Low-temperature rain and snow freezing disaster risk prediction method and system based on big data

Abstract

The invention relates to a low-temperature rain and snow freezing disaster risk prediction method and system based on big data, and the method comprises the steps: collecting and preprocessing multi-source heterogeneous data, and obtaining a standardized data set, the multi-source heterogeneous data including meteorological data, traffic data and power data; extracting dynamic features in the meteorological data according to the standardized data set, and screening the dynamic features to obtain an optimized feature set; performing feature extraction on the optimized feature set to obtain a spatial-temporal feature vector; fusing the traffic data and the power data to obtain a comprehensive influence factor; combining the spatial-temporal feature vector with the comprehensive influence factor to obtain a disaster risk feature matrix; and inputting the disaster risk feature matrix into a risk prediction model, and outputting a disaster risk level and probability distribution, the risk prediction model being constructed based on a long short-term memory network. According to the method, accurate risk prediction of low-temperature rain and snow freezing disaster situations is realized, and powerful decision support is provided for disaster prevention and reduction.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a method and system for predicting the risk of low-temperature rain, snow and freezing disasters based on big data. Background Art

[0002] Cold, rain, snow, and freezing weather are major natural disasters that impact the socioeconomic landscape and public safety. Their prediction and assessment are crucial for minimizing losses and safeguarding people's livelihoods. Traditional methods rely primarily on single data sources and simple statistical models, making it difficult to integrate data from multiple sources. Furthermore, their prediction accuracy is insufficient to cope with the dynamic changes in complex meteorological and geographical conditions. This makes real-time disaster risk assessment and accurate early warning a pressing challenge.

[0003] In the field of disaster prediction, the core challenge lies in how to effectively integrate multi-source heterogeneous data and extract key features from them. The formats and dimensions of meteorological, transportation, electricity, geographical and other data vary greatly, making it difficult to unify traditional processing methods, resulting in inefficient information fusion. The difficulty of data integration further affects the accuracy of feature extraction, especially dynamic features directly related to disaster risk, such as the temperature drop rate and the rain-snow conversion coefficient, which are difficult to accurately capture through a single algorithm. The quality of these feature extraction directly determines the performance of the prediction model. If they cannot be accurately extracted, the model will find it difficult to reflect the temporal and spatial evolution of the disaster, which will lead to distorted risk level predictions.

[0004] Therefore, how to efficiently integrate multi-source heterogeneous data, accurately extract dynamic features related to low-temperature rain, snow and freezing disasters, and build a high-precision risk prediction model based on this has become a key research issue. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for predicting the risk of low-temperature rain, snow and freezing disasters based on big data, which can realize accurate risk prediction of low-temperature rain, snow and freezing disaster scenarios and provide powerful decision-making support for disaster prevention and mitigation.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data, including:

[0008] Collecting multi-source heterogeneous data and preprocessing it to obtain a standardized data set, wherein the multi-source heterogeneous data includes meteorological data, traffic data, and power data;

[0009] extracting dynamic features from the meteorological data according to the standardized data set, screening the dynamic features, and obtaining an optimized feature set;

[0010] Performing feature extraction on the optimized feature set to obtain a spatiotemporal feature vector;

[0011] fusing the traffic data with the power data to obtain a comprehensive impact factor;

[0012] Combining the spatiotemporal characteristic vector and the comprehensive influencing factor to obtain a disaster risk characteristic matrix;

[0013] The disaster risk characteristic matrix is ​​input into a risk prediction model, and the disaster risk level and probability distribution are output, wherein the risk prediction model is constructed based on a long short-term memory network and obtained through training with a training set, and the training set includes a historical disaster risk characteristic matrix.

[0014] Optionally, collect and preprocess multi-source heterogeneous data to obtain a standardized dataset, including:

[0015] Acquire multi-source heterogeneous data from meteorological, transportation, and power data sources through preset protocols, and parse them using standard interfaces and protocols to obtain the original data set;

[0016] If the original data set contains missing values ​​or outliers, the data set is cleaned using mean filling and K-nearest neighbor algorithm to obtain a cleaned data set;

[0017] Based on the cleaned data set, a format conversion algorithm is used to map the formats of different data sources into a unified structure to obtain a data set with consistent format;

[0018] If the dimensions of the datasets with the same format are not uniform, the numerical data are normalized using a normalization algorithm to obtain a dataset with uniform dimensions.

[0019] The standardized data set is obtained by using a data set with unified dimensions and using data verification rules to determine data integrity and consistency.

[0020] Optionally, based on the standardized data set, extracting dynamic features from the meteorological data includes: performing dimensionality reduction processing on the standardized data set through a principal component analysis algorithm, extracting principal components, and obtaining dynamic features including the temperature drop rate and the rain-snow conversion coefficient.

[0021] Optionally, screening the dynamic features to obtain an optimized feature set includes:

[0022] If the cumulative variance contribution rate of the principal component reaches the preset threshold, the corresponding principal component is retained and the preliminary feature set is obtained;

[0023] Using a cluster analysis algorithm to group the preliminary feature set, determine the correlation between features, and obtain feature grouping results;

[0024] From the feature grouping results, features that have a preset correlation with the temperature drop rate and the rain-snow conversion coefficient are selected to construct the optimized feature set.

[0025] Optionally, performing feature extraction on the optimized feature set to obtain a spatiotemporal feature vector includes:

[0026] Inputting the optimized feature set into a convolutional neural network, performing spatiotemporal feature extraction on the optimized feature set, and obtaining initial spatiotemporal features;

[0027] If the resolution of the initial spatiotemporal features is lower than a preset threshold, enhancing the resolution of the feature map by upsampling to obtain an enhanced feature map;

[0028] The pooling layer of a convolutional neural network is used to extract key spatiotemporal features from the enhanced feature map to generate the spatiotemporal feature vector.

[0029] Optionally, fusing the traffic data with the power data to obtain a comprehensive impact factor includes:

[0030] Extracting key traffic characteristics and grid load regional distribution data based on the traffic data and the power data, respectively, wherein the regional distribution data includes real-time power fluctuations of different power supply nodes;

[0031] The key traffic characteristics are fused with the regional distribution data of the power grid load by a weighted average method to obtain the comprehensive impact factor.

[0032] Optionally, extracting key traffic characteristics and grid load regional distribution data separately includes: inputting the traffic data and the power data distribution into a random forest algorithm to extract the key traffic characteristics and grid load regional distribution data.

[0033] Optionally, outputting the disaster risk level and probability distribution includes: if the shortest distance between one area and another area is less than a preset threshold, merging the areas into the same rain, snow and freezing process;

[0034] If the overlap rate of meteorological stations in the rain, snow and freezing area on the day and the rain, snow and freezing area on the previous day is greater than the preset value, then the two days are the same rain, snow and freezing event; otherwise, they are two separate rain, snow and freezing events.

[0035] The present invention also provides a system implemented according to a method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data, comprising: a data acquisition module, a data processing module, a first fusion module, a second fusion module, and a risk prediction module;

[0036] The data acquisition module is used to collect and preprocess multi-source heterogeneous data to obtain a standardized data set, wherein the multi-source heterogeneous data includes meteorological data, traffic data, and power data;

[0037] The data processing module is configured to extract dynamic features from the meteorological data based on the standardized data set, screen the dynamic features to obtain an optimized feature set, perform feature extraction on the optimized feature set, and obtain a spatiotemporal feature vector;

[0038] The first fusion module is configured to fuse the traffic data with the power data to obtain a comprehensive impact factor;

[0039] The second fusion module is used to combine the spatiotemporal feature vector and the comprehensive impact factor to obtain a disaster risk feature matrix;

[0040] The risk prediction module is used to input the disaster risk characteristic matrix into a risk prediction model and output the disaster risk level and probability distribution, wherein the risk prediction model is constructed based on a long short-term memory network and obtained through training with a training set, and the training set includes a historical disaster risk characteristic matrix.

[0041] The beneficial effects of the present invention are as follows: the present invention obtains multi-source heterogeneous data such as meteorological, transportation, and electricity through a preset standardized protocol, and uses a data cleaning and format conversion algorithm to unify the data format. Principal component analysis is used to extract key features. The random forest algorithm is combined to analyze key traffic characteristics and regional distribution characteristics of power grid loads, and a disaster risk feature matrix is ​​obtained through weighted fusion. Finally, a risk prediction model is constructed based on a long-short-term memory network, and the gradient boosting algorithm is used to optimize the model parameters. The present invention achieves accurate risk prediction for low-temperature rain, snow, and freezing disaster scenarios, providing powerful decision-making support for disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 This is a flowchart of a method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1:

[0047] like Figure 1 As shown, this embodiment provides a method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data, including:

[0048] Collect and preprocess multi-source heterogeneous data to obtain standardized data sets. Multi-source heterogeneous data includes meteorological data, traffic data, and power data.

[0049] According to the standardized data set, dynamic features in meteorological data are extracted, and the dynamic features are screened to obtain the optimized feature set;

[0050] Perform feature extraction on the optimized feature set to obtain the spatiotemporal feature vector;

[0051] Integrate traffic data with power data to obtain comprehensive impact factors;

[0052] Combine the spatiotemporal characteristic vectors and comprehensive impact factors to obtain the disaster risk characteristic matrix;

[0053] The disaster risk characteristic matrix is ​​input into the risk prediction model, and the disaster risk level and probability distribution are output. The risk prediction model is constructed based on the long short-term memory network and obtained through training with a training set, which includes the historical disaster risk characteristic matrix.

[0054] Specifically, meteorological, transportation, and power data are crucial for predicting, assessing, and preventing the risks of freezing rain, snow, and freezing weather. By integrating and analyzing this data, we can more accurately predict the occurrence of disasters, assess their intensity, and assess their impact. This provides scientific decision-making support for relevant departments and the public, helping to mitigate potential losses.

[0055] In this embodiment, the spatiotemporal feature vector is combined with the comprehensive impact factor through a weighted fusion algorithm. Assuming that the normalized value of vehicle flow at a certain moment is 0.8, the grid load is 0.6, and the comprehensive impact factor is 0.75, the weighted calculation is (0.7×0.8+0.3×0.6=0.74).

[0056] Furthermore, we collect and preprocess multi-source heterogeneous data to obtain standardized data sets, including:

[0057] Acquire multi-source heterogeneous data from meteorological, transportation, and power data sources through preset protocols, and parse them using standard interfaces and protocols to obtain the original data set;

[0058] If the original data set contains missing values ​​or outliers, mean filling and K-nearest neighbor algorithm are used to clean the data set to obtain the cleaned data set;

[0059] Based on the cleaned data set, a format conversion algorithm is used to map the formats of different data sources into a unified structure to obtain a data set with consistent format;

[0060] If the dimensions of the datasets with the same format are not uniform, the numerical data are normalized using a normalization algorithm to obtain a dataset with uniform dimensions.

[0061] Through the unified dimension data set, data verification rules are used to judge the data integrity and consistency to obtain a standardized data set.

[0062] Specifically, in one possible implementation, heterogeneous data from multiple sources, such as meteorological, transportation, and electricity, must be acquired through standard protocols, such as the WMO standard for meteorological data, the GTFS format for transportation data, and the IEC61850 protocol for electricity data. Temperature and humidity data are obtained from meteorological stations via HTTP, real-time vehicle flow data is obtained from transportation departments, and grid load data is obtained from power companies. The original datasets may contain missing values, such as when a weather station does not record a temperature at a certain hour, or outliers, such as when traffic flow suddenly increases to an unreasonable value. To address missing values, mean imputation can be used to address missing temperatures in meteorological data for a certain hour. This imputation is performed by taking the average temperature of the two preceding and following hours. For example, if the temperature is 25°C in the preceding hour and 27°C in the following hour, the imputation is 26°C. Anomaly detection uses the K-nearest neighbor algorithm, with K=5, to calculate the Euclidean distance between a data point and its neighbors. If the traffic flow in a certain hour is 5,000 vehicles, which far exceeds the neighboring mean of 1,000 vehicles, the data point is marked as an anomaly and removed. This cleansing process ensures data reliability and provides high-quality input for subsequent analysis. Format conversion requires mapping the heterogeneous formats of different data sources into a unified structure. ETL tools are used to convert meteorological data in CSV format, traffic data in JSON format, and power data in XML format into a unified JSON structure, including fields such as timestamp, value, and unit. Dimensional normalization normalizes the different dimensions of numerical data. Temperature (°C), vehicle volume (vehicles / hour), and grid load (MW) have different dimensions, so minimum-maximum normalization is used to map the data to the [0, 1] range.

[0063] Furthermore, based on the standardized data set, the dynamic features in the meteorological data are extracted, including: performing dimensionality reduction processing on the standardized data set through a principal component analysis algorithm, extracting the principal components, and obtaining dynamic features including the temperature drop rate and the rain-snow conversion coefficient.

[0064] Specifically, principal component analysis combines the original variables into new variables through linear transformation, preserving the data's key information. Assuming the standardized data contains 10 variables, such as temperature, precipitation, and humidity, principal component analysis can extract two to three principal components, explaining over 80% of the variance. The rate of temperature drop can be inferred from the time-dependent variables in the principal components, while the rain-to-snow conversion coefficient is related to the combined variables of precipitation and temperature.

[0065] Furthermore, dynamic features are screened to obtain an optimized feature set including:

[0066] If the cumulative variance contribution rate of the principal component reaches the preset threshold, the corresponding principal component is retained and the preliminary feature set is obtained;

[0067] Use cluster analysis algorithm to group the preliminary feature set, determine the correlation between features, and obtain feature grouping results;

[0068] From the feature grouping results, features that have a preset correlation with the temperature drop rate and the rain-snow conversion coefficient are selected to construct an optimized feature set.

[0069] Specifically, if the cumulative variance contribution reaches a preset threshold of 85%, these principal components are retained to form a preliminary feature set. This dimensionality reduction method reduces data redundancy and improves analysis efficiency. Based on this preliminary feature set, a cluster analysis algorithm is used to determine the correlation between features and obtain feature grouping results.

[0070] In one possible implementation, the K-means clustering algorithm is used to group features into three groups, representing temperature-related, precipitation-related, and comprehensive meteorological characteristics. Hypothetically, analysis reveals that the temperature drop rate is highly correlated with nighttime temperature and humidity, and the rain-to-snow conversion coefficient is correlated with precipitation and air pressure. The clustering results clearly demonstrate the correlation between these features. This grouping helps identify key features and improves the targeted nature of subsequent modeling. From the resulting feature grouping, features highly correlated with the temperature drop rate and the rain-to-snow conversion coefficient are extracted to construct an optimized feature set. It should be noted that the optimized feature set only retains features with a correlation with the target variable greater than 0.7 to reduce interference from irrelevant variables. This screening method ensures that the feature set focuses on core dynamic features. Modeling of the optimized feature set is performed using a regression analysis algorithm to determine the prediction model for the dynamic features and obtain the final feature set. If the variance contribution rate of the preliminary feature set falls below the preset threshold of 0.85, a recursive feature elimination algorithm is used to filter out features highly correlated with disaster risk to obtain the optimized feature set.

[0071] Furthermore, feature extraction is performed on the optimized feature set to obtain the spatiotemporal feature vector, including:

[0072] The optimized feature set is input into the convolutional neural network, and the spatiotemporal features of the optimized feature set are extracted to obtain the initial spatiotemporal features;

[0073] If the resolution of the initial spatiotemporal features is lower than a preset threshold, the feature map resolution is enhanced by upsampling to obtain an enhanced feature map;

[0074] The pooling layer of the convolutional neural network is used to extract key spatiotemporal features from the enhanced feature map and generate a spatiotemporal feature vector.

[0075] Specifically, convolutional neural networks are used to extract spatiotemporal features from the reduced-dimensionality feature set. The reduced-dimensionality feature set can be viewed as two-dimensional grid data, with each grid representing a feature value at a specific time and location. The convolutional neural network scans the grid through the convolution kernel, capturing the spatial distribution patterns of features such as rainfall and slope, and generating initial spatiotemporal features. The initial spatiotemporal features may show that a certain area has a higher risk due to a combination of high rainfall and steep slopes. If the resolution of the initial spatiotemporal features is low, such as 64x64 pixels, which is lower than the preset threshold of 128x128 pixels, the resolution is increased to 128x128 pixels through an upsampling operation (such as bilinear interpolation) to obtain enhanced features. This enhancement can improve the detail expression capability of the feature map. The pooling layer extracts key spatiotemporal features from the enhanced feature map. The pooling operation focuses on the features of high-risk areas through maximum pooling or average pooling.

[0076] Furthermore, traffic data and power data are integrated to obtain comprehensive impact factors including:

[0077] Key traffic characteristics and regional distribution data of power grid load are extracted based on traffic data and power data, respectively. The regional distribution data includes real-time power fluctuations at different power supply nodes.

[0078] The key traffic characteristics are integrated with the regional distribution data of the power grid load through the weighted average method to obtain the comprehensive impact factors.

[0079] Furthermore, respectively extracting key traffic characteristics and grid load regional distribution data includes: inputting traffic data and power data distribution into a random forest algorithm to extract key traffic characteristics and grid load regional distribution data.

[0080] Specifically, the random forest algorithm processes traffic data to identify trends in road capacity. If the prediction error for the trend is less than a preset threshold, feature selection is performed on the traffic data to identify key traffic characteristics, including traffic volume, average speed, and road congestion index. The random forest algorithm randomly samples and splits these characteristics to generate multiple decision trees. Each tree predicts capacity changes, and the average is ultimately used to identify trends. The random forest algorithm also processes power data to identify regional distribution data on grid loads.

[0081] Furthermore, training the risk prediction model includes: inputting the historical disaster risk feature matrix into the risk prediction model, outputting the predicted risk level and probability distribution, and optimizing the model parameters using back propagation.

[0082] Specifically, a risk prediction model is used to process time series data, outputting a preliminary risk level and probability distribution. If the preliminary risk level exceeds a preset threshold, a secondary analysis of the feature matrix is ​​performed to determine whether model parameters need to be adjusted. Based on the results of the secondary analysis, the risk prediction model is updated to obtain an optimized risk level and probability distribution. Based on the optimized probability distribution, a confidence interval for the risk level is calculated to determine the final output.

[0083] When processing new data, the risk prediction model inputs the feature matrix for the past 30 days and outputs a daily risk value. Assuming a risk value ranging from 0 to 1, it is mapped to three levels: low, medium, and high, with thresholds of 0.3 and 0.7, respectively. The model might predict a risk value of 0.75 on a given day, classifying it as high risk. The probability distribution shows a 60% probability of high risk, a 30% probability of medium risk, and a 10% probability of low risk. This output provides a quantitative basis for decision-making, facilitating prioritization of high-risk areas.

[0084] In one possible implementation, when updating the model, the learning rate is adjusted to 0.001, and the number of hidden layer units is increased to 150 to enhance the capture of complex patterns. After retraining, the model outputs an optimized risk level. On a given day, the risk value drops to 0.65, still considered high risk, but the probability distribution is adjusted to 50% high risk, 40% medium risk, and 10% low risk, more evenly reflecting the influence of multiple factors. This optimization improves the robustness of the forecast.

[0085] In this embodiment, the accuracy and recall of the risk prediction model are evaluated by a cross-validation algorithm to obtain model performance indicators. The data is divided by a cross-validation method to obtain a training set and a test set. The K-fold cross-validation algorithm is used to divide the data set into K subsets, and K-1 subsets are iteratively used as training sets, and the remaining 1 subset is used as a test set to obtain K division results. The test set of each division is predicted by the risk prediction model to obtain the classification result. The accuracy and recall are calculated based on the classification results to obtain the performance indicators of each division. If the accuracy or recall is lower than the preset threshold, the performance indicator of the division is recorded to obtain an abnormal performance set. The comprehensive performance indicator of the model is obtained by averaging the performance indicators of the K divisions. Based on the comparison of the comprehensive performance indicator and the abnormal performance set, the stability of the model is judged to obtain the final evaluation result.

[0086] Furthermore, the output of disaster risk level and probability distribution includes:

[0087] If the shortest distance between one area and another area is less than a preset threshold, the areas are merged into the same rain, snow and freezing process; the distance threshold in this embodiment is 350 km.

[0088] If the overlap rate of meteorological stations in the rain, snow and freezing area on the day is ≥30% with that in the rain, snow and freezing area on the previous day, then the two days are the same rain, snow and freezing event; otherwise, they are two separate rain, snow and freezing events.

[0089] Example 2:

[0090] A low-temperature rain, snow and freezing disaster risk prediction system based on big data, comprising: a data acquisition module, a data processing module, a first fusion module, a second fusion module, and a risk prediction module;

[0091] The data acquisition module is used to collect and preprocess multi-source heterogeneous data to obtain a standardized data set. The multi-source heterogeneous data includes meteorological data, traffic data, and power data.

[0092] The data processing module is used to extract dynamic features from meteorological data based on the standardized data set, screen the dynamic features to obtain an optimized feature set, perform feature extraction on the optimized feature set, and obtain a spatiotemporal feature vector;

[0093] The first fusion module is used to fuse traffic data with power data to obtain comprehensive impact factors;

[0094] The second fusion module is used to combine the spatiotemporal feature vectors and the comprehensive impact factors to obtain the disaster risk feature matrix;

[0095] The risk prediction module is used to input the disaster risk feature matrix into the risk prediction model and output the disaster risk level and probability distribution. The risk prediction model is constructed based on the long short-term memory network and obtained through training with a training set, which includes the historical disaster risk feature matrix.

[0096] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data, characterized in that: include: Collecting multi-source heterogeneous data and preprocessing it to obtain a standardized data set, wherein the multi-source heterogeneous data includes meteorological data, traffic data, and power data; extracting dynamic features from the meteorological data according to the standardized data set, screening the dynamic features, and obtaining an optimized feature set; Performing feature extraction on the optimized feature set to obtain a spatiotemporal feature vector; fusing the traffic data with the power data to obtain a comprehensive impact factor; Combining the spatiotemporal characteristic vector and the comprehensive influencing factor to obtain a disaster risk characteristic matrix; The disaster risk characteristic matrix is ​​input into a risk prediction model, and the disaster risk level and probability distribution are output, wherein the risk prediction model is constructed based on a long short-term memory network and obtained through training with a training set, and the training set includes a historical disaster risk characteristic matrix.

2. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 1 is characterized in that: Collecting multi-source heterogeneous data and preprocessing it to obtain a standardized data set includes: Acquire multi-source heterogeneous data from meteorological, transportation, and power data sources through preset protocols, and parse them using standard interfaces and protocols to obtain the original data set; If the original data set contains missing values ​​or outliers, the data set is cleaned using mean filling and K-nearest neighbor algorithm to obtain a cleaned data set; Based on the cleaned data set, a format conversion algorithm is used to map the formats of different data sources into a unified structure to obtain a data set with consistent format; If the dimensions of the datasets with the same format are not uniform, the numerical data are normalized using a normalization algorithm to obtain a dataset with uniform dimensions. The standardized data set is obtained by using a data set with unified dimensions and using data verification rules to determine data integrity and consistency.

3. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 1 is characterized in that: Extracting dynamic features from the meteorological data based on the standardized data set includes: performing dimensionality reduction processing on the standardized data set through a principal component analysis algorithm, extracting principal components, and obtaining dynamic features including a temperature drop rate and a rain-snow conversion coefficient.

4. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 3 is characterized in that: Screening the dynamic features to obtain an optimized feature set includes: If the cumulative variance contribution rate of the principal component reaches the preset threshold, the corresponding principal component is retained and the preliminary feature set is obtained; Using a cluster analysis algorithm to group the preliminary feature set, determine the correlation between features, and obtain feature grouping results; From the feature grouping results, features that have a preset correlation with the temperature drop rate and the rain-snow conversion coefficient are selected to construct the optimized feature set.

5. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 1 is characterized in that: Extracting features from the optimized feature set to obtain a spatiotemporal feature vector includes: Inputting the optimized feature set into a convolutional neural network, performing spatiotemporal feature extraction on the optimized feature set, and obtaining initial spatiotemporal features; If the resolution of the initial spatiotemporal features is lower than a preset threshold, enhancing the resolution of the feature map by upsampling to obtain an enhanced feature map; The pooling layer of a convolutional neural network is used to extract key spatiotemporal features from the enhanced feature map to generate the spatiotemporal feature vector.

6. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 1 is characterized in that: The traffic data and the power data are integrated to obtain comprehensive impact factors, including: Extracting key traffic characteristics and grid load regional distribution data based on the traffic data and the power data, respectively, wherein the regional distribution data includes real-time power fluctuations of different power supply nodes; The key traffic characteristics are fused with the regional distribution data of the power grid load by a weighted average method to obtain the comprehensive impact factor.

7. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 6 is characterized in that: Extracting key traffic characteristics and grid load regional distribution data separately includes: inputting the traffic data and the power data distribution into a random forest algorithm to extract the key traffic characteristics and grid load regional distribution data.

8. The method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to claim 1 is characterized in that: Outputting the disaster risk level and probability distribution includes: if the shortest distance from one area to another area is less than a preset threshold, the areas are merged into the same rain, snow and freezing process; If the rain, snow and freezing area on the day overlaps with the rain, snow and freezing area on the previous day and the overlap rate of meteorological stations is greater than the preset value, then the two days are the same rain, snow and freezing event; otherwise, they are two separate rain, snow and freezing events.

9. A system for implementing a method for predicting the risk of low-temperature rain, snow and freezing disasters based on big data according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, data processing module, first fusion module, second fusion module, risk prediction module; The data acquisition module is used to collect and preprocess multi-source heterogeneous data to obtain a standardized data set, wherein the multi-source heterogeneous data includes meteorological data, traffic data, and power data; The data processing module is configured to extract dynamic features from the meteorological data based on the standardized data set, screen the dynamic features to obtain an optimized feature set, perform feature extraction on the optimized feature set, and obtain a spatiotemporal feature vector; The first fusion module is configured to fuse the traffic data with the power data to obtain a comprehensive impact factor; The second fusion module is used to combine the spatiotemporal feature vector and the comprehensive impact factor to obtain a disaster risk feature matrix; The risk prediction module is used to input the disaster risk characteristic matrix into a risk prediction model and output the disaster risk level and probability distribution, wherein the risk prediction model is constructed based on a long short-term memory network and obtained through training with a training set, and the training set includes a historical disaster risk characteristic matrix.

Images