Deep learning-based drainage basin abnormal rainfall event identification monitoring and weather behavior analysis method and system

Abstract

The invention discloses a deep learning-based drainage basin abnormal rainfall event identification monitoring and weather behavior analysis method and system, and the method comprises the steps: obtaining the summer rainfall original data of a drainage basin, and carrying out the preprocessing; the abnormal rainfall event is divided into different subspace modes through a self-organizing mapping machine learning method; qualitatively analyzing a space-time evolution rule and an interaction mechanism of water vapor conveying characteristics and atmospheric circulation configuration in abnormal rainfall according to each subspace mode; evaluating the influence of different meteorological factors on extreme weather conditions through a Bayesian method, and quantitatively exploring abnormal rainfall event change factors; the system comprises a data preprocessing module, a self-organizing mapping mode classification module, a water vapor conveying and atmospheric circulation influence mechanism module and a Bayesian method meteorological factor influence analysis module. According to the method, accurate identification and monitoring of the drainage basin abnormal rainfall event are realized, the internal characteristics of rainfall data can be effectively learned, and the identification capability of the abnormal rainfall mode is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of extreme hydrological event analysis, and particularly to a method and system for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning. Background Art

[0002] As an important branch in the field of artificial intelligence, deep learning has achieved remarkable results in many fields with its powerful data processing ability and pattern recognition ability. Against the background of intensified global climate change, abnormal rainfall events in basins are characterized by strong suddenness, high disaster-causing potential, and great prediction difficulty, which pose a severe challenge to the traditional hydrometeorological monitoring technology system. The current mainstream methods for monitoring extreme hydrological events mainly rely on statistical analysis of meteorological station observation data and numerical weather prediction models. These methods have problems such as relying on expert experience, insufficient ability to handle nonlinear problems, low data utilization efficiency, poor real-time performance, and difficulty in fusing multi-source data. When dealing with the challenges of frequent extreme weather under the background of climate change, traditional methods have gradually revealed significant technical limitations. Applying deep learning technology to the identification and monitoring of abnormal rainfall events in basins and weather behavior analysis is expected to break through the limitations of traditional methods, improve the accuracy of identifying abnormal rainfall events, and reveal the causes of abnormal rainfall events. Therefore, it is very necessary to invent a method for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning. Summary of the Invention

[0003] Object of the Invention: The object of the present invention is to provide a method and system for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning.

[0004] To achieve the above technical object and reach the above technical effect, the present invention is realized through the following technical solutions:

[0005] In the first aspect, the present invention provides a method for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning, including:

[0006] Obtain the original summer rainfall data of the basin, perform preprocessing, construct a data set, and divide the data set into a training set and a validation set;

[0007] Input the deep learning autoencoder network with an embedded multi-head attention mechanism to train and identify abnormal rainfall events in the target basin; the self-organizing mapping machine learning method divides abnormal precipitation events into different subspace patterns;

[0008] Qualitatively analyze the spatio-temporal evolution law and interaction mechanism of water vapor transport characteristics and atmospheric circulation configuration in abnormal precipitation according to each subspace pattern;

[0009] Evaluate the impacts of different meteorological factors on extreme weather conditions through Bayesian methods, and quantitatively explore the factors affecting abnormal rainfall events.

[0010] Furthermore, the deep learning autoencoder network embedded with a multi-head attention mechanism includes an encoder, a decoder, a skip connection structure, and a multi-head attention mechanism module:

[0011] The encoder compresses the training data to achieve feature learning of the network. The original input data size (H, W, C1) represents height, width, and the number of channels. The input convolutional module Conv2Dblock and the Maxpooling2D layer perform four downsamplings to capture significant local and global representations.

[0012] Among them, the Conv2D convolution operation is (f * g)(u * v) = ∑ i ∑ j f(i, j)g(u, v), the function f(i, j) is called the pixel value at the position (i, j) in the input function image, g(u, v) is called the filter or convolution kernel, and the Dropout layer randomly zeros out convolutional features to optimize the model;

[0013] Pooling MaxPooling2D, that is, in the image matrix, it slides from left to right and from top to bottom according to the 2 * 2 window size, and takes the maximum value of the pixels in the indexed area, so that the convolutional features H and W are compressed to half of the original size in turn;

[0014] Input the multi-head attention mechanism module to learn the spatial heterogeneity of abnormal rainfall data, and perform layer normalization on the input feature map P5; divide the normalized feature map into local windows of K × K, and each window contains M × M spatial positions, where Perform the following operations:

[0015] (i) Flatten the features within the window into a sequence of dimension (M 2 , C);

[0016] (ii) Generate query matrix Q, key matrix K, and value matrix V through learnable linear projections;

[0017] (iii) Calculate the attention weights with relative position biases;

[0018]

[0019] Among them, d k = C / num_heads, num_heads is the number of heads, the input features are divided into multiple heads, and each head independently performs attention calculation. B is a learnable bias matrix based on the relative positions of the pixels within the window;

[0020] (iv) Concatenate the results of multi-head attention and restore the dimension through linear projection;

[0021] After that, perform a residual connection, that is, add the window attention output and the original feature map P5 element by element; perform secondary normalization on the residual result and input it into a multi-layer perceptron unit, which includes two fully connected layers and a GELU activation function:

[0022] MLP(x) = W2(GELU(W1x + B)) + B

[0023] W1 ∈ R {c×4c} ,W2 ∈ R {4c×c} ,both are learnable parameters, x is the input feature value, and the output feature image P6 is obtained, whose height, width, and number of channels are the same as P5. where erf(·) is the error function.

[0024] Skip connection and decoder, that is, restore the training data to the original scale, and share the encoder learning parameters through the skip connection to realize the training of the autoencoder network. Its structure is successively input, deconvolution layer Conv2Dtranspose, and convolution module Conv2Dblock.

[0025] where Conv2Dtranspose performs the upsampling operation, where S is the stride, that is, the interval of the convolution kernel moving pixel values, which is 2 here, and p is the pixel value filled to ensure that the size is restored to twice the current size. The output result is fused with the encoder layer through a Concatenate layer, that is, the two feature maps are stacked in the channel number dimension to double the number of channels, and then input into the convolution module Conv2Dblock. Repeat the above operations until the original size of the data is finally restored.

[0026] Furthermore, the self-organizing mapping machine learning method divides abnormal precipitation events into different subspace patterns, including:

[0027] Input the time data x corresponding to the abnormal rainfall time series into the self-organizing mapping machine learning method, and its formula is:

[0028]

[0029] where, w i is the weight vector of the i-th neuron, and n is the dimension of the data; the neurons of SOM will be randomly initialized, and each input value x j is calculated with the corresponding weight vector w ij The network calculates the Euclidean distance between it and the weight vectors of each neuron and finds the node most similar to the input vector;

[0030] The weight update formula is:

[0031] w i (t + 1) = w i (t) + α(t)·h ci (t)·(x - w i (t))

[0032] where w i (t) is the weight vector to be updated for the i-th neuron, α(t) is the learning rate, h ci (t) is the neighborhood function, representing the degree of weight update of neighboring neurons, x is the input data, w i (t + 1) is the updated weight vector of the i-th neuron;

[0033] Until the weight vector of the network tends to be stable, output the subspace patterns of different abnormal rainfall events for abnormal precipitation events.

[0034] Furthermore, the method for evaluating the influence of different meteorological factors on extreme weather conditions through the Bayesian method and quantitatively exploring the factors causing changes in abnormal rainfall events includes:

[0035] Qualitatively analyze the spatio-temporal evolution law and interaction mechanism of water vapor transport characteristics and atmospheric circulation configuration in abnormal precipitation, and determine the meteorological factors with a mechanism affecting basin-wide abnormal rainfall; the Bayesian test method is that the probability p of abnormal rainfall in the subspace pattern is the ratio of abnormal rainfall events to the total observed time series rainfall events; by assuming the prior distribution of α, calculate the posterior distribution of β from another distribution with the same different parameters, α(s + 1) and β(n + 1), the number of abnormal rainfall events (s) and the date of the observed time series (n) come from a sample of a Bernoulli process, and use s and n as sufficient statistics to analyze the Bernoulli process sample:

[0036]

[0037] where s represents the number of abnormal rainfall events under the positive (negative) correlation influence of meteorological factors, n represents the time series rainfall events under the positive (negative) correlation influence of meteorological factors, constant represents a constant term, where Γ(*) represents the gamma function, and B(s + 1, n - s + 1) represents the beta function.

[0038] In a second aspect, the present invention provides a system for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning, including:

[0039] A data preprocessing module for obtaining the original summer rainfall data of the basin, performing preprocessing, constructing a data set, and dividing the data set into a training set and a validation set;

[0040] An abnormal rainfall event recognition and clustering module is used to train a deep learning autoencoder network with a multi-head attention mechanism to recognize abnormal rainfall events in the target basin; the self-organizing mapping machine learning method clusters abnormal precipitation events into different subspace patterns;

[0041] A water vapor transport and atmospheric circulation influence mechanism module is used to qualitatively analyze the spatio-temporal evolution law and interaction mechanism of water vapor transport characteristics and atmospheric circulation configuration in abnormal precipitation according to each subspace pattern;

[0042] A Bayesian method for analyzing the influence of meteorological factors module is used to evaluate the influence of different meteorological factors on extreme weather conditions through the Bayesian method and quantitatively explore the changing factors of abnormal rainfall events.

[0043] Furthermore, the abnormal rainfall event recognition and clustering module is used to train a deep learning autoencoder network with a multi-head attention mechanism to recognize abnormal rainfall events in the target basin; the self-organizing mapping machine learning method clusters abnormal precipitation events into different subspace patterns;

[0044] The deep learning autoencoder network with a multi-head attention mechanism mainly includes an encoder, a decoder, a skip connection structure, and a multi-head attention mechanism module:

[0045] The encoder compresses the training data to achieve feature learning of the network. The original input data size (H, W, C1) represents height, width, and number of channels. The input convolutional module Conv2Dblock and Maxpooling2D layer perform four downsamplings to capture significant local and global representations.

[0046] Among them, the Conv2D convolution operation is (f * g)(u * v) = ∑ i ∑ j f(i, j)g(u, v), the function f(i, j) is called the pixel value at the position (i, j) in the input function image, and g(u, v) is called the filter or convolution kernel. The Dropout layer randomly sets zero convolution features to optimize the model;

[0047] Pooling MaxPooling2D, that is, in the image matrix, slide from left to right and from top to bottom according to the 2 * 2 window size, and take the maximum value of the pixels in the indexed area, so that the convolution features H and W are compressed to half of the original size in turn;

[0048] The input multi-head attention mechanism module is used to learn the spatial heterogeneity of abnormal rainfall data, and perform layer normalization on the input feature map P5; the normalized feature map is divided into local windows of K × K, and each window contains M × M spatial positions, where Perform the following operations:

[0049] (i) Flatten the features within the window into a sequence of dimension (M 2 , C);

[0050] (ii) Generate query matrix Q, key matrix K, and value matrix V through learnable linear projections;

[0051] (iii) Calculate attention weights with relative position biases;

[0052]

[0053] where d k = C / num_heads, num_heads is the number of heads, the input features are split into multiple heads, and each head performs attention calculation independently. B is a learnable bias matrix based on the relative positions of pixels within the window;

[0054] (iv) Concatenate the multi-head attention results and restore the dimension through a linear projection;

[0055] After that, perform a residual connection, that is, add the window attention output and the original feature map P5 element-wise; perform layer normalization on the residual result, and input it into a multi-layer perceptron unit, which includes two fully connected layers and a GELU activation function:

[0056] MLP(x) = W2(GELU(W1x + B)) + B

[0057] W1 ∈ R {c×4c} , W2 ∈ R {4c×c} , both are learnable parameters, x is the input feature value, and the output feature image P6 is obtained, whose height, width, and number of channels are the same as those of P5, where erf(·) is the error function.

[0058] Skip connection and decoder, that is, restore the training data to the original scale, and share the encoder learning parameters through the skip connection to realize the training of the autoencoder network. Its structure is sequentially input, transposed convolutional layer Conv2Dtranspose, and convolutional module Conv2Dblock,

[0059] where Conv2Dtranspose performs an upsampling operation where S is the stride, that is, the pixel value interval at which the convolutional kernel moves, here it is 2, and p is the pixel value for padding to ensure that the size is restored to twice the current size. The output result is fused with the encoder layer through a Concatenate layer, that is, the two feature maps are stacked together in the channel dimension to double the number of channels, and input into the convolutional module Conv2Dblock. Repeat the above operations until the original size of the data is finally restored.

[0060] Furthermore, the abnormal rainfall event recognition and clustering module is used to input a deep learning autoencoder network embedded with a multi-head attention mechanism to train and recognize abnormal rainfall events in the target basin; the self-organizing mapping machine learning method divides abnormal precipitation events into different subspace patterns, including:

[0061] Input the corresponding time data x of the abnormal rainfall time series into the self-organizing mapping machine learning method, and its formula is:

[0062]

[0063] where w i is the weight vector of the i-th neuron, and n is the dimension of the data; the neurons of SOM will be randomly initialized, and each input value x j , and the corresponding weight vector w ij The network calculates the Euclidean distance between it and the weight vectors of each neuron, and finds the node most similar to the input vector;

[0064] The weight update formula is:

[0065] w i (t + 1) = w i (t) + α(t).h ci (t).(x - w i (t))

[0066] where w i (t) is the weight vector to be updated of the i-th neuron, α(t) is the learning rate, h ci (t) is the neighborhood function, indicating the weight update degree of neighboring neurons, x is the input data, and w i (t + 1) is the updated weight vector of the i-th neuron;

[0067] Until the weight vectors of the network tend to be stable, output different subspace patterns of abnormal rainfall events for abnormal precipitation events.

[0068] Furthermore, the Bayesian method evaluation module is used to evaluate the influence of different meteorological factors on extreme weather conditions through the Bayesian method, and quantitatively explore the change factors of abnormal rainfall events; including:

[0069] Qualitatively analyze the spatio-temporal evolution law of the water vapor transport characteristics and the atmospheric circulation configuration in abnormal precipitation and their interaction mechanism, and determine the meteorological factors with the influence mechanism of basin abnormal rainfall; The Bayesian test method is that the probability p of the subspace mode abnormal rainfall is the ratio of the abnormal rainfall event to the total observed time series rainfall event; By assuming the prior distribution of α, calculate the posterior distribution of β from another different parameter with the same distribution, α(s + 1) and β(n + 1), the number of abnormal rainfall events (s) and the observation time series date (n) come from a Bernoulli process sample. Using s and n as sufficient statistics, analyze the Bernoulli process sample:

[0070]

[0071] Among them, s represents the number of abnormal rainfall events under the positive (negative) correlation influence of meteorological factors, n represents the time series rainfall events under the positive (negative) correlation influence of meteorological factors, constant represents the constant term, where Γ(*) represents the gamma function, and B(s + 1, n - s + 1) represents the beta function.

[0072] In a third aspect, the present invention provides an electronic device, including a memory, a processor, and a program stored on the memory and executable on the processor. When the program is executed, the steps of the method for identifying and monitoring basin abnormal rainfall events and analyzing weather behavior based on deep learning are implemented.

[0073] In a fourth aspect, the present invention provides a computer-readable storage medium, hereinafter simply referred to as a storage medium. The storage medium stores a computer program, which is designed to implement the steps of the method for identifying and monitoring basin abnormal rainfall events and analyzing weather behavior based on deep learning when running.

[0074] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: By training an autoencoder network with embedded multi-head attention on the input preprocessed data, the present invention realizes the accurate identification and monitoring of basin abnormal rainfall events. The network architecture is divided into an encoder, a decoder, a skip connection, and a multi-head attention module; The encoder compresses the training data to achieve feature learning, the decoder and the skip connection restore the data to the original scale, and the skip connection shares the encoder parameters to improve the local prediction ability. The multi-head attention module aims at the spatial heterogeneity of rainfall data, captures multi-scale and multi-faceted correlations by splitting features and independently calculating the attention weights of each head, enhances the global spatial understanding and local abnormal rainfall sensitivity; The autoencoder network shows significant advantages in processing abnormal rainfall data. Through the collaborative work of the encoder, decoder, skip connection structure, and multi-head attention mechanism module, the network can effectively learn the internal features of rainfall data and enhance the ability to identify abnormal rainfall patterns. Description of the Drawings

[0075] Figure 1 It is a flow chart of the method of the present invention;

[0076] Figure 2 It is a structural diagram of an autoencoder network embedded with multi-head attention;

[0077] Figure 3 It is a schematic diagram of different abnormal rainfall patterns divided by the self-organizing mapping machine learning method;

[0078] Figure 4 It is a schematic diagram of using Bayesian test to evaluate the influence of different meteorological factors on extreme weather conditions;

[0079] Figure 5 It is a structural schematic diagram of the system of the present invention. Specific implementation manners

[0080] The technical solution of the present invention will be further described below with reference to the accompanying drawings. Those skilled in the art will understand that the purposes and advantages that can be achieved by the present invention are not limited to the specific beneficial effects described above, and the above and other purposes that the present invention can achieve will be more clearly understood according to the following detailed description.

[0081] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in combination with the embodiments disclosed in the present invention can be implemented in hardware, software, or a combination of both. Specifically, whether to execute in a hardware or software manner depends on the specific application and design and tree conditions of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0082] The mention of "embodiment" in the present invention means that the specific features, structures, or characteristics described in combination with the embodiments may be included in at least one embodiment of the present invention. The occurrence of this phrase at various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art understand explicitly and implicitly that the embodiments described herein can be combined with other embodiments.

[0083] Embodiment 1

[0084] As Figure 1 shown, a method for identifying and monitoring abnormal rainfall events and analyzing weather behaviors in a river basin based on deep learning provided by the present invention includes the following steps:

[0085] Step 1: Obtain the original summer rainfall data of the river basin, subtract the seasonal cycle, remove short-term weather events, and normalize to obtain the input data.

[0086] The original data is generally obtained from meteorological stations in various places and is observational data obtained using various interpolation methods. The commonly used cdo (Climate Data Operators), an open-source command-line tool widely used in the market for processing and analyzing climate data, is used to perform operations such as data merging, cropping, and interpolation to match the original rainfall data of the target watershed time series required for the study.

[0087] Data preprocessing, including data preparation and normalization, is required to generate suitable input data for the autoencoder technology. Among them, data preparation includes calculating the climatological daily average value of the original rainfall data over a fixed period as the seasonal cycle, and obtaining the daily average precipitation anomaly value by subtracting the seasonal cycle from the original data. Calculate the 5-day running mean anomaly to remove short-term weather events.

[0088] Finally, the anomaly data at each grid point is normalized to obtain the input data for the autoencoder network.

[0089] The normalization formula is:

[0090]

[0091] where p represents the pixel value or the original rainfall data value at any position, and p r is the normalized pixel value or the original rainfall data value. P min is the minimum value of the pixel value or the minimum value of the original rainfall data value, and p max is the maximum value of the pixel value or the maximum value of the original rainfall data value

[0092] Step 2: Input the deep learning autoencoder network with multi-head attention mechanism for training to identify abnormal rainfall events in the target watershed;

[0093] It should be noted that the attention mechanism is a technology widely used in the field of deep learning, especially in natural language processing (NLP) and computer vision (CV). Its core idea is to optimize the calculation process and improve the detection efficiency of some features by adjusting the weight distribution, introducing new calculation methods, and analyzing the calculation context vector, so as to improve the performance of the model. The multi-head attention mechanism used in this application can capture the semantic associations in different subspaces of the input sequence simultaneously by running multiple independent attention heads in parallel. By integrating the results of multiple attention heads, the diversity of data can be captured from different perspectives. Affected by regional geographical features and the complexity of the microclimate system, meteorological drought shows more closely related spatio-temporal variations in local areas. In view of this, adopting a multi-head attention mechanism (Multi-Head-Attention) module with local attention ability and shifted window mechanism can more accurately capture and effectively analyze these inherent spatio-temporal structure features, thus enhancing the model's understanding and generalization ability for complex sequence tasks, and building an abnormal rainfall recognition model for the target basin based on an improved autoencoder neural network. It should be noted that the abnormal rainfall recognition model for the target basin can also be based on other versions of autoencoders with encoding-decoding capabilities, which are not limited here. After building the abnormal rainfall recognition model for the target basin, it is trained and debugged to obtain an abnormal rainfall recognition model with more accurate predictions.

[0094] As Figure 2 shown, a deep learning autoencoder network embedding a multi-head attention mechanism is constructed and the model is trained using the preprocessed dataset. The specific model details are as follows;

[0095] The autoencoder network structure includes an encoder, a decoder, a skip connection structure, and a multi-head attention mechanism module.

[0096] The input data enters, and the encoder outputs that compress features of 5 different scales are used to achieve skip connection. Each layer includes a ConV2D Block and a MaxPooling layer. The ConV2D Block consists of a ConV2D convolutional layer, a Dropout packet loss layer, and a ConV2D convolutional layer. The feature map size is successively reduced to

[0097] of the previous layer. The input feature map P5 enters the multi-head attention mechanism module and performs layer normalization processing; then the normalized feature map is divided into local windows of K×K, and each window contains M×M spatial positions, where the following operations will be performed:

[0098] (i) Flatten the features within the window into a sequence of dimension (M 2 , C5);

[0099] (ii) Generate query matrix Q, key matrix K, and value matrix V through learnable linear projections;

[0100] (iii) Calculate attention weights with relative position biases:

[0101]

[0102] where d k = C5 / num_heads, and B is a learnable bias matrix based on the relative positions of pixels within the window;

[0103] (iv) Concatenate the multi-head attention results and restore the dimension through linear projection;

[0104] After that, perform a residual connection, that is, add the window attention output and the original feature map P5 element-wise; perform layer normalization on the residual result and input it into a multi-layer perceptron unit, which includes two fully connected layers and a GELU activation function:

[0105] MLP(x) = W2(GELU(W1x + b1)) + b2

[0106] W1 ∈ R {c×4c} and W1 ∈ R {4c×c} are both learnable parameters. Thus, the output feature image P6 is obtained, whose height, width, and number of channels are the same as those when inputting the attention module.

[0107] Then, the decoder restores to the original data size by upsampling 5 different-scale features. Each layer includes a ConV2Dtranspose deconvolution layer and a ConV2D Block. It should be noted that the feature images restored by deconvolution upsampling will be fused with the features of the same size output by the encoder ConV2D Block through a skip connection Concatenate layer, that is, the two feature maps are stacked in the channel dimension until the input data size is restored, which is the complete autoencoder network architecture.

[0108] Step 3: The self-organizing mapping machine learning method divides abnormal precipitation events into different subspace patterns.

[0109] The core idea of the SOM algorithm is to perform clustering analysis on the input high-dimensional data by simulating the self-organization process of the human brain's nervous system and map it onto a low-dimensional grid structure. This process not only preserves the topological structure of the original data but also intuitively shows the similarities and differences between data. The implementation of the SOM algorithm mainly relies on two core mechanisms: competitive learning and cooperative adjustment.

[0110] In the step, the input layer is the rainfall of the abnormal precipitation time series, and the output layer is a two-dimensional grid. Each neuron on the grid represents a clustering center. It should be noted that for different subspace pattern clustering requirements, parameters including but not limited to the grid size (m_neurons x n_neurons), learning rate, neighborhood function, etc. can be modified. As Figure 3 shown, in the embodiment, a 2×3 SOM grid is selected from several different n×m grid analyses for the classification of abnormal precipitation, and the following schematic diagrams of different abnormal rainfall patterns are obtained.

[0111] Step 4: Qualitatively analyze the spatio-temporal evolution law and interaction mechanism of the water vapor transport characteristics and atmospheric circulation configuration in abnormal precipitation according to each subspace pattern

[0112] Taking the ERA5 dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF; https: / / cds.climate.copemicus.eu / ) as an example, according to the time and space distribution characteristics of the abnormal precipitation types divided by each subspace pattern, analyze the differences and changes of key elements such as water vapor sources, water vapor transport paths, and water vapor transport intensities under different subspace patterns. Specifically, by extracting the reanalysis data of the abnormal rainfall events corresponding to the subspace patterns in the Era5 data, analyze the changes in the structure, intensity, and position of atmospheric circulation systems (such as monsoons, westerly belts, subtropical highs, etc.) under different subspace patterns to explore the influence mechanism on the precipitation process and precipitation types. On this basis, combined with the interaction mechanism between water vapor transport characteristics and atmospheric circulation configuration, specifically analyze the potential impact of water vapor transport on the stability and dynamic changes of the atmospheric circulation, and how the atmospheric circulation configuration regulates the path and intensity of water vapor transport.

[0113] Step 5: Use the Bayesian method to evaluate the influence of different meteorological factors on extreme weather conditions and quantitatively explore the changing factors of abnormal rainfall events.

[0114] As Figure 4As shown in the figure, several meteorological factors with the influence mechanism of basin abnormal rainfall are selected through qualitative analysis of different subspace abnormal rainfall patterns. The time series of PDO and AMO from 1961 to 2019 are selected, which are taken from the monthly scale index of the Tokyo Climate Center (http: / / ds.data.jma.go.jp / tcc / tcc / products / elnino / decadal / pdo.html) and the NOAA Earth System Research Laboratory (https: / / ps1.noaa.gov / data / correlation / amon.us.long.data) respectively. For ENSO, we use the annual value of the Southern Oscillation Index (SOI) of the monthly index of the Australian Bureau of Meteorology (http: / / www.bom.gov.au / climate / current / soi2.shtml), which is calculated based on the pressure difference between Tahiti and Darwin.

[0115] By assuming the prior distribution of α, the posterior distribution of β is calculated from another different parameter with the same distribution. α(s + 1) and β(n + 1), the number of abnormal rainfall events (s) and the date of the observed time series (n) come from a sample of the Bernoulli process. Using s and n as sufficient statistics, the Bernoulli process sample is analyzed.

[0116]

[0117] Among them, s represents the number of abnormal rainfall events under the positive (negative) correlation influence of meteorological factors, n represents the time series rainfall events under the positive (negative) correlation influence of meteorological factors, constant represents the constant term, where Γ(*) represents the gamma function, and B(s + 1, n - s + 1) represents the beta function.

[0118] As Figure 4 shown in the figure, the probability density functions representing each meteorological factor (such as S01+, PDO+, AMO+, etc.) are plotted. The Beta distribution parameters corresponding to each meteorological factor are significantly different. S01- and PDO+ may be more likely to trigger abnormal rainfall events with higher probabilities, while the Anomalies distribution reflects the stability of systematic anomalies. By quantifying the distribution characteristics, it can provide a basis for the parameter optimization of climate prediction models.

[0119] Example Two

[0120] In addition, based on the same inventive concept, as Figure 5 shown in the figure, the embodiment of the present invention also provides a system for identifying and monitoring basin abnormal rainfall events and analyzing weather behavior based on deep learning, including:

[0121] A data preprocessing module, which is used to obtain the original summer rainfall data of the basin, perform preprocessing, construct a data set, and divide the data set into a training set and a validation set;

[0122] An abnormal rainfall event recognition and clustering module, which is used to input a deep learning autoencoder network with a multi-head attention mechanism to train and recognize abnormal rainfall events in the target basin; the self-organizing mapping machine learning method clusters abnormal precipitation events into different subspace patterns;

[0123] A water vapor transport and atmospheric circulation influence mechanism module, which is used to qualitatively analyze the spatio-temporal evolution law and interaction mechanism of the water vapor transport characteristics and atmospheric circulation configuration in abnormal precipitation according to each subspace pattern;

[0124] A Bayesian method for analyzing the influence of meteorological factors module, which is used to evaluate the influence of different meteorological factors on extreme weather conditions through the Bayesian method and quantitatively explore the changing factors of abnormal rainfall events.

[0125] The abnormal rainfall event recognition and clustering module is used to input a deep learning autoencoder network with a multi-head attention mechanism to train and recognize abnormal rainfall events in the target basin; the self-organizing mapping machine learning method clusters abnormal precipitation events into different subspace patterns;

[0126] The deep learning autoencoder network with a multi-head attention mechanism mainly includes an encoder, a decoder, a skip connection structure, and a multi-head attention mechanism module:

[0127] The encoder compresses the training data to realize the feature learning of the network. The original input data size (H, W, C1) represents height, width, and number of channels. The input convolutional module Conv2Dblock and the Maxpooling2D layer perform four downsamplings to capture significant local and global representations.

[0128] Among them, the Conv2D convolution operation is (f*g)(u*v) = ∑ i ∑ j f(i, j)g(u, v). The function f(i, j) is called the pixel value at the position (i, j) in the input function image, and g(u, v) is called the filter or convolution kernel. The Dropout layer randomly sets zero convolution features to optimize the model;

[0129] Pooling MaxPooling2D, that is, in the image matrix, it slides from left to right and from top to bottom according to the 2*2 window size, and indexes the maximum value of the pixels in the area, so that the convolution features H and W are compressed to half of the original size in turn;

[0130] Input the multi - head attention mechanism module to learn the spatial heterogeneity of abnormal rainfall data, and perform layer normalization on the input feature map P5; divide the normalized feature map into local windows of K×K, where each window contains M×M spatial positions, where Perform the following operations:

[0131] (i) Flatten the features within the window into a sequence of dimension (M 2 , C);

[0132] (ii) Generate query matrix Q, key matrix K, and value matrix V through learnable linear projections;

[0133] (iii) Calculate the attention weights with relative position biases;

[0134]

[0135] where d k = C / num_heads, num_heads is the number of heads, the input features are divided into multiple heads, and each head performs attention calculation independently. B is a learnable bias matrix based on the relative positions of pixels within the window;

[0136] (iv) Concatenate the multi - head attention results and restore the dimension through a linear projection;

[0137] After that, perform a residual connection, that is, add the window attention output and the original feature map P5 element - by - element; perform secondary normalization on the residual result and input it into a multi - layer perceptron unit, which contains two fully - connected layers and a GELU activation function:

[0138] MLP(x) = W2(GELU(W1x + B))+B

[0139] W1 ∈ R {c×4c} , W2 ∈ R {4c×c} , both are learnable parameters, x is the input feature value, and the output feature image P6 is obtained, whose height, width, and number of channels are the same as P5, where erf(·) is the error function.

[0140] Skip connections and decoders, that is, the restoration of the training data to the original scale, and sharing the encoder learning parameters through skip connections to achieve the training of the auto - encoder network. Its structure is successively input, transposed convolutional layer Conv2Dtranspose, and convolutional module Conv2Dblock,

[0141] where Conv2Dtranspose performs The upsampling operation, where S is the stride, i.e., the pixel value interval by which the convolutional kernel moves, here it is 2, and p is the pixel value for padding to ensure the size is restored to twice the current size. The output result is skip-connected to the encoder layer and fused by the Concatenate layer, that is, the two feature maps are stacked together in the channel dimension to double the number of channels. Then it is input into the convolutional module Conv2Dblock, and the above operations are repeated until the original size of the data is finally restored.

[0142] The abnormal rainfall event recognition and clustering module is used to input and train a deep learning autoencoder network embedded with a multi-head attention mechanism to recognize abnormal rainfall events in the target basin; the self-organizing mapping machine learning method divides abnormal precipitation events into different subspace patterns, including:

[0143] Input the corresponding time data x of the abnormal rainfall time series into the self-organizing mapping machine learning method, and its formula is:

[0144]

[0145] where, w i is the weight vector of the i-th neuron, and n is the dimension of the data; the neurons of SOM will be randomly initialized, and each input value x j , and the corresponding weight vector w ij The network calculates the Euclidean distance between it and the weight vectors of each neuron and finds the node most similar to the input vector;

[0146] The weight update formula is:

[0147] w i (t + 1) = w i (t) + α(t).h ci (t).(x - w i (t))

[0148] where, w i (t) is the weight vector to be updated of the i-th neuron, α(t) is the learning rate, h ci (t) is the neighborhood function, indicating the degree of weight update of neighboring neurons, x is the input data, and w i (t + 1) is the updated weight vector of the i-th neuron;

[0149] Until the weight vectors of the network tend to be stable, different abnormal rainfall event subspace patterns of abnormal precipitation events are output.

[0150] The Bayesian method evaluation module is used to evaluate the impact of different meteorological factors on extreme weather conditions through the Bayesian method and quantitatively explore the change factors of abnormal rainfall events; including:

[0151] Qualitatively analyze the spatio-temporal evolution law of the water vapor transport characteristics and the atmospheric circulation configuration in abnormal precipitation and their interaction mechanism, and determine the meteorological factors with the influence mechanism of basin abnormal rainfall; the Bayesian test method is that the probability p of the subspace mode abnormal rainfall is the ratio of the abnormal rainfall event to the total observed time series rainfall events; by assuming the prior distribution of α, calculate the posterior distribution of β from another different parameter with the same distribution, α(s + 1) and β(n + 1), the number of abnormal rainfall events (s) and the observation time series date (n) come from a Bernoulli process sample, and use s and n as sufficient statistics to analyze the Bernoulli process sample:

[0152]

[0153] Among them, s represents the number of abnormal rainfall events under the positive (negative) correlation influence of meteorological factors, n represents the time series rainfall events under the positive (negative) correlation meteorological factor influence, constant represents the constant term, where Γ(*) represents the gamma function, and B(s + 1, n - s + 1) represents the beta function.

[0154] Example Three

[0155] In addition, based on the same inventive concept, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behaviors based on deep learning are implemented.

[0156] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in the memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following technologies well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGA), field programmable gate arrays (FPGA), etc.

[0157] Example Four

[0158] In addition, based on the same inventive concept, an embodiment of the present invention further provides a storage medium storing a computer program, and the computer program is designed to implement the steps of the method for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behaviors based on deep learning when running.

[0159] Those of ordinary skill in the art can understand that all or part of the steps carried out in the methods of the above embodiments can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0160] In addition, in each of the embodiments of the present invention, the functional units can be integrated into one processing module, or each unit can exist physically alone, or two or more units can be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0161] The above-mentioned computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium well-known in the art.

[0162] The above specific description further details the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above is only a specific embodiment of the present invention and is not used to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying and monitoring abnormal rainfall events and analyzing weather behavior in a river basin based on deep learning, characterized in that, Including: Obtain the original data of summer rainfall in the basin, conduct preprocessing, construct a dataset, and divide the dataset into a training set and a validation set; Input the deep learning autoencoder network with an embedded multi-head attention mechanism to train and identify abnormal rainfall events in the target basin; the self-organizing mapping machine learning method divides the abnormal precipitation events into different subspace patterns; Qualitatively analyze the spatio-temporal evolution law and interaction mechanism of water vapor transport characteristics and atmospheric circulation configuration in abnormal rainfall according to each subspace pattern; Evaluate the influence of different meteorological factors on extreme weather conditions through the Bayesian method, and quantitatively explore the changing factors of abnormal rainfall events.

2. The method for identifying and monitoring abnormal rainfall events and analyzing weather behaviors in a river basin based on deep learning according to claim 1, wherein The deep learning autoencoder network with an embedded multi-head attention mechanism includes an encoder, a decoder, a skip connection structure, and a multi-head attention mechanism module: The encoder compresses the training data to achieve the feature learning of the network. The original input data size (H, W, C1) represents height, width, and number of channels. The input convolutional module Conv2Dblock and the Maxpooling2D layer perform four downsamplings to capture significant local and global representations. Among them, the Conv2D convolution operation is (f * g)(u * v) = ∑ i ∑ j f(i, j)g(u, v), the function f(i, j) is called the pixel value at the position (i, j) in the input function image, g(u, v) is called the filter or convolution kernel, and the Dropout layer randomly sets convolution features to zero to optimize the model; Pooling MaxPooling2D, that is, in the image matrix, from left to right and from top to bottom, slide according to the 2*2 window size, and take the maximum value of the pixels in the indexed area, so that the convolutional features H and W are successively compressed to half of the original size; Input the multi-head attention mechanism module to learn the spatial heterogeneity of abnormal rainfall data, and perform layer normalization on the input feature map P5; divide the normalized feature map into local windows of K×K, and each window contains M×M spatial positions, where Perform the following operations: (i) Flatten the features within the window into a sequence of dimension (M 2 , C); (ii) Generate query matrix Q, key matrix K, and value matrix V through learnable linear projections; (iii) Calculate the attention weights with relative position biases; where d k = C / num_heads, num_heads is the number of heads, the input features are split into multiple heads, and each head performs attention calculation independently. B is a learnable bias matrix based on the relative positions of pixels within the window; (iv) Concatenate the multi-head attention results and restore the dimension through linear projection; Then perform a residual connection, that is, add the window attention output and the original feature map P5 element by element; perform secondary normalization on the residual result and input it into a multi-layer perceptron unit, which includes two fully connected layers and a GELU activation function: MLP(x) = W2(GELU(W1x + B)) + B W1 ∈ R {c×4c} ,W2 ∈ R {4c×c} are both learnable parameters, x is the input feature value, and the output feature image P6 is obtained, whose height, width, and number of channels are the same as those of P5. where erf(·) is the error function. The skip connection and the decoder, that is, restore the training data to the original scale, and share the encoder learning parameters through the skip connection to realize the training of the autoencoder network. Its structure is successively input, the transposed convolutional layer Conv2Dtranspose, and the convolutional module Conv2Dblock. Among them, Conv2Dtranspose performs the upsampling operation. Here, S is the stride, that is, the pixel value interval for the convolutional kernel to move, which is 2 here, and p is the pixel value for padding to ensure that the size is restored to twice the current size. The output result is connected in a skip connection with the encoder layer, and the Concatenate layer is fused, that is, the two feature maps are stacked together in the channel number dimension to double the channel number, and then input into the convolutional module Conv2Dblock. Repeat the above operations until the original size of the data is finally restored.

3. The method for identifying and monitoring abnormal rainfall events and analyzing weather behaviors in a river basin based on deep learning according to claim 1, wherein, The self-organizing mapping machine learning method divides the abnormal precipitation events into different subspace patterns, including: Input the time data x corresponding to the abnormal rainfall time series into the self-organizing mapping machine learning method, and its formula is: where wi is the weight vector of the i-th neuron and n is the dimension of the data; the neurons of the SOM are randomly initialized, and each input value x j , and the corresponding weight vector w ij The network calculates the Euclidean distance between it and the weight vectors of each neuron to find the node most similar to the input vector; The weight update formula is: w i (t + 1) = w i( t) + α(t).h ci (t).(x - w i (t)) where, w i (t) is the weight vector to be updated for the i-th neuron, α(t) is the learning rate, h ci (t) is the neighborhood function, indicating the degree of weight update of neighboring neurons, x is the input data, w i (t + 1) is the weight vector updated for the i-th neuron; Until the weight vectors of the network tend to be stable, output different subspace patterns of abnormal rainfall events for abnormal precipitation events.

4. The method for identifying and monitoring abnormal rainfall events and analyzing weather behaviors in a river basin based on deep learning according to claim 1, wherein The evaluation of the influence of different meteorological factors on extreme weather conditions through the Bayesian method and the quantitative exploration of the changing factors of abnormal rainfall events include: Qualitatively analyze the spatio-temporal evolution law of the water vapor transport characteristics and the atmospheric circulation configuration in abnormal precipitation and their interaction mechanism, and determine the meteorological factors with the influence mechanism of basin abnormal rainfall; The Bayesian test method is that the probability p of abnormal rainfall in the subspace mode is the ratio of the abnormal rainfall event to the total observed time series rainfall event; By assuming the prior distribution of α, calculate the posterior distribution of β from another distribution with the same different parameters, α(s + 1) and β(n + 1), where the number of abnormal rainfall events (s) and the date of the observed time series (n) come from a sample of the Bernoulli process. Use s and n as sufficient statistics to analyze the Bernoulli process sample: Among them, s represents the number of abnormal rainfall events under the positive (negative) correlation influence of meteorological factors, n represents the time series rainfall events under the positive (negative) correlation influence of meteorological factors, constant represents the constant term, where Γ(*) represents the gamma function, and B(s + 1, n - s + 1) represents the beta function.

5. A system for identifying and monitoring abnormal rainfall events and analyzing weather behavior in a river basin based on deep learning, characterized in that, Including: A data preprocessing module for obtaining the original summer rainfall data of the basin, preprocessing it, constructing a data set, and dividing the data set into a training set and a validation set; An abnormal rainfall event recognition and clustering module for training and recognizing abnormal rainfall events in the target basin by inputting a deep learning autoencoder network embedded with a multi-head attention mechanism; The self-organizing mapping machine learning method clusters abnormal precipitation events into different subspace modes; A water vapor transport and atmospheric circulation influence mechanism module for qualitatively analyzing the spatio-temporal evolution law of the water vapor transport characteristics and the atmospheric circulation configuration in abnormal precipitation and their interaction mechanism according to each subspace mode; A Bayesian method for analyzing the meteorological factor influence module for evaluating the influence of different meteorological factors on extreme weather conditions by the Bayesian method and quantitatively exploring the factors affecting abnormal rainfall events.

6. The system for identifying and monitoring abnormal rainfall events and analyzing weather behaviors in a river basin based on deep learning according to claim 5, characterized in that, The abnormal rainfall event recognition and clustering module is used to train and recognize abnormal rainfall events in the target basin by inputting a deep learning autoencoder network embedded with a multi-head attention mechanism; The self-organizing mapping machine learning method clusters abnormal precipitation events into different subspace modes; The deep learning autoencoder network embedded with a multi-head attention mechanism mainly includes an encoder, a decoder, a skip connection structure, and a multi-head attention mechanism module: The encoder compresses the training data to realize the feature learning of the network. The original input data size (H, W, C1) represents the height, width, and number of channels. The input convolutional module Conv2Dblock and the Maxpooling2D layer perform four times of downsampling to capture significant local and global representations. Among them, the Conv2D convolution operation is (f * g)(u * v) = ∑ i ∑ j f(i, j)g(u, v), the function f(i, j) is called the pixel value at the position (i, j) in the input function image, g(u, v) is called the filter or convolution kernel, and the Dropout layer randomly sets convolution features to zero to optimize the model; Pooling MaxPooling2D means that in the image matrix, it slides from left to right and from top to bottom according to the 2*2 window size, and takes the maximum value of the pixels in the indexed area, so that the convolutional features H and W are successively compressed to half of the original size; Input the multi-head attention mechanism module to learn the spatial heterogeneity of abnormal rainfall data, and perform layer normalization on the input feature map P5; divide the normalized feature map into local windows of K×K, and each window contains M×M spatial positions, where Perform the following operations: (i) Flatten the features within the window into a sequence of dimension (M 2 , C); (ii) Generate query matrix Q, key matrix K, and value matrix V through learnable linear projections; (iii) Calculate the attention weights with relative position biases; where d k = C / num_heads, num_heads is the number of heads, the input features are split into multiple heads, and each head performs attention calculation independently. B is a learnable bias matrix based on the relative positions of pixels within the window; (iv) Concatenate the results of multi-head attention and restore the dimension through linear projection; Subsequently, a residual connection is performed, that is, the window attention output is added to the original feature map P5 element by element; the residual result is subjected to secondary normalization and input into a multi-layer perceptron unit, which includes two fully connected layers and a GELU activation function: MLP(x) = W2(GELU(W1x + B)) + B W1 ∈ R {c×4c} ,W2 ∈ R {4c×c} Both are learnable parameters, and x is the input feature value. Then the output feature image P6 is obtained, whose height, width, and number of channels are the same as those of P5. where erf(·) is the error function. Skip connection and decoder, that is, the restoration of the training data to the original scale, and the encoder learning parameters are shared through the skip connection to realize the training of the autoencoder network. Its structure is sequentially input, transposed convolutional layer Conv2Dtranspose, and convolutional module Conv2Dblock. Among them, the Conv2Dtranspose performs the upsampling operation, where S is the stride, that is, the pixel value interval by which the convolutional kernel moves, here it is 2, and p is the pixel value for padding to ensure that the size is restored to twice the current size. The output result is connected in a skip connection with the encoder layer and fused by the Concatenate layer, that is, the two feature maps are stacked together in the channel dimension to double the number of channels, and then input into the convolutional module Conv2Dblock. Repeat the above operations until the original size of the data is finally restored.

7. The system for identifying and monitoring abnormal rainfall events and analyzing weather behaviors in a river basin based on deep learning according to claim 5, wherein The abnormal rainfall event recognition and clustering module is used to input and train a deep learning autoencoder network embedded with a multi-head attention mechanism to identify abnormal rainfall events in the target basin. The self-organizing mapping machine learning method divides abnormal precipitation events into different subspace patterns, including: The corresponding time data x of the abnormal rainfall time series is input into the self-organizing mapping machine learning method, and its formula is: where, w i is the weight vector of the i-th neuron, and n is the dimension of the data; the neurons of the SOM are randomly initialized, and for each input value x j , and the corresponding weight vector w ij The network calculates the Euclidean distance between it and the weight vectors of each neuron, and finds the node that is most similar to the input vector; The weight update formula is: w i (t + 1)= w i (t)+α(t).h ci (t).(x - w i (t)) Among them, w i (t) is the weight vector to be updated for the i-th neuron, α(t) is the learning rate, h ci (t) is the neighborhood function, representing the degree of weight update of neighboring neurons, x is the input data, w i (t + 1) is the weight vector updated for the i-th neuron; Until the weight vectors of the network tend to be stable, different abnormal rainfall event subspace patterns of abnormal precipitation events are output.

8. The system for identifying and monitoring abnormal rainfall events and analyzing weather behavior in a river basin based on deep learning according to claim 5, wherein The Bayesian method evaluation module is used to evaluate the influence of different meteorological factors on extreme weather conditions through the Bayesian method and quantitatively explore the changing factors of abnormal rainfall events; including: Qualitatively analyze the spatio-temporal evolution law and interaction mechanism of the water vapor transport characteristics and atmospheric circulation configuration in abnormal rainfall, and determine the meteorological factors with the influence mechanism of basin abnormal rainfall; the Bayesian test method is that the abnormal rainfall probability p of the subspace pattern is the ratio of the abnormal rainfall event to the total observed time series rainfall event; by assuming the α prior distribution, the β posterior distribution is calculated from another different parameter with the same distribution, α(s + 1) and β(n + 1), the number of abnormal rainfall events (s) and the date of the observed time series (n) come from a Bernoulli process sample, and s and n are used as sufficient statistics to analyze the Bernoulli process sample: Among them, s represents the number of abnormal rainfall events under the positive (negative) correlation influence of meteorological factors, n represents the time series rainfall events under the positive (negative) correlation influence of meteorological factors, constant represents a constant term, where Γ(*) represents the gamma function, and B(s + 1, n - s + 1) represents the beta function.

9. An electronic device, comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it realizes the method for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning according to any one of claims 1-4.

10. A storage medium stores a computer program, characterized in that, The computer program is designed to realize the method for identifying and monitoring abnormal rainfall events in a basin and analyzing weather behavior based on deep learning according to any one of claims 1 to 4 when running.

Images