Airport severe convective weather early recognition method based on multi-source heterogeneous meteorological data

Abstract

The invention provides an airport severe convective weather early recognition method based on multi-source heterogeneous meteorological data, and belongs to the technical field of meteorological prediction. Comprising the following steps: S1, multi-source heterogeneous data description and sample data extraction; used sample data is mainly concentrated before severe convection weather; using a severe convection weather record obtained by manual observation to extract data 2 hours before the beginning of severe convection weather; s2, constructing a model architecture; the architecture of the model comprises a time sequence feature extraction module, a spatial feature extraction module, a multi-modal feature fusion module and a classification module. The design of each module is to be able to process different types of data and to capture relevant features in both temporal and spatial dimensions. According to the invention, timeliness and accuracy of airport severe convection weather automatic identification are improved; multi-modal data fusion provides more comprehensive information; and identification of various types of severe convection weather is supported.

Description

Technical Field

[0001] The invention provides an early identification method for severe convective weather at an airport based on multi-source heterogeneous meteorological data, belonging to the technical field of meteorological forecasting. Background Art

[0002] In the civil aviation system, the impacts of thunderstorms, lightning, strong winds, and short-term heavy rainfall in severe convective weather on flight vary. Pilots often adopt different response strategies for different types of severe convective weather. Therefore, being able to accurately identify and predict the types of severe convective weather has become an urgent issue to effectively improve aviation production efficiency and ensure flight safety, and it has important guiding significance for flight safety assurance.

[0003] At present, the main means of civil aviation meteorological support rely on weather radar, satellite data and conventional meteorological observation data. Although there are a variety of equipment, the amount of data is also huge, and these data are independent and have the problem of mismatch in time and space scales. Although deep learning and traditional machine learning have made certain progress in meteorological data analysis, from the actual business operation of various large, medium and small airports, no algorithm can achieve the best results in all meteorological scenarios of severe convective weather identification and warning. Therefore, it can be said that the existing technology still has significant deficiencies, and the accuracy, real-time and relevance of severe convective weather identification and warning in airports need to be further improved.

[0004] Most of the existing severe convection extrapolation technologies in the world are based on single-scale research, with simple model levels, single data use or improper fusion leading to information loss or misjudgment. Secondly, the formation and development of thunderstorm clouds are the result of the interaction of multiple meteorological factors. Directly using models such as transformers makes it a major challenge to extract useful features from raw data. Therefore, it is necessary to propose an optimization method that is more in line with the airport's severe convective weather support business, provide a more powerful automatic recognition capability for airport severe convective monitoring and early warning, and be able to effectively and comprehensively apply multi-source heterogeneous meteorological observation data to extract more effective information for severe convective nowcasting and early warning.

[0005] "Research on classification and identification of severe convective weather based on LightGBM algorithm" (Zhang Bingxiang, Li Guocui, Liu Liping, Li Zhe, Wang Congmei, Wang Liping. Hail weather radar identification algorithm based on fuzzy logic [J]. Journal of Applied Meteorology, 2014, 25(04): 415-426.) The LightGBM model constructed by using C-band radar echo products and ground observation data from three regions in Gansu is used to classify and identify three major types of severe convective weather [hail, thunderstorms, and short-term heavy precipitation (short-strong)]. This technology has the following disadvantages:

[0006] 1. The quality of the LightGBM model depends heavily on the number of samples in the training set, and a larger training set is required for modeling;

[0007] 2. The classification of severe convective and non-severe convective weather in this technology depends on ground observation data, and the thunderstorm observation in the data belongs to manual observation. In the manual observation specifications, thunderstorm events, thunderstorm directions and durations must still be recorded when thunder is heard. In fact, lightning can be observed within a range of 40-50km, and thunder can be heard within a range of 20km. This leads to errors in the longitude and latitude of severe convection observed on the ground, and the distribution of ground observation stations is uneven.

[0008] 3. The output of this technology is mainly for identifying the types of severe convective weather, rather than extracting the range of convective weather, which cannot meet the flight needs of civil aviation routes.

[0009] "Severe Convective Weather Identification Algorithm Based on Deep Neural Network" (Wang Xing, Lv Jingjing, Wang Luyao, Wang Hui, Zhan Shaowei. Severe Convective Weather Identification Algorithm Based on Deep Neural Network [J]. Science, Technology and Engineering, 2021, 21(07): 2737-2746.) is an algorithm developed by the National Experimental Teaching Demonstration Center of Atmospheric Science and Environmental Meteorology of Nanjing University of Information Science and Technology. They proposed an intelligent recognition model for severe convective weather based on deep neural network, and constructed a deep neural network model based on CNN. Through iterative training of a large number of sample "data pairs", the radar echo image and the optical flow image representing the echo movement path are used as input. Through the self-learning of the neural network, the function mapping relationship between the radar image and "whether severe convective weather has occurred" is sought; and the label for determining "whether severe convective weather has occurred" is calculated according to certain rules based on the hourly precipitation and maximum wind speed in the ground meteorological observation data. This technical solution uses technologies such as data set enhancement, cost function optimization and model generalization performance optimization to solve the problem of imbalanced training samples and avoid the problem of the model training process falling into local extreme values. This technology has the following disadvantages:

[0010] 1. This technology uses the one-hour cumulative precipitation and one-hour maximum wind speed in ground meteorological observation data to establish a label for whether severe convective weather occurs or not. Although it does not have the error caused by direct use of manual observations like the first technology, there is a large error in defining severe convective weather using only two factors, which leads to a high false alarm rate of this technology;

[0011] 2. The algorithm cannot meet the needs of airport operation support and is designed based on local meteorological data. Summary of the invention

[0012] The present invention aims to solve the problem of "early identification" of severe convective weather at airports. Therefore, a model is provided, which is constructed by deep learning, and the multi-source heterogeneous meteorological data is enhanced to be integrated, so as to identify severe convective weather at an early stage, and the meteorological data from different sources and time scales are effectively integrated through data fusion technology, and a variety of machine learning models are used to accurately identify the initial severe convective weather in the airport, and through model integration optimization, the problems of low warning accuracy and poor timeliness existing in the prior art are solved. This technology focuses on the spatial range of the medium-γ scale, that is, the identification of severe convective events within 20km centered on the airport. The present invention mainly identifies three types of phenomena in severe convective weather: thunderstorms, short-term heavy rainfall and strong winds. These three types of phenomena can occur individually or at the same time, so the problem to be solved belongs to a multi-classification problem.

[0013] The early identification method of severe convective weather at airports based on multi-source heterogeneous meteorological data includes:

[0014] S1, multi-source heterogeneous data description and sample data extraction;

[0015] By using data fusion technology, combining data from different detection equipment and time and space scales, integrating meteorological information from different sources, eliminating noise and time and space deviations between data, and providing a high-quality data foundation for subsequent severe convective weather identification.

[0016] In the early warning stage before an event occurs, the sample data used is mainly concentrated before the occurrence of severe convective weather. Using the severe convective weather records obtained by manual observation, the data 2 hours before the onset of severe convective weather is extracted. The meteorological characteristics during this period can usually reflect the evolution trend of the weather system and possible abnormal changes.

[0017] S2, build model architecture;

[0018] The architecture of the model will include a temporal feature extraction module, a spatial feature extraction module, a multimodal feature fusion module, and a classification module. Each module should be designed to be able to process different types of data and capture relevant features in both the temporal and spatial dimensions.

[0019] S2.1 Data preprocessing

[0020] Airport ground automatic observation data: First, the time series data of the airport meteorological station is standardized to improve the convergence speed of the model.

[0021] Data extraction: Air pressure, temperature, wind direction and speed, and relative humidity are automatically detected. In addition, based on the physical mechanism of the occurrence and development of severe convective weather, the ground 3-hour pressure change ΔP3 and 24-hour pressure change ΔP are calculated and extracted. 24 , 3 hours temperature change ΔT3, 24 hours temperature change ΔT24 And 10-minute wind speed pulsation value u'.

[0022] ΔP3=P(t)-P(t-3)

[0023] ΔP 24 =P(t)-P(t-24)

[0024] ΔT3=T(t)-T(t-3)

[0025] ΔT 24 =T(t)-T(t-24)

[0026]

[0027] Radar data: The radar basic reflectivity map and radial velocity map at 14 different elevation angles (the scanning setting is 0-20° pitch angle, a total of 14 elevation angles) are used. The radar reflectivity map and radial velocity map at these 14 elevation angles are regarded as spatiotemporal data of different scales. Each radar map has been cut into 400*400 pixels centered on the target airport and a range of 50*50km. The radar image at each time step will be used as the input feature, and the data shape is (12,28,400,400).

[0028] Cloud data: The same standardization is performed. The satellite image data of the three infrared bands are selected and merged into a three-channel image for input into the convolutional network. The shape is (12, 200, 200, 3). In this way, the convolution operation will be performed along the space (height and width) and the data of each channel will be processed at the same time. Each pixel will have 3 different values, representing the information of three different infrared bands.

[0029] S2.2. Feature extraction;

[0030] (1) Ground meteorological element text data-time series feature extraction (1-D TCN)

[0031] The temporal convolutional network (TCN) is used to extract temporal features from the time series data of airport ground meteorological elements. The one-dimensional convolution layer effectively captures the dependency of ground meteorological data in the time dimension, and the Flatten operation is used to convert multidimensional data into a one-dimensional feature vector.

[0032] (2) Radar reflectivity and satellite cloud image data - spatiotemporal feature extraction (3-D CNN)

[0033] Radar reflectivity, radial velocity map and satellite cloud image data - spatiotemporal feature extraction (3-D CNN)

[0034] Original multi-elevation radar reflectivity map and radial velocity map: The present invention uses a convolutional neural network (CNN) to extract spatiotemporal features. Radar maps at multiple elevations are regarded as input data at different levels. Each elevation corresponds to a specific spatial layer, and these radar maps represent information at different altitudes. By inputting these radar maps at different elevations into the convolutional neural network, the model can learn how to effectively fuse the spatiotemporal features of each elevation layer and extract the key information that is most conducive to the target weather phenomenon.

[0035] Multi-band infrared cloud images: Use 3-D CNN to extract the temporal and spatial features of infrared cloud images, use standard convolutional layers to extract texture and structural features in the cloud images, and extract cross-band related information between different bands.

[0036] Radar and satellite cloud image two-dimensional optical flow vector map - spatiotemporal feature extraction (3-D CNN): Use the Farneback optical flow method to extract motion features from the basic reflectivity map of 14 elevation angles and the satellite cloud map of 3 infrared bands, capture the motion information of pixels in the radar and satellite cloud maps, and form a two-dimensional optical flow vector map. The optical flow method is used to extract spatial information and capture the motion characteristics in the time dimension at the same time;

[0037] The optical flow features extracted by CNN, the spatial features of radar images / satellite cloud images extracted by CNN, and the temporal features of single-point meteorological elements extracted by TCN are spliced ​​together to obtain all-round features including space, time, and motion information.

[0038] S2.3 Transformer multimodal feature fusion;

[0039] The Transformer model is used as the core tool for feature fusion. Transformer automatically identifies the part that contributes most to the current task by calculating the correlation between each input feature and other features.

[0040] S2.4 Cross-Attention mechanism optimizes the fusion process of multimodal data;

[0041] The Cross-Attention mechanism is introduced on the basis of the Transformer model to optimize the fusion process of multimodal data. By adaptively adjusting the attention weights between different data sources, it promotes information interaction and joint representation learning between modalities.

[0042] S2.5 Classification;

[0043] (1) Multi-classification model

[0044] Use the fully connected layer to map the fused spatiotemporal features to the target category. Train using multi-label classification and use the sigmoid activation function to process independent predictions for each category.

[0045] Loss function: Use the multi-label binary cross entropy loss function (binarycrossentropy), which enables the model to identify each category independently and supports simultaneous identification of multiple categories.

[0046] (2) Output layer

[0047] Each output node corresponds to the probability value of a weather event. When the value is greater than the set threshold, the event is considered to have occurred.

[0048] (3) Threshold adjustment and multi-label evaluation

[0049] A threshold is set based on the recognition probability of each category to determine whether a specific weather event has occurred. An independent probability threshold is set for each category, and this threshold is adjusted according to the specific situation to achieve better classification results.

[0050] S2.6 Model training and evaluation;

[0051] (1) Training

[0052] The Adam optimizer is used to optimize the loss function of the model.

[0053] (2) Evaluation indicators

[0054] The model evaluation uses four indicators to evaluate its performance in multi-label classification tasks, including multi-label accuracy, F1-score, ROC-AUC, and risk assessment (TS).

[0055] The following is the specific indicator evaluation formula:

[0056] A. Multi-label Accuracy

[0057] Multi-label accuracy is a measure of the proportion of correct predictions for each label. It calculates whether each label is predicted correctly and then takes the average.

[0058]

[0059] in:

[0060] N is the number of labels;

[0061] y′ i is the prediction result of the model;

[0062] y i is the actual label;

[0063] 1 is the indicator function, which is 1 if the prediction is correct and 0 otherwise.

[0064] B.F1-Score

[0065] F1-Score is the harmonic mean of precision and recall. F1-Score comprehensively considers the precision and recall of the model.

[0066]

[0067] in:

[0068] P stands for precision: a measure of how many of the predicted positive examples are actually positive examples.

[0069] R stands for recall: a measure of how many of the actual positive examples are predicted to be positive.

[0070] C.ROC-AUC

[0071] ROC-AUC measures the performance of the classifier under all possible classification thresholds. The larger the area under the AUC curve, the better the classification performance.

[0072]

[0073] in:

[0074] TPR: Recall rate, which indicates the ratio of true positive examples to all actual positive examples.

[0075] FPR: It indicates the ratio of false positive examples to all actual negative examples.

[0076] The value of AUC is between 0 and 1. The closer the AUC is to 1, the better the performance of the model.

[0077] D. Risk Assessment TS

[0078] Risk assessment is used to measure the accuracy of model predictions, taking into account the model's true positives, false positives, and false negatives.

[0079]

[0080] in:

[0081] TP is a true example.

[0082] FP is a false positive.

[0083] FN is a false negative example.

[0084] For the fields of aviation safety and weather warning, building this multi-classification early identification model for severe convective weather based on multimodal data and deep learning can bring beneficial effects in many aspects.

[0085] The technical effects of the present invention are as follows:

[0086] 1. Improve the timeliness and accuracy of automatic identification of severe convective weather at airports

[0087] Early warning: Through real-time analysis of multi-source information such as ground meteorological elements, radar data, infrared cloud images, etc., the model can identify the occurrence trend of severe convective weather (such as thunderstorms, strong winds, short-term heavy precipitation, etc.) in advance, and even issue warnings in the early stages of meteorological events. This is crucial to improving the timeliness and accuracy of warnings, especially for aviation, transportation and people's livelihood security.

[0088] Avoid danger in advance: For airport operations, early warning of different types of severe convective weather can provide a time window for airport dispatchers and pilots, thereby avoiding or reducing safety risks caused by sudden weather. Airports usually rely on multiple data sources such as ground meteorological stations, radars, and satellite cloud images to monitor the weather. Fusion of these data sources through deep learning models can help to more accurately judge weather conditions, especially for early identification of complex severe convective weather events, helping airlines or air traffic control departments to adjust flight plans and airport operation strategies in a timely manner, thereby reducing accidents and delays caused by weather.

[0089] 2. Multimodal data fusion provides more comprehensive information

[0090] Information complementarity: Different types of data such as ground meteorological elements, radar basic reflectivity, radial velocity maps, and infrared cloud maps detect relevant information about the atmospheric environment from different directions, and each detection has its own limitations. For example, satellite cloud maps detect from top to bottom and can only obtain information from the top layer, while radar detects from bottom to top and has blind spots and can only obtain information within a certain height and distance range. Through multimodal data fusion, the model can comprehensively utilize these complementary information to obtain more comprehensive and accurate weather identification results, thereby making up for the shortcomings of a single data source.

[0091] 3. Support multiple categories of severe convective weather identification

[0092] Multi-label classification capability: The model can simultaneously identify multiple types of severe convective weather (such as thunderstorms, short-term heavy precipitation, high winds, etc.), rather than predicting a single weather event. This multi-label classification capability makes the model more flexible and comprehensive, and can make more accurate decisions in complex weather environments. For example, a severe thunderstorm may be accompanied by multiple weather phenomena such as high winds and heavy precipitation. The model can capture these different weather characteristics at the same time, thereby providing decision makers with more comprehensive information. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 This is a diagram of the architecture of the airport severe convective weather early identification model of the present invention. DETAILED DESCRIPTION

[0094] The specific technical solution of the present invention is explained in conjunction with embodiments.

[0095] The early identification method of severe convective weather at airports based on multi-source heterogeneous meteorological data includes:

[0096] S1. Description of multi-source heterogeneous data and sample data extraction; as shown in Table 1.

[0097] Table 1 Multi-source heterogeneous data table

[0098]

[0099] By utilizing data fusion technology and combining data from different detection equipment and time and space scales, this system can efficiently integrate meteorological information from different sources, eliminate noise and time and space deviations between data, and provide a high-quality data foundation for subsequent severe convective weather identification.

[0100] Sample extraction: The focus of the present invention is on the "early identification" of severe convective weather, so it is mainly the early warning stage before the event occurs, and the sample data used is mainly concentrated before the severe convective weather occurs. Use the severe convective weather records obtained by artificial observation to extract the data 2 hours before the start of severe convective weather. The meteorological characteristics during this period can usually reflect the evolution trend and possible abnormal changes of the weather system. Generally speaking, the early warning signals of meteorological patterns (such as temperature changes, humidity changes, wind speed, etc.) will fluctuate significantly in the hours before the arrival of severe convective weather. Therefore, a total of 2 hours of sample data is extracted for each severe convective weather process (the advance amount can be appropriately increased).

[0101] S2, build model architecture;

[0102] The goal of this invention is to extract spatiotemporal features from multi-source, heterogeneous data and perform multi-label weather classification warning recognition. The main architecture of the model will include a temporal feature extraction module, a spatial feature extraction module, a multimodal feature fusion module, and a classification module. Each module should be designed to be able to process different types of data and capture relevant features in both the temporal and spatial dimensions, such as Figure 1 .

[0103] S2.1 Data preprocessing

[0104] Airport ground automatic observation data (one-dimensional time series): First, the time series data of the airport meteorological station is standardized to improve the convergence speed of the model.

[0105] Data extraction: Air pressure, temperature, wind direction and speed, and relative humidity are automatically detected. In addition, based on the physical mechanism of the occurrence and development of severe convective weather, the ground 3-hour pressure change ΔP3 and 24-hour pressure change ΔP are calculated and extracted. 24 , 3 hours temperature change ΔT3, 24 hours temperature change ΔT 24 And 10-minute wind speed pulsation value u'.

[0106] ΔP3=P(t)-P(t-3)

[0107] ΔP 24 =P(t)-P(t-24)

[0108] ΔT3=T(t)-T(t-3)

[0109] ΔT 24 =T(t)-T(t-24)

[0110]

[0111] Radar data: Many previous studies or inventions often use combined reflectivity or CAPPI images or data of a certain altitude layer to build deep learning models for identification and prediction of severe convective weather. The present invention uses radar basic reflectivity maps and radial velocity maps at 14 different elevation angles (scanning setting 0-20° pitch angle for a total of 14 elevation angles). The radar reflectivity maps and radial velocity maps at these 14 elevation angles are regarded as spatiotemporal data of different scales. Each radar map has been cut into 400*400 pixels centered on the target airport and a range of 50*50km. The radar image (reflectivity, radial velocity) at each time step will be used as input features, and the data shape is (12,28,400,400).

[0112] Cloud image data: Standardization is also performed. In the present invention, the satellite cloud image data has been cut into 200*200 pixels centered on the target airport and a range of 100*100km. The present invention selects satellite image data of three infrared bands and merges them into a three-channel image for input into the convolutional network with a shape of (12,200,200,3). In this way, the convolution operation will be performed along the space (height and width) while processing the data of each channel. Each pixel will have 3 different values, representing three different infrared band information.

[0113] S2.2. Feature extraction;

[0114] (1) Ground meteorological element text data-time series feature extraction (1-D TCN);

[0115] The method of the present invention uses a temporal convolutional network (TCN) to extract time series features from the time series data of airport ground meteorological elements. The dependency of ground meteorological data in the time dimension is effectively captured through a one-dimensional convolutional layer (1-D Convolutional Layer), and the multidimensional data is converted into a one-dimensional feature vector using a Flatten operation. TCN has significant advantages, especially when dealing with long-term dependencies. With its parallel processing capabilities and the characteristics of extended convolution, it can efficiently model complex time series data. In this module, the selection of hyperparameters such as convolution kernel size and expansion rate has an important influence on the extraction effect of time series features.

[0116] (2) Radar reflectivity and satellite cloud image data - spatiotemporal feature extraction (3-D CNN);

[0117] Radar reflectivity, radial velocity map and satellite cloud image data - spatiotemporal feature extraction (3-D CNN):

[0118] Original multi-elevation radar reflectivity map and radial velocity map: The present invention uses convolutional neural network (CNN) to extract spatiotemporal features, because a three-dimensional convolutional neural network (3D CNN) is used here, which can extract complex spatiotemporal features in both time and space dimensions. At each time step, the radar map (including reflectivity and radial velocity) is used as input data and processed through a 3D convolution layer to capture the local spatial features in the data and its dynamic characteristics over time.

[0119] Specifically, radar images at multiple elevation angles are regarded as input data at different levels. Each elevation angle corresponds to a specific spatial layer, and these radar images represent information at different altitudes. By inputting these radar images at different elevation angles into the convolutional neural network, the model can learn how to effectively integrate the spatiotemporal features of each elevation layer and extract the key information that is most conducive to the target weather phenomenon (such as severe convective weather, strong winds, etc.).

[0120] In this process, the 3D convolution operation can simultaneously perform convolution along the spatial dimension (such as the width and height of the image) and the temporal dimension (i.e., the changes at different time steps), thereby extracting richer spatiotemporal features. Compared with traditional two-dimensional convolution or single time step processing methods, this method has stronger spatiotemporal modeling capabilities and can better capture the complex dynamic patterns in radar data.

[0121] In addition, with the input of multiple elevation angle data, the model can capture meteorological information at different altitudes, which is particularly important for analyzing meteorological phenomena such as wind speed changes and precipitation distribution at different altitudes in the atmosphere. Through the deep learning capabilities of the convolutional layer, the model will automatically identify which elevation angle features have a high relevance when predicting the target weather phenomenon, and learn how to effectively integrate these multi-level information to improve prediction accuracy.

[0122] Multi-band infrared cloud images: Use 3-D CNN to extract the temporal and spatial features of infrared cloud images, use standard convolutional layers to extract texture and structural features in cloud images (such as cloud top temperature, cloud height, etc.), and extract cross-band related information between different bands.

[0123] Radar and satellite cloud image 2D optical flow vector map - spatiotemporal feature extraction (3-D CNN):

[0124] Although 3D CNN has extracted spatiotemporal features from radar images and satellite cloud images, considering that it extracts global spatiotemporal features, it learns the local features of radar images and satellite cloud images in the time and space dimensions, but it focuses more on capturing the spatial distribution and change trends of these features. Therefore, on this basis, the present invention uses the Farneback optical flow method to extract motion features from the basic reflectivity images of 14 elevation angles and the satellite cloud images of 3 infrared bands, captures the motion information of pixels in the radar images and satellite cloud images, and forms a two-dimensional optical flow vector map. The optical flow method is specifically used to extract the motion vector of objects or weather systems in images, that is, the displacement of the target in the image. This motion information is crucial for the analysis of strong convective systems at the medium-γ scale. In this method, the optical flow method not only extracts spatial information, but also captures the motion features in the time dimension. This information cannot be directly provided by the 3D CNN spatiotemporal feature extraction.

[0125] The reason why the Farneback method is selected among many optical flow methods is that the specific problem to be solved by the present invention is taken into consideration. The Farneback method is based on local polynomial fitting of the image, and it estimates the optical flow by constructing a polynomial model of the image. The Farneback method adopts a multi-scale (or pyramid) strategy to process optical flow fields at different scales, which can better handle large-scale motion (such as fast motion) and small-scale details in the image. It is suitable for processing large-scale and small-scale motions, and at the same time has good robustness for non-smooth optical flow fields and can handle fast motion. Therefore, it is more suitable for the problem to be solved by the present invention.

[0126] The motion vector extracted by the optical flow method is often a relatively "low-level" information, representing simple pixel-level motion. However, CNN can learn more abstract features through convolutional layers, such as the overall shape of cloud movement and the direction of storm progress. Through multiple layers of convolution, CNN can gradually transform local information in the motion vector into high-level spatiotemporal patterns. Use 3-D convolutional neural network (CNN) to extract features from the two-dimensional motion vector extracted by optical flow. CNN can extract local spatial patterns, motion trajectories and other information in multiple convolutional layers and pooling layers. Here we mainly focus on extracting motion-related spatial features.

[0127] Although the motion information extracted by the optical flow method may have some overlap with the spatiotemporal changes extracted by 3-D CNN, the optical flow method can supplement the detailed dynamic information that CNN+TCN cannot directly obtain by capturing specific motion patterns (such as the motion trajectory of the storm). CNN+TCN is more about high-level feature extraction of images and time series, while the optical flow method directly extracts the specific direction and speed of movement from the image sequence, which may help improve the prediction accuracy of radar maps or cloud maps before severe convection occurs.

[0128] The optical flow features extracted by CNN, the spatial features of radar images / satellite cloud images extracted by CNN, and the temporal features of single-point meteorological elements extracted by TCN are spliced ​​together. In this way, all-round features containing space, time, and motion information can be obtained.

[0129] S2.3 Transformer multimodal feature fusion;

[0130] After extracting various features, the next step is feature fusion. Different data sources (meteorological elements, radar maps, cloud maps) have different time scales and spatial structures, so feature fusion needs to consider the information integration between different modalities. How to effectively fuse them is also the key to solving the problem.

[0131] In order to solve this problem, the present invention adopts the Transformer model as the core tool for feature fusion. The Transformer architecture, especially its self-attention mechanism (Self-Attention), shows great advantages in processing multi-source heterogeneous data. The self-attention mechanism can dynamically assign different attention weights to different data sources according to the importance of each feature in the data, thereby effectively extracting the key features in each modality. Specifically, the Transformer automatically identifies the part that contributes most to the current task by calculating the correlation (ie "attention") between each input feature and other features. For example, the local spatial features in radar images may be highly correlated with the time series features in meteorological elements. The Transformer can automatically weight this information through the self-attention mechanism, thereby achieving more accurate information fusion.

[0132] Transformer's multi-head attention mechanism further enhances the model's expressiveness. Through the parallel calculation of multiple attention heads, the model can focus on different data features in different subspaces, which is especially important for processing complex meteorological data. Ultimately, this adaptive feature fusion enables the key information of each modality to be fully integrated in the same space, thereby providing a more comprehensive and accurate feature representation for the subsequent early detection and analysis of severe convective weather.

[0133] S2.4 Cross-Attention mechanism optimizes the fusion process of multimodal data;

[0134] In order to further improve the effect of multi-source data fusion, this paper introduces the cross-attention mechanism based on the Transformer model, aiming to optimize the fusion process of multimodal data. The cross-attention mechanism is particularly suitable for processing heterogeneous data sources, and can effectively promote information interaction and joint representation learning between modalities by adaptively adjusting the attention weights between different data sources.

[0135] In the traditional self-attention mechanism, the calculation of attention weights usually only depends on the information within a single modality, while the cross-attention mechanism further enhances the model's spatiotemporal perception by modeling the relationship between different modalities. Specifically, cross-attention can help the model identify the spatiotemporal dependencies between the time series characteristics of meteorological elements and the spatial characteristics of radar images. For example, changes in meteorological elements may affect spatial features such as precipitation intensity and wind speed in radar images, and vice versa. The cross-attention mechanism dynamically adjusts the weights between the modalities, allowing the model to capture this mutual influence, thereby improving the understanding and prediction of complex meteorological phenomena.

[0136] The working principle of cross-attention usually includes the operations of query, key, and value. In the process of fusing meteorological elements and radar images, the features of meteorological elements are used as query vectors, and the features of radar images are used as keys and values. By calculating the similarity between the query and the key, the cross-attention mechanism can dynamically adjust the degree of attention to the value according to its similarity, thereby extracting the cross-modal information most relevant to the current task. This mechanism can effectively handle the heterogeneity between multiple modalities and ensure that the key information of different types of data is reasonably weighted during the fusion process.

[0137] In addition, the cross-attention mechanism can also achieve more complex inter-modal dependency modeling through multi-level attention calculations. In this way, the model can not only focus on the features in a single modality, but can also cross the boundaries between multiple modalities for deep learning, thereby improving the prediction accuracy in complex severe convective weather phenomena. Therefore, the present invention first uses Transformer to perform preliminary spatiotemporal feature fusion, which can reduce the computational complexity of the subsequent cross-attention mechanism, making the model more focused when learning the relationship between modalities.

[0138] S2.5 Classification;

[0139] (1) Multi-classification model

[0140] The fused spatiotemporal features are mapped to the target category using a fully connected layer. Since severe convective weather events may occur simultaneously, the model can be trained using a multi-label classification method, using a sigmoid activation function to process independent predictions for each category instead of the traditional softmax.

[0141] Loss function: The present invention uses a multi-label binary cross entropy loss function (binary cross entropy), which enables the model to independently identify each category (thunderstorm, short-term heavy rainfall, strong wind, etc.) and supports simultaneous identification of multiple categories.

[0142] (2) Output layer

[0143] Each output node corresponds to the probability value of a weather event. When the value is greater than the set threshold, the event is considered to have occurred.

[0144] (3) Threshold adjustment and multi-label evaluation

[0145] In the present invention, a threshold is set according to the recognition probability of each category in order to determine whether a specific weather event occurs. The present invention sets an independent probability threshold for each category, and this threshold can also be adjusted according to specific circumstances to obtain a better classification effect.

[0146] S2.6 Model training and evaluation;

[0147] (1) Training

[0148] The Adam optimizer is used in the present invention to optimize the loss function of the model.

[0149] (2) Evaluation indicators

[0150] In the present invention, the above model evaluation finally uses four indicators to evaluate its performance in the multi-label classification task, including multi-label accuracy, F1-score, ROC-AUC and risk assessment (TS).

[0151] The model uses different data sources and feature extraction modules to effectively extract spatiotemporal features from time series data, radar images, and infrared cloud images, and fuses multimodal features through a cross-attention mechanism. Ultimately, through multi-label classification, the model is able to predict different types of severe convective weather (such as thunderstorms, short-term heavy precipitation, and strong winds) and can handle situations where they occur simultaneously.

[0152] The following is the specific indicator evaluation formula:

[0153] E. Multi-label Accuracy

[0154] Multi-label accuracy (also called multi-label precision) is a measure of the proportion of correct predictions for each label. It calculates whether each label is correctly predicted (that is, whether the model correctly classifies each label) and then takes the average.

[0155]

[0156] in:

[0157] N is the number of tags (three tags in the present invention: thunderstorm, short-term heavy rainfall, and strong wind).

[0158] y′ i It is the prediction result of the model (for example, predicting whether there will be thunderstorms, short-term heavy rainfall or strong winds at a certain moment).

[0159] y i are actual labels (e.g., actual occurrences of thunderstorms, short-term heavy precipitation, and high winds).

[0160] 1 is the indicator function, which is 1 if the prediction is correct and 0 otherwise.

[0161] 2-F1-Score

[0162] F1-Score is the harmonic mean of precision and recall, and is often used to evaluate unbalanced datasets. F1-Score comprehensively considers the precision and recall of the model.

[0163]

[0164] in:

[0165] P stands for precision: a measure of how many of the predicted positive examples (for example, predicting a thunderstorm) are true positive examples.

[0166] R stands for recall: it measures how many of the actual positive examples are predicted to be positive.

[0167] 3-ROC-AUC

[0168] ROC-AUC (Receiver Operating Characteristic-Area Under Curve) is an indicator for comprehensively evaluating the performance of a classifier. It measures the performance of the classifier under all possible classification thresholds. The larger the AUC (area under the curve) value, the better the classification performance.

[0169]

[0170] in:

[0171] TPR (True Positive Rate): also known as recall rate, which indicates the ratio of true positive examples to all actual positive examples.

[0172] FPR (False Positive Rate): indicates the ratio of false positive examples to all actual negative examples.

[0173] The value of AUC is between 0 and 1. The closer the AUC is to 1, the better the performance of the model.

[0174] 4-Threat Score (TS)

[0175] Risk assessment (also called threat score) is used to measure the accuracy of model predictions, especially when predicting rare events. It takes into account the model's true positives, false positives, and false negatives.

[0176]

[0177] in:

[0178] TP is a true example.

[0179] FP is a false positive.

[0180] FN is a false negative example.

[0181] At present, most of the research is focused on the identification or extrapolation of severe convective weather, while the research in the field of early identification of severe convective weather itself is relatively small and mainly focuses on large-scale or regional areas. Secondly, although the existing technology can achieve the prediction of weather events to a certain extent, it has problems such as low accuracy, high underreporting and false alarms, and cannot effectively solve the current problems of civil aviation meteorological support. Traditional meteorological data models mostly rely on a single data source, lack comprehensiveness, and are difficult to provide sufficiently high accuracy when facing complex and changeable severe convective weather. Existing machine learning methods often have underreporting or false alarms when dealing with these weather phenomena, especially under extreme weather conditions.

[0182] To address this problem, this technical solution uses multimodal data fusion, combined with ground meteorological data, radar reflectivity, infrared cloud images and other meteorological data, to give full play to the advantages of each data source, and uses deep learning methods (3D CNN) combined with transformer and cross-attention mechanism to strengthen the fusion of spatiotemporal features of multi-source heterogeneous meteorological data, effectively capture the early signals of severe convective weather, significantly improve the accuracy of prediction, and reduce missed reports and false reports. Especially in the sub-field of early identification of severe convective weather, the innovative technology of this solution can make more accurate warnings at the early stage of meteorological events, greatly improving timeliness.

[0183] Although there are some similar technologies in the field of severe convective weather identification, targeted customized research and development is still needed to achieve higher accuracy and solve the problems of false reporting and missed reporting. Existing general technologies alone are difficult to effectively respond to complex meteorological changes and early prediction needs. Therefore, this solution can achieve relatively accurate early identification of severe convective weather by comprehensively using multimodal data and deep learning technology, filling the gap in existing technologies and having high practical application value.

[0184] This customized research and development technology, which is specifically designed for the early identification of severe convective weather, has great application potential, especially in improving the accuracy of meteorological warnings, ensuring aviation safety, and improving disaster response efficiency.

Claims

1. An early identification method for severe convective weather at an airport based on multi-source heterogeneous meteorological data, characterized in that: include: S1, multi-source heterogeneous data description and sample data extraction; By using data fusion technology, combining data from different detection equipment and time and space scales, integrating meteorological information from different sources, eliminating noise and time and space deviations between data, and providing a high-quality data foundation for subsequent severe convective weather identification; In the early warning stage before an event occurs, the sample data used is mainly concentrated before severe convective weather occurs; Using the severe convective weather records obtained by manual observation, we extracted the data 2 hours before the onset of severe convective weather. The meteorological characteristics during this period can usually reflect the evolution trend of the weather system and possible abnormal changes. S2, build model architecture; The model architecture will include a temporal feature extraction module, a spatial feature extraction module, a multimodal feature fusion module, and a classification module; each module should be designed to be able to process different types of data and capture relevant features in both the temporal and spatial dimensions; It includes the following sub-steps: S2.1 Data Preprocessing S2.2, feature extraction; (1) Ground meteorological element text data - time series feature extraction; (2) Radar reflectivity and satellite cloud image data - temporal and spatial feature extraction; S2.3, Transformer multimodal feature fusion; The Transformer model is used as the core tool for feature fusion. The Transformer automatically identifies the part that contributes most to the current task by calculating the correlation between each input feature and other features. S2.4, Cross-Attention mechanism optimizes the fusion process of multimodal data; The Cross-Attention mechanism is introduced on the basis of the Transformer model to optimize the fusion process of multimodal data. By adaptively adjusting the attention weights between different data sources, it promotes information interaction and joint representation learning between modalities. S2.5, classification; (1) Multi-classification model Use a fully connected layer to map the fused spatiotemporal features to the target category; train using a multi-label classification method and use a sigmoid activation function to process independent predictions for each category; Loss function: Use multi-label binary cross entropy loss function, which enables the model to identify each category independently and supports simultaneous identification of multiple categories; (2) Output layer Each output node corresponds to the probability value of a weather event. When the value is greater than the set threshold, the event is considered to have occurred. (3) Threshold adjustment and multi-label evaluation A threshold is set based on the recognition probability of each category in order to determine whether a specific weather event has occurred; an independent probability threshold is set for each category, and this threshold is adjusted according to the specific situation to achieve the classification effect; S2.6 Model training and evaluation; (1) Training The Adam optimizer is used to optimize the loss function of the model; (2) Evaluation indicators The model evaluation uses four indicators to evaluate its performance in the multi-label classification task, including multi-label accuracy, F1-score, ROC-AUC, and risk assessment.

2. The method for early identification of severe convective weather at an airport based on multi-source heterogeneous meteorological data according to claim 1 is characterized in that: S2.1 Data preprocessing, including: Airport ground automatic observation data: First, the time series data of the airport weather station is standardized to improve the convergence speed of the model; Data extraction: Air pressure, temperature, wind direction and speed, and relative humidity are automatically detected and obtained; in addition, based on the physical mechanism of the occurrence and development of severe convective weather, the ground 3-hour pressure change ΔP3 and 24-hour pressure change ΔP are calculated and extracted. 24 , 3 hours temperature change ΔT3, 24 hours temperature change ΔT 24 and 10-minute wind speed pulsation value u'; ΔP3=P(t)-P(t-3) ΔP 24 =P(t)-P(t-24) ΔT3=T(t)-T(t-3) ΔT 24 =T(t)-T(t-24) Radar data: The radar basic reflectivity map and radial velocity map at 14 different elevation angles (the scan setting is 0-20° pitch angle, a total of 14 elevation angles) are used. The radar reflectivity map and radial velocity map at these 14 elevation angles are regarded as spatiotemporal data of different scales. Each radar map has been cut into 400*400 pixels centered on the target airport and a range of 50*50km. The radar image at each time step will be used as the input feature, and the data shape is (12,28,400,400); Cloud map data: The same standardization process is performed; satellite image data of three infrared bands are selected and merged into a three-channel image for input into the convolutional network with a shape of (12, 200, 200, 3). In this way, the convolution operation will be performed along the space and the data of each channel will be processed simultaneously; each pixel will have 3 different values, representing the information of three different infrared bands.

3. The method for early identification of severe convective weather at an airport based on multi-source heterogeneous meteorological data according to claim 1 is characterized in that: S2.2 specifically includes: (1) Ground meteorological element text data - time series feature extraction; The temporal convolutional network (TCN) is used to extract temporal features from the time series data of airport ground meteorological elements. The dependency of ground meteorological data in the time dimension is effectively captured through a one-dimensional convolutional layer, and the multi-dimensional data is converted into a one-dimensional feature vector using the Flatten operation. (2) Radar reflectivity and satellite cloud image data - temporal and spatial feature extraction; Radar reflectivity, radial velocity map and satellite cloud image data - temporal and spatial feature extraction; Original multi-elevation radar reflectivity map and radial velocity map: Convolutional neural network (CNN) is used to extract spatiotemporal features. Radar maps at multiple elevation angles are regarded as input data at different levels. Each elevation angle corresponds to a specific spatial layer. These radar maps represent information at different altitude layers. By inputting these radar maps at different elevation angles into the convolutional neural network, the model can learn how to effectively fuse the spatiotemporal features of each elevation layer and extract the key information that is most conducive to the target weather phenomenon. Multi-band infrared cloud images: Use 3-D CNN to extract the temporal and spatial features of infrared cloud images, use standard convolutional layers to extract texture and structural features in the cloud images, and extract cross-band related information between different bands; Two-dimensional optical flow vector map of radar and satellite cloud images - extraction of spatiotemporal features: Use the Farneback optical flow method to extract motion features from the basic reflectivity maps of 14 elevation angles and satellite cloud images of 3 infrared bands, capture the motion information of pixels in radar and satellite cloud images, and form a two-dimensional optical flow vector map; use the optical flow method to extract spatial information and capture the motion features in the time dimension at the same time; The optical flow features extracted by CNN, the spatial features of radar images / satellite cloud images extracted by CNN, and the temporal features of single-point meteorological elements extracted by TCN are spliced ​​to obtain all-round features including space, time, and motion information.

4. The method for early identification of severe convective weather at an airport based on multi-source heterogeneous meteorological data according to claim 1 is characterized in that: Evaluation index formula in S2.6: A. Multi-label Accuracy MLA Multi-label accuracy is a metric used to evaluate the performance of multi-label classification models; it calculates whether each label is predicted correctly and then takes the average; in: N is the number of labels; y′ i is the prediction result of the model; y i is the actual label; 1 is the indicator function, which is 1 if the prediction is correct and 0 otherwise. B.F1-Score F1-Score is the harmonic mean of precision and recall. F1-Score comprehensively considers the precision and recall of the model. in: P stands for precision: it measures how many of the predicted positive examples are actually positive examples; R stands for recall: it measures how many of the actual positive examples are predicted to be positive; The value of F1-Score is between 0 and 1, where 1 indicates perfect precision and recall and 0 indicates the worst performance. C.ROC-AUC ROC-AUC measures the performance of the classifier under all possible classification thresholds; the larger the area under the AUC curve, the better the classification performance; AUC=∫0 1 TPR(FPR)dFPR in: TPR: Recall rate, which indicates the ratio of true positive examples to all actual positive examples; FPR: indicates the ratio of false positive examples to all actual negative examples; The value of AUC is between 0 and 1. The closer the AUC is to 1, the better the performance of the model. D. Risk Assessment TS Risk assessment is used to measure the accuracy of the model’s predictions, taking into account the model’s true positives, false positives, and false negatives; in: TP is a true example; FP is a false positive; FN is a false negative example.

Images