Typhoon path and strength prediction method and device based on TAM-CL model and storage medium

Abstract

The invention relates to a typhoon path and intensity prediction method and device based on a neural network model and a storage medium, and the method comprises the steps: collecting data, and carrying out the preprocessing of the data, and obtaining the preprocessed data; constructing and training a neural network model; and inputting the preprocessed data into the neural network model for prediction to obtain tropical cyclone prediction intensity data and tropical cyclone prediction path data. Through the data preprocessing method, the method is superior to the existing data preprocessing method in the aspect of reducing the prediction error of the baseline model; the spatial-temporal characteristics of the data are extracted through the three-dimensional convolution and the long-short-term memory network; by introducing an attention mechanism, the long-term prediction capability of the model is enhanced, so that the accuracy of tropical cyclone path long-time prediction and tropical cyclone strength long-time prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to a typhoon path and intensity prediction method, device and storage medium based on a TAM-CL model. Background Art

[0002] In recent years, deep learning has been increasingly applied to typhoon prediction problems. In order to make longer-term predictions, some researchers treat the movement of tropical cyclones as a time series to add previous information to the input and use recurrent neural networks (RNN) to handle this problem. The movement of tropical cyclones occurs on spatial and temporal scales, so some researchers have tried to use the products of numerical models as input to build spatiotemporal models based on ConvLSTM, such as the deep typhoon tracker for tropical cyclone tracking. The intensity of a tropical cyclone is usually defined as the maximum wind speed or minimum pressure at the center of the tropical cyclone, but the specific definition varies in different ocean regions. There are many factors that affect the intensity change of tropical cyclones, such as sea-air interactions and large-scale atmospheric environment. However, the above theory is also difficult to fully explain, so both numerical models and statistical dynamic models are difficult to give accurate intensity predictions. How to improve the accuracy of intensity predictions to provide more accurate disaster warnings is an urgent problem that countries around the world need to solve.

[0003] The prediction of tropical cyclones is a complex problem. Its complexity lies in its variable evolution process, many of which are still unclear. The evolution of tropical cyclones is closely related to atmospheric variables, so we need to grasp the relationship between atmospheric variables in three-dimensional space and the relationship between them and tropical cyclones. Some researchers have tried to extract the characteristics of variables through algorithms such as CNN, and have achieved certain results. But in addition, tropical cyclone prediction is still a time series prediction problem. The surrounding environmental variables in the past period of time will have an impact on the future period of time. The strengthening or weakening of tropical cyclone intensity is the result of the joint action of atmospheric variables over a period of time. Therefore, some researchers have tried to solve it through recurrent neural networks such as RNN and LSTM, but these methods have ignored the spatial characteristics of atmospheric variables to a certain extent. Summary of the invention

[0004] The present invention provides a typhoon path and intensity prediction method, device and storage medium based on a neural network model, aiming to solve at least one of the technical problems existing in the prior art.

[0005] The technical solution of the present invention relates to a typhoon path and intensity prediction method based on a neural network model, the method comprising the following steps:

[0006] S100, collecting data and preprocessing the data to obtain preprocessed data, wherein the preprocessed data at least includes isobaric surface temperature, isobaric surface potential height, isobaric surface relative humidity, ground temperature, ground lift index, and longitude and latitude wind speed;

[0007] S200, build and train neural network models;

[0008] S300, inputting the preprocessed data into a neural network model for prediction, and obtaining tropical cyclone prediction intensity data and tropical cyclone prediction path data.

[0009] Further, step S100 includes:

[0010] S110, based on linear interpolation, clean the data and complete the missing data;

[0011] S120, converting the longitude and latitude position data of the tropical cyclone center into vector data, wherein the vector data includes relative direction data and distance data between the current time and the previous time of the tropical cyclone center location;

[0012] S130, cutting redundant data by sliding window method, cutting out atmospheric variable matrix data of a square area with an actual geographical range of 200 kilometers with the center position coordinates of the tropical cyclone at the current time as the center;

[0013] S140. Perform data balance processing, randomly undersample tropical depression, tropical storm and severe tropical storm data to reduce the input samples of tropical depression, tropical storm and severe tropical storm, and randomly oversample typhoon, severe typhoon and super typhoon data to increase the input samples of typhoon, severe typhoon and super typhoon.

[0014] Further, in step S200, the construction of the neural network model includes the following steps:

[0015] S210, construct a single-layer convolutional long short-term memory network layer;

[0016] S220, stacking multiple convolutional long short-term memory network layers to form a multi-layer convolutional long short-term memory network layer;

[0017] S230, connecting the residual neural network to the output of the multi-layer convolutional long short-term memory network layer;

[0018] S240. Connect the attention layer to the output of the residual neural network.

[0019] Further, in step S210, the single-layer convolutional long short-term memory network layer is composed of n cells connected in chronological order, the n neural network cells represent n time steps respectively, and the n neural network cells are connected in series. The output hidden layer and state layer of the neural network cells in each layer are connected in sequence with the input hidden layer and input state layer of the next neural network cell in the current layer.

[0020] Furthermore, the calculation formula for a single cell is:

[0021] i t =σ(W xi *x t +W hi *h t-1 +b i ),

[0022] f t =σ(W xf *x t +W hf *h t-1 +b if ),

[0023]

[0024] O t =σ(W xo *x t +W ho *h t-1 +b o ),

[0025]

[0026] Among them, i t represents the input gate vector, f t represents the forget gate vector, c t represents the cell state vector, O t represents the output gate vector, h t represents the hidden state vector, x t represents the input feature vector, W represents the weight matrix, b represents the bias term, σ represents the sigmoid function, represents the Hadamard product and * represents the convolution operation.

[0027] Further, in step S220, the stacked convolutional long short-term memory network layer includes w layers of convolutional long short-term memory network layers, and the output hidden layer of the neural network cells in each layer is connected to the input hidden layer of the neural network cells at the corresponding position in the previous layer.

[0028] Further, in step S230, the residual neural network includes a first convolutional layer, a first pooling layer, a first residual block, a second residual block, a third residual block, a fourth residual block and a second pooling layer connected in sequence,

[0029] The size of the first convolutional layer is 7*7*7, the size of the convolution kernel is 64, and the step size is 2;

[0030] The size of the first pooling layer is 3*3*3;

[0031] The first residual block includes a first sub-block and a second sub-block connected in sequence, the number of convolution kernels of the first sub-block is 64, the size of the convolution kernel is 3*3*3, and the step length is 2, the number of convolution kernels of the second sub-block is 64, the size of the convolution kernel is 3*3*3, and the step length is 2;

[0032] The second residual block includes a third sub-block and a fourth sub-block connected in sequence, the number of convolution kernels of the third sub-block is 128, the size of the convolution kernel is 3*3*3, and the step length is 2, and the number of convolution kernels of the fourth sub-block is 128, the size of the convolution kernel is 3*3*3, and the step length is 2;

[0033] The third residual block includes a fifth sub-block and a sixth sub-block connected in sequence, the number of convolution kernels of the fifth sub-block is 256, the size of the convolution kernel is 3*3*3, and the step length is 2, and the number of convolution kernels of the sixth sub-block is 256, the size of the convolution kernel is 3*3*3, and the step length is 2;

[0034] The fourth residual block includes a seventh sub-block and an eighth sub-block connected in sequence, the number of convolution kernels of the seventh sub-block is 512, the size of the convolution kernel is 3*3*3, and the step size is 2, the number of convolution kernels of the eighth sub-block is 512, the size of the convolution kernel is 3*3*3, and the step size is 2.

[0035] Further, in step S240, the attention layer includes an encoder and a decoder, the input of the encoder is the input X1, X2, X3 and X4 corresponding to the four time steps, the hidden states output at each time step are h1, h2, h3, h4, and the output of the encoder is C1, C2 and C3. At the same time, C1, C2 and C3 are the inputs of each step in the decoder, which are linear combinations of hidden states, λ ij , i = 1, 2, 3, 4, j = 1, 2, 3 are weight coefficients, y1, y2, y3 are the outputs of the decoder, corresponding to the predicted values ​​of each future time step, respectively.

[0036] The calculation formulas for C1, C2 and C3 are:

[0037]

[0038] Among them, the coefficient λ ij First, the hidden layer state h of i Convolution operation is performed to extract features, and then a fully connected layer is used for output. The calculation formula is:

[0039] α ij =γ ij tanh(W ij ·Conv(h i )+b ij )

[0040]

[0041] Among them, γ ij , W ij and b ij are all parameters in the fully connected layer, α ij is the weight parameter of the i-th input corresponding to the j-th output.

[0042] Furthermore, the present invention also proposes a typhoon path and intensity prediction device based on a neural network model, which is used to implement the above-mentioned typhoon path and intensity prediction method based on a neural network model, and the device includes:

[0043] A data acquisition device, used for collecting and preprocessing data;

[0044] The data prediction device is used to predict the intensity and track of a tropical cyclone, and the data prediction device is electrically connected to the data acquisition device.

[0045] Furthermore, the present invention also relates to a computer-readable storage medium having program instructions stored thereon, wherein the program instructions implement the above-mentioned method when executed by a processor.

[0046] Compared with the prior art, the present invention has the following characteristics.

[0047] The present invention is superior to existing data preprocessing methods in reducing baseline model prediction errors through a data preprocessing method; the spatiotemporal characteristics of the data are extracted through three-dimensional convolution and long short-term memory networks; and the long-term prediction capability of the model is enhanced by introducing an attention mechanism, thereby improving the accuracy of long-term predictions of tropical cyclone paths and tropical cyclone intensity. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the typhoon path and intensity prediction method based on the neural network model.

[0049] Figure 2 A flow chart of collecting data and preprocessing the data to obtain preprocessed data in a typhoon path and intensity prediction method based on a neural network model.

[0050] Figure 3 A flowchart for constructing a neural network model in the typhoon path and intensity prediction method based on the neural network model.

[0051] Figure 4 This is a schematic diagram of sliding window processing of data in the typhoon path and intensity prediction method based on the neural network model.

[0052] Figure 5 This is a schematic diagram of the network structure of the long short-term memory network LSTM in the typhoon path and intensity prediction method based on the neural network model.

[0053] Figure 6 Schematic diagram of the stacked convolutional long short-term memory network layers in the typhoon path and intensity prediction method based on the neural network model.

[0054] Figure 7 Schematic diagram of the structure of the residual block in the typhoon path and intensity prediction method based on the neural network model.

[0055] Figure 8 This is a schematic diagram of the structure of the residual neural network in the typhoon path and intensity prediction method based on the neural network model.

[0056] Fig. 9 This is a schematic diagram of the structure of the Encoder-Decoder framework with an attention mechanism in the typhoon path and intensity prediction method based on the neural network model. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The concept, specific structure and technical effects of the present invention will be clearly and completely described below in combination with the embodiments and drawings to fully understand the purpose, scheme and effect of the present invention.

[0059] It should be noted that, unless otherwise specified, when a feature is referred to as being "fixed" or "connected" to another feature, it may be directly fixed or connected to another feature, or it may be indirectly fixed or connected to another feature. The singular forms "a", "said" and "the" used herein are also intended to include the plural forms, unless the context clearly indicates otherwise. In addition, unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art. The terms used in this specification are intended only to describe specific embodiments and are not intended to limit the invention. The term "and / or" used herein includes any combination of one or more of the related listed items.

[0060] It should be understood that although the terms first, second, third, etc. may be used to describe various elements in the present disclosure, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, without departing from the scope of the present disclosure, the first element may also be referred to as the second element, and similarly, the second element may also be referred to as the first element. The use of any and all examples or exemplary language ("for example", "such as", etc.) provided herein is intended only to better illustrate embodiments of the present invention, and unless otherwise required, will not impose limitations on the scope of the present invention. In addition, the industry term "posture" used herein refers to the position and attitude of a certain element relative to a spatial coordinate system.

[0061] Reference Figures 1 to 9 The embodiment of the present invention provides a typhoon path and intensity prediction method based on a neural network model, referring to Figure 1 , the method specifically comprises the following steps:

[0062] S100, collecting data and preprocessing the data to obtain preprocessed data, wherein the preprocessed data at least includes isobaric surface temperature, isobaric surface potential height, isobaric surface relative humidity, ground temperature, ground lift index, and longitude and latitude wind speed;

[0063] S200, build and train neural network models;

[0064] S300, inputting the preprocessed data into a neural network model for prediction, and obtaining tropical cyclone prediction intensity data and tropical cyclone prediction path data.

[0065] Compared with the prior art, the present invention has the following characteristics.

[0066] The present invention is superior to existing data preprocessing methods in reducing baseline model prediction errors through a data preprocessing method; the spatiotemporal characteristics of the data are extracted through three-dimensional convolution and long short-term memory networks; and the long-term prediction capability of the model is enhanced by introducing an attention mechanism, thereby improving the accuracy of long-term predictions of tropical cyclone paths and tropical cyclone intensity.

[0067] Specifically, for step S100, the data set used in the present invention is the global atmospheric reanalysis data jointly produced by the National Center for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR). The content of this data set includes isobaric surface data, ground data and flux data, and has the advantages of many elements, wide spatial range, long time range, high spatiotemporal resolution, etc. The basic information of tropical cyclones uses the HURSAT-B1 best typhoon path of the National Oceanic and Atmospheric Administration (NOAA), which contains the longitude and latitude of all tropical cyclones in the world from their generation to extinction every 3 hours, the maximum average wind speed near the center, the minimum pressure, and the basic information such as typhoon number, year, and region since 1876.

[0068] Further, refer to Figure 2 , step S100 includes:

[0069] S110, based on linear interpolation, clean the data and complete the missing data;

[0070] Specifically, for the missing values ​​x that appear in the reanalysis data ij , i and j represent their positions, and we will use the missing value

[0071] The average value of all non-missing values ​​in the 5×5 matrix centered is filled, and the corresponding formula is:

[0072]

[0073] Among them, take x ab For missing values ​​x ij All non-missing values ​​in a 5×5 matrix centered on ij The number of all non-missing values ​​in the 5×5 matrix centered at x^ ij To fill in the missing value, this method can make the data transition smoother and restore the real variable value as much as possible.

[0074] Since the path and intensity of tropical cyclones do not change dramatically in most cases, linear interpolation is used to fill in the missing values ​​in the tropical cyclone optimal path data. The corresponding formula is:

[0075]

[0076] USA_WIND t is the missing value of the maximum wind speed near the center of the tropical cyclone, USA_WIND t-1 and USA_WIND t+1 is the non-missing value at two time points before and after the missing value. If there are more than two consecutive missing values, they are filled in according to the linear difference:

[0077]

[0078] Among them, USA_WIND t+i , i = 1, 2, ..., n-1 is a number of consecutive missing values ​​of the maximum wind speed near the center of the tropical cyclone, USA_WIND t and USA_WIND t+n are the two nearest non-missing values ​​before and after the missing value.

[0079] For the location data in the best tropical cyclone track dataset, we first uniformly express the latitude of tropical cyclones in the form of -90°-90° and longitude in the form of 0°-360°, which is consistent with the storage format of the reanalysis data. For the missing values ​​of longitude and latitude and the maximum average wind speed near the center, linear interpolation is used to fill in the missing values. The formula is as follows:

[0080]

[0081]

[0082] Among them, USA_LAT t 、USA_LON t are the missing values ​​of the longitude and latitude of the center of the tropical cyclone at time t, USA_LAT t-1 、USA_LON t-1 and USA_LAT t+1 、USA_LON t+1 are the two non-missing values ​​of the center position of the tropical cyclone at time t-1 and time t+1 respectively. If there are more than two consecutive missing values, they are filled in according to the linear difference:

[0083]

[0084]

[0085] Among them, USA_LAT t+1 、USA_LON t+1 , i=1, 2...n-1 are several consecutive missing values ​​of the longitude and latitude of the center of the tropical cyclone at time t+i, USA_LAT t 、USA_LAT t+n and USA_LON t 、USA_LON t+n They are the two nearest non-missing values ​​before and after the missing value of the corresponding longitude and latitude.

[0086] S120, converting the longitude and latitude position data of the tropical cyclone center into vector data, wherein the vector data includes relative direction data and distance data between the current time and the previous time of the tropical cyclone center location;

[0087] Specifically, the longitude and latitude data of the tropical cyclone center are converted into the relative direction and distance between the current time step and the previous time step, thereby eliminating the influence of the tropical cyclone center position at the previous moment on the center position at the current moment. The corresponding calculation formula is:

[0088]

[0089]

[0090] Among them, d is the displacement of the center of the tropical cyclone at the current moment relative to the center of the tropical cyclone at the previous moment, α is the angle of the center of the tropical cyclone at the current moment relative to the center of the tropical cyclone at the previous moment, lat1 and lon1 are the longitude and latitude of the center of the tropical cyclone at the previous moment, lat2 and lon2 are the longitude and latitude of the center of the tropical cyclone at the current moment, and R is the radius of the earth. In practice, the average value of the radius of the earth is about 6371 kilometers for calculation.

[0091] S130, cutting redundant data by sliding window method, cutting out atmospheric variable matrix data of a square area with an actual geographical range of 200 kilometers with the center position coordinates of the tropical cyclone at the current time as the center;

[0092] Specifically, the diameter of a tropical cyclone is usually hundreds of kilometers, and a few can reach thousands of kilometers. In order to remove redundant data and improve computational efficiency, we will crop the NCEP / NCAR reanalysis dataset, that is, crop an atmospheric variable matrix of size 18×18 with the center coordinates of the tropical cyclone at the current time as the center, corresponding to the actual geographical range of a square area with a length and width of 2,000 kilometers. This not only greatly reduces the scale of the original global variables and removes a large amount of redundant data, but also focuses the model's attention on the tropical cyclone itself and the surrounding area, which not only improves the efficiency of training, but also improves the accuracy of the model.

[0093] Generally speaking, the larger the amount of data, the stronger the generalization ability of the model after training. The input of the model is a set of atmospheric environmental variables corresponding to several time steps. In order to make full use of the existing data, we will use a sliding window method to read the data. Figure 4 , the sliding window operation is the case where the time step is 4, where t1, t2, t3...t n are the n time nodes with records in the entire life cycle of a tropical cyclone, x1, x2, x3, ..., x m are m atmospheric variables, y ij t i Time variable x j The input time step of the model is set to k (k<n), so for the tropical cyclone data, each input is as follows,

[0094] The first input data is: y1, y2.......y k ,

[0095] The second input data is: y2, y3......y k+1 , ......

[0097] The last input data is: y n-k+1 ,y n-k+2 ......y n ,

[0098] where y j =(y j1 ,y j2 ,y j3 ......y jm ).

[0099] In addition, since the dimensions of the data are significantly different, normalization can improve the convergence speed of the model on the one hand, and avoid gradient explosion on the other.

[0100] S140. Perform data balance processing, randomly undersample tropical depression, tropical storm and severe tropical storm data to reduce the input samples of tropical depression, tropical storm and severe tropical storm, and randomly oversample typhoon, severe typhoon and super typhoon data to increase the input samples of typhoon, severe typhoon and super typhoon.

[0101] Specifically, sample imbalance will affect the results of the model, because the majority of values ​​have a high frequency of occurrence, and the model may pay more attention to the majority of values ​​during the training process, while ignoring the minority of values, resulting in poor prediction of the minority of values. In a realistic environment, the proportion of tropical depressions far exceeds that of tropical cyclones of several other levels, which will lead to poor generalization of the model and the inability to accurately predict the evolution of typhoons, strong typhoons and super typhoons, which have strong destructive power and are the focus of our prediction. Therefore, the present invention alleviates the poor prediction effect caused by data imbalance by random oversampling and random undersampling. The specific operation is to reduce the input samples of tropical depressions, tropical storms and strong tropical storms by random undersampling before each round of training, and increase the input samples of typhoons, strong typhoons and super typhoons by random oversampling. Randomly selecting samples in each round can greatly reduce the risk of overfitting of the model. The number of occurrences of tropical cyclones of different levels after equilibrium processing in the data set becomes relatively balanced.

[0102] Further, refer to Figure 3 In step S200, the construction of the neural network model includes the following steps:

[0103] S210, construct a single-layer convolutional long short-term memory network layer;

[0104] S220, stacking multiple convolutional long short-term memory network layers to form a multi-layer convolutional long short-term memory network layer;

[0105] S230, connecting the residual neural network to the output of the multi-layer convolutional long short-term memory network layer;

[0106] S240. Connect the attention layer to the output of the residual neural network.

[0107] Specifically, based on the Convolution Long Short-Term Memory (ConvLSTM), a Temporal Attention Mechanism ConvLSTM (TAM-CL) model with a temporal attention mechanism is proposed. Compared with the traditional LSTM model, the ConvLSTM, which is more sensitive to spatiotemporal relationships, is used as the encoder and the two-dimensional convolution is changed to a three-dimensional convolution. It is hoped that the spatial features of the reanalysis data can be extracted through three-dimensional convolution, and the temporal features between the data can be extracted using the structure of the long short-term memory network. In order to improve the shortcomings of the existing model's poor long-term prediction ability, an attention mechanism is also introduced so that the model can focus on the most critical time nodes. Finally, this paper uses residual neural networks and batch normalization in the decoder to improve the stability and ease of use of the training process.

[0108] The experimental part includes tropical cyclone path prediction experiment and intensity prediction experiment. We set corresponding loss functions, evaluation indicators and model parameters for the two different experiments. By comparing the impact of different parameter settings on the experimental results, we selected the best model for comparative analysis with the baseline method to verify the effectiveness of the TAM-CL model.

[0109] Typhoon problems mainly include typhoon generation prediction, typhoon intensity determination, typhoon path prediction, typhoon intensity prediction, typhoon wind and rain prediction, and storm surge prediction. Although many mechanisms of typhoons are still not completely clear, it is certain that the structure, intensity, and movement trajectory of typhoons are closely related to the atmospheric and oceanic variables around the current moment, such as atmospheric humidity, wind speed, sea level temperature, sea level pressure, etc. Since the generation and dissipation of a typhoon is a time series process, its evolution process is not only affected by the current environmental variables, but also by the environmental variables over a period of time in the past, so many problems in typhoon prediction can be regarded as time series prediction problems. The corresponding function of the model is as follows:

[0110]

[0111] in, It is the tth n+1 The output results corresponding to each time step can be typhoon position coordinates, typhoon intensity, etc. is a vector of environmental variables corresponding to the ti-th time step. It is an atmospheric or oceanic environmental variable.

[0112] Further, in step S210, the single-layer convolutional long short-term memory network layer is composed of n cells connected in chronological order, the n neural network cells represent n time steps respectively, and the n neural network cells are connected in series. The output hidden layer and state layer of the neural network cells in each layer are connected in sequence with the input hidden layer and input state layer of the next neural network cell in the current layer.

[0113] Further, refer to Figure 5 , Figure 5 The network structure of the long short-term memory network LSTM, the convolutional long short-term memory network ConvLSTM changes the fully connected structure in the neuron structure to a convolutional structure, that is, Figure 5 The three locations circled in red can better extract the characteristics of the variable space and the relationship between different variables. The calculation formula for a single cell is:

[0114] i t =σ(W xi *x t +W hi *h t-1 +b i ),

[0115] f t =σ(W xf *x t +W hf *h t-1 +b if ),

[0116]

[0117] O t =σ(W xo *x t +W ho *h t-1 +b o ),

[0118]

[0119] Among them, i t represents the input gate vector, f t represents the forget gate vector, c t represents the cell state vector, O t represents the output gate vector, h t represents the hidden state vector, x t represents the input feature vector, W represents the weight matrix, b represents the bias term, σ represents the sigmoid function, represents the Hadamard product and * represents the convolution operation.

[0120] Further, in step S220, the stacked convolutional long short-term memory network layer includes w layers of convolutional long short-term memory network layers, and the output hidden layer of the neural network cells in each layer is connected to the input hidden layer of the neural network cells at the corresponding position in the previous layer.

[0121] Reference Figure 6 , the above is the operation of a single cell in ConvLSTM. Multiple cells are connected in time sequence to form a single-layer ConvLSTM network. In order to enhance the expressive power of the network, multiple layers of networks are usually stacked, x1, x2, ..., x n They represent the input of n time steps, which are input to the first layer of the network model. (0)0 , c (0)0) , (h (1)0 , c (1)0 ),...,(h (0w) , c (0w) ) are the initial hidden layer and state layer for each network layer. (ik) , c (ik) is the k-th layer, the hidden layer and state layer of the output of the i-th time step, h (1w) ,h (2w) , ..., h (nw) is the final output of each time step. n Input the first layer of network structure and get the output h of the first layer (0)1 ,h (0)2 , ..., h (0)n The output of the first layer is used as the input of the second layer, so that the output of the second layer h is obtained. (1)1 ,h (1)2 , ..., h (1)n And so on, we finally get the output h of the last layer. (1w) ,h (2w) , ..., h (nw) At this time, the last time step of the last layer of the network is h (nw) Contains all the information in the past time step, which is a collection of spatial and temporal characteristics of atmospheric variables.

[0122] The decoder decodes the existing information to obtain the final output.

[0123] Further, in step S230, the residual neural network includes a first convolutional layer, a first pooling layer, a first residual block, a second residual block, a third residual block, a fourth residual block and a second pooling layer connected in sequence,

[0124] The size of the first convolutional layer is 7*7*7, the size of the convolution kernel is 64, and the step size is 2;

[0125] The size of the first pooling layer is 3*3*3;

[0126] The first residual block includes a first sub-block and a second sub-block connected in sequence, the number of convolution kernels of the first sub-block is 64, the size of the convolution kernel is 3*3*3, and the step length is 2, the number of convolution kernels of the second sub-block is 64, the size of the convolution kernel is 3*3*3, and the step length is 2;

[0127] The second residual block includes a third sub-block and a fourth sub-block connected in sequence, the number of convolution kernels of the third sub-block is 128, the size of the convolution kernel is 3*3*3, and the step length is 2, and the number of convolution kernels of the fourth sub-block is 128, the size of the convolution kernel is 3*3*3, and the step length is 2;

[0128] The third residual block includes a fifth sub-block and a sixth sub-block connected in sequence, the number of convolution kernels of the fifth sub-block is 256, the size of the convolution kernel is 3*3*3, and the step length is 2, and the number of convolution kernels of the sixth sub-block is 256, the size of the convolution kernel is 3*3*3, and the step length is 2;

[0129] The fourth residual block includes a seventh sub-block and an eighth sub-block connected in sequence, the number of convolution kernels of the seventh sub-block is 512, the size of the convolution kernel is 3*3*3, and the step size is 2, the number of convolution kernels of the eighth sub-block is 512, the size of the convolution kernel is 3*3*3, and the step size is 2.

[0130] Specifically, when the number of layers of a neural network is too large, gradient explosion or vanishing problems are likely to occur, making model training difficult. This problem can be solved to a certain extent by batch normalization (BN) and network initialization methods. In addition, it will cause network degradation, that is, as the number of network layers increases, network optimization becomes more difficult, and its training error and prediction error become larger. The present invention solves this problem by adding a residual structure.

[0131] Reference Figure 7 In theory, the effect of a deeper network model should be no weaker than that of a shallow network model. In fact, we only need to set the extra layers of the deeper network model to the identity mapping. However, it is actually difficult for a deep network model to fit the identity mapping. Therefore, the residual block transforms the objective function from F(x) to F(x)+x through an identity mapping of cross-layer connections, so that the deep network can fit the identity function to achieve an effect that is no weaker than that of the shallow network model.

[0132] In a specific embodiment, referring to Figure 8, using an 18-layer residual neural network, the input of the model first undergoes a convolution operation with 64 convolution kernels of size 7×7×7 and a stride of 2, and a maximum pooling operation of size 3×3×3. It is then input into four types of residual blocks, each with two. The size of the convolution kernel in the residual block is 3×3×3, the stride is 2, and the number of convolution kernels is 64, 128, 256, and 512 respectively. Finally, an average pooling is performed, and the result is input into the fully connected layer and the final result is output. Depending on the prediction problem, the output results are also different. When it is necessary to predict the future intensity of a typhoon, the output result is the maximum wind speed near its center, which is a one-dimensional vector; when predicting the typhoon's trajectory, the output result is a two-dimensional vector, the elements of which are the direction and distance of its movement in the next time step.

[0133] Specifically, refer to Figure 8 For the convenience of explanation, the following describes the parameters of the initial model settings. The number of ConvLSTM layers is initially set to 3. When the number of layers is less than 3, the expression ability is weak and overfitting is prone to occur; when the number of layers is greater than 3, there are too many parameters and it is often difficult to train. The time step is set to 4, and one time step is 6 hours, that is, the position of the tropical cyclone in 6 hours is predicted based on the atmospheric variables in the past 6 hours, 12 hours, 18 hours and the current time. The convolution kernel size in Conv3D of each layer of ConvLSTM is 3×3×3, the step size is 1×1×1, the padding is 1×1×1, and the number of convolution kernels is 16. The activation function uses the Sigmoid function. The convolution kernel size of the first layer in the residual neural network is 7×7×7, the number of convolution kernels is 64, and the step size is 2×2×2; the first and second residual blocks have the same structure, in which there are two convolution layers, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of convolution kernels is 64; the third and fourth residual blocks have the same structure, in which there are two convolution layers, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of convolution kernels is 128; the fifth and sixth residual blocks have the same structure, in which there are two convolution layers, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of convolution kernels is 256; the first and second residual blocks have the same structure, in which there are two convolution layers, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of convolution kernels is 512; finally, a 1×2 vector is output through the fully connected layer.

[0134] Further, in step S240, the attention layer includes an encoder and a decoder, the input of the encoder is the input X1, X2, X3 and X4 corresponding to the four time steps, the hidden states output at each time step are h1, h2, h3, h4, and the output of the encoder is C1, C2 and C3. At the same time, C1, C2 and C3 are the inputs of each step in the decoder, which are linear combinations of hidden states, λ ij, i = 1, 2, 3, 4, j = 1, 2, 3 are weight coefficients, y1, y2, y3 are the outputs of the decoder, corresponding to the predicted values ​​of each future time step, respectively.

[0135] The calculation formulas for C1, C2 and C3 are:

[0136]

[0137] Among them, the coefficient λ ij First, the hidden layer state h of i Convolution operation is performed to extract features, and then a fully connected layer is used for output. The calculation formula is:

[0138] α ij =γ ij tanh(W ij ·Conv(h i )+b ij )

[0139]

[0140] Among them, γ ij , W ij and b ij are all parameters in the fully connected layer, α ij is the weight parameter of the i-th input corresponding to the j-th output.

[0141] Specifically, refer to Fig. 9 , because it is necessary to predict the situation in the next 24h, 48h and 72h respectively, their dependence on the past is completely different. The basic idea of ​​the attention mechanism is to match the input information of each time step with the corresponding output information, calculate the weight corresponding to each input respectively, and then perform weighted summation of the input sequence information according to the weight. x1, x2, x3, x4 are the inputs corresponding to the 4 time steps of the model, and the hidden states of each time step output are h1, h2, h3, h4 respectively. C1, C2, C3 are the inputs of each step in the decoder, which are linear combinations of hidden states, λ ij , i = 1, 2, 3, 4, j = 1, 2, 3 are weight coefficients, y1, y2, y3 are the outputs of the decoder, corresponding to the predicted values ​​of each future time step respectively.

[0142] To α ij Use the softmax function to ensure that all weights add up to 1, and get the final attention weight λ ij Finally, we get C1, C2, and C3. After the attention mechanism is applied, the input can learn the dependency weights of predictions of different durations on the past. After removing some redundant information, it is more accurate in making long-term predictions.

[0143] Furthermore, the present invention also proposes a typhoon path and intensity prediction device based on a neural network model, which is used to implement the above-mentioned typhoon path and intensity prediction method based on a neural network model, and the device includes:

[0144] A data acquisition device, used for collecting and preprocessing data;

[0145] The data prediction device is used to predict the intensity and track of a tropical cyclone, and the data prediction device is electrically connected to the data acquisition device.

[0146] Furthermore, the present invention also relates to a computer-readable storage medium having program instructions stored thereon, wherein the program instructions implement the above-mentioned method when executed by a processor.

[0147] It should be appreciated that the method steps in the embodiments of the present invention can be implemented or implemented by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer readable memory. The method can use standard programming techniques. Each program can be implemented in a high-level process or object-oriented programming language to communicate with a computer system. However, if necessary, the program can be implemented in an assembly or machine language. In any case, the language can be a compiled or interpreted language. In addition, the program can be run on a programmed ASIC for this purpose.

[0148] Furthermore, the operations of the processes described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The processes described herein (or variations and / or combinations thereof) may be performed under the control of one or more computer systems configured with executable instructions, and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that is executed collectively on one or more processors, by hardware, or a combination thereof. The computer program includes a plurality of instructions that may be executed by one or more processors.

[0149] Further, the method can be implemented in any type of computing platform that is operably connected to a suitable computer, including but not limited to a personal computer, a minicomputer, a mainframe, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or in communication with a charged particle tool or other imaging device, etc. Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, an optical read and / or write storage medium, a RAM, a ROM, etc., so that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the process described herein. In addition, the machine-readable code, or portions thereof, can be transmitted via a wired or wireless network. When such media includes instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor, the invention described herein includes these and other different types of non-transitory computer-readable storage media. When programmed according to the methods and techniques of the present invention, the present invention can also include the computer itself.

[0150] The computer program can be applied to input data to perform the functions described herein, thereby converting the input data to generate output data stored in a non-volatile memory. The output information can also be applied to one or more output devices such as a display. In a preferred embodiment of the present invention, the converted data represents physical and tangible objects, including specific visual depictions of physical and tangible objects produced on the display.

[0151] The above is only a preferred embodiment of the present invention. The present invention is not limited to the above implementation. As long as the technical effect of the present invention is achieved by the same means, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention. Within the scope of protection of the present invention, its technical scheme and / or implementation method may have various modifications and changes.

Claims

1. A typhoon path and intensity prediction method based on a neural network model, characterized in that: The method comprises the following steps: S100, collecting data and preprocessing the data to obtain preprocessed data, wherein the preprocessed data at least includes isobaric surface temperature, isobaric surface potential height, isobaric surface relative humidity, ground temperature, ground lift index, and longitude and latitude wind speed; S200, build and train neural network models; S300, inputting the preprocessed data into a neural network model for prediction, and obtaining tropical cyclone prediction intensity data and tropical cyclone prediction path data.

2. The typhoon path and intensity prediction method based on the neural network model according to claim 1 is characterized in that: Step S100 includes: S110, based on linear interpolation, clean the data and complete the missing data; S120, converting the longitude and latitude position data of the tropical cyclone center into vector data, wherein the vector data includes relative direction data and distance data between the current time and the previous time of the tropical cyclone center location; S130, cutting redundant data by sliding window method, cutting out atmospheric variable matrix data of a square area with an actual geographical range of 200 kilometers with the center position coordinates of the tropical cyclone at the current time as the center; S140. Perform data balance processing, randomly undersample tropical depression, tropical storm and severe tropical storm data to reduce the input samples of tropical depression, tropical storm and severe tropical storm, and randomly oversample typhoon, severe typhoon and super typhoon data to increase the input samples of typhoon, severe typhoon and super typhoon.

3. The typhoon path and intensity prediction method based on a neural network model according to claim 1, characterized in that: In step S200, the construction of the neural network model includes the following steps: S210, construct a single-layer convolutional long short-term memory network layer; S220, stacking multiple convolutional long short-term memory network layers to form a multi-layer convolutional long short-term memory network layer; S230, connecting the residual neural network to the output of the multi-layer convolutional long short-term memory network layer; S240. Connect the attention layer to the output of the residual neural network.

4. The typhoon path and intensity prediction method based on the neural network model according to claim 3 is characterized in that: In step S210, a single-layer convolutional long short-term memory network layer is composed of n cells connected in chronological order, the n neural network cells represent n time steps respectively, and the n neural network cells are connected in series. The output hidden layer and state layer of the neural network cells in each layer are connected in sequence with the input hidden layer and input state layer of the next neural network cell in the current layer.

5. The typhoon path and intensity prediction method based on the neural network model according to claim 4 is characterized in that: The calculation formula for a single cell is: i t =σ(W xi *x t +W hi *h t-1 +b i ), f t =σ(W xf *x t +W hf *h t-1 +b if ), O t =σ(W xo *x t +W ho *h t-1 +b o ), Among them, i t represents the input gate vector, f t represents the forget gate vector, c t represents the cell state vector, O t represents the output gate vector, h t represents the hidden state vector, x t represents the input feature vector, W represents the weight matrix, b represents the bias term, σ represents the sigmoid function, ° represents the Hadamard product, and * represents the convolution operation.

6. The typhoon path and intensity prediction method based on a neural network model according to claim 3, characterized in that: In step S220, the stacked convolutional long short-term memory network layer includes w layers of convolutional long short-term memory network layers, and the output hidden layer of the neural network cells in each layer is connected to the input hidden layer of the neural network cells at the corresponding position in the previous layer.

7. The typhoon path and intensity prediction method based on a neural network model according to claim 1, characterized in that: In step S230, the residual neural network includes a first convolutional layer, a first pooling layer, a first residual block, a second residual block, a third residual block, a fourth residual block, and a second pooling layer connected in sequence. The size of the first convolutional layer is 7*7*7, the size of the convolution kernel is 64, and the step size is 2; The size of the first pooling layer is 3*3*3; The first residual block includes a first sub-block and a second sub-block connected in sequence, the number of convolution kernels of the first sub-block is 64, the size of the convolution kernel is 3*3*3, and the step length is 2, the number of convolution kernels of the second sub-block is 64, the size of the convolution kernel is 3*3*3, and the step length is 2; The second residual block includes a third sub-block and a fourth sub-block connected in sequence, the number of convolution kernels of the third sub-block is 128, the size of the convolution kernel is 3*3*3, and the step length is 2, and the number of convolution kernels of the fourth sub-block is 128, the size of the convolution kernel is 3*3*3, and the step length is 2; The third residual block includes a fifth sub-block and a sixth sub-block connected in sequence, the number of convolution kernels of the fifth sub-block is 256, the size of the convolution kernel is 3*3*3, and the step length is 2, and the number of convolution kernels of the sixth sub-block is 256, the size of the convolution kernel is 3*3*3, and the step length is 2; The fourth residual block includes a seventh sub-block and an eighth sub-block connected in sequence, the number of convolution kernels of the seventh sub-block is 512, the size of the convolution kernel is 3*3*3, and the step size is 2, the number of convolution kernels of the eighth sub-block is 512, the size of the convolution kernel is 3*3*3, and the step size is 2.

8. The typhoon path and intensity prediction method based on a neural network model according to claim 1, characterized in that: In step S240, the attention layer includes an encoder and a decoder. The input of the encoder is the input X1, X2, X3 and X4 corresponding to the four time steps. The hidden states output at each time step are h1, h2, h3, h4 respectively. The output of the encoder is C1, C2 and C3. At the same time, C1, C2 and C3 are the inputs of each step in the decoder, which are linear combinations of hidden states. ij ,i=1,2,3,4,j=1,2,3 are weight coefficients, y1,y2,y3 are the outputs of the decoder, corresponding to the predicted values ​​of each future time step, respectively. The calculation formulas for C1, C2 and C3 are: Among them, the coefficient λ ij First, the hidden layer state h of i Convolution operation is performed to extract features, and then a fully connected layer is used for output. The calculation formula is: α ij =γ ij tanh(W ij ·Conv(h i )+b ij ) Among them, γ ij , W ij and b ij are all parameters in the fully connected layer, α ij is the weight parameter of the i-th input corresponding to the j-th output.

9. A typhoon path and intensity prediction device based on a neural network model, used to implement a typhoon path and intensity prediction method based on a neural network model as claimed in any one of claims 1 to 8, characterized in that: The device comprises: A data acquisition device, used for collecting and preprocessing data; The data prediction device is used to predict the intensity and track of a tropical cyclone, and the data prediction device is electrically connected to the data acquisition device. 10 . A computer-readable storage medium having program instructions stored thereon, wherein the program instructions are executed by a processor to implement the method according to claim 1 .

Images