Extreme gale identification method based on time-space contrast learning and physical constraint fusion

Abstract

The invention discloses an extreme gale identification method based on time-space comparative learning and physical constraint fusion, and relates to the technical field of atmospheric science. The method comprises the following steps: preprocessing minute-level meteorological monitoring data to obtain preprocessed data; performing screening based on the preprocessed data to obtain sample data in space-time contrast learning; the sample data comprises an anchor sample, a positive sample and a negative sample; performing data enhancement on the anchor sample, inputting the enhanced anchor sample, the positive sample and the negative sample into a double-branch feature extraction network, and performing training by using a contrast loss function; and on the basis of the trained double-branch feature extraction network, an extreme gale classification head is introduced, optimization is performed by using a cross entropy loss function, and an extreme gale recognition model is obtained and recognized. According to the method, the defects of a traditional method in noise robustness and small sample generalization ability can be overcome, and therefore the precision and stability of extreme gale recognition are improved.

Description

Technical Field

[0001] This invention relates to the field of atmospheric science and technology, and in particular to an extreme wind identification method based on the fusion of spatiotemporal comparative learning and physical constraints. Background Technology

[0002] In the fields of atmospheric science and meteorological forecasting, extreme wind identification, as a meteorological task with significant environmental and social impact, has received widespread attention. However, due to the diversity and complexity of wind speed variation patterns, existing technologies struggle to perform in-depth feature mining and effective correlation of meteorological data, exhibiting significant limitations.

[0003] Current research on extreme wind identification often suffers from limitations in its ability to represent complex patterns or its inadequate capture of implicit relationships between data. There is a lack of efficient methods for modeling the complex nonlinear relationships between multidimensional meteorological elements and wind speed. Existing numerical weather prediction models and traditional machine learning methods, when capturing key features influencing wind speed and their interactions, often fail to effectively mine deep similarities and differences in the data, thus limiting a comprehensive understanding and prediction of wind speed variation patterns. These shortcomings make it difficult to accurately identify extreme winds under complex and variable meteorological conditions.

[0004] Furthermore, existing technologies lack the ability to effectively utilize unlabeled data and discriminate features when dealing with the diversity of spatiotemporal wind speed variation patterns. They also fail to fully leverage the inherent structural information of large amounts of historical meteorological data, thus affecting the accuracy and reliability of prediction results. Therefore, developing a new method capable of deeply mining the characteristics of meteorological data and accurately predicting wind speed has become an urgent task in current research. Summary of the Invention

[0005] The purpose of this invention is to provide an extreme wind identification method based on the fusion of spatiotemporal contrastive learning and physical constraints, aiming to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following solution: An extreme wind identification method based on the fusion of spatiotemporal contrastive learning and physical constraints includes: Acquire minute-level meteorological monitoring data and preprocess the meteorological monitoring data to remove data noise, thereby obtaining preprocessed data; Based on the preprocessed data, sample data for spatiotemporal contrastive learning is obtained through filtering. The sample data includes anchor samples, positive samples, and negative samples. The positive samples are selected using a dual constraint of spatiotemporal proximity and physical consistency. The anchor samples are augmented with data, and the augmented anchor samples, the positive samples, and the negative samples are input into a dual-branch feature extraction network for training using a contrastive loss function; the data augmentation includes temporal surround augmentation and amplification synergistic augmentation. Based on the trained dual-branch feature extraction network, an extreme wind classification head is introduced and optimized using the cross-entropy loss function to obtain an extreme wind recognition model. The extreme wind recognition model is then used to identify the test data.

[0007] Optionally, the meteorological monitoring data includes at least a 10-meter wind speed component, sea level pressure, and a 2-meter temperature component.

[0008] Optionally, the specific process of the preprocessing includes: First, data was filtered according to the geographical scope of the target area and the peak periods of extreme winds. At the same time, the timestamps of all data were unified to UTC+8 time zone to ensure consistency of time base. Then, outliers were repaired according to meteorological and physical laws, and linear interpolation was used to complete missing data to ensure data continuity. Next, core features directly related to extreme winds were selected, and wind direction was converted into a direction cosine vector to avoid the problem of angular periodicity. All core features were standardized to eliminate the influence of dimensional differences. Finally, the processed data were uniformly organized into a CSV format of "site ID-timestamp-feature vector".

[0009] Optionally, the step of filtering based on the preprocessed data to obtain sample data for spatiotemporal contrastive learning specifically includes: Using a 15-minute sliding time window and a 5-minute sliding step, 15-step sequences are slidably extracted from the single-site data in the preprocessed data as anchor samples. The anchor samples include multiple training batches of data, each training batch consisting of samples from the same site, season, and sequence length, ensuring that the samples within a batch have consistent spatiotemporal background and physical standards. In the time dimension, samples with an overlap of ≥60% with the time window of the anchor sample are selected from the entire dataset to ensure that both are affected by the same extreme weather system at the same time. In the spatial dimension, station data with a straight-line distance of ≤5km from the anchor sample station are selected to avoid the physical process being disconnected due to excessive distance, thus obtaining positive samples. First, all positive samples are removed. Then, feature matrices are constructed for the remaining samples and the anchor sample respectively. The feature difference between the two is quantified using matrix cosine similarity. The 30% of samples with the lowest similarity are selected as negative samples. The feature matrix of the anchor sample is a 15×5 matrix. The 15 rows correspond to a 15-minute sequence, and the 5 columns correspond to the core features of smoothed wind speed, air pressure change rate, wind direction cosine vector, 2-meter temperature, and sea level air pressure.

[0010] Optionally, the constraint of physical consistency is verified by dual indicators: wind speed temporal trend and pressure system synergy. First, calculate the smoothed wind speed sequence over 15 minutes between the candidate sample and the anchor sample. Pearson correlation coefficient The calculation method is as follows: in, For anchor sample number Smooth wind speed for minutes, For candidate samples Smooth wind speed for minutes, The average wind speed over 15 minutes for the anchor sample. The candidate sample's 15-minute average wind speed, if This indicates that the wind speed changes of both devices are highly consistent across minutes, achieving synchronous growth and reaching peak values ​​simultaneously. Then, the average air pressure change rate of both devices over 15 minutes is calculated simultaneously. relative deviation The calculation method is as follows: in, The average rate of change of air pressure over 15 minutes for the anchor sample. The average rate of change of air pressure over 15 minutes for the candidate sample; if the candidate sample meets the following conditions... and If it meets the physical consistency requirement, it is determined to be a positive sample. .

[0011] Optionally, the formula for calculating the matrix cosine similarity is: in, This represents taking the trace of the result after multiplying the transpose of the anchor sample matrix with the candidate sample matrix. The F-norm of the anchor sample feature matrix is ​​represented by the F-norm. This represents the F-norm of the feature matrix of the candidate sample.

[0012] Optionally, the contrastive loss function is expressed as: in, This is a function for calculating vector similarity. For temperature hyperparameters, z i The enhanced anchor sample is the feature vector after processing by the encoder and projector head. z pos The feature vector of a positive sample after processing by the encoder and projector head. zneg The feature vector is the result of processing negative samples by the encoder and projector. p denoted as , where n is the number of positive samples and n is the number of negative samples.

[0013] Optionally, the structure of the extreme wind classification head consists of two stacked fully connected layers and a Softmax layer: the first fully connected layer has an input dimension of 256 and an output dimension of 128, uses the ReLU activation function, and is subject to batch normalization; the second fully connected layer has an input dimension of 128 and an output dimension of 2.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses an extreme wind identification method based on the fusion of spatiotemporal contrastive learning and physical constraints. The method includes preprocessing minute-level meteorological monitoring data to obtain preprocessed data; filtering the preprocessed data to obtain sample data for spatiotemporal contrastive learning; the sample data includes anchor samples, positive samples, and negative samples; data augmentation of the anchor samples; and inputting the augmented anchor samples, the positive samples, and the negative samples into a bi-branch feature extraction network for training using a contrastive loss function; based on the trained bi-branch feature extraction network, an extreme wind classification head is introduced and optimized using a cross-entropy loss function to obtain an extreme wind identification model, which is then used for identification. This invention addresses the shortcomings of traditional methods in terms of noise robustness and small-sample generalization ability, thereby improving the accuracy and stability of extreme wind identification. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the extreme wind identification method in this embodiment; Figure 2 This is a schematic diagram of the extreme wind identification model structure in this embodiment. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The purpose of this invention is to provide an extreme wind identification method based on the fusion of spatiotemporal contrastive learning and physical constraints, aiming to solve or improve at least one of the above-mentioned technical problems.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figures 1-2 As shown, this invention provides an extreme wind identification method based on the fusion of spatiotemporal contrastive learning and physical constraints. By constructing a temporal contrastive learning framework, STC-WindNet, adapted to the characteristics of extreme winds, and designing a positive-negative sample construction strategy and a temporal encoder oriented towards the spatiotemporal evolution of wind speed data, accurate identification of extreme winds is achieved. This method first analyzes the periodicity, non-stationarity, and spatiotemporal dependence characteristics of wind speed data, and constructs meteorologically significant positive-negative sample pairs by combining meteorological and physical laws. Then, a temporal encoder is used to extract spatiotemporal pattern features related to extreme winds. The feature space is optimized through a contrastive loss function, causing similar wind speed patterns to cluster and dissimilar patterns to separate. Finally, the pre-trained encoder parameters are frozen to preserve the learned meteorological and physical consistency, and only the extreme wind classification head parameters are fine-tuned, ultimately achieving accurate identification of extreme winds. Specific steps include: Step 1: Data Collection and Preprocessing: This approach uses a minute-level meteorological monitoring dataset as input. This dataset contains core physical quantities such as 10-meter wind speed u10 / v10 components, sea level pressure msl, and 2-meter temperature t2m, covering observational data for a certain period in the target area. Preprocessing is carried out through a four-stage core process: First, data is filtered according to the geographical range of the target area and the peak periods of extreme winds, while unifying the timestamps of all data to UTC+8 time zone to ensure consistent time references. Then, outliers are corrected based on meteorological and physical laws, and linear interpolation is used to complete missing data to ensure data continuity. Next, core features directly related to extreme winds are selected, wind direction is converted into a direction cosine vector to avoid angular periodicity issues, and all core features are standardized to eliminate the influence of dimensional differences. Finally, the processed data is uniformly organized into a CSV format of "site ID-timestamp-feature vector". This preprocessing effectively removes data noise while preserving the minute-level time-series details of extreme winds, providing a high-quality data foundation for subsequent comparative sample construction.

[0021] Step 2: Contrast Sample Construction Mechanism: Based on the preprocessed minute-level data, anchor samples, positive samples, and negative samples are selected for contrastive learning, as follows: Anchor sample generation: with a sliding time window of 15 minutes. The sliding step is 5 minutes. A 15-step sequence of data from a single site is selected as the anchor sample. Samples from the same site, season, and sequence length are grouped into the same training batch to ensure that samples within the batch have a consistent spatiotemporal background and physical standards, thus avoiding cross-scene differences from interfering with sample comparison.

[0022] Positive sample selection: Positive samples are selected using a dual constraint of spatiotemporal proximity and physical consistency, fully utilizing the fine spatiotemporal and physical characteristics of minute-level data. In the temporal dimension, samples with an overlap of ≥60% with the anchor sample's time window are selected from the entire dataset to ensure that both are affected by the same extreme weather system concurrently. In the spatial dimension, data from stations with a straight-line distance of ≤5km from the anchor sample's station are selected to avoid discrepancies in physical processes due to excessive distance. Regarding physical consistency, verification is performed using dual indicators of wind speed temporal trend and pressure system synergy: First, a smoothed 15-minute wind speed sequence between the candidate sample and the anchor sample is calculated. Pearson correlation coefficient The calculation method is as follows: in, For anchor sample number Smooth wind speed for minutes, For candidate samples Smooth wind speed for minutes, , The two are the 15-minute average wind speeds, respectively. This indicates that the wind speed changes of both devices are highly consistent in terms of minute-level variation, meaning they increase in speed and reach their peak values ​​simultaneously. The average air pressure change rate of both devices over a 15-minute period is also calculated. relative deviation The calculation method is as follows: Extreme winds are driven by strong pressure gradients, if This indicates that the physical driving source strengths of the two are consistent. If the candidate samples satisfy... , If it meets the physical consistency requirement, it will be used as a positive sample. .

[0023] Negative sample screening: First, all positive sample candidate sets that meet the criteria of "spatiotemporal proximity and physical consistency" are eliminated to prevent negative samples from mixing with positive sample features at the source, ensuring the essential difference between negative samples and anchor samples. Then, a negative sample candidate pool is determined from the remaining samples. For each sample in the candidate pool and the anchor sample, feature matrices are constructed from their minute-level data. The anchor sample is a 15×5 dimensional matrix. The 15 rows correspond to a 15-minute sequence, and the 5 columns correspond to the core features of smoothed wind speed, air pressure change rate, wind direction cosine vector, 2-meter temperature, and sea level air pressure. The candidate samples are matrices of the same dimension. The feature difference between the two is quantified using matrix cosine similarity, and the calculation formula is as follows: in, This means taking the trace of the result after multiplying the anchor sample matrix by the transposed matrix and the candidate sample matrix. This calculation method can comprehensively reflect the overall differences of multiple features under minute-level time series, avoid the one-sidedness of single feature comparison, and select the 30% of samples with the lowest similarity as negative samples.

[0024] Step 3: Construct a feature extraction framework, which is a two-branch network structure. Each branch includes an encoder and a projector head. The training process is as follows: To improve model robustness and avoid compromising the spatiotemporal-physical consistency of samples, only anchor samples are trained. Data augmentation was performed, generating two augmented samples for comparative training. Two augmentation methods were selected: time-series augmentation and amplification-based co-augmentation. Anchor samples were randomly processed, and the augmented samples were obtained after two data augmentations. and , is represented as: , Next, using the encoder Feature extraction is performed on the enhanced samples to generate a high-dimensional feature vector, represented as: in, The encoder network can be replaced by any base network, such as ConvLSTM or Transformer networks. This is the output of the average pooling layer of the network. This work uses a ConvLSTM network, and the final fully connected layer output has a dimension of 256. Then, the projection head... Mapping the feature representation vector to a lower-level feature space is represented as: The projection head consists of a two-layer fully connected perceptron structure. Then, the feature representation vectors are mapped to a unit hypersphere space using L2 normalization to improve the stability of contrastive learning. The formula is as follows: Finally, positive and negative samples are selected using the positive and negative sample selection method in step 2, and the contrastive loss between standardized features is calculated. in, This is a function for calculating vector similarity. Temperature is a hyperparameter that is optimized so that the distance between positive samples becomes closer and the distance between negative samples becomes farther during training, thereby enabling the encoder to learn feature representations without label supervision.

[0025] Step 4: Extreme Wind Recognition Framework: After the contrastive learning pre-training is completed, this stage constructs the extreme wind recognition module. This framework employs a feature sharing and task-specific fine-tuning strategy: the pre-trained encoder is frozen to retain the representational power of general meteorological features, and an extreme wind classification head is introduced for end-to-end optimization. For the minute-level sequences to be recognized, which have undergone preprocessing and whose format is consistent with the anchor samples... Process it according to the following steps: First, process the input sequence Input to shared encoder Output high-dimensional feature vectors ; then Input category header The predicted labels are obtained as follows: The classification head consists of two stacked fully connected layers and a Softmax layer: The first fully connected layer has an input dimension of 256 and an output dimension of 128, uses the ReLU activation function, and applies batch normalization (BN). The second fully connected layer has an input dimension of 128 and an output dimension of 2. The final outputs are "Extreme Wind" (label). ) and "non-extreme gales" (label) The probability of "extreme wind" is calculated, where the threshold for "extreme wind" is determined by calculating the 95th percentile of the smoothed historical 15-minute average wind speed of the target area over the past three years. If the 15-minute average wind speed to be predicted... If the extreme wind prediction probability output by the classification head is greater than 0.5, and the event is classified as an extreme wind event, the identification is complete. The classification head of this module is optimized using the cross-entropy loss function, as shown in the following formula: in, For real labels, To predict probabilities, This represents the number of samples.

[0026] Step 5: A phased optimization strategy is adopted during model training: First, the model encoder is optimized using an unsupervised contrastive learning loss function. The goal is to learn a general and robust feature representation that can effectively distinguish different weather patterns, especially those with or without extreme events. On this basis, the model is fine-tuned according to the extreme wind recognition task. The parameters of the extreme wind classification head are optimized using the cross-entropy loss function to focus on the discrimination task between extreme and non-extreme winds, avoiding physical feature shifts caused by multi-task interference. At the same time, the physical feature preservation capability of the encoder is relied upon to ensure the rationality and accuracy of the recognition results.

[0027] During the model evaluation phase, a series of metrics are used to comprehensively measure the model's predictive performance and physical plausibility. First, classification evaluation metrics such as accuracy, recall, and F1 score are employed to quantify the model's accuracy and robustness in identifying extreme events such as hurricanes, thunderstorms, and tornadoes. Simultaneously, the proportion of samples identified as extreme winds by the model that simultaneously meet the thresholds for both the core physical driving source of extreme winds and the wind direction stability characteristic threshold is statistically analyzed. This quantifies the degree of fit between the identification results and the physical mechanisms, preventing the model from outputting false identification results without physical basis. Furthermore, to evaluate the effectiveness of contrastive learning pre-training, cosine similarity in the feature space is calculated to measure the clustering degree of similar samples. Finally, spatial visualization techniques are used to visually display the extreme event identification results in different geographical regions, verifying the model's spatial generalization ability, analyzing its performance differences under different terrain and climatic conditions, and explaining potential identification biases.

[0028] In summary, this invention first proposes a meteorological-driven dynamic comparison sample construction mechanism. Anchor samples are defined as continuous time window wind speed sequences that characterize wind field evolution. Positive samples are dynamically selected based on the dual criteria of physical consistency and spatiotemporal consistency of wind field evolution. Negative samples are selected from those with the lowest cosine similarity to anchor samples within a batch. This ensures that the feature space simultaneously captures the consistency and heterogeneity of the same meteorological patterns, embedding physical constraints at the sample level to provide an accurate sample foundation for subsequent identification.

[0029] Secondly, a dual-branch feature extraction framework adapted to extreme wind identification is designed. The encoder extracts the temporal features related to extreme winds, maps them to the unit hyperspherical space through the projection head, and then optimizes the feature discrimination power with the contrast loss function. This enables similar extreme wind weather patterns to cluster in the feature space and separate other weather patterns, thereby strengthening the model's ability to capture the core physical features of extreme winds and reflecting the physical constraint adaptability from the feature learning level.

[0030] Finally, a physical constraint-oriented single-task fine-tuning architecture is established, freezing the pre-trained encoder parameters to fully preserve the key physical laws governing extreme wind formation. Physical constraints have already been incorporated into sample construction and feature optimization during the pre-training phase. The fine-tuning phase focuses solely on parameter fine-tuning for the extreme wind classification branch, using a classification loss function to focus on the accurate optimization of the extreme wind recognition task, avoiding accuracy loss due to multi-task interference. Simultaneously, the physical feature preservation capability of the pre-trained encoder ensures the physical plausibility of the recognition results.

[0031] It is evident that this method innovatively integrates meteorological consistency constraints into the entire comparative learning process by fusing temporal comparative learning with physical constraints. This avoids the impact of additional constraints on model adaptability during the fine-tuning stage and ensures from the source that the extreme wind identification results conform to atmospheric physical laws, significantly improving the credibility of the identification results and avoiding false identifications without physical basis. At the same time, it effectively solves the shortcomings of traditional extreme wind identification models in terms of noise robustness and small sample generalization ability, providing core technical support for the accurate identification of extreme winds.

[0032] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0033] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying extreme winds based on the fusion of spatiotemporal contrastive learning and physical constraints, characterized in that, include: Acquire minute-level meteorological monitoring data and preprocess the meteorological monitoring data to remove data noise, thereby obtaining preprocessed data; Based on the preprocessed data, sample data for spatiotemporal contrastive learning is obtained through filtering. The sample data includes anchor samples, positive samples, and negative samples. The positive samples are selected using a dual constraint of spatiotemporal proximity and physical consistency. The anchor sample is augmented with data, and the augmented anchor sample, the positive sample, and the negative sample are input into a dual-branch feature extraction network for training using a contrastive loss function. The data enhancement includes time-series surround enhancement and amplification synergy enhancement; Based on the trained dual-branch feature extraction network, an extreme wind classification head is introduced and optimized using the cross-entropy loss function to obtain an extreme wind recognition model. The extreme wind recognition model is then used to identify the test data.

2. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion as described in claim 1, characterized in that, The meteorological monitoring data includes at least a 10-meter wind speed component, sea level air pressure, and a 2-meter temperature component.

3. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion as described in claim 1, characterized in that, The specific process of the preprocessing includes: First, data was filtered according to the geographical scope of the target area and the peak periods of extreme winds. At the same time, the timestamps of all data were unified to UTC+8 time zone to ensure consistency of time base. Then, outliers were repaired according to meteorological and physical laws, and linear interpolation was used to complete missing data to ensure data continuity. Next, core features directly related to extreme winds were selected, and wind direction was converted into a direction cosine vector to avoid the problem of angular periodicity. All core features were standardized to eliminate the influence of dimensional differences. Finally, the processed data were uniformly organized into a CSV format of "site ID-timestamp-feature vector".

4. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion as described in claim 1, characterized in that, The step of filtering based on the preprocessed data to obtain sample data for spatiotemporal contrastive learning specifically includes: Using a 15-minute sliding time window and a 5-minute sliding step, 15-step sequences are slidably extracted from the single-site data in the preprocessed data as anchor samples. The anchor samples include multiple training batches of data, each training batch consisting of samples from the same site, season, and sequence length, ensuring that the samples within a batch have consistent spatiotemporal background and physical standards. In the time dimension, samples with an overlap of ≥60% with the time window of the anchor sample are selected from the entire dataset to ensure that both are affected by the same extreme weather system at the same time. In the spatial dimension, station data with a straight-line distance of ≤5km from the anchor sample station are selected to avoid the physical process being disconnected due to excessive distance, thus obtaining positive samples. First, all positive samples are removed. Then, feature matrices are constructed for the remaining samples and the anchor sample respectively. The feature difference between the two is quantified using matrix cosine similarity. The 30% of samples with the lowest similarity are selected as negative samples. The feature matrix of the anchor sample is a 15×5 matrix. The 15 rows correspond to a 15-minute sequence, and the 5 columns correspond to the core features of smoothed wind speed, air pressure change rate, wind direction cosine vector, 2-meter temperature, and sea level air pressure.

5. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion according to claim 4, characterized in that, The constraint of physical consistency is verified by dual indicators: wind speed temporal trend and air pressure system synergy. First, calculate the smoothed wind speed sequence over 15 minutes between the candidate sample and the anchor sample. Pearson correlation coefficient The calculation method is as follows: in, For anchor sample number Smooth wind speed for minutes, For candidate samples Smooth wind speed for minutes, The average wind speed over 15 minutes for the anchor sample. The candidate sample's 15-minute average wind speed, if This indicates that the wind speed changes of both devices are highly consistent across minutes, achieving synchronous growth and reaching peak values ​​simultaneously. Then, the average air pressure change rate of both devices over 15 minutes is calculated simultaneously. relative deviation The calculation method is as follows: in, The average rate of change of air pressure over 15 minutes for the anchor sample. The average rate of change of air pressure over 15 minutes for the candidate sample; if the candidate sample meets the following conditions... and If it meets the physical consistency requirement, it is determined to be a positive sample. .

6. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion according to claim 4, characterized in that, The formula for calculating the matrix cosine similarity is: in, This represents taking the trace of the result after multiplying the transpose of the anchor sample matrix with the candidate sample matrix. The F-norm of the anchor sample feature matrix is ​​represented by the F-norm. This represents the F-norm of the feature matrix of the candidate sample.

7. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion according to claim 1, characterized in that, The contrastive loss function is expressed as follows: in, This is a function for calculating vector similarity. For temperature hyperparameters, z i To enhance the feature vectors of the anchor samples after processing by the encoder and projector, z pos The feature vector of a positive sample after processing by the encoder and projector head. z neg The feature vector is the result of processing negative samples by the encoder and projector. p denoted as , where n is the number of positive samples and n is the number of negative samples.

8. The extreme wind identification method based on spatiotemporal contrastive learning and physical constraint fusion according to claim 1, characterized in that, The structure of the extreme wind classification head consists of two stacked fully connected layers and a Softmax layer: the first fully connected layer has an input dimension of 256 and an output dimension of 128, uses the ReLU activation function, and adds batch normalization; the second fully connected layer has an input dimension of 128 and an output dimension of 2.

Images