Tropical cyclone maximum continuous wind speed prediction method based on deep learning

Abstract

The invention discloses a tropical cyclone maximum continuous wind speed prediction method based on deep learning. The tropical cyclone maximum continuous wind speed prediction method comprises the following steps: 1) acquiring the maximum continuous wind speed data of a tropical cyclone; 2) dividing the data into a training set and a test set; 3) inputting the training set into a one-dimensional convolutional neural network for training; 4) inputting the feature data output by the one-dimensional convolutional neural network into a long short-term memory neural network for training; 5) determining optimal parameters; 6) obtaining a prediction result of the training set; 7) comparing training effects of the model and a traditional model, and performing error comparison on a prediction result and a test set; and 8) inputting the test set into the model to obtain a future prediction result. The method takes the historical data of the maximum continuous wind speed of the tropical cyclone as an input value, can achieve real-time prediction, has higher prediction precision, can adapt to meteorological data under the influence of the El Nino phenomenon, and has important significance for prediction research of the maximum continuous wind speed of the tropical cyclone under the influence of the El Nino phenomenon.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary technical field of computer and meteorology, and in particular to a method for predicting maximum sustained wind speed of tropical cyclones based on deep learning. Background Art

[0002] Atmospheric circulation is a multi-scale, nonlinear system that is difficult to accurately predict. Complex climate conditions and the lack of long-term observation data make it even more challenging. Traditional meteorological forecasting methods mainly rely on numerical weather forecast models, which are calculated based on physical equations and historical meteorological data. However, these methods have certain limitations when dealing with complex meteorological phenomena, especially in capturing nonlinear relationships and local features. In recent years, the rapid development of deep learning technology has provided new ideas for meteorological forecasting. Deep learning models can automatically extract features and establish complex nonlinear mapping relationships through large-scale data training, thereby improving the accuracy and reliability of forecasts. Deep learning methods have strong data feature extraction and fitting capabilities, and have a strong ability to solve nonlinear problems. At present, deep learning technology has made great progress in wind speed forecasting, remote sensing image recognition and other fields. At the same time, deep learning also has significant advantages in processing spatiotemporal data, especially for solving various forecasting problems. The relationship between the wind speed of tropical cyclones and various meteorological factors is usually highly nonlinear. Deep learning models, especially deep neural networks, can effectively capture these complex nonlinear relationships. At present, some studies have attempted to apply deep learning to meteorological data analysis and wind speed forecasting, but most methods still face challenges such as data scarcity and insufficient model generalization capabilities. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the prior art and provide a tropical cyclone maximum sustained wind speed prediction method based on deep learning. The method is based on a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory neural network (LSTM). The method takes the historical data of the maximum sustained wind speed of a tropical cyclone as input and outputs the prediction result. The method can automatically learn and extract complex meteorological characteristics and wind speed patterns to achieve real-time prediction. The method is suitable for rapidly changing meteorological conditions and can adapt to different types of meteorological data. The method has higher prediction accuracy than traditional statistical methods and physical models.

[0004] The object of the present invention is achieved by: a method for predicting the maximum sustained wind speed of a tropical cyclone based on deep learning, comprising the following steps:

[0005] Step 1) Collect historical data on maximum sustained wind speed of tropical cyclones, which comes from the Tropical Cyclone Data Center of the China Meteorological Administration;

[0006] Step 2) constructing training data, merging the wind speed data processed in step 1) into a data matrix in chronological order, using the first 70% of the data as a training set and the last 30% of the data as a test set;

[0007] Step 3) using the training set data obtained in step 2) as input of a one-dimensional convolutional neural network to extract and reduce the data features;

[0008] Step 4) using the feature data obtained in step 3) as input data of the long short-term memory neural network to obtain the prediction result of the training set;

[0009] Step 5) Determine the network model parameters through the training data; Input the test set data into the one-dimensional convolutional neural network-long short-term memory neural network model, fold it into small batches of data through the sequence of the one-dimensional convolutional neural grid, flatten the multidimensional array into a one-dimensional array, and then input it into the long short-term memory neural network, adjust its own parameters through continuous iterative training of a large amount of data, and finally fit the trained data through the one-dimensional convolutional neural network-long short-term memory neural network model network, connect the output of the folding layer with the input of the flattening layer through the fully connected layer, and output the final result of the network through the regression layer;

[0010] Step 6) The best parameters obtained through training in step 5) are used as the final parameters of the model, and the test set is input into the one-dimensional convolutional neural network-long short-term memory neural network model. The obtained prediction result is the final prediction result. The fully connected layer uses a single hidden layer deep neural network as the output layer of the one-dimensional convolutional neural network-long short-term memory neural network model to fit and predict the data. The output result is the predicted value of the maximum sustained wind speed of the tropical cyclone;

[0011] Step 7) Compare the training effects of this model and the traditional single model, and compare the prediction results with the test set for error;

[0012] Step 8) Input the test set data into the one-dimensional convolutional neural network-long short-term memory neural network model to obtain a prediction of the maximum sustained wind speed of future tropical cyclones.

[0013] As a further limitation of the present invention, the historical data of maximum sustained wind speed of tropical cyclones in step 1) include the name of the tropical cyclone, the historical year of its formation, the maximum sustained wind speed, and the longitude and latitude; the historical year of tropical cyclone formation is determined according to the average value SSTA of the annual average sea level temperature anomaly in the equatorial central and eastern Pacific Ocean; when the average value of the annual average sea level temperature anomaly is greater than or equal to 0.5°C, it is defined as an El Niño year; the year in which the average value of the annual average sea level temperature anomaly is less than or equal to -0.5°C is defined as a La Niña year, and other years are defined as normal years.

[0014] As a further limitation of the present invention, the one-dimensional convolutional neural network in step 3) includes a convolution layer and a pooling layer, and the calculation formula of the one-dimensional convolution is as follows:

[0015]

[0016] In the formula is the kth convolution mapping of layer l, f is the activation function, N is the number of input convolution mappings, * is the convolution operation, is the weight of the i-th operation of the k-th convolution kernel in layer l, is the bias of the kth convolution kernel corresponding to layer l;

[0017] Add a pooling layer after the convolution layer. For sequence data, for sequence x, select the maximum pooling method to process the pooling layer. Repeat the maximum pooling operation for each continuous vector with window r, and get the maximum feature sequence. Formula (2) means taking the vector arrive The maximum value of

[0018]

[0019] in, It is the k+r-1th convolution mapping of layer l.

[0020] As a further limitation of the present invention, the long short-term memory neural network LSTM in step 4) introduces an input gate, an output gate and a forget gate in the hidden layer, and adds a memory unit, and continuously adjusts its parameters through training with a large amount of data, and captures the temporal relationship from the data extracted from the one-dimensional convolutional neural network, thereby realizing dynamic modeling of time series data; when the maximum sustained wind speed data is input into the hidden layer of the long short-term memory neural network, it is firstly subjected to nonlinear transformation through the input gate, and then combined with the memory unit state processed by the forget gate to generate a new memory unit state; finally, it is processed by a nonlinear function and multiplied by the result of the current information state after the nonlinear function processing to obtain the output of the hidden layer; the functional relationship between the variables is:

[0021] f t =sigmoid(w fx x t +w fs s t-1 +b t ) (3)

[0022] c t =tanh(w cx x t +w cs s t-1 +b c ) (4)

[0023] o t =sigmoid(w ox x t +w os s t-1 +b c ) (5)

[0024] c t =tanh(c t ) (6)

[0025] In formula (3), f t is the output signal of the forget gate. The forget gate outputs a value between 0 and 1 through the Sigmoid function, which determines how much information is retained from the unit at the previous moment to the current moment. fx is the weight of hidden layer feedback, x t is the input vector at time t, s t-1 is the output value of the long short-term memory neural network LSTM at time t-1, b t is the threshold of the forget gate neuron;

[0026] In formula (4), c t is the output signal of the input gate. The input gate determines the information of the input memory cell. First, a Sigmoid layer is used to determine which values ​​need to be updated, and then a tanh layer is used to create a new candidate cell state vector. cx is the weight of hidden layer feedback, b c is the hidden layer neuron threshold;

[0027] In formula (5), o t is the output signal of the output gate at time t. The output gate determines how much information is output to the long short-term memory neural network LSTM. A Sigmoid layer determines which parts of the cell state will be output. The cell state is then processed by the tanh function and multiplied by the output of the output gate to obtain the final output. ox is the weight of the output layer feedback; in formula (6), c t is the preliminary information to be output to the hidden layer state s; the memory unit c at time t t After being filtered by the forget gate and the input gate, and then filtered by the output gate, the hidden state s is obtained. t :

[0028] s t =o t c t (7).

[0029] As a further limitation of the present invention, the step 5) specifically includes: first, defining a folding layer in a one-dimensional convolutional neural network, folding the sequence of the input one-dimensional convolutional neural grid into small batches of data, and adding a normalization layer after the convolution layer to accelerate training, adding an expansion layer and a smoothing layer after the pooling layer, flattening the multidimensional array into a one-dimensional array, inputting it into the long short-term memory neural network, adjusting its own parameters through continuous iterative training of a large amount of data, adding a Dropout layer after the long short-term memory neural network, and finally fitting the trained data through a one-dimensional convolutional neural network-long short-term memory neural network, connecting the output of the folding layer with the input of the flattening layer through a fully connected layer, and outputting the final result of the network through a regression layer; the training of the one-dimensional convolutional neural network-long short-term memory neural network model uses an Adam optimizer to train the neural network, adopts a time-expanded reverse error propagation algorithm to train the expanded network, and shuffles the data after each training cycle.

[0030] The present invention adopts the above technical scheme, and compared with the prior art, the beneficial effects are as follows: the present invention uses the historical data of the maximum sustained wind speed of tropical cyclones as input values ​​and outputs prediction results, can automatically learn and extract complex meteorological characteristics and wind speed patterns, can provide higher prediction accuracy than traditional statistical methods and physical models, can quantitatively analyze the moment-to-moment variation characteristics of tropical cyclones and accurately predict the trend of tropical cyclones, and has strong stability and robustness.

[0031] The present invention can provide a higher prediction accuracy than a traditional statistical model by combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory neural network (LSTM). Especially when the amount of data is large, the method can realize real-time prediction, is suitable for rapidly changing meteorological conditions, and can provide timely information support for meteorological warning and decision-making; the traditional prediction method relies on manual feature extraction, while the one-dimensional convolutional neural network-long short-term memory neural network model (1D-CNN-LSTM) can automatically learn features, reduce the need for manual intervention, and is more flexible and efficient when processing complex data; and can adapt to meteorological data under the influence of the El Nino phenomenon. La Nina, El Nino and normal years have different effects on meteorological conditions. Separate training can reduce such interference, enhance the learning effect of the model, and enable the model to focus on data features under a specific climate mode, thereby improving the accuracy of the prediction, which is of great significance for the study of the maximum sustained wind speed prediction of tropical cyclones under the influence of the El Nino phenomenon. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The figure is a flow chart of the prediction method of the present invention.

[0033] Figure 2It is a structural schematic diagram of the one-dimensional convolutional neural network-long short-term memory neural network module 1D-CNN-LSTM model of the present invention.

[0034] Figure 3 This is the basic structure diagram of the one-dimensional convolutional neural network 1D-CNN model of the present invention.

[0035] Figure 4 This is a comparison chart of the predictions of the maximum sustained wind speed of tropical cyclones that land in China in normal years by the present invention and the traditional model.

[0036] Figure 5 This is a comparison chart of the predictions of the maximum sustained wind speed of tropical cyclones that land in China in El Nino years by the present invention and the traditional model.

[0037] Figure 6 This is a comparison chart of the predictions of the maximum sustained wind speed of tropical cyclones that land in China in La Niña years by the present invention and the traditional model.

[0038] Figure 7 This is a forecast chart of the maximum sustained wind speed of tropical cyclones that land in China in a normal year according to the present invention.

[0039] Figure 8 This is a forecast map of the maximum sustained wind speed of tropical cyclones that land in China during El Nino years according to the present invention.

[0040] Fig. 9 This is a forecast map of the maximum sustained wind speed of tropical cyclones that land in China in La Niña years according to the present invention. DETAILED DESCRIPTION

[0041] like Figure 1 A tropical cyclone maximum sustained wind speed prediction method based on deep learning is shown, comprising the following steps:

[0042] Step 1) Collect historical data on maximum sustained wind speed of tropical cyclones, which comes from the Tropical Cyclone Data Center of the China Meteorological Administration;

[0043] The historical data of maximum sustained wind speed of tropical cyclones include the name of tropical cyclones, historical year of formation, maximum sustained wind speed, longitude and latitude; the historical year of tropical cyclone formation is determined according to the average value of the annual average sea level temperature anomaly SSTA in the equatorial central and eastern Pacific. When the average value of the annual average sea level temperature anomaly is greater than or equal to 0.5℃, it is defined as an El Niño year; the year when the average value of the annual average sea level temperature anomaly is less than or equal to -0.5℃ is defined as a La Niña year, and other years are defined as normal years;

[0044] In this example, the maximum sustained wind speed data of tropical cyclones that landed in China during 1980-2022 are first collected. The data comes from the Tropical Cyclone Data Center of the China Meteorological Administration (tcdata.typhoon.org.cn). This data set records the name, generation time, maximum sustained wind speed, longitude and latitude of tropical cyclones generated in the northwest Pacific Ocean. The time interval between adjacent records is 6 hours. These data are collected according to the key areas for El Niño monitoring ( 3.4 area, 120 ~ 170 ° W, 5 ° N ~ 5 ° S) annual average sea level temperature anomaly (SSTA) average value to determine the cold and warm event years of the ENSO cycle, when the SSTA average value ≥ 0.5 ° C, it is defined as an El Niño year (i.e. a warm year); the year with an SSTA average value of less than or equal to -0.5 ° C is defined as a La Niña year (i.e. a cold year), and the rest of the years are defined as normal years. There are 15 years in total, namely 1982, 1987, 1991, 1992, 1997, 2002, 2015, 2020, 2021, 2022; There are 10 years with La Nina, namely 1984, 1985, 1988, 1989, 1999, 2000, 2007, 2008, 2011 and 2019; there are 23 normal years, namely 1980, 1981, 1983, 1986, 1990, 1993, 1994, 1995, 1996, 1998, 2001, 2003, 2004, 2005, 2006, 2009, 2010, 2012, 2013, 2014, 2016, 2017 and 2018. The total number of tropical cyclones generated in 12 La Nina years is 49, the total number of tropical cyclones generated in 8 El Nino years is 71, and the total number of tropical cyclones generated in 23 normal years is 167.

[0045] Step 2) constructing training data, merging the wind speed data processed in step 1) into a data matrix in chronological order, taking the first 70% of the data as a training set, and the last 30% of the data as a test set.

[0046] Step 3) The training set data obtained in step 2) is used as the input of a one-dimensional convolutional neural network. The one-dimensional convolutional neural network 1D-CNN model is mainly composed of a convolution layer and a pooling layer to extract and reduce the data features, such as Figure 3As shown in the figure, the training parameters are set as follows: the time lag order is 1:8, the number of batch samples is 24, the maximum number of iterations is 60, and the learning rate is 0.005. Taking a normal year as an example, after the data is preprocessed, the training set records contain 167 wind speed data. After division, the number of data in the training set is 116. The first layer of the 1D-CNN grid is the convolution layer, and the convolution kernel size of the filter is defined as 10. By defining a filter, the neural network can learn a single feature in the first layer. The output of the first layer is a 1*106 matrix, and each column of the output matrix contains the weight of a filter. The calculation formula of the one-dimensional convolution is as follows:

[0047]

[0048] In the formula is the kth convolution mapping of layer l, f is the activation function, N is the number of input convolution mappings, * is the convolution operation, is the weight of the i-th operation of the k-th convolution kernel in layer l, is the bias of the kth convolution kernel corresponding to layer l;

[0049] In order to simplify the output and avoid data overfitting, a pooling layer is added after the convolution layer. For sequence data, for sequence x, the maximum pooling method is selected to process the pooling layer. The maximum pooling operation is repeated for each continuous vector with a window of r, that is, the maximum feature sequence is obtained. Formula (2) means taking the vector arrive The maximum value of

[0050]

[0051] in, It is the k+r-1th convolution mapping of layer l;

[0052] After the pooling layer is the Dropout layer, which randomly assigns zero weights to neurons in the network. A ratio of 0.5 is selected, which means that 50% of the neurons will have zero weights. Through the Dropout layer, the network is less sensitive to small changes in the data, so it can further improve the accuracy of processing unseen data. The output of this layer is still a 1*106 matrix.

[0053] Step 4) using the feature data obtained in step 3) as input data of the long short-term memory neural network to obtain the prediction result of the training set;

[0054] The long short-term memory neural network LSTM introduces input gates, output gates and forget gates in the hidden layer, and adds memory units to improve the modeling ability of time series data. Through the training of a large amount of data, its parameters are continuously adjusted to capture the temporal relationship from the data extracted from the one-dimensional convolutional neural network, thereby realizing the dynamic modeling of time series data. When the maximum sustained wind speed data is input into the hidden layer of the long short-term memory neural network, it is first transformed nonlinearly through the input gate, and then combined with the memory unit state processed by the forget gate to generate a new memory unit state. Finally, it is processed by a nonlinear function and multiplied by the result of the current information state after the nonlinear function processing to obtain the output of the hidden layer. The functional relationship between the variables is:

[0055] f t =sigmoid(w fx x t +w fs s t-1 +b t ) (3)

[0056] c t =tanh(w cx x t +w cs s t-1 +b c ) (4)

[0057] o t =sigmoid(w ox x t +w os s t-1 +b c ) (5)

[0058] c t =tanh(c t ) (6)

[0059] In formula (3), f t is the output signal of the forget gate. The forget gate outputs a value between 0 and 1 through the Sigmoid function, which determines how much information is retained from the unit at the previous moment to the current moment. fx is the weight of hidden layer feedback, x t is the input vector at time t, s t-1 is the output value of the long short-term memory neural network LSTM at time t-1, b t is the threshold of the forget gate neuron;

[0060] In formula (4), c tis the output signal of the input gate. The input gate determines the information of the input memory cell. First, a Sigmoid layer is used to determine which values ​​need to be updated, and then a tanh layer is used to create a new candidate cell state vector. cx is the weight of hidden layer feedback, b c is the hidden layer neuron threshold;

[0061] In formula (5), o t is the output signal of the output gate at time t. The output gate determines how much information is output to the long short-term memory neural network LSTM. A Sigmoid layer determines which parts of the cell state will be output. The cell state is then processed by the tanh function and multiplied by the output of the output gate to obtain the final output. ox is the weight of the output layer feedback; in formula (6), c t is the preliminary information to be output to the hidden layer state s; the memory unit c at time t t After being filtered by the forget gate and the input gate, and then filtered by the output gate, the hidden state s is obtained. t :

[0062] s t =o t c t (7).

[0063] Step 5) Determine the network model parameters through training data; input the test set data into the one-dimensional convolutional neural network-long short-term memory neural network model, such as Figure 2 As shown; first, a folding layer is defined in the one-dimensional convolutional neural network, and the sequence of the input one-dimensional convolutional neural network 1D-CNN is folded into small batches of data, and a normalization layer is added after the convolution layer to accelerate training. An unfolding layer and a smoothing layer are added after the pooling layer to flatten the multidimensional array into a one-dimensional array and input it into the long short-term memory neural network LSTM. The parameters are adjusted through continuous iterative training of a large amount of data. A Dropout layer is added after the long short-term memory neural network LSTM, and the ratio is selected as 0.25. Finally, the trained data is fitted by the one-dimensional convolutional neural network-long short-term memory neural network 1D-CNN-LSTM. The output of the folding layer is connected to the input of the flattening layer through the fully connected layer, and the final result of the network is output through the regression layer; the training of the one-dimensional convolutional neural network-long short-term memory neural network model 1D-CNN-LSTM uses the Adam optimizer to train the neural network, and the time-expanded reverse error propagation algorithm is used to train the unfolded network, and the data is shuffled after each training cycle; to enhance the generalization ability of the model. The final training parameters are: time lag order is 1:8, batch size is 24, maximum number of iterations is 60, and learning rate is 0.005.

[0064] Step 6) The best parameters obtained through training in step 5) are used as the final parameters of the model, and the test set is input into the one-dimensional convolutional neural network-long short-term memory neural network model. The obtained prediction result is the final prediction result. The fully connected layer uses a single hidden layer deep neural network as the output layer of the one-dimensional convolutional neural network-long short-term memory neural network model to fit and predict the data. The output result is the predicted value of the maximum sustained wind speed of the tropical cyclone; the output result is the predicted value at time t.

[0065] Step 7) Compare the training effects of this model with those of the traditional single model, and compare the prediction results with the test set for error; Figure 4-6 As shown in the figure, the training effect of the single model CNN grid and the 1D-CNN-LSTM model are compared; the prediction results are compared with the test set for error, and the root mean square error (RMSE) is introduced to evaluate the prediction performance of this method, which is defined as follows:

[0066]

[0067] In formula (8), N is the number of samples, K is i is the predicted value, k i is the true value of the test set. The RMSE range is [0, +∞). The smaller the value, the smaller the error of the model. The RMSE of the training set prediction in a normal year is 0.271; The RMSE of the annual training set prediction is 0.274; The RMSE for the annual training set prediction is 0.203.

[0068] Step 8) Input the test set data into the 1D-CNN-LSTM model to obtain the prediction of the maximum sustained wind speed of tropical cyclones landing in China in the future. The prediction results are as follows: Figure 7-9 As shown in the figure, it can be found that the prediction results for normal years have two peaks, which means that super typhoons with higher wind speeds may occur in normal years in the future. In El Nino years, the wind speed tends to decrease, while in La Nina years, the wind speed tends to increase in the short term.

[0069] The present invention establishes a deep learning model based on 1D-CNN and LSTM network, takes the historical data of maximum sustained wind speed of tropical cyclones as input values, outputs prediction results, can automatically learn and extract complex meteorological features and wind speed patterns, can realize real-time prediction, is suitable for rapidly changing meteorological conditions, and can adapt to different types of meteorological data, and has higher prediction accuracy than traditional statistical methods and physical models. 1D-CNN can effectively extract local features in time series data, capture local patterns and trends of wind speed changes, thereby improving the accuracy of prediction. LSTM is good at processing time series data and can remember long-term dependencies. The present invention combines 1D-CNN and LSTM to provide higher prediction accuracy than traditional statistical models, especially when the amount of data is large. The method can realize real-time prediction, is suitable for rapidly changing meteorological conditions, and can provide timely information support for meteorological warning and decision-making. Since the influence of La Nina years, El Nino years and normal years on meteorological conditions is different, the present invention trains them separately to reduce such interference, enhance the learning effect of the model, and enable the model to focus on the data features in a specific climate mode and predict the changing trend of the maximum sustained wind speed of tropical cyclones under different climate modes. Traditional prediction methods usually rely on manual feature extraction, while 1D-CNN-LSTM can automatically learn features, reducing the need for manual intervention and being more flexible and efficient when processing complex data. It can also adapt to meteorological data under the influence of El Niño. La Niña, El Niño and normal years have different impacts on meteorological conditions. Training them separately can reduce this interference, enhance the learning effect of the model, and enable the model to focus on the data features under a specific climate mode, thereby improving the accuracy of the prediction, which is of great significance for the study of the maximum sustained wind speed prediction of tropical cyclones under the influence of El Niño.

[0070] The present invention is not limited to the above-mentioned embodiments. On the basis of the technical solution disclosed in the present invention, technicians in this field can make some substitutions and deformations to some technical features therein according to the disclosed technical content without creative labor, and these substitutions and deformations are all within the protection scope of the present invention.

Claims

1. A tropical cyclone maximum sustained wind speed prediction method based on deep learning, characterized in that: The following steps are involved: Step 1) Collect historical data on maximum sustained wind speed of tropical cyclones, which comes from the Tropical Cyclone Data Center of the China Meteorological Administration; Step 2) constructing training data, merging the wind speed data processed in step 1) into a data matrix in chronological order, using the first 70% of the data as a training set and the last 30% of the data as a test set; Step 3) using the training set data obtained in step 2) as input of a one-dimensional convolutional neural network to extract and reduce the data features; Step 4) using the feature data obtained in step 3) as input data of the long short-term memory neural network to obtain the prediction result of the training set; Step 5) Determine the network model parameters through the training data; Input the test set data into the one-dimensional convolutional neural network-long short-term memory neural network model, fold it into small batches of data through the sequence of the one-dimensional convolutional neural grid, flatten the multidimensional array into a one-dimensional array, and then input it into the long short-term memory neural network, adjust its own parameters through continuous iterative training of a large amount of data, and finally fit the trained data through the one-dimensional convolutional neural network-long short-term memory neural network model network, connect the output of the folding layer with the input of the flattening layer through the fully connected layer, and output the final result of the network through the regression layer; Step 6) The best parameters obtained through training in step 5) are used as the final parameters of the model, and the test set is input into the one-dimensional convolutional neural network-long short-term memory neural network model. The obtained prediction result is the final prediction result. The fully connected layer uses a single hidden layer deep neural network as the output layer of the one-dimensional convolutional neural network-long short-term memory neural network model to fit and predict the data. The output result is the predicted value of the maximum sustained wind speed of the tropical cyclone; Step 7) Compare the training effects of this model and the traditional single model, and compare the prediction results with the test set for error; Step 8) Input the test set data into the one-dimensional convolutional neural network-long short-term memory neural network model to obtain a prediction of the maximum sustained wind speed of future tropical cyclones.

2. The method for predicting the maximum sustained wind speed of a tropical cyclone based on deep learning according to claim 1, characterized in that: The historical data of maximum sustained wind speed of tropical cyclones in step 1) include the name of the tropical cyclone, the historical year of its formation, the maximum sustained wind speed, and the longitude and latitude; the historical year of tropical cyclone formation is determined according to the average value SSTA of the annual average sea level temperature anomaly in the equatorial central and eastern Pacific, and when the average value of the annual average sea level temperature anomaly is greater than or equal to 0.5°C, it is defined as an El Niño year; the year in which the average value of the annual average sea level temperature anomaly is less than or equal to -0.5°C is defined as a La Niña year, and other years are defined as normal years.

3. The method for predicting maximum sustained wind speed of a tropical cyclone based on deep learning according to claim 1, characterized in that: The one-dimensional convolutional neural network in step 3) includes a convolution layer and a pooling layer. The calculation formula of the one-dimensional convolution is as follows: In the formula is the kth convolution mapping of layer l, f is the activation function, N is the number of input convolution mappings, * is the convolution operation, is the weight of the i-th operation of the k-th convolution kernel in layer l, is the bias of the kth convolution kernel corresponding to layer l; Add a pooling layer after the convolution layer. For sequence data, for sequence x, select the maximum pooling method to process the pooling layer. Repeat the maximum pooling operation for each continuous vector with window r, and get the maximum feature sequence. Formula (2) means taking the vector arrive The maximum value of in, It is the k+r-1th convolution mapping of layer l.

4. The method for predicting maximum sustained wind speed of a tropical cyclone based on deep learning according to claim 1, characterized in that: The long short-term memory neural network LSTM described in step 4) introduces an input gate, an output gate and a forget gate in the hidden layer, and adds a memory unit. Its parameters are continuously adjusted through training with a large amount of data, and the temporal relationship is captured from the data extracted from the one-dimensional convolutional neural network, thereby realizing dynamic modeling of time series data; when the maximum sustained wind speed data is input into the hidden layer of the long short-term memory neural network, it is firstly transformed nonlinearly through the input gate, and then combined with the memory unit state processed by the forget gate to generate a new memory unit state; finally, it is processed by a nonlinear function and multiplied by the result of the current information state after the nonlinear function processing to obtain the output of the hidden layer; the functional relationship between the variables is: f t =sigmoid(w fx x t +w fs s t-1 +b t ) (3) c t =tanh(w cx x t +w cs s t-1 +b c ) (4) o t =sigmoid(w ox x t +w os s t-1 +b c ) (5) c t =tanh(c t ) (6) In formula (3), f t is the output signal of the forget gate. The forget gate outputs a value between 0 and 1 through the Sigmoid function, which determines how much information is retained from the unit at the previous moment to the current moment. fx is the weight of hidden layer feedback, x t is the input vector at time t, s t-1 is the output value of the long short-term memory neural network LSTM at time t-1, b t is the threshold of the forget gate neuron; In formula (4), c t is the output signal of the input gate. The input gate determines the information of the input memory cell. First, a Sigmoid layer is used to determine which values ​​need to be updated, and then a tanh layer is used to create a new candidate cell state vector. cx is the weight of hidden layer feedback, b c is the hidden layer neuron threshold; In formula (5), o t is the output signal of the output gate at time t. The output gate determines how much information is output to the long short-term memory neural network LSTM. A Sigmoid layer determines which parts of the cell state will be output. The cell state is then processed by the tanh function and multiplied by the output of the output gate to obtain the final output. ox is the weight of the output layer feedback; in formula (6), c t is the preliminary information to be output to the hidden layer state s; Memory unit c at time t t After being filtered by the forget gate and the input gate, and then filtered by the output gate, the hidden state s is obtained. t : s t =o t c t (7)。 5. The method for predicting maximum sustained wind speed of a tropical cyclone based on deep learning according to claim 1, characterized in that: The step 5) specifically includes: first, defining a folding layer in a one-dimensional convolutional neural network, folding the sequence of the input one-dimensional convolutional neural grid into small batches of data, adding a normalization layer after the convolution layer to accelerate training, adding an unfolding layer and a smoothing layer after the pooling layer, flattening the multidimensional array into a one-dimensional array, inputting it into the long short-term memory neural network, adjusting its own parameters through continuous iterative training of a large amount of data, adding a Dropout layer after the long short-term memory neural network, and finally fitting the trained data through the one-dimensional convolutional neural network-long short-term memory neural network to connect the output of the folding layer with the input of the flattening layer through a fully connected layer, and outputting the final result of the network through a regression layer; the training of the one-dimensional convolutional neural network-long short-term memory neural network model uses the Adam optimizer to train the neural network, adopts the time-expanded reverse error propagation algorithm to train the unfolded network, and shuffles the data after each training cycle.

Images