Typhoon intensity forecasting method, device and equipment

Abstract

The present invention relates to the field of weather forecast technology, and specifically to a typhoon intensity forecasting method, device, and equipment. The method obtains forecast data of various meteorological elements output by an AI meteorological large model for existing typhoons, obtains its observation data, and obtains a screening area for forecast factors; within the area, the meteorological elements are screened to obtain forecast factors, and a forecast factor library is constructed; the forecast factor data and the observed intensity data are preprocessed to obtain a data set; a convolutional and long-short-term memory neural network optimization model is constructed, and a self-attention mechanism is introduced; the model is trained using the data set; the output of the AI meteorological large model is processed using the trained model to generate a target typhoon intensity forecast result. The technical solution of the present invention can effectively improve the forecast of typhoon intensity by the meteorological large model, accurately and efficiently, performs targeted optimization on typhoon intensity, effectively grasps the time-series variation characteristics of typhoon intensity, and improves the meteorological large model's ability to forecast typhoon intensity.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and in particular to a typhoon intensity forecasting method, device and equipment. Background Art

[0002] Typhoons, as extremely destructive weather systems formed in tropical oceans, can trigger strong winds, heavy rains, and storm surges when they make landfall, often causing significant casualties and property losses in coastal areas. Against the backdrop of global warming, the frequency and intensity of typhoons are becoming increasingly correlated with climate change and extreme weather events, making typhoon-related research increasingly important. The Northwest Pacific is the most active region for tropical cyclone formation in the world, contributing more than one-third of high-intensity typhoons. Accurately predicting typhoon intensity in this region, especially during the rapid intensification phase, has become a core requirement for disaster prevention and mitigation in coastal cities.

[0003] Currently, typhoon intensity prediction methods are mainly divided into three categories: numerical dynamic models, statistical regression models, and deep learning models. Numerical dynamic models are based on partial differential equations of atmospheric motion and have certain advantages in forecasting large-scale and mesoscale weather systems. However, the rapid intensification of typhoons involves complex nonlinear interactions between environmental variables. The chaotic nature of the atmosphere makes it difficult for numerical models to fully capture these nonlinear characteristics, limiting prediction accuracy. Statistical regression models establish predictive relationships by correlating the motion trajectory of the typhoon center with historical characteristics. Although they have the advantages of high computational efficiency and low energy consumption, they generally ignore the modeling of physical processes and are particularly difficult to characterize the complex dynamic mechanisms of atmospheric interactions.

[0004] In recent years, deep learning technology, which trains AI models based on atmospheric dynamics data, has demonstrated potential to surpass traditional methods in capturing nonlinear correlations between environmental variables, achieving significantly better forecasting accuracy than numerical dynamics and statistical regression models. For example, recurrent neural networks (RNNs) are being used to automatically extract knowledge about typhoon structural domains; three-dimensional convolutional neural networks (3D-CNNs) are being used to integrate three-dimensional atmospheric and oceanic structural features for intensity forecasting; and the spatiotemporal attention fusion network (SAF-Net) is being designed to capture both two-dimensional and three-dimensional typhoon structural information. Furthermore, with the accumulation of AI technology and meteorological big data, data-driven weather modeling has become a trend, with several advanced models achieving initial success. AI models are not only used for forecasting but also support typhoon predictability research through their efficient self-optimization capabilities.

[0005] However, typhoon intensity is affected by complex multi-scale environmental factors, and existing AI models still face the following challenges: First, although the accuracy of typhoon path prediction has been improved through innovations in three-dimensional deep neural network architecture, its intensity prediction performance still lags behind traditional numerical models; second, the chaotic characteristics of the atmosphere and the coupling effects of multiple physical fields during the rapid intensification of typhoons lead to insufficient model generalization capabilities; third, existing methods do not have sufficient depth in exploring the dynamic correlations of key forecasting factors, making it difficult to quantify their contribution to intensity mutations. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a typhoon intensity forecasting method, device and equipment to solve the problem that the AI model in the existing technology still has shortcomings when facing typhoon prediction.

[0007] According to a first aspect of an embodiment of the present invention, a typhoon intensity forecasting method is provided, comprising:

[0008] Obtain forecast data on various meteorological elements output by the AI meteorological model for existing typhoons, obtain the observed intensity data and observed location data of existing typhoons, calculate the typhoon center location based on the observed location data, and use the preset area around the typhoon center as the meteorological model forecast factor screening area;

[0009] In the screening area, the forecast data of various meteorological elements that affect typhoon intensity are screened to obtain forecast factors, a characteristic vector corresponding to each forecast factor is generated, and a forecast factor library is generated based on the characteristic vectors; the forecast factor data and the observed intensity data are preprocessed to obtain a meteorological element data set;

[0010] Constructing a convolutional and long short-term memory neural network optimization model, introducing a self-attention mechanism for preset key variables in the model; and training the model using the meteorological element dataset;

[0011] Obtain forecast data on various meteorological elements of the target typhoon output by the AI meteorological model, input it into the trained convolutional and long short-term memory neural network optimization models, and generate the target typhoon intensity forecast results.

[0012] Preferably, the constructed convolutional and long short-term memory neural network optimization model includes: an input layer, a key variable self-attention layer, a convolutional layer, a long short-term memory network layer, a feature fusion layer, and an output layer;

[0013] The input layer is used to receive the forecast factors and observation intensity data in the meteorological element dataset;

[0014] The key variable self-attention layer is used to enhance the features of the preset key variables in the predictor. The weights of the preset key variables are calculated through the self-attention mechanism and weighted fused with the original features.

[0015] Use convolutional layers to extract local spatiotemporal features of input data, including one-dimensional convolution operations, batch normalization, activation functions, and maximum pooling operations to reduce data dimensions and enhance feature expression capabilities;

[0016] The long short-term memory network layer is used to capture the temporal dependency of the input data, the bidirectional long short-term memory network is used to extract the temporal features, and the final temporal feature representation is output;

[0017] The feature fusion layer is used to fuse the features extracted by the convolutional layer and the long short-term memory network layer, and output the fused feature representation;

[0018] The output layer is used to map the fused feature representation output by the feature fusion layer to the final predicted value as the typhoon intensity prediction value.

[0019] Preferably, when training the model, at least the following steps are included:

[0020] The sliding time window test method is used to enrich the length of the training data set; the model generalization performance and model robustness are improved through the regularization process; the Adam optimizer is used to backpropagate the model to update the model parameters.

[0021] Preferably, when the model is trained using the meteorological element dataset, the method further includes:

[0022] The meteorological element dataset is divided into a training set and a validation set according to a preset ratio;

[0023] The model is trained using the training set. After each training cycle, the model prediction performance is evaluated using the validation set to calculate the validation loss.

[0024] If the validation loss does not decrease within the preset training period, the model training is stopped and the optimal model parameters are saved.

[0025] Preferably, after the model training is completed, the following steps are further included:

[0026] Generate predictions of existing typhoons using trained convolutional and long short-term memory neural network optimization models, and calculate the mean absolute error, root mean square error, and anomaly correlation coefficient between the predictions and observed values of existing typhoons;

[0027] The model performance was comprehensively evaluated using mean absolute error, root mean square error, and anomaly correlation coefficient;

[0028] Based on the evaluation results, the model parameters are optimized and adjusted.

[0029] Preferably, calculating the typhoon center location based on the observed location data includes:

[0030] Determine the a priori grid points based on the typhoon's longitude and latitude in the observation location data, and limit the longitude and latitude range of the typhoon center based on the a priori grid points;

[0031] Calculate the distance between the grid point and the estimated center; calculate the relative vorticity based on the grid point wind speed field;

[0032] For grid points within a radius of 200 km from the estimated center that meet the distance requirement of less than 200 km, the vorticity-weighted average coordinates are calculated as the typhoon center positioning result.

[0033] Preferably, the predictive factors obtained by screening include:

[0034] Typhoon fullness, maximum wind speed at the typhoon center, minimum air pressure at the typhoon center, sea surface temperature, vertical wind shear, 200hPa average divergence and temperature, and 6-hour speed change at the typhoon center.

[0035] Preferably, preprocessing the predictor data and the observed intensity data comprises:

[0036] Performing data cleaning on the predictor data and the observed intensity data;

[0037] The cleaned data is standardized using the following formula:

[0038]

[0039] in, is the original data matrix of specific meteorological elements, for The average value of for The standard deviation of is the standardized meteorological element data matrix.

[0040] According to a second aspect of an embodiment of the present invention, a typhoon intensity forecasting device is provided, comprising:

[0041] The regional delineation module is used to obtain the forecast data of various meteorological elements output by the AI meteorological large model for existing typhoons, obtain the observed intensity data and observed location data of existing typhoons, calculate the typhoon center location based on the observed location data, and use the preset area around the typhoon center as the meteorological model forecast factor screening area;

[0042] A meteorological element processing module is configured to screen the forecast data of various meteorological elements that affect typhoon intensity in the screening area to obtain forecast factors, generate a eigenvector corresponding to each forecast factor, and generate a forecast factor library based on the eigenvectors; and pre-process the forecast factor data and the observed intensity data to obtain a meteorological element dataset;

[0043] A model building and training module is used to build convolutional and long short-term memory neural network optimization models, introduce a self-attention mechanism to preset key variables in the model, and train the model using the meteorological element dataset;

[0044] The typhoon intensity forecast module is used to obtain the forecast data of various meteorological elements of the target typhoon output by the AI meteorological model, input it into the trained convolution and long short-term memory neural network optimization model, and generate the target typhoon intensity forecast result.

[0045] According to a third aspect of an embodiment of the present invention, a typhoon intensity forecasting device is provided, comprising:

[0046] A main controller, and a memory connected to the main controller;

[0047] The memory stores program instructions;

[0048] The main controller is used to execute program instructions stored in the memory and perform any of the above methods.

[0049] The technical solutions provided by the embodiments of the present invention may have the following beneficial effects:

[0050] It is understood that the technical solution presented in this invention, in its first aspect, can obtain forecast data for various meteorological elements output by the AI meteorological large-scale model for existing typhoons, obtain observation data from these elements, and derive a screening area for forecast factors. Within this area, meteorological elements are screened to obtain forecast factors, and a forecast factor library is constructed. This results in more comprehensive data coverage, more thorough data mining, and higher forecast accuracy.

[0051] Secondly, a convolutional and long short-term memory neural network optimization model was constructed, and a self-attention mechanism was introduced to the preset key variables in the model to better capture the temporal variation characteristics of typhoon intensity, thereby improving the ability of large meteorological models to forecast typhoon intensity and having extremely strong application value.

[0052] Thirdly, the constructed model introduces a convolutional neural network to extract the local spatiotemporal feature optimization model of the input data, which reduces the data dimension and enhances the feature expression ability, helps to solve the gradient vanishing problem in deep neural networks, allows for the effective construction of deeper networks, and allows gradients to be back-propagated more directly to shallower layers, alleviating the gradient vanishing problem and reducing the overfitting problem, while improving the model's signal extraction ability, feature learning ability, and robustness.

[0053] Through the above three aspects, the convolution and long short-term memory neural network optimization model constructed in this application can effectively improve the forecast of typhoon intensity by the large meteorological model, accurately and efficiently, carry out targeted optimization of typhoon intensity, effectively grasp the time series change characteristics of typhoon intensity, and improve the ability of the large meteorological model to forecast typhoon intensity.

[0054] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0056] Figure 1 is a schematic diagram showing the steps of a typhoon intensity forecasting method according to an exemplary embodiment;

[0057] Figure 2 is a schematic diagram of a forecast area screening and key factor dynamic enhancement mechanism according to an exemplary embodiment;

[0058] Figure 3 It is a structural diagram of a neural network based on self-attention convolution and long short-term memory according to an exemplary embodiment. DETAILED DESCRIPTION

[0059] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0060] In one embodiment, Figure 1 The present invention is a schematic diagram showing the steps of a typhoon intensity forecasting method according to an exemplary embodiment, which provides a typhoon intensity forecasting method, including:

[0061] Step S11: Obtain forecast data of various meteorological elements output by the AI meteorological large model for the existing typhoon, obtain the observed intensity data and observed position data of the existing typhoon, calculate the typhoon center position based on the observed position data, and use the preset area around the typhoon center as the meteorological model forecast factor screening area.

[0062] Step S12: In the screening area, the forecast data of various meteorological elements that affect the typhoon intensity are screened to obtain forecast factors, a characteristic vector corresponding to each forecast factor is generated, and a forecast factor library is generated based on the characteristic vector; the forecast factor data and the observed intensity data are preprocessed to obtain a meteorological element data set.

[0063] Step S13: construct a convolutional and long short-term memory neural network optimization model, introduce a self-attention mechanism for preset key variables in the model; and train the model using the meteorological element dataset.

[0064] Step S14: Obtain forecast data of various meteorological elements of the target typhoon output by the AI meteorological large model, input it into the trained convolution and long short-term memory neural network optimization model, and generate the target typhoon intensity forecast result.

[0065] It is understood that the technical solution presented in this invention, in its first aspect, can obtain forecast data for various meteorological elements output by the AI meteorological large-scale model for existing typhoons, obtain observation data from these elements, and derive a screening area for forecast factors. Within this area, meteorological elements are screened to obtain forecast factors, and a forecast factor library is constructed. This results in more comprehensive data coverage, more thorough data mining, and higher forecast accuracy.

[0066] Secondly, a convolutional and long short-term memory neural network optimization model was constructed, and a self-attention mechanism was introduced to the preset key variables in the model to better capture the temporal variation characteristics of typhoon intensity, thereby improving the ability of large meteorological models to forecast typhoon intensity and having extremely strong application value.

[0067] Thirdly, the constructed model introduces a convolutional neural network to extract the local spatiotemporal feature optimization model of the input data, which reduces the data dimension and enhances the feature expression ability, helps to solve the gradient vanishing problem in deep neural networks, allows for the effective construction of deeper networks, and allows gradients to be back-propagated more directly to shallower layers, alleviating the gradient vanishing problem and reducing the overfitting problem, while improving the model's signal extraction ability, feature learning ability, and robustness.

[0068] Through the above three aspects, the convolution and long short-term memory neural network optimization model constructed in this application can effectively improve the forecast of typhoon intensity by the large meteorological model, accurately and efficiently, carry out targeted optimization of typhoon intensity, effectively grasp the time series change characteristics of typhoon intensity, and improve the ability of the large meteorological model to forecast typhoon intensity.

[0069] In specific practice, step S11: forecast data of various meteorological elements is gridded meteorological elements that affect the occurrence and development of typhoons, mainly divided into humidity, temperature, horizontal wind speed, potential height of the upper air layer (such as 850 hPa, 500 hPa, 200 hPa) and 10m longitudinal wind and 10m zonal wind of the near-ground layer, sea surface temperature and lowest sea level pressure.

[0070] Obtaining the observed intensity and location data of existing typhoons is mainly to obtain the observed data of existing typhoons in the tropical cyclone optimal track dataset corresponding to the output of the large meteorological model, including: the longitude and latitude of the typhoon center and other data representing the typhoon's location, as well as physical quantities representing the typhoon's intensity. Among them, the meteorological elements use the area of ±5° around the typhoon center as the meteorological model forecast factor screening area. Forecast area screening, see Figure 2 .

[0071] It should be noted that in typhoon intensity prediction research, accurate determination of the typhoon center is the core foundation for building a prediction model. This embodiment uses the potential vorticity field weighted centroid method to determine the typhoon center. The main steps include:

[0072] Step S21: Determine the priori grid points ( , ), and the longitude and latitude range of the typhoon center is limited according to the prior grid points.

[0073] The limited latitude and longitude range of the typhoon center is ( , ).

[0074] Step S22: Calculate the distance between the grid point and the estimated center; and calculate the relative vorticity based on the grid point wind speed field.

[0075] In the calculation grid ( , ) and the estimated center (km), the formula is as follows:

[0076]

[0077]

[0078] in, , , is the grid point position.

[0079] When calculating relative vorticity based on the grid wind speed field, the grid wind speed fields U and V in the screening area are used to accurately characterize the rotation characteristics of the tropical cyclone and calculate the relative vorticity. :

[0080]

[0081] in and are local Cartesian coordinates.

[0082] Step S23: For the estimated center radius of 200km, the distance The grid points (km) are smaller than 200 km (to make the positioning results closer to the core of the physical field of the typhoon eye area), and the vorticity-weighted average coordinates are calculated as the typhoon center positioning results.

[0083] The formula for calculating the vorticity weighted mean coordinate is as follows:

[0084]

[0085]

[0086] Final output ( , ) is the typhoon center positioning result, is the grid point position.

[0087] In practice, step S12 involves screening meteorological variables that influence typhoon intensity (particularly the rapid intensification process) within the prediction factor screening area determined in step S11, based on the key dynamic and thermal mechanisms of the typhoon's rapid intensification process. It should be noted that the resulting prediction factors include: typhoon fullness, maximum wind speed at the typhoon center, minimum pressure at the typhoon center, sea surface temperature, vertical wind shear, 200hPa average divergence and temperature, and 6-hour speed variation at the typhoon center. For each of these variables, corresponding eigenvectors are generated, thereby constructing a dimensionless prediction factor library. The following is the specific calculation method for each variable.

[0088] Regarding Tropical Cyclone Fullness (TCF): TCF is an important indicator for assessing the structural integrity of a typhoon. It can directly reflect the morphological characteristics of a typhoon, especially the distribution range of its maximum wind speed. The structural integrity of a typhoon plays a vital role in its intensity development. A structurally complete typhoon has a more uniform wind speed distribution and more concentrated energy. These characteristics create favorable conditions for the rapid increase in typhoon intensity. The calculation formula for typhoon fullness is as follows:

[0089]

[0090] in, is the typhoon's maximum wind speed radius, is the preset wind speed threshold radius. When the TCF value is greater than 0.8, it indicates that the typhoon has the structural integrity required for high-intensity development.

[0091] Regarding sea surface temperature: It is a core indicator for measuring typhoon energy supply. To accurately obtain the sea surface temperature conditions within the typhoon path area, this embodiment uses an interpolation method to match the sea surface temperature grid points with the meteorological model output forecast factor screening area, thereby dynamically updating the sea surface temperature (SST) spatial distribution field within the typhoon path area and further deriving the average sea surface temperature field for the corresponding screening area. The specific implementation method of interpolation is as follows:

[0092] When the grid point spacing is greater than or equal to the horizontal resolution of the grid reanalysis data, the nearest neighbor method is used for interpolation; when the grid point spacing is less than the horizontal resolution of the grid reanalysis data, the bilinear interpolation method is used. The bilinear interpolation formula is as follows:

[0093]

[0094] in Z ij Location The location of the grid points is , and the reanalysis data of the four adjacent grid points are 、 、 、 , The location is , The location is , The location is , The location is , to eliminate the missing effects of surface ocean parameters in meteorological model outputs, and are the grid point positions respectively.

[0095] The following is the calculation method of the dynamic environment factor:

[0096] Maximum wind speed in the center of a typhoon: the maximum value of the average wind speed within a radius of 200km with the center of the typhoon as the origin.

[0097] Typhoon center minimum pressure: the minimum sea level pressure at the grid point where the typhoon center is located.

[0098] Vertical wind shear: zonal wind based on the 850 hPa and 200 hPa isobaric surfaces ( u ) and meridional wind ( v ) Difference calculation:

[0099]

[0100] in, , , and represent the average zonal wind speed at 200 hPa and 850 hPa, and the average meridional wind speed at 200 hPa and 850 hPa, respectively.

[0101] 200hPa average divergence: The average divergence of the 200hPa isobaric surface within 200 km of the typhoon center:

[0102]

[0103] in, represents the 200 hPa average zonal wind speed, It represents the average meridional wind speed at 200 hPa.

[0104] 200hPa average temperature: the average temperature of the 200 hPa isobaric surface within 200 km of the typhoon center.

[0105] Typhoon center 6-hour speed change: the rate of change of the maximum wind speed in the typhoon center in the past 6 hours.

[0106] It should be noted that the preprocessing of the prediction factor data and the observed intensity data includes:

[0107] The forecast factor data and the observed intensity data are cleaned, specifically by identifying and removing incomplete, incorrect, inaccurate or irrelevant parts of the data.

[0108] The cleaned data is standardized using the following formula:

[0109]

[0110] in, is the original data matrix of specific meteorological elements, for The average value of for The standard deviation of is the standardized meteorological element data matrix.

[0111] In practice, in step S13, a convolutional and long short-term memory neural network optimization model is constructed, including: input layer, key variable self-attention layer, convolution layer, long short-term memory network layer, feature fusion layer and output layer. For the structure diagram of self-attention convolutional and long short-term memory neural network, please refer to Figure 3 .

[0112] The input layer is used to receive forecast factors and observation intensity data from the meteorological element dataset.

[0113] The input layer receives the meteorological element forecast factor tensor obtained in step S12 , and typhoon observation intensity series , Where is the time series length, The input layer performs dimension alignment and splicing operations on multi-source data to generate initial fusion features. .

[0114] The key variable self-attention layer is used to enhance the features of the preset key variables in the predictor. The weights of the preset key variables are calculated through the self-attention mechanism and weighted fused with the original features.

[0115] Dynamic enhancement mechanism of key factors Figure 2 The key variable attention layer adopts the temporal self-attention mechanism module, selecting vertical wind shear, sea surface temperature, typhoon fullness and sea surface pressure. The meteorological factors that are significantly correlated with typhoon intensity are dynamically screened and enhanced. The input feature tensor is set to ,in is the batch size, is the time series length, is the feature dimension. The key factor enhancement weight is calculated through the temporal attention mechanism :

[0116]

[0117] Extract key factors from input features, the shape is [B, T, ], use Linear to set the time series long graph T to 16, extract time series features, use the hyperbolic tangent activation function Tanh to introduce nonlinearity, and enhance the model feature expression ability. The specific formula is:

[0118]

[0119] in, is the base of natural logarithms, is the element in T after the linear transformation.

[0120] Map feature dimensions back to key factor quantities , while generating normalized weights for the key factors at each time step .

[0121] For global gating weights :

[0122]

[0123] Extract feature factors from input features, including all features, with a shape of [B, T, d]. Set the time series long graph T to 64 to extract global features. Use the hyperbolic tangent activation function Tanh to introduce nonlinearity to enhance the model's feature expression capabilities. Generate global gating weights, , used to scale the full feature matrix, where is the Sigmoid activation function.

[0124] At the same time, key factors are enhanced :

[0125]

[0126] Key factors extracted from input features , the shape is [B, T, d], where B is the batch size, T is the number of time series, and d is the feature dimension. Indicates the key factor enhancement weight, generated by the temporal attention mechanism, using 1+ Weight the key factors, enhance their importance, and output the enhanced key factors , Represents element-wise multiplication by channel.

[0127] The key factors after replacement Replace back to the original feature matrix In, get , perform feature replacement, use Scale the full feature matrix to retain key information and suppress noise features. , used for subsequent spatiotemporal feature extraction and prediction.

[0128]

[0129] The convolutional layer is used to extract the local spatiotemporal features of the input data, including one-dimensional convolution operation, batch normalization, activation function and maximum pooling operation to reduce the data dimension and enhance the feature expression ability.

[0130] The convolutional layer (spatial-temporal feature extraction layer) is used to capture the local spatial patterns and multi-scale feature interactions of meteorological factors.

[0131] It uses a deep convolution module and a bidirectional LSTM module, where:

[0132] The convolutional layer unit contains two sets of 1D convolutional layers with a convolution kernel size of 5. The module input is the enhanced feature matrix , the output is The specific process is as follows:

[0133] Each convolution unit contains the following operations:

[0134]

[0135] The convolution kernel size is 5, the padding is 2, and the time series length T is kept unchanged. The output .

[0136] Perform batch normalization (BatchNorm):

[0137]

[0138] Normalize the channel dimension of the convolution output to accelerate training convergence.

[0139] Use Gaussian Error Linear Unit (GELU) The activation function introduces nonlinearity and enhances the feature expression capability.

[0140]

[0141] Perform maximum pooling with a pooling kernel size of 2 and a step size of 2, halving the time series length, and the output is . And randomly take Dropout for neurons , that is, randomly shielding 40% of the neurons to prevent overfitting. The dropout layer randomly "discards" the output results of the neurons in the fully connected layer according to a certain probability. That is, only the output results of some neurons in the fully connected layer are considered during training. The introduction of the dropout layer helps to make the network easier to train and better extract features.

[0142]

[0143] in, Represents convolution operation, BN Represents a batch normalization operation, using Perform global average pooling on the output of the convolutional layer to reduce the dimension and extract global information.

[0144] The LSTM module uses a bidirectional long short-term memory network (Bi-directional LSTM) as the core module, which captures the front-end dependency of time series data through a bidirectional gating mechanism. First, the convolution output Transpose the time series to the time series input format To adapt to LSTM input. The bidirectional temporal dependency model is:

[0145]

[0146] is the transposed convolution output, Represents a bidirectional LSTM network, which contains two LSTM layers, one processing the forward sequence and the other processing the reverse sequence. and are the hidden state and the cell state respectively. The hidden state update formula is:

[0147]

[0148] is the hidden state of the previous time step, Indicates that at time step t The input data, and are the weight matrices from input to hidden and hidden to hidden, respectively.

[0149] Output of Bidirectional LSTM , where hidden_size is the number of hidden units in the LSTM, 2 represents the concatenation of bidirectional hidden states, and the hidden units are composed of multiple fully connected layers. The input layer integrates information from the simulation results and passes it through the fully connected layers, which helps effectively extract important feature information from high dimensions and capture complex nonlinear relationships in the data. The output of each fully connected layer is calculated as follows:

[0150]

[0151] in, is the weight matrix, is the input vector, is the bias vector, is the output vector, Represents the activation function (Leaky Rectified Linear Unit).

[0152] The long short-term memory network layer is used to capture the temporal dependency of the input data, the bidirectional long short-term memory network is used to extract the temporal features, and the final temporal feature representation is output.

[0153] The feature fusion layer is used to fuse the features extracted by the convolutional layer and the long short-term memory network layer, and output the fused feature representation.

[0154] Hidden units , through linear transformation , normalization , activation function and random dropout , the spatiotemporal features are fused as the output of the dynamic feature fusion layer.

[0155] The output layer is used to map the fused feature representation output by the feature fusion layer to the final predicted value as the typhoon intensity prediction value.

[0156] Take the feature fusion layer as input and transform it linearly Map it to the final predicted value as the output typhoon intensity prediction value. The calculation process is:

[0157]

[0158] pass Map 128 dimensions to 1 dimension, output .

[0159] It should be noted that when training the model, it at least includes: using a sliding time window test method to enrich the length of the training data set, with a window length of 24 observation hours and a step length of 6 hours.

[0160] To prevent overfitting, the model generalization performance and robustness are improved through regularization. The model is back-propagated using the Adam (Adaptive Moment Estimation) optimizer to update the model parameters.

[0161] Its parameter update rule is:

[0162]

[0163] in and Respectively represent the parameter sets before and after the model parameter update, 、 They are the first-order moment and second-order moment estimates after bias correction, and the parameters are added Avoid gradients of 0.

[0164] In a specific embodiment, when the model is trained using the meteorological element dataset, the method further includes:

[0165] The meteorological element dataset is divided into a training set and a validation set according to a preset ratio; the model is trained using the training set, and after each training cycle, the model prediction performance is evaluated using the validation set to calculate the validation loss; if the validation loss does not decrease within the preset training cycles, the model training is stopped and the optimal model parameters are saved.

[0166] In practice, gradient clipping (threshold 1.0) is used to prevent gradient explosion. After training the model on the training set, the prediction performance on the validation set is evaluated and the model parameters are adjusted in a timely manner. The specific steps are to evaluate the prediction performance on the validation set after every 6 rounds of training and calculate the loss function:

[0167]

[0168] in represents the loss of the validation set, N is the number of samples in the validation set, Indicates the model i The predicted value of the sample, Indicates the i The true value of the samples.

[0169] Dynamically control the training cycle. If the validation loss does not decrease for 10 consecutive epochs, terminate the training and save the optimal model parameters to ensure the model is efficient and accurate.

[0170] It should be noted that after the model training is completed, it also includes:

[0171] Use the trained convolutional and long short-term memory neural network optimization models to generate predictions of existing typhoons, and calculate the mean absolute error, root mean square error, and anomaly correlation coefficient between the predicted values and the observed values of existing typhoons; use the mean absolute error, root mean square error, and anomaly correlation coefficient to comprehensively evaluate the model performance; and optimize and adjust the model parameters based on the evaluation results.

[0172] Compute the mean absolute error between the predicted and observed values:

[0173]

[0174] in is the total number of prediction times, For the The predicted typhoon intensity value of the time period, For the The observed typhoon intensity value for a certain period of time.

[0175] Compute the root mean square error between the predicted and observed values:

[0176]

[0177] The parameter definition is the same as above.

[0178] Compute the unusual correlation coefficient between predicted and observed values:

[0179]

[0180] in, is the time series average of the predicted typhoon intensity values, is the time series average of the observed typhoon intensity values. Based on the calculation results of the above indicators, a comprehensive evaluation of the model performance is conducted and the model parameters are optimized and adjusted according to the evaluation results.

[0181] In practice, when executing step S14, the convolutional and long short-term memory neural network optimization models trained in step S13 are used to select a target typhoon. Based on the various meteorological element forecast data output by the AI meteorological model, the typhoon intensity forecast results for the target area are generated, as follows:

[0182] The AI meteorological model outputs various meteorological element forecast data within the target time. Substitute it into the convolution and long short-term memory neural network optimization model trained in step S13 to generate a typhoon intensity forecast result feature vector; use the standardized inverse process to generate the target typhoon intensity forecast result.

[0183] In another embodiment, a typhoon intensity forecasting device is provided, comprising:

[0184] The regional delineation module is used to obtain the forecast data of various meteorological elements output by the AI meteorological large model for existing typhoons, obtain the observed intensity data and observed location data of existing typhoons, calculate the typhoon center location based on the observed location data, and use the preset area around the typhoon center as the meteorological model forecast factor screening area;

[0185] A meteorological element processing module is configured to screen the forecast data of various meteorological elements that affect typhoon intensity in the screening area to obtain forecast factors, generate a eigenvector corresponding to each forecast factor, and generate a forecast factor library based on the eigenvectors; and pre-process the forecast factor data and the observed intensity data to obtain a meteorological element dataset;

[0186] A model building and training module is used to build convolutional and long short-term memory neural network optimization models, introduce a self-attention mechanism to preset key variables in the model, and train the model using the meteorological element dataset;

[0187] The typhoon intensity forecast module is used to obtain the forecast data of various meteorological elements of the target typhoon output by the AI meteorological model, input it into the trained convolution and long short-term memory neural network optimization model, and generate the target typhoon intensity forecast result.

[0188] In another embodiment, a typhoon intensity forecasting device is provided, comprising:

[0189] A main controller, and a memory connected to the main controller;

[0190] The memory stores program instructions;

[0191] The main controller is used to execute program instructions stored in the memory and perform any of the above methods.

[0192] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0193] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" is at least two.

[0194] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0195] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0196] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and 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 embodiment.

[0197] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0198] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0199] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0200] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A typhoon intensity forecasting method, characterized in that: include: Obtain forecast data on various meteorological elements output by the AI meteorological model for existing typhoons, obtain the observed intensity data and observed location data of existing typhoons, calculate the typhoon center location based on the observed location data, and use the preset area around the typhoon center as the meteorological model forecast factor screening area; In the screening area, the forecast data of various meteorological elements that affect the typhoon intensity are screened to obtain forecast factors, a characteristic vector corresponding to each forecast factor is generated, and a forecast factor library is generated based on the characteristic vectors; Preprocessing the prediction factor data and the observed intensity data to obtain a meteorological element data set; wherein the prediction factors obtained by screening include: typhoon richness, maximum wind speed at the typhoon center, minimum air pressure at the typhoon center, sea surface temperature, vertical wind shear, 200hPa average divergence and temperature, and 6-hour speed change at the typhoon center; Constructing a convolutional and long-short-term memory neural network optimization model, introducing a self-attention mechanism for preset key variables in the model, the convolutional and long-short-term memory neural network optimization model comprising: an input layer, a key variable self-attention layer, a convolutional layer, a long-short-term memory network layer, a feature fusion layer, and an output layer, utilizing the key variable self-attention layer to enhance the features of the preset key variables in the prediction factors, calculating the weights of the preset key variables through the self-attention mechanism, and performing weighted fusion with the original features; and training the model using the meteorological element dataset; Obtain forecast data on various meteorological elements of the target typhoon output by the AI meteorological model, input it into the trained convolutional and long short-term memory neural network optimization models, and generate the target typhoon intensity forecast results.

2. The typhoon intensity forecasting method according to claim 1, characterized in that: Also includes: The input layer is used to receive the forecast factors and observation intensity data in the meteorological element dataset; Use convolutional layers to extract local spatiotemporal features of input data, including one-dimensional convolution operations, batch normalization, activation functions, and maximum pooling operations to reduce data dimensions and enhance feature expression capabilities; The long short-term memory network layer is used to capture the temporal dependency of the input data, the bidirectional long short-term memory network is used to extract the temporal features, and the final temporal feature representation is output; The feature fusion layer is used to fuse the features extracted by the convolutional layer and the long short-term memory network layer, and output the fused feature representation; The output layer is used to map the fused feature representation output by the feature fusion layer to the final predicted value as the typhoon intensity prediction value.

3. The typhoon intensity forecasting method according to claim 1, characterized in that: When training a model, at least: The sliding time window test method is used to enrich the length of the training data set; the model generalization performance and model robustness are improved through the regularization process; the Adam optimizer is used to backpropagate the model to update the model parameters.

4. The typhoon intensity forecasting method according to claim 1, characterized in that: When the model is trained using the meteorological element dataset, the method further includes: The meteorological element dataset is divided into a training set and a validation set according to a preset ratio; The model is trained using the training set. After each training cycle, the model prediction performance is evaluated using the validation set to calculate the validation loss. If the validation loss does not decrease within the preset training period, the model training is stopped and the optimal model parameters are saved.

5. The typhoon intensity forecasting method according to claim 1, characterized in that: After the model training is completed, it also includes: Generate predictions of existing typhoons using trained convolutional and long short-term memory neural network optimization models, and calculate the mean absolute error, root mean square error, and anomaly correlation coefficient between the predictions and observed values of existing typhoons; The model performance was comprehensively evaluated using mean absolute error, root mean square error, and anomaly correlation coefficient; Based on the evaluation results, the model parameters are optimized and adjusted.

6. The typhoon intensity forecasting method according to claim 1, characterized in that: The typhoon center location is calculated based on the observed location data, including: Determine the a priori grid points based on the typhoon's longitude and latitude in the observation location data, and limit the longitude and latitude range of the typhoon center based on the a priori grid points; Calculate the distance between the grid point and the estimated center; calculate the relative vorticity based on the grid point wind speed field; For grid points within a radius of 200 km from the estimated center that meet the distance requirement of less than 200 km, the vorticity-weighted average coordinates are calculated as the typhoon center positioning result.

7. The typhoon intensity forecasting method according to claim 1, characterized in that: Preprocessing the predictor data and the observed intensity data includes: Performing data cleaning on the predictor data and the observed intensity data; The cleaned data is standardized using the following formula: in, is the original data matrix of specific meteorological elements, for The average value of for The standard deviation of is the standardized meteorological element data matrix.

8. A typhoon intensity forecasting device, characterized in that: include: The regional delineation module is used to obtain the forecast data of various meteorological elements output by the AI meteorological large model for existing typhoons, obtain the observed intensity data and observed location data of existing typhoons, calculate the typhoon center location based on the observed location data, and use the preset area around the typhoon center as the meteorological model forecast factor screening area; A meteorological element processing module is used to screen the forecast data of various meteorological elements that affect the typhoon intensity in the screening area to obtain forecast factors, generate a characteristic vector corresponding to each forecast factor, and generate a forecast factor library based on the characteristic vectors; Preprocessing the prediction factor data and the observed intensity data to obtain a meteorological element data set; wherein the prediction factors obtained by screening include: typhoon richness, maximum wind speed at the typhoon center, minimum air pressure at the typhoon center, sea surface temperature, vertical wind shear, 200hPa average divergence and temperature, and 6-hour speed change at the typhoon center; A model construction and training module is used to construct a convolutional and long-short-term memory neural network optimization model, and introduce a self-attention mechanism for preset key variables in the model. The convolutional and long-short-term memory neural network optimization model includes: an input layer, a key variable self-attention layer, a convolution layer, a long-short-term memory network layer, a feature fusion layer, and an output layer. The key variable self-attention layer is used to enhance the features of the preset key variables in the prediction factors. The weights of the preset key variables are calculated through the self-attention mechanism and weightedly fused with the original features. The model is trained using the meteorological element dataset. The typhoon intensity forecast module is used to obtain the forecast data of various meteorological elements of the target typhoon output by the AI meteorological model, input it into the trained convolution and long short-term memory neural network optimization model, and generate the target typhoon intensity forecast result.

9. A typhoon intensity forecasting device, characterized in that: include: A main controller, and a memory connected to the main controller; The memory stores program instructions; The main controller is used to execute program instructions stored in the memory and perform the method according to any one of claims 1 to 7.

Images