Prediction Method for the Formation of Severe Typhoons under the Condition of Imbalanced Data Based on Transfer Learning

Abstract

The present application relates to a method for predicting the formation of severe typhoons under the condition of imbalanced data based on transfer learning. The method includes: First, obtaining the historical tropical cyclone best track data set and the corresponding environmental variable data set, and constructing a balanced data set and an imbalanced data set from the obtained data sets; Then, constructing a pre-learning model for severe typhoons, training and fitting the pre-learning model for severe typhoons using the balanced data set, and transferring the prior knowledge in the trained pre-learning model for severe typhoons to the re-learning model for imbalanced severe typhoon samples, and then training and fitting the re-learning model for imbalanced severe typhoon samples using the imbalanced data set, so as to obtain a trained prediction framework for typhoon formation; Finally, using the trained prediction framework for typhoon formation to predict whether a severe typhoon will form, and obtaining a typhoon formation prediction result. Using this method can improve the prediction accuracy of severe typhoon samples under the condition of imbalanced data.

Description

Technical Field

[0001] This application relates to the cross - technical field of computers and meteorology / oceanography, and particularly to a method for predicting the formation of severe typhoons under unbalanced data conditions based on transfer learning. Background Art

[0002] A typhoon, as a tropical cyclone with a strength greater than 64 kt, is a low - pressure system with a very low central pressure, significant inward airflow convergence in the lower layer, and mainly outward - divergent airflow at the top. Tropical cyclones (including severe typhoons) are high - impact weather phenomena with serious disasters in meteorological oceanography. In recent years, extreme weather and climate events such as severe typhoons have become increasingly frequent, not only causing huge economic losses to coastal areas but also greatly endangering people's lives and property safety.

[0003] To reduce the losses caused by severe typhoon disasters, the most important way is to accurately predict severe typhoons, provide a decision - making basis for governments at all levels, and ensure the effectiveness and timeliness of typhoon prevention and disaster - resistance measures. Therefore, the accurate prediction of severe typhoons is an important research topic faced by major countries in the world. At present, there are mainly three methods for predicting the formation of tropical cyclones: the method based on numerical prediction, the method based on statistical prediction, and the method based on machine learning.

[0004] The method based on numerical prediction is widely used in major typhoon / hurricane warning and forecasting centers around the world and plays a very important role in providing forecast and warning services for society. Some classic forecasting systems are, for example, the Hurricane Weather Research and Forecasting Model HWRF of the National Oceanic and Atmospheric Administration of the United States, the GFS (Global Forecast System) of the National Centers for Environmental Prediction of the United States, and the numerical model IFS of the European Centre for Medium - Range Weather Forecasts, etc. They forecast future weather conditions by directly integrating from the initial state and have high accuracy and interpretability. The second method is the method based on statistical prediction, which relies on numerical weather prediction and establishes a forecasting model considering the changes in the future atmospheric environment and ocean conditions. Common statistical forecasting models include the Statistical Tropical Cyclone Intensity Prediction Scheme (STIPS) and the Statistical Hurricane Intensity Prediction Model (SHIPS) used in the operational operation of the National Hurricane Center of the United States. The third is the method based on machine learning, which can predict tropical cyclones based on a data - driven model without considering physical laws.

[0005] Although existing severe typhoon prediction methods have achieved relatively accurate prediction results for severe typhoons, there are still many deficiencies. Numerical-based methods have drawbacks such as inaccurate vortex initialization, incomplete expression of complex physical processes, relatively coarse model resolution, and high computational costs. Statistical-based methods are often helpless in the face of massive data, struggling to extract key and effective forecast information and getting lost in the messy data. Compared with the first two methods, the forecast model based on machine learning is a pure data-driven model. It can ignore the inaccurate physical mechanisms in the existing meteorological forecasting process and has great advantages in capturing the nonlinear relationship between forecast factors and forecast targets. On the other hand, machine learning-related algorithms can well learn the spatial relationship between variables, greatly improving the classification accuracy of the model and thus increasing the accuracy of model prediction. However, existing methods do not consider the imbalance in severe typhoon observation samples, which makes the unbalanced sample data have a greater impact on the model's prediction, resulting in the prediction accuracy of severe typhoon samples being much lower than that of ordinary typhoon samples. If only the artificial method is used to balance the samples, the model cannot be well applied to the prediction of severe typhoons with actual imbalance phenomena. Therefore, how to provide a method to reduce the impact of unbalanced sample data on severe typhoon prediction at the model level is a technical issue that highly concerns those skilled in the art. Summary of the Invention

[0006] Based on this, in view of the above technical problems, it is necessary to provide a method for predicting the formation of severe typhoons under unbalanced data conditions based on transfer learning, which can reduce the impact of unbalanced sample data on severe typhoon prediction.

[0007] A method for predicting the formation of severe typhoons under unbalanced data conditions based on transfer learning, the method comprising:

[0008] Obtain the official historical tropical cyclone best track dataset, screen the historical tropical cyclone best track dataset to obtain historical tropical cyclone best track data records and their corresponding environmental variable datasets, and construct a balanced dataset for training a severe typhoon pre-learning model and an unbalanced dataset for training an unbalanced severe typhoon sample re-learning model according to the historical tropical cyclone best track data records and environmental variable datasets;

[0009] Construct a prediction framework for severe typhoon formation under imbalanced data conditions. The prediction framework for severe typhoon formation includes a pre-learning model for severe typhoon and a re-learning model for imbalanced severe typhoon samples. Among them, the pre-learning model for severe typhoon includes an atmospheric three-dimensional convolutional module, a sea surface two-dimensional convolutional module, an ocean hydrology three-dimensional convolutional module, a first long short-term memory network module, and a first fully connected layer. The re-learning model for imbalanced severe typhoon samples includes the atmospheric three-dimensional convolutional module, the sea surface two-dimensional convolutional module, and the ocean hydrology three-dimensional convolutional module saved and loaded from the pre-learning model for severe typhoon, as well as a second long short-term memory network module and a second fully connected layer.

[0010] Train the prediction framework for severe typhoon formation according to the balanced dataset and the imbalanced dataset to obtain a trained prediction framework for severe typhoon formation.

[0011] Perform typhoon formation prediction on the prediction data according to the trained prediction framework for severe typhoon formation to obtain the typhoon formation prediction result.

[0012] Among them, the training steps of the prediction framework for severe typhoon formation specifically include:

[0013] Divide the balanced dataset and the imbalanced dataset into a balanced data training set, a balanced data test set, an imbalanced data training set, and an imbalanced data test set according to a certain ratio.

[0014] Use the binary cross-entropy loss function, the balanced data training set, and the balanced data test set to train and evaluate the performance of the pre-learning model for severe typhoon, and select the pre-learning model for severe typhoon with the best evaluation effect on the balanced data test set as the trained pre-learning model for severe typhoon.

[0015] Transfer the prior knowledge in the trained pre-learning model for severe typhoon to the re-learning model for imbalanced severe typhoon samples, and use the pre-constructed loss function for imbalanced severe typhoon samples, the imbalanced data training set, and the imbalanced data test set to train and evaluate the performance of the re-learning model for imbalanced severe typhoon samples, and select the re-learning model for imbalanced severe typhoon samples with the best evaluation effect on the imbalanced data test set as the trained re-learning model for imbalanced severe typhoon samples.

[0016] In one embodiment, obtain the official historical tropical cyclone best track dataset, screen the historical tropical cyclone best track dataset, and obtain the historical tropical cyclone best track data records and their corresponding environmental variable datasets, including:

[0017] Obtain the official historical tropical cyclone best track dataset, divide the historical tropical cyclone best track dataset according to the formation sea area of tropical cyclones, and obtain the historical tropical cyclone best track data records containing different years within each sea area; among them, each historical tropical cyclone best track data record includes the time from the formation to the dissipation of the tropical cyclone, the central longitude and latitude position, the minimum central pressure, and the maximum central wind speed.

[0018] Divide the historical tropical cyclone best track data records into a severe typhoon sample dataset C1 and a common typhoon sample dataset C2 according to the maximum central wind speed, and obtain the corresponding severe typhoon sample environmental variable dataset and common typhoon sample environmental variable dataset according to C1 and C2 and common typhoon sample environmental variable dataset

[0019] In one embodiment, obtain the corresponding severe typhoon sample environmental variable dataset and common typhoon sample environmental variable dataset including:

[0020] Randomly sort the severe typhoon records in C1. For the nth i severe typhoon record, assign the serial number of the severe typhoon record point in n i as According to the time and central longitude and latitude position recorded by the th record point, download the corresponding atmospheric variable data, sea surface variable data, and ocean hydrological variable data from the official weather forecast center website, and package and combine the downloaded atmospheric variable data, sea surface variable data, and ocean hydrological variable data to construct the severe typhoon sample environmental variable dataset as wherein, represents the severe typhoon sample atmospheric variable dataset, represents the severe typhoon sample sea surface variable dataset, represents the severe typhoon sample ocean hydrological dataset;

[0021] Randomly sort the common typhoon records in C2. For the nth j common typhoon record, assign the serial number of the common typhoon record point in n j as According to the time and central longitude and latitude position recorded by the th record point, download the corresponding atmospheric variable data, sea surface variable data, and ocean hydrological variable data from the official weather forecast center website, and package and combine the downloaded atmospheric variable data, sea surface variable data, and ocean hydrological variable data to construct the common typhoon sample environmental variable dataset as wherein, Denote the ordinary typhoon sample atmospheric variable dataset, Denote the ordinary typhoon sample sea surface variable dataset, Denote the ordinary typhoon sample ocean hydrology dataset.

[0022] In one embodiment, a balanced dataset for training a strong typhoon pre-learning model and an unbalanced dataset for training an unbalanced strong typhoon sample re-learning model are constructed according to the historical tropical cyclone best track data records and environmental variable datasets, including:

[0023] According to the strong typhoon sample dataset C1 in the historical tropical cyclone best track data records and its corresponding strong typhoon sample environmental variable dataset Construct the first strong typhoon dataset R for training the strong typhoon pre-learning model T1 and the second strong typhoon dataset R for training the unbalanced strong typhoon sample re-learning model T2 ;

[0024] According to the ordinary typhoon sample dataset C2 in the historical tropical cyclone best track data records and its corresponding ordinary typhoon sample environmental variable dataset Construct the first ordinary typhoon dataset R for training the ordinary typhoon pre-learning model F1 and the second ordinary typhoon dataset R for training the unbalanced ordinary typhoon sample re-learning model F2 ;

[0025] According to R T1 and R F1 Construct the balanced dataset R1 for training the strong typhoon pre-learning model, and according to R T2 and R F2 Construct the unbalanced dataset R2 for training the unbalanced strong typhoon sample re-learning model.

[0026] In one embodiment, according to the strong typhoon sample dataset C1 in the historical tropical cyclone best track data records and its corresponding strong typhoon sample environmental variable dataset Construct the first strong typhoon dataset R for training the strong typhoon pre-learning model T1 and the second strong typhoon dataset R for training the unbalanced strong typhoon sample re-learning model T2 , including:

[0027] Arbitrarily select the nth i strong typhoon record from C1 and select the corresponding strong typhoon sample environmental variable data from . Number the first strong typhoon record point with a wind speed greater than 84 kt in the nth i strong typhoon record as Select the prediction time step as k and the regression time step as b. Let k = T / 6, where T is the prediction time and is a positive integer multiple of 6, and b is any positive integer between 1 and k.

[0028] If then select from select the last k + b strong typhoon recording points after numbering, and package their corresponding strong typhoon sample environmental data as a strong typhoon input sample; if then perform data screening on the next strong typhoon record.

[0029] Package all the strong typhoon input samples obtained after screening all the strong typhoon records in C1 to obtain a strong typhoon input data set for training the strong typhoon formation prediction framework, and assign the label 1 to all the strong typhoon input samples to obtain a strong typhoon output data set for training the strong typhoon formation prediction framework.

[0030] Extract an equal amount of a part of the samples from the strong typhoon input data set and the strong typhoon output data set to form the first strong typhoon data set R T1 for training the strong typhoon pre-learning model, and form the second strong typhoon data set R T2 for training the unbalanced strong typhoon sample re-learning model from the remaining data in the strong typhoon input data set and the strong typhoon output data set.

[0031] In one embodiment, according to the ordinary typhoon sample data set C2 and its corresponding ordinary typhoon sample environmental variable data set in the historical tropical cyclone best track data record construct the first ordinary typhoon data set R F1 for training the ordinary typhoon pre-learning model and the second ordinary typhoon data set R F2 for training the unbalanced ordinary typhoon sample re-learning model, including:

[0032] Arbitrarily select the nth j ordinary typhoon record from C2 and select the corresponding ordinary typhoon sample environmental variable data from The number of the ordinary typhoon recording point with the first maximum intensity of the nth j ordinary typhoon record is

[0033] If then select from select the last k + b ordinary typhoon recording points after numbering, and package their corresponding ordinary typhoon sample environmental data as an ordinary typhoon input sample; if then perform data screening on the next ordinary typhoon record.

[0034] All the ordinary typhoon input samples obtained after screening the ordinary typhoon records used in C2 are packaged to obtain an ordinary typhoon input dataset for training the strong typhoon formation prediction framework; all the ordinary typhoon input samples are assigned the label 0 to obtain an ordinary typhoon output dataset for training the strong typhoon formation prediction framework;

[0035] Extract from the ordinary typhoon input dataset and the ordinary typhoon output dataset samples with the same quantity as the first strong typhoon dataset R T1 to form the first ordinary typhoon dataset R for training the ordinary typhoon pre-learning model F1 and the remaining data in the ordinary typhoon input dataset and the ordinary typhoon output dataset are used to form the second ordinary typhoon dataset R for training the unbalanced ordinary typhoon sample re-learning model F2 .

[0036] In one embodiment, after constructing the balanced dataset R1 for training the strong typhoon pre-learning model according to R T1 and R F1 , and constructing the unbalanced dataset R2 for training the unbalanced strong typhoon sample re-learning model according to R T2 and R F2 , it further includes:

[0037] Fill the missing values and non-numerical values in the balanced dataset R1 and the unbalanced dataset R2 with 0, and use the method of maximum-minimum normalization to normalize the processed R1 and R2 respectively to obtain the normalized balanced dataset and the normalized unbalanced dataset.

[0038] In one embodiment, before constructing the strong typhoon formation prediction framework under unbalanced data conditions, it further includes: setting up a deep learning model environment and constructing the strong typhoon formation prediction framework under unbalanced data conditions in the deep learning model environment.

[0039] In one embodiment, the construction steps of the strong typhoon pre-learning model include:

[0040] Combine the atmospheric three-dimensional convolution module, the sea surface two-dimensional convolution module and the ocean hydrology three-dimensional convolution module according to the connection function, connect the combined module with the first long short-term memory network module, connect the first fully connected layer after the first long short-term memory network module, and select the sigmoid function as the activation function to complete the construction of the strong typhoon pre-learning model;

[0041] Among them, the atmospheric three-dimensional convolution module in the severe typhoon pre-learning model is used to learn atmospheric variable data, including an input layer, a three-dimensional convolution layer, a three-dimensional max pooling layer, an unfolding layer, and a fully connected layer; the sea surface two-dimensional convolution module in the severe typhoon pre-learning model is used to learn sea surface variable data, including an input layer, a two-dimensional convolution layer, a two-dimensional max pooling layer, an unfolding layer, and a fully connected layer; the ocean hydrology three-dimensional convolution module is used to learn ocean hydrology variable data, including a three-dimensional convolution layer, a three-dimensional max pooling layer, an unfolding layer, and a fully connected layer; the first long short-term memory network module is used to classify time series; the first fully connected layer is used for data output.

[0042] In one embodiment, the steps for constructing the unbalanced severe typhoon sample re-learning model include:

[0043] Obtain the loaded and saved atmospheric three-dimensional convolution module, sea surface two-dimensional convolution module, and ocean hydrology three-dimensional convolution module in the severe typhoon pre-learning model and combine them through a connection function. Connect the combined module with the second long short-term memory network module, connect the second fully connected layer after the second long short-term memory network module, and select the sigmoid function as the activation function to complete the construction of the unbalanced severe typhoon sample re-learning model; among them, when obtaining the loaded and saved atmospheric three-dimensional convolution module, sea surface two-dimensional convolution module, and ocean hydrology three-dimensional convolution module in the severe typhoon pre-learning model, import the parameters of the atmospheric three-dimensional convolution module, sea surface two-dimensional convolution module, and ocean hydrology three-dimensional convolution module in the trained severe typhoon pre-learning model.

[0044] In one embodiment, the binary cross-entropy loss function, balanced data training set, and balanced data test set are used to train and evaluate the performance of the severe typhoon pre-learning model, and the severe typhoon pre-learning model with the best evaluation effect on the balanced data test set is selected as the trained severe typhoon pre-learning model, including:

[0045] Initialize the number of loops, number of training epochs, batch size, and learning rate during the training of the severe typhoon pre-learning model;

[0046] Call the deep learning library to compile, fit, and evaluate the severe typhoon pre-learning model, and record and save the binary cross-entropy loss function corresponding to the balanced data training set and the evaluation metrics corresponding to the balanced data test set during the training process of the severe typhoon pre-learning model; among them, the evaluation metrics include the area under the ROC curve, the area under the PR curve, and the harmonic mean of precision and recall; the binary cross-entropy loss function is expressed as

[0047]

[0048] In the formula, represents the prediction result, Y t+6k represents the true label value, Yt+6k = 1 represents a positive sample, Y t+6k = 0 represents a negative sample;

[0049] During training, the number of loops, the number of training epochs, the batch size, and the learning rate are adjusted in sequence. Record the parameters of each adjustment and the corresponding model training results, and select the strong typhoon pre-learning model with the best evaluation effect on the balanced data test set from the adjustment records as the trained strong typhoon pre-learning model.

[0050] In one embodiment, transfer the prior knowledge in the trained strong typhoon pre-learning model to the unbalanced strong typhoon sample re-learning model, and use the pre-constructed unbalanced strong typhoon sample loss function, unbalanced data training set, and unbalanced data test set to train and evaluate the performance of the unbalanced strong typhoon sample re-learning model. Select the unbalanced strong typhoon sample re-learning model with the best evaluation effect on the unbalanced data test set as the trained unbalanced strong typhoon sample re-learning model, including:

[0051] Transfer the prior knowledge in the trained strong typhoon pre-learning model to the unbalanced strong typhoon sample re-learning model, and initialize the number of loops, the number of training epochs, the batch size, and the learning rate during the training of the unbalanced strong typhoon sample re-learning model;

[0052] Call the deep learning library to compile, fit, and evaluate the unbalanced strong typhoon sample re-learning model, and call the deep learning library to freeze the parameters of some layers in the unbalanced strong typhoon sample re-learning model;

[0053] Call the deep learning library to record and save the unbalanced strong typhoon sample loss function corresponding to the unbalanced data training set and the evaluation metrics corresponding to the unbalanced data test set during the training of the unbalanced strong typhoon sample re-learning model; where the unbalanced strong typhoon sample loss function is expressed as

[0054]

[0055] In the formula, represents the prediction result, Y t+6k represents the true label value, Y t+6k = 1 represents a positive sample, Y t+6k = 0 represents a negative sample, γ represents the weight factor for focusing on difficult-to-classify misclassified samples, and α represents the balance factor for balancing the imbalance ratio of positive and negative samples;

[0056] During training, the number of loops, the number of training epochs, the batch size, and the learning rate are adjusted in sequence. Record the parameters of each adjustment and the corresponding model training results, and select the unbalanced severe typhoon sample re-learning model with the best evaluation effect on the unbalanced data test set from the adjustment records as the trained unbalanced severe typhoon sample re-learning model.

[0057] The above-mentioned method for predicting the formation of severe typhoons under unbalanced data based on transfer learning first obtains the official historical tropical cyclone best track dataset and its corresponding environmental variable dataset, and constructs a balanced dataset and an unbalanced dataset from the obtained datasets; then constructs a severe typhoon pre-learning model, and uses the constructed balanced dataset to train and fit the severe typhoon pre-learning model, and transfers the prior knowledge in the trained severe typhoon pre-learning model to the unbalanced severe typhoon sample re-learning model, and then uses the unbalanced dataset to train and fit the unbalanced severe typhoon sample re-learning model, so as to obtain a trained prediction framework for the formation of severe typhoons; finally, use the trained prediction framework for the formation of severe typhoons under unbalanced data based on transfer learning to predict whether a severe typhoon will form, and obtain the typhoon formation prediction result.

[0058] The beneficial effects of adopting the present invention are as follows:

[0059] 1. The present invention can conveniently construct and predict models using existing public datasets and deep learning frameworks. The present invention first obtains a public dataset from an official website and constructs a dataset and a model, then obtains prior knowledge through a pre-learning model, and then obtains a prediction result through a re-learning model. Therefore, the cost is small and it is easy to implement.

[0060] 2. The present invention constructs a deep learning framework for predicting the formation of severe typhoons under unbalanced data conditions by utilizing the ability of transfer learning to obtain prior knowledge and applying the knowledge learned from balanced data to the prediction of unbalanced data. Specifically, the framework is trained using tropical cyclone data and atmospheric variables, sea surface variables, and ocean hydrological variables in the past ten years or so. Therefore, the trained framework can effectively improve the prediction accuracy of the formation of severe typhoons under unbalanced data conditions.

[0061] 3. An unbalanced severe typhoon sample loss function is designed for the unbalanced severe typhoon sample re-learning model for unbalanced data. By assigning different loss weights to severe typhoon samples and ordinary typhoon samples, the optimization of the unbalanced data problem is realized, and the prediction accuracy of the formation of severe typhoons can be successfully improved. Brief Description of the Drawings

[0062] Figure 1 It is a schematic diagram of a prediction framework for the formation of severe typhoons under unbalanced data conditions in an embodiment;

[0063] Figure 2 Schematic diagram of the model architecture of the strong typhoon pre-learning model as prior knowledge in an embodiment. Specific implementation manner

[0064] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0065] First step, obtain the official historical tropical cyclone best track dataset, screen the historical tropical cyclone best track dataset, obtain the historical tropical cyclone best track data records and their corresponding environmental variable datasets, and construct a balanced dataset for training the strong typhoon pre-learning model and an unbalanced dataset for training the unbalanced strong typhoon sample re-learning model according to the historical tropical cyclone best track data records and environmental variable datasets. The specific steps include:

[0066] 1.1 Obtain the IBTrACS (International Best Track Archive for Climate Stewardship, translated as International Best Archive for Climate Stewardship) global tropical cyclone best track dataset certified by the World Meteorological Organization from 1979 to 2016 from the official website (https: / / www.ncdc.noaa.gov / ibtracs / ), and divide the obtained tropical cyclone dataset according to the formation sea area. Each region contains historical tropical cyclone best track data records of different years, and each record contains all the times from the formation to the dissipation of the tropical cyclone, the central longitude and latitude positions, the central minimum pressure, and the central maximum wind speed.

[0067] 1.2 For any region Q1, by checking each tropical cyclone record in Q1, the records with a central maximum wind speed reaching 84 kt are classified into the strong typhoon sample dataset, and the records with a central maximum wind speed reaching 64 kt but not reaching 84 kt are classified into the ordinary typhoon sample dataset. Thus, the strong typhoon sample dataset C1 and the ordinary typhoon sample dataset C2 are constructed, and the remaining sample records are discarded.

[0068] 1.3 According to the time, central longitude and latitude of the strong typhoon records in the strong typhoon sample dataset C1, download and organize the data of atmospheric variables, sea surface variables, and ocean hydrological variables corresponding to the records to obtain the strong typhoon sample environmental variable dataset for the training framework The specific steps include:

[0069] 1.3.1 Randomly sort the strong typhoon records in C1. For any strong typhoon record n among themi , the serial number of the severe typhoon recording point is According to the time, central longitude and latitude position recorded at this recording point, download the atmospheric variables and sea surface variable data in the ERA-Interim reanalysis data from the ECMWF (European Centre for Medium-Range Weather Forecasts) website, and download the ocean hydrological variable data from the Hybrid Coordinate Ocean Model dataset.

[0070] 1.3.2 Determine the download range and resolution of the environmental variable data according to the central longitude and latitude position. The download range is a surrounding longitude and latitude range of 10°×10° centered on the longitude and latitude position of the recording point, representing longitude and latitude respectively. According to the resolution provided by the official website, the highest resolution that can be obtained for downloading atmospheric variable data and sea surface variable data is 0.125°×0.125°, and the highest resolution that can be obtained for downloading ocean hydrological variable data is 0.16°×0.16°.

[0071] 1.3.3 Set the pressure layer heights of the atmospheric variable data at the i nth recording point of the nth severe typhoon record to be 1000 / 975 / 925 / 850 / 800 / 700 / 600 / 500 / 400 / 300 / 200 / 100 hPa. Set the atmospheric variables at the nth recording point of the nth severe typhoon record related to the development of severe typhoons to be zonal wind (abbreviated as u), meridional wind (abbreviated as v), geopotential height (abbreviated as z), temperature (abbreviated as T), and relative humidity (abbreviated as Rh). i nth recording point of the nth severe typhoon record to be zonal wind (abbreviated as u), meridional wind (abbreviated as v), geopotential height (abbreviated as z), temperature (abbreviated as T), and relative humidity (abbreviated as Rh).

[0072] 1.3.4 Set the sea surface variable data at the i nth recording point of the nth severe typhoon record to be sea surface temperature (abbreviated as sst), and select the sea surface variable data downloaded to be related to the development of severe typhoons. nth recording point of the nth severe typhoon record to be

[0073] 1.3.5 Set the i nth recording point of the nth severe typhoon record to be The ocean depths at which ocean hydrological variable data are recorded are 100 / 90 / 80 / 70 / 60 / 50 / 45 / 40 / 35 / 30 / 25 / 20 / 15 / 12 / 10 / 8 / 6 / 4 / 2 m respectively. For the nth i strong typhoon record downloaded, the ocean hydrological variable data at the

[0074] th recording point are seawater temperature (SeawaterTemperature, abbreviated as st), eastward water velocity (Eastward Water Velocity, abbreviated as water_u), and northward water velocity (Northward Water Velocity, abbreviated as water_v). The downloaded ocean hydrological variable data related to the development of strong typhoons are selected. Strong typhoon sample sea surface variable dataset and strong typhoon sample ocean hydrological variable dataset These datasets are combined to form the strong typhoon sample environmental variable dataset corresponding to the strong typhoon sample dataset C1, which is

[0075] 1.4 According to the method of constructing the strong typhoon sample environmental variable dataset based on the strong typhoon sample dataset C1 a normal typhoon sample environmental variable dataset is constructed according to the normal typhoon sample dataset C2 Specifically, the normal typhoon records in C2 are randomly sorted. For the nth j normal typhoon record among them, the serial number of the normal typhoon recording point in n j is According to the time and the central longitude and latitude position recorded at the th recording point, the corresponding atmospheric variable data, sea surface variable data, and ocean hydrological variable data are downloaded from the official weather forecast center website, and the downloaded atmospheric variable data, sea surface variable data, and ocean hydrological variable data are packaged and combined to construct the normal typhoon sample environmental variable dataset as Among them, represents the normal typhoon sample atmospheric variable dataset, represents the normal typhoon sample sea surface variable dataset, represents the normal typhoon sample ocean hydrological dataset.

[0076] 1.5 According to datasets C1, C2 and Construct a balanced dataset R1 for training a strong typhoon pre-learning model and an unbalanced dataset R2 for training an unbalanced strong typhoon sample re-learning model. The specific steps are as follows:

[0077] 1.5.1 According to C1 and Construct R T1 and R T2 , the method is:

[0078] 1.5.1.1 Select the nth i strong typhoon record from C1, and select the corresponding atmospheric variable, sea surface variable, and ocean hydrological variable data from . Denote the record point number of the first record with a strength greater than 84 kt in the nth i strong typhoon record as Select the specified prediction time step k and regression time step b. Let k = T / 6, where T is the prediction time and is a positive integer multiple of 6, and b is any positive integer between 1 and k.

[0079] 1.5.1.2 If then select the k + b strong typhoon record points after the numbered one, and package their corresponding strong typhoon sample environmental data as a strong typhoon input sample. If then perform data screening on the next strong typhoon record.

[0080] 1.5.1.3 Package all the strong typhoon input samples obtained after screening the strong typhoon records used in C1 to obtain a strong typhoon input dataset for training the strong typhoon formation prediction framework. Label all the strong typhoon input samples with 1 to form a strong typhoon output dataset for framework training. Extract an equal amount from the strong typhoon input dataset and the strong typhoon output dataset to form the first strong typhoon dataset R T1 for training the strong typhoon pre-learning model, and the remaining data forms the second strong typhoon dataset R T2 for training the unbalanced strong typhoon sample re-learning model.

[0081] 1.5.2 According to C2 and Construct R F1 and R F2 , the method is:

[0082] 1.5.2.1 Select the nth j ordinary typhoon record from C2, and select the corresponding atmospheric variable, sea surface variable, and ocean hydrological variable data from . Denote the record point number of the maximum value of the first strength in the nth j ordinary typhoon record as Select the prediction time step \(k\) and the regression time step \(b\). Let \(k = T / 6\), where \(T\) is the prediction time and is a positive integer multiple of 6, and \(b\) is any positive integer between 1 and \(k\).

[0083] 1.5.2.2 If then select pick the last \(k + b\) recorded points after numbering, and pack their corresponding general typhoon sample environmental data as a general typhoon input sample. If then perform data screening on the next general typhoon record.

[0084] 1.5.2.3 After screening all the general typhoon records used in C2, pack all the obtained general typhoon input samples to get the general typhoon input dataset for framework training. Label all the general typhoon input samples with 0 to form the general typhoon output dataset for framework training. Extract samples with the same quantity as dataset \(R\) from the general typhoon input dataset and the general typhoon output dataset T1 to form the first general typhoon dataset \(R_1\) for training the strong typhoon pre - learning model F1 , and the remaining data forms the second general typhoon dataset \(R_2\) for training the unbalanced strong typhoon sample re - learning model F2 .

[0085] 1.6 Combine the constructed \(R_1\) T1 and \(R_2\) F1 to form the balanced dataset \(R_1\) for training the strong typhoon pre - learning model, and combine the constructed \(R_2\) T2 and \(R_3\) F2 to form the unbalanced dataset \(R_2\) for training the unbalanced strong typhoon sample re - learning model.

[0086] 1.7 Perform normalization processing on the constructed \(R_1\) and \(R_2\) respectively. First, fill the missing values and Nan values (non - numerical values) in the dataset with 0, and then use the method of maximum - minimum normalization for processing. The processing formula is as follows:

[0087]

[0088] where, \(X'\) * represents the normalized dataset, \(X\) represents the dataset before normalization, \(X_{\min}\) min represents the minimum value in the dataset, and \(X_{\max}\) max represents the maximum value in the dataset.

[0089] Second step, as Figure 1 shown, construct the strong typhoon formation prediction framework under unbalanced data conditions. The specific steps include:

[0090] 2.1 Build a deep learning model environment. Install the Anaconda software on a server with Tesla V100 GPU cards, and use the conda manager to install the TensorFlow 2.4 environment package and the corresponding Keras deep learning library.

[0091] 2.2 Build a strong typhoon pre-learning model on the built deep learning model environment. The specific steps are as follows:

[0092] 2.2.1 Use the sequential model in Keras to build a 3D CNN for learning atmospheric variables, also known as the atmospheric three-dimensional convolution module. The network in the atmospheric three-dimensional convolution module uses an input layer, several three-dimensional convolutional layers (Conv3D layers), a three-dimensional max pooling layer (MaxPooling3D layer), a flattening layer (Flatten layer), and a fully connected layer (Dense layer). Set 3 Conv3D layers, 1 MaxPooling3D layer, 1 Flatten layer, and 1 Dense layer. The first and second Conv3D layers select 64 convolutional kernels, each with a size of 5×5×1 and a stride of 2×2×1, and select 'ReLU' as the activation function. The third Conv3D layer selects 128 convolutional kernels, each with a size of 5×5×1 and a stride of 2×2×1, and selects 'ReLU' as the activation function. The size of the pooling window of the MaxPooling3D layer is 5×5×1, and the stride is 2×2×1. The number of neurons in the Dense layer is 100, and L2 regularization with a parameter of 0.01 is added to prevent the model from overfitting during training.

[0093] 2.2.2 Build the 2D CNN module for learning sea surface variables in the sequential model in Keras, also known as the two-dimensional sea surface convolution module. The network in the 2D CNN module uses an input layer, a two-dimensional convolutional layer (Conv2D layer), a two-dimensional max pooling layer (MaxPooling2D), a flattening layer (Flatten layer), and a fully connected layer (Dense layer). Set 3 Conv2D layers, 1 MaxPooling2D layer, 1 Flatten layer, and 1 Dense layer. The first Conv2D layer selects 32 convolutional kernels, each with a size of 5×5 and a stride of 2×2, and selects the 'ReLU' function as the activation function. The second Conv2D layer selects 64 convolutional kernels, each with a size of 5×5 and a stride of 2×2, and selects the 'ReLU' function as the activation function. The third Conv2D layer selects 128 convolutional kernels, each with a size of 5×5 and a stride of 2×2, and selects the 'ReLU' function as the activation function. The size of the pooling window in the MaxPooling2D layer is 5×5 and the stride is 2×2. The Dense layer adds L2 regularization with a parameter of 0.01, and the number of neurons is 100.

[0094] 2.2.3 Build the 3D CNN module for learning oceanographic variables in the model, also known as the three-dimensional oceanographic convolution module. The network in the 3D CNN module uses an input layer, a three-dimensional convolutional layer (Conv3D layer), a three-dimensional max pooling layer (MaxPooling3D layer), a flattening layer (Flatten layer), and a fully connected layer (Dense layer). The first and second Conv3D layers select 64 convolutional kernels, each with a size of 3×3×1 and a stride of 2×2×1, and select 'ReLU' as the activation function. The third Conv3D layer selects 128 convolutional kernels, each with a size of 3×3×1 and a stride of 2×2×1, and selects 'ReLU' as the activation function. The size of the pooling window in the MaxPooling3D layer is 3×3×1 and the stride is 2×2×1. The number of neurons in the Dense layer is 100, and L2 regularization with a parameter of 0.01 is added to prevent the model from overfitting during training.

[0095] 2.2.4 Build the first long short-term memory network module for classifying time series in the model. The LSTM module is implemented by selecting the LSTM (long short-term memory network) layer in the Keras library, and the number of neurons is 100.

[0096] 2.2.5 Combine the constructed 3DCNN module and 2DCNN module using the Concatenate function, connect the combined module with the first long short-term memory network module, and add a first fully connected layer at the end of the model as the output of the model. The overall connected model is the strong typhoon pre-learning model. Among them, the number of neurons in the last fully connected layer is 1, and the sigmoid function is selected as the activation function. The formula is:

[0097]

[0098] 2.3 Construct an imbalanced strong typhoon sample re-learning model on the established deep learning model environment. The specific steps include:

[0099] 2.3.1 As Figure 2 shown, load the 3DCNN module and 2DCNN module of the saved strong typhoon pre-learning model, and import the parameters of the 3DCNN module and 2DCNN module in the trained strong typhoon pre-learning model.

[0100] 2.3.2 Connect the loaded module with the second long short-term memory network module. The second long short-term memory network module is implemented using the LSTM layer in the Keras library, with 100 neurons, and L2 regularization is added.

[0101] 2.3.3 Add a second fully connected layer, with the number of neurons in the fully connected layer being 1 and the sigmoid function being selected as the activation function.

[0102] The third step is to train the strong typhoon formation prediction framework according to the balanced dataset and the imbalanced dataset to obtain the trained strong typhoon formation prediction framework. The specific steps include:

[0103] 3.1 Randomly shuffle the constructed balanced dataset R1 and imbalanced dataset R2 and divide them into a balanced data training set R 11 and an imbalanced data training set R 21 , as well as a balanced data test set R 12 and an imbalanced data test set R 22 .

[0104] 3.2 Use the divided balanced data training set R 11 and balanced data test set R 12 to train the strong typhoon pre-learning model. The method is:

[0105] 3.2.1 Select the binary cross-entropy loss function as the loss function of the strong typhoon pre-learning model. The specific form of the loss function is as follows:

[0106]

[0107] Among them, represents the prediction result of the model, Y t+6k represents the true label value, Y t+6k = 1 represents a positive sample, Y t+6k = 0 represents a negative sample.

[0108] 3.2.2 Select the ROC_AUC, PR_AUC, and F1 values of the model as the evaluation metrics of the model. ROC_AUC refers to the area under the ROC curve. The abscissa of the ROC curve is the false positive rate The ordinate is the true positive rate It is defined as follows: ROC_AUC = ∫ROC.

[0109] PR_AUC refers to the area under the PR curve. The abscissa of the PR curve is the recall rate The ordinate is the precision It is defined as follows: PR_AUC = ∫PR.

[0110] F1 refers to the harmonic mean of the precision and the recall rate. It is defined as follows:

[0111] 3.2.3 Use the binary cross-entropy loss function, balance the data training set R 11 and the balanced data test set R 12 to train the strong typhoon pre-learning model, and use the ROC_AUC, PR_AUC, and F1 values as evaluation metrics for evaluation. The method is:

[0112] 3.2.3.1 Initialize the number of loops, the number of training epochs, the batch size (batchsize), and the learning rate lr. Take the atmospheric variable data X 11 in the balanced data training set R P and the ocean hydrological variable data X O as the inputs of the 3DCNN module respectively, and the outputs are the atmospheric variable feature vector V P and the ocean hydrological variable feature vector V O , and the vector length is the number of neurons in the fully connected layer in 3DCNN, both of which are 100. Take the sea surface variable data X 11 in the training set R S as the input of the 2DCNN module, and the output is the sea surface variable feature vector V S , and the vector length is the number of neurons in the fully connected layer of 2DCNN, which is 100. The input of the first long short-term memory network module is V = [V P , V S , V O, the length of the input vector is 300.

[0113] 3.2.3.2 Call the built-in model.compile function in Keras to compile the strong typhoon pre-learning model. Call the built-in model.fit function in Keras to fit the strong typhoon pre-learning model. Call the built-in model.evaluate function in Keras to evaluate the strong typhoon pre-learning model.

[0114] 3.2.3.3 During the training process of the strong typhoon pre-learning model, record the loss value loss of the balanced data training set and the ROC_AUC, PR_AUC, and F1 values of the balanced data test set. Call the built-in model.save function in Keras to save the results of the trained model this time. By adjusting the parameters such as the number of loops, the number of training epochs, the batch size (batchsize), and the learning rate lr in sequence, and record the adjusted parameters each time and the model training results.

[0115] 3.2.3.3 Select the model with the best evaluation effect on the balanced data test set from the adjustment records for saving. Use it as the prior knowledge for migrating to the re-learning model of unbalanced strong typhoon samples.

[0116] 3.3 Use the divided unbalanced data training set R 21 and the unbalanced data test set R 22 to train the re-learning model of unbalanced strong typhoon samples. The method is:

[0117] 3.3.1 Construct the loss function of unbalanced strong typhoon samples as the loss function of the re-learning model of unbalanced strong typhoon samples. The specific form is as follows:

[0118]

[0119] In the formula, represents the prediction result, Y t+6k represents the true label value, Y t+6k =1 represents the positive sample, Y t+6k =0 represents the negative sample, γ represents the weight factor used to focus on difficult-to-classify misclassified samples, and α represents the balance factor used to balance the imbalance ratio of positive and negative samples. In the experiment, the larger the value of γ, the more the model focuses on difficult-to-classify misclassified samples. The value of α is moderately weighted according to the positive and negative sample ratio of the training model. The training process parameters are set as γ = 2 and α = 0.9.

[0120] 3.3.2 Similar to that described in 3.2.2, select the ROC_AUC, PR_AUC, and F1 values of the model as the evaluation indicators of the model.

[0121] 3.3.3 Using the loss function of unbalanced severe typhoon samples and the unbalanced data training set R 21 and the unbalanced data test set R 22 Train the unbalanced severe typhoon sample re - learning model, and use ROC_AUC, PR_AUC and F1 value as evaluation indicators for evaluation. The method is as follows:

[0122] 3.3.3.1 Initialize the number of loops, number of training epochs, batch size (batchsize), and learning rate lr. Load the 3DCNN module and 2DCNN module of the pre - learning model of the severe typhoon with the best training effect that has been saved. Connect the loaded model with the initialized second long short - term memory network module.

[0123] 3.2.3.2 Call the built - in model.compile function of Keras to compile the unbalanced severe typhoon sample re - learning model. Call the built - in model.fit function of Keras to fit the unbalanced severe typhoon sample re - learning model. Call the built - in model.evaluate function of Keras to evaluate the unbalanced severe typhoon sample re - learning model. Use the trainable parameter in the Keras layer to freeze the parameters of some layers of the loaded model.

[0124] 3.2.3.3 During the training process of the re - learning model, record the loss value loss of the unbalanced data training set and the ROC_AUC, PR_AUC and F1 values of the unbalanced data test set. Call the built - in model.save function of Keras to save the results of this training model. Adjust the parameters such as the number of loops, number of training epochs, batch size (batchsize), and learning rate lr in turn, and record the adjusted parameters and results each time.

[0125] 3.2.3.3 Select the model with the best evaluation effect on the unbalanced data test set from the adjustment records for saving. The evaluation effect of its test set is the training effect of the severe typhoon formation prediction framework under the condition of unbalanced data based on transfer learning.

[0126] The fourth step is to perform typhoon formation prediction on the prediction data according to the trained severe typhoon formation prediction framework to obtain the typhoon formation prediction result. The specific steps include:

[0127] 4.1 The central position of the target severe typhoon needs to be in the same sea area as the tropical cyclone records used for framework training, and select the same prediction time step k and retrospective time step b as the training framework. Download the atmospheric variables, sea surface variables, and ocean hydrological variable data with the longitude and latitude range of L°×W° and the resolution of U°×U° at the target typhoon center position.

[0128] 4.2 Download the atmospheric variable, sea surface variable, and ocean hydrological variable data at times t, t-6, t-12, ..., t-6b one by one according to the current time t and the central longitude and latitude position. Package the data at these times according to the variable type respectively to obtain the predicted input data of atmospheric variables Predicted input data of sea surface variables And the predicted input data of ocean hydrological variables Combine the three types of variable data to form the predicted environmental variable data X t .

[0129] 4.3 Input the constructed predicted data X t Into the trained framework for prediction, and realize whether a strong typhoon can form k hours after time t.

[0130] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as these combinations of technical features do not conflict, they should be considered to be within the scope described in this specification.

[0131] The above-described embodiments only represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for predicting the formation of severe typhoons under the condition of imbalanced data based on transfer learning, characterized in that, The method includes: Obtaining the official historical tropical cyclone best track dataset, screening the historical tropical cyclone best track dataset, obtaining historical tropical cyclone best track data records and their corresponding environmental variable datasets, and constructing a balanced dataset for training a severe typhoon pre-learning model and an unbalanced dataset for training an unbalanced severe typhoon sample re-learning model according to the historical tropical cyclone best track data records and environmental variable datasets; Constructing a severe typhoon formation prediction framework under unbalanced data conditions, where the severe typhoon formation prediction framework includes a severe typhoon pre-learning model and an unbalanced severe typhoon sample re-learning model; among them, the severe typhoon pre-learning model includes an atmospheric three-dimensional convolution module, a sea surface two-dimensional convolution module, an ocean hydrology three-dimensional convolution module, a first long short-term memory network module, and a first fully connected layer; the unbalanced severe typhoon sample re-learning model includes the atmospheric three-dimensional convolution module, the sea surface two-dimensional convolution module, and the ocean hydrology three-dimensional convolution module loaded and saved in the severe typhoon pre-learning model, as well as a second long short-term memory network module and a second fully connected layer; Training the severe typhoon formation prediction framework according to the balanced dataset and the unbalanced dataset to obtain a trained severe typhoon formation prediction framework; Performing typhoon formation prediction on prediction data according to the trained severe typhoon formation prediction framework to obtain a typhoon formation prediction result; Among them, the training steps of the severe typhoon formation prediction framework specifically include: Dividing the balanced dataset and the unbalanced dataset into a balanced data training set, a balanced data test set, an unbalanced data training set, and an unbalanced data test set according to a certain ratio; Using a binary cross-entropy loss function, the balanced data training set, and the balanced data test set to train and evaluate the performance of the severe typhoon pre-learning model, and selecting the severe typhoon pre-learning model with the best evaluation effect on the balanced data test set as the trained severe typhoon pre-learning model; Transferring the prior knowledge in the trained severe typhoon pre-learning model to the unbalanced severe typhoon sample re-learning model, and using the pre-constructed unbalanced severe typhoon sample loss function, the unbalanced data training set, and the unbalanced data test set to train and evaluate the performance of the unbalanced severe typhoon sample re-learning model, and selecting the unbalanced severe typhoon sample re-learning model with the best evaluation effect on the unbalanced data test set as the trained unbalanced severe typhoon sample re-learning model.

2. The method according to claim 1, wherein Obtaining the official historical tropical cyclone best track dataset, screening the historical tropical cyclone best track dataset, obtaining historical tropical cyclone best track data records and their corresponding environmental variable datasets, including: Obtaining the official historical tropical cyclone best track dataset, dividing the historical tropical cyclone best track dataset according to the formation sea area of the tropical cyclone, and obtaining historical tropical cyclone best track data records containing different years in each sea area; among them, each historical tropical cyclone best track data record includes the time from the formation to the dissipation of the tropical cyclone, the central longitude and latitude position, the central minimum pressure, and the central maximum wind speed. Divide the historical tropical cyclone best track data records into a severe typhoon sample data set according to the maximum wind speed at the center and a common typhoon sample data set , and according to the and obtain the corresponding severe typhoon sample environmental variable data set and the common typhoon sample environmental variable data set .

3. The method according to claim 2, characterized in that, According to the said and obtain the corresponding dataset of environmental variables of severe typhoon samples and the dataset of environmental variables of ordinary typhoon samples , including: Randomly sort the severe typhoon records in . For the th severe typhoon record, assign the serial number of the severe typhoon record point in as . According to the time and the central longitude and latitude position recorded in , download the corresponding atmospheric variable data, sea surface variable data, and ocean hydrological variable data from the official weather forecast center website, and package and combine the downloaded atmospheric variable data, sea surface variable data, and ocean hydrological variable data to construct a severe typhoon sample environmental variable dataset as ; where represents the severe typhoon sample atmospheric variable dataset, represents the severe typhoon sample sea surface variable dataset, and represents the severe typhoon sample ocean hydrological dataset; Randomly sort the ordinary typhoon records in . For the -th ordinary typhoon record, assign the serial number of the ordinary typhoon record point in as . According to the time and the central longitude and latitude position recorded at the -th recording point, download the corresponding atmospheric variable data, sea surface variable data and ocean hydrological variable data from the official weather forecast center website, and package and combine the downloaded atmospheric variable data, sea surface variable data and ocean hydrological variable data to construct an ordinary typhoon sample environmental variable data set as ; where represents the ordinary typhoon sample atmospheric variable data set, represents the ordinary typhoon sample sea surface variable data set, represents the ordinary typhoon sample ocean hydrological data set.

4. The method according to claim 3, wherein Construct a balanced dataset for training a strong typhoon pre-learning model and an unbalanced dataset for training an unbalanced strong typhoon sample re-learning model based on the historical tropical cyclone best track data records and environmental variable datasets, including: According to the severe typhoon sample data set in the historical tropical cyclone best track data record and its corresponding severe typhoon sample environmental variable data set , construct a first severe typhoon data set for training a severe typhoon pre-learning model and a second severe typhoon data set for training an unbalanced severe typhoon sample re-learning model ; According to the general typhoon sample data set in the historical tropical cyclone best track data record and its corresponding general typhoon sample environmental variable data set , construct a first general typhoon data set for training a general typhoon pre-learning model and a second general typhoon data set for training an unbalanced general typhoon sample re-learning model ; According to the said and construct a balanced data set for training a strong typhoon pre-learning model ; according to the said and construct an imbalanced data set for training an imbalanced strong typhoon sample re-learning model .

5. The method according to claim 4, characterized in that According to the dataset of severe typhoon samples in the historical tropical cyclone best track data record and its corresponding dataset of environmental variables of severe typhoon samples , construct the first severe typhoon dataset for training the severe typhoon pre-learning model and the second severe typhoon dataset for training the re-learning model of unbalanced severe typhoon samples , including: Select any from the to obtain the th strong typhoon record and select the corresponding strong typhoon sample environmental variable data from the . Denote the record point number of the first strong typhoon with an intensity greater than 84 kt in the th strong typhoon record as ; select the prediction time step as k and the regression time step as b. Let k = T / 6, where T is the prediction time and is a positive integer multiple of 6, and b is any positive integer between 1 and k. If , then select from after numbering records of strong typhoons, and package the corresponding environmental data of strong typhoon samples as an input sample of strong typhoon; If , then perform data screening on the next record of strong typhoon;​ For all of the All of the severe typhoon input samples obtained by screening all the severe typhoon records in are packaged to obtain a severe typhoon input data set for training a severe typhoon formation prediction framework. All the severe typhoon input samples are labeled with 1 to obtain a severe typhoon output data set for training a severe typhoon formation prediction framework; Extract an equal amount of a part of the samples from the severe typhoon input dataset and the severe typhoon output dataset to form the first severe typhoon dataset for training the severe typhoon pre-learning model , and form the second severe typhoon dataset for training the unbalanced severe typhoon sample re-learning model with the remaining data in the severe typhoon input dataset and the severe typhoon output dataset .

6. The method according to claim 5, wherein According to the ordinary typhoon sample data set in the historical tropical cyclone best track data record and its corresponding ordinary typhoon sample environmental variable data set , construct a first ordinary typhoon data set for training an ordinary typhoon pre-learning model and a second ordinary typhoon data set for training an unbalanced ordinary typhoon sample re-learning model , including: From the arbitrarily select the nth ordinary typhoon record and from the select the corresponding ordinary typhoon sample environmental variable data. Take the ordinary typhoon record point number of the first intensity maximum value of the nth ordinary typhoon record as ; If , then select from after numbering ordinary typhoon recording points, and package the corresponding ordinary typhoon sample environmental data as an ordinary typhoon input sample; if , then perform data screening on the next ordinary typhoon record;​ For the described All the ordinary typhoon input samples obtained by screening the ordinary typhoon records used in are packaged to obtain an ordinary typhoon input data set for training the strong typhoon formation prediction framework; all the ordinary typhoon input samples are assigned the label 0 to obtain an ordinary typhoon output data set for training the strong typhoon formation prediction framework; Extract samples with the same quantity from the ordinary typhoon input dataset and the ordinary typhoon output dataset to form the first ordinary typhoon dataset for training the ordinary typhoon pre-learning model and form the second ordinary typhoon dataset for training the unbalanced ordinary typhoon sample re-learning model with the remaining data in the ordinary typhoon input dataset and the ordinary typhoon output dataset . .

7. The method according to claim 6, wherein When constructing a balanced data set for training a strong typhoon pre-learning model according to the said and and constructing an unbalanced data set for training an unbalanced strong typhoon sample re-learning model according to the said and and , it further includes: After that, it further includes: Fill the missing values and non-numeric values in the balanced dataset and the imbalanced dataset with 0, and use the maximum-minimum normalization method to normalize the processed and respectively, to obtain the normalized balanced dataset and the normalized imbalanced dataset.

8. The method according to claim 1, characterized in that, Before constructing a strong typhoon formation prediction framework under unbalanced data conditions, it also includes: Build a deep learning model environment, and construct a strong typhoon formation prediction framework under unbalanced data conditions in the deep learning model environment.

9. The method according to claim 1, wherein The construction steps of the strong typhoon pre-learning model include: Combine the atmospheric three-dimensional convolutional module, the sea surface two-dimensional convolutional module, and the ocean hydrology three-dimensional convolutional module according to the connection function, connect the combined module with the first long short-term memory network module, connect a first fully connected layer after the first long short-term memory network module, and select the sigmoid function as the activation function to complete the construction of the strong typhoon pre-learning model; Among them, the atmospheric three-dimensional convolutional module in the strong typhoon pre-learning model is used to learn atmospheric variable data, including an input layer, a three-dimensional convolutional layer, a three-dimensional max pooling layer, a flattening layer, and a fully connected layer; the sea surface two-dimensional convolutional module in the strong typhoon pre-learning model is used to learn sea surface variable data, including an input layer, a two-dimensional convolutional layer, a two-dimensional max pooling layer, a flattening layer, and a fully connected layer; the ocean hydrology three-dimensional convolutional module in the strong typhoon pre-learning model is used to learn ocean hydrology variable data, including a three-dimensional convolutional layer, a three-dimensional max pooling layer, a flattening layer, and a fully connected layer; the first long short-term memory network module is used to classify time series; the first fully connected layer is used for data output.

10. The method according to claim 1, wherein The construction steps of the unbalanced strong typhoon sample re-learning model include: Obtain the loaded and saved atmospheric three-dimensional convolutional module, sea surface two-dimensional convolutional module, and ocean hydrology three-dimensional convolutional module in the strong typhoon pre-learning model and combine them through the connection function, connect the combined module with the second long short-term memory network module, connect a second fully connected layer after the second long short-term memory network module, and select the sigmoid function as the activation function to complete the construction of the unbalanced strong typhoon sample re-learning model; among them, when obtaining the loaded and saved atmospheric three-dimensional convolutional module, sea surface two-dimensional convolutional module, and ocean hydrology three-dimensional convolutional module in the strong typhoon pre-learning model, import the parameters of the atmospheric three-dimensional convolutional module, sea surface two-dimensional convolutional module, and ocean hydrology three-dimensional convolutional module in the trained strong typhoon pre-learning model.

11. The method according to claim 1, wherein Use the binary cross-entropy loss function, the balanced data training set, and the balanced data test set to train and evaluate the performance of the strong typhoon pre-learning model, and select the strong typhoon pre-learning model with the best evaluation effect on the balanced data test set as the trained strong typhoon pre-learning model, including: Initialize the number of epochs, number of training rounds, batch size, and learning rate during the training of the strong typhoon pre-learning model; Call a deep learning library to compile, fit, and evaluate the strong typhoon pre-learning model, and record and save the binary cross-entropy loss function corresponding to the balanced data training set and the evaluation metrics corresponding to the balanced data test set during the training process of the strong typhoon pre-learning model; wherein, the evaluation metrics include the area under the ROC curve, the area under the PR curve, and the harmonic mean of precision and recall; the binary cross-entropy loss function is expressed as Wherein, represents the prediction result, represents the true label value, represents the positive sample, represents the negative sample; During training, adjust the number of loops, the number of training epochs, the batch size, and the learning rate in sequence, record the parameters of each adjustment and the corresponding model training results, and select the strong typhoon pre-learning model with the best evaluation effect on the balanced data test set from the adjustment records as the trained strong typhoon pre-learning model.

12. The method according to claim 1, wherein Transfer the prior knowledge in the trained strong typhoon pre-learning model to the unbalanced strong typhoon sample re-learning model, and use the pre-constructed unbalanced strong typhoon sample loss function, unbalanced data training set, and unbalanced data test set to train and evaluate the performance of the unbalanced strong typhoon sample re-learning model, and select the unbalanced strong typhoon sample re-learning model with the best evaluation effect on the unbalanced data test set as the trained unbalanced strong typhoon sample re-learning model, including: Transfer the prior knowledge in the trained strong typhoon pre-learning model to the unbalanced strong typhoon sample re-learning model, and initialize the number of loops, the number of training epochs, the batch size, and the learning rate during the training of the unbalanced strong typhoon sample re-learning model. Call a deep learning library to compile, fit, and evaluate the unbalanced strong typhoon sample re-learning model, and call the deep learning library to freeze the parameters of some layers in the unbalanced strong typhoon sample re-learning model. Call the deep learning library to record and save the unbalanced strong typhoon sample loss function corresponding to the unbalanced data training set and the evaluation metrics corresponding to the unbalanced data test set during the training process of the unbalanced strong typhoon sample re-learning model; wherein, the unbalanced strong typhoon sample loss function is expressed as In the formula, represents the prediction result, represents the true label value, represents the positive sample, represents the negative sample, γ represents the weight factor for focusing on misclassified samples that are difficult to classify, and α represents the balance factor for balancing the imbalance ratio between positive and negative samples; During training, adjust the number of loops, the number of training epochs, the batch size, and the learning rate in sequence, record the parameters of each adjustment and the corresponding model training results, and select the unbalanced strong typhoon sample re-learning model with the best evaluation effect on the unbalanced data test set from the adjustment records as the trained unbalanced strong typhoon sample re-learning model.

Images