to generate multi-scale text feature maps. The number of multi-scale text feature maps is the same as the number of convolutional kernels. Global max pooling is performed on the multi-scale text feature maps to extract salient features from each feature map and generate text semantic vectors. The dimension of the text semantic vectors is positively correlated with the number of multi-scale text feature maps.

8. The method according to claim 1, characterized in that, The process involves inputting the warning description vector set and the meteorological situation evolution data output by the meteorological situation prediction system into the spatiotemporal inference network of the large model to model the dynamic correlation between changes in meteorological elements and user risk perception, thereby obtaining a warning priority assessment sequence ranked according to the degree of user risk exposure, including: The meteorological situation evolution data is analyzed, and the predicted values ​​and trends of meteorological elements at different time points are extracted and combined in chronological order to form a meteorological situation time series. The length of the time series is equal to the number of time intervals in the prediction period. The warning description vector set is matched with the meteorological situation time series by timestamp matching to establish the association between the warning elements and the meteorological situation data at the corresponding time nodes, and a spatiotemporal association feature matrix is ​​generated, where the rows of the matrix correspond to the time nodes and the columns correspond to the warning elements. Spatial dimension feature enhancement is performed on the spatiotemporal correlation feature matrix, the contribution weight of each early warning element in the user's risk perception is calculated, a spatial attention weight matrix is ​​generated, and the correlation feature matrix is ​​weighted by the weight matrix to obtain the spatial enhancement feature matrix. The spatial augmentation feature matrix is ​​dynamically modeled in the time dimension. The dependency relationship of meteorological elements over time is captured by multi-layer convolution processing to generate a time evolution feature vector. The dimension of the evolution feature vector is consistent with the column dimension of the spatial augmentation feature matrix. The time evolution feature vector is correlated with the risk perception sensitivity sub-vector in the personalized parameter set to form a model, and the matching degree between each early warning element and the user's risk perception is calculated to generate a risk perception matching degree vector. Based on the element values ​​in the risk perception matching vector, the warning description vectors in the warning description vector set are sorted, and a warning priority evaluation sequence is generated in descending order of element values. The order of vectors in the sequence reflects the user's attention to the warning information.

9. The method according to claim 8, characterized in that, The step of enhancing the spatial dimension features of the spatiotemporal correlation feature matrix, calculating the contribution weight of each early warning element in the user's risk perception, and generating a spatial attention weight matrix includes: The spatiotemporal correlation feature matrix is ​​subjected to feature compression processing. The time dimension of the matrix is ​​compressed by global average pooling, while the spatial dimension features are preserved to generate a spatial feature vector. The dimension of the spatial feature vector is consistent with the column dimension of the correlation feature matrix. The spatial feature vector is input into a two-layer fully connected network. The first fully connected network performs dimensionality reduction on the spatial feature vector to generate an intermediate vector. The second fully connected network performs dimensionality increase on the intermediate vector to generate a spatial weight vector with the same dimension as the spatial feature vector. The spatial weight vector is multiplied element-wise with the risk perception sensitivity sub-vector in the personalized parameter set to adjust the weight values ​​of each early warning element, enhance the element weights corresponding to high-value dimensions in the risk perception sensitivity sub-vector, and generate the adjusted spatial weight vector. The adjusted spatial weight vector is normalized so that the value range of each element in the weight vector is within a preset range, so as to avoid the weight values ​​being too large or too small and affecting subsequent processing. The normalized spatial weight vector is copied and expanded to generate a spatial attention weight matrix with the same size as the spatiotemporal correlation feature matrix. The weight row vector corresponding to each time node in the matrix is ​​the same as the normalized spatial weight vector. The spatial attention weight matrix is ​​used to perform element-wise multiplication on the spatiotemporal correlation feature matrix to enhance the feature values ​​of early warning elements with weight values ​​greater than the average weight value, and weaken the feature values ​​of early warning elements with weight values ​​less than the average weight value, thereby generating a spatial enhancement feature matrix.

10. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, which is loaded and executed by the processor to implement the weather warning data push method based on large model decision-making as described in any one of claims 1 to 9.

Images