CN115587641B - Multi-generator-based environment factor-guided typhoon multi-trend prediction method - Google Patents

Abstract

The invention discloses a typhoon multi-trend prediction method guided by environmental factors based on multiple generators, which comprises the steps of constructing a typhoon multi-trend prediction model, wherein the typhoon multi-trend prediction model comprises a generation network and a judgment network, the generation network comprises an encoding module and a decoding module, the encoding module comprises a two-dimensional data encoder, a one-dimensional data encoder and an environmental data encoder, the two-dimensional data, the one-dimensional data and the environmental data of typhoon are respectively subjected to feature extraction, the decoding module comprises a generator selector and a plurality of same generators, the generator selector predicts a probability sequence selected by each generator according to the features extracted by the encoding module, selects the generators according to the obtained probability sequences to perform typhoon prediction, and outputs the predicted typhoon sequences. The method extracts features from various heterogeneous meteorological data and gives out reasonable and possible typhoon trends, and provides more references for meteorological forecasting.

Description

Multi-generator-based environment factor-guided typhoon multi-trend prediction method

Technical Field

The application belongs to the technical field of meteorological disaster forecasting, and particularly relates to a typhoon multi-trend prediction method based on environmental factor guidance of multiple generators.

Background

Typhoons are powerful and complex weather systems also known as tropical cyclones, hurricanes or tropical storms. It typically develops in tropical, subtropical and temperate regions and can bring a lot of precipitation and balance global heat. But strong typhoons can cause fatal dangers to ships and work platforms at sea. When the user logs in, the system can cause a plurality of natural disasters, such as strong wind, storm surge, flood and the like. These natural disasters cause serious economic losses and casualties. In order to resist natural disasters caused by typhoon, it is necessary to predict the future trend of typhoon track and strength accurately in advance. Since the development of typhoon is influenced by a large number of factors, the prediction of typhoon is a very difficult task, and it is very necessary to establish a typhoon system which can capture the environmental information of typhoon and can give reasonable prediction trend.

The current authoritative meteorological institution usually uses a numerical prediction method to predict the typhoon development, and the method can simulate various influence factors possibly suffered by the typhoon development process to ensure the accuracy of self prediction. Typhoon is a very complicated weather system, and the simulation process of typhoon numerical prediction not only needs the support of a supercomputer, but also takes a lot of time. Therefore, the typhoon prediction method only needs a plurality of GPU deep learning to be widely concerned by people, and the current methods based on most of the deep learning can only give a determined prediction trend, which is not in line with the complexity and uncertainty of typhoon. Typhoon prediction methods that give a definite prediction alone are not accepted by meteorologists because the reference meaning of their prediction results is too small. In addition, the current typhoon prediction method based on deep learning does not pay attention to the importance of feature extraction of environmental factors of typhoons on typhoon prediction.

Disclosure of Invention

The application aims to provide a multi-generator-based environmental factor guided typhoon multi-trend prediction method to overcome a plurality of problems of the existing deep learning-based typhoon prediction method, such as: the problem that reliable multi-trend typhoon prediction results cannot be given, the problem that typhoon environment information characteristics cannot be well extracted and the like.

In order to achieve the purpose, the technical scheme of the application is as follows:

a multi-generator-based environmental factor guided typhoon multi-trend prediction method comprises the following steps:

the method comprises the steps of constructing a typhoon multi-trend prediction model, wherein the typhoon multi-trend prediction model comprises a generation network and a judgment network, the generation network comprises an encoding module and a decoding module, the encoding module comprises a two-dimensional data encoder, a one-dimensional data encoder and an environment data encoder, feature extraction is respectively carried out on typhoon two-dimensional data, one-dimensional data and environment data, the decoding module comprises a generator selector and a plurality of same generators, the generator selector predicts a probability sequence selected by each generator according to features extracted by the encoding module, the generators are selected according to the obtained probability sequences to carry out typhoon prediction, and predicted typhoon sequences are output; the discrimination network is used for judging the probability value of the authenticity of the input typhoon sequence;

and acquiring a training data set, training the typhoon multi-trend prediction model, and predicting the typhoon by adopting the trained typhoon multi-trend prediction model.

Further, the two-dimensional data encoder comprises a 3D U-net convolutional neural network and a full connection layer, the one-dimensional data encoder comprises two full connection layers and a plurality of LSTM units, and the environment data encoder comprises a full connection layer and a convolutional layer;

the two-dimensional data encoder is used for extracting spatial features of the two-dimensional data and generating the two-dimensional data of a future preset time point, the one-dimensional data encoder is used for extracting time sequence features of the one-dimensional data and the two-dimensional data, and the environment data encoder is used for extracting environment data features;

the input of the one-dimensional data encoder is fusion data of spatial characteristics of the one-dimensional data and the two-dimensional data.

Further, the one-dimensional data comprises typhoon longitude and latitude and central wind speed and air pressure, and the two-dimensional data comprises potential height data of a preset isobaric surface.

Further, the generator selector comprises a plurality of fully-connected layers, and the generator comprises a plurality of identical LSTM units and a fully-connected layer;

the output of the environment data encoder and the output of the one-dimensional data encoder are input to a generator selector after splicing and fusion, and the generator selector outputs the probability of each generator being selected;

and the output of the environment data encoder, the output of the one-dimensional data encoder and the output characteristic of the two-dimensional data encoder are spliced and input into each generator after splicing.

Further, the discrimination network includes two fully-connected layers, a plurality of LSTM units, and another fully-connected layer, which are connected in sequence.

According to the multi-generator-based environmental factor-guided typhoon multi-trend prediction method, the data required for prediction comprise one-dimensional data, two-dimensional grid data and additional environmental data. The one-dimensional data are longitude and latitude of typhoon, wind speed and air pressure data. The two-dimensional grid data contains potential height data of 500hPa representing a typhoon pressure structure. The environmental data comprise months, typhoon moving speed, typhoon moving direction in the past 24 hours, typhoon intensity change in the past 24 hours, the range of high pressure of the subtropical zone and the position of the typhoon, and can extract characteristics from various heterogeneous meteorological data and give out reasonable and multiple possible typhoon trends, thereby providing more references for meteorological forecast. The method and the device extract the characteristics of the typhoon and the environmental characteristics of the typhoon from the heterogeneous meteorological data, can select a plurality of reasonable generators for typhoon prediction by utilizing the characteristics, and finally provide a plurality of possible future typhoon trends.

Drawings

FIG. 1 is a flow chart of a typhoon multi-trend prediction method of the present application;

FIG. 2 is a structural diagram of a typhoon multi-trend prediction model of the present application.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

In one embodiment, as shown in fig. 1, there is provided a multi-generator-based environmental factor guided typhoon multi-trend prediction method, including:

s1, building a typhoon multi-trend prediction model, wherein the typhoon multi-trend prediction model comprises a generation network and a judgment network, the generation network comprises an encoding module and a decoding module, the encoding module comprises a two-dimensional data encoder, a one-dimensional data encoder and an environment data encoder, feature extraction is respectively carried out on typhoon two-dimensional data, one-dimensional data and environment data, the decoding module comprises a generator selector and a plurality of identical generators, the generator selector predicts a probability sequence selected by each generator according to the features extracted by the encoding module, selects the generators according to the obtained probability sequences to carry out typhoon prediction, and outputs the predicted typhoon sequences; and the discrimination network is used for calculating the true probability value of the predicted typhoon sequence.

The typhoon multi-trend prediction model constructed in the embodiment is a generation countermeasure network framework including a generation network and a discrimination network. Wherein, the generation network is composed of two modules: an encoding module and a decoding module.

The encoding module consists of three components: a two-dimensional data encoder, a one-dimensional data encoder, and an environmental data encoder.

The two-dimensional data encoder comprises a 3D-Unet convolutional neural network and a full connection layer, and is used for extracting spatial features of the two-dimensional data and generating the two-dimensional data of a future preset time point. The number of the predetermined time points depends on the length of the predicted time. The number of input channels and the number of output channels of the 3D-Unet are both 1, the weight value of the whole 3D-Unet convolution neural network is W _ U _1, and the weight value is defined as a floating point type variable without bias. The two-dimensional data features output by the 3D-Unet convolutional neural network are extracted again through a full connection layer (embedding), the number of the neurons of the full connection layer is 32, the weight value is W _2D, the floating point type variable is defined, and the floating point type variable is defined when the bias value is b _ 2D.

The 3D-Unet convolutional neural network has two outputs, the upward arrow in FIG. 2 refers to the spatial characteristics of the two-dimensional data extracted by the 3D-Unet, and the rightward arrow outputs the two-dimensional data for generating a preset time point in the future. 3D-Unet is a well-established technique in the art and will not be described herein. Two-dimensional data (represented as two-dimensional data characteristics in fig. 2) at a future preset time point passes through a full connection layer, and then output characteristics of a two-dimensional data encoder are output.

The one-dimensional data encoder consists of two fully-connected layers and several identical LSTM units. In this embodiment, the number of the two full connection layer neurons is set to 32 and 64, the weights are W _1 and W _2, and are defined as floating point type variables, and the offsets are b _1 and b _2, and are defined as floating point type variables. The number of hidden layer neurons of the LSTM unit is set to be 64, the weight value is W _ L _1, the number is defined as a floating point type variable, and no bias exists. The number of LSTM units is set to 8.

The environment data encoder is composed of a plurality of fully connected layers and a plurality of convolutional layers. In this embodiment, the number of the fully connected layers is set to 12, the first 9 fully connected layers are used for feature extraction of different environment information, the number of neurons thereof is 16, and 9 weights are W _3 to W _12, and are defined as floating point type variables. The 9 offsets are b _3 to b _12, respectively, and are defined as floating point type variables. The number of convolutional layers is set to 1, the size of the convolutional kernel is 3, the number of the convolutional kernels is 1, the convolution step length is 1, and the filling width is 1. The last three fully-connected layers are used for fusing different environment characteristics, the number of the neurons is 104, 104 and 32 respectively, the weights are W _13, W _14and W _15respectively, and the variables are defined as floating point type variables. The offsets are b _13 to b _15, respectively, and are defined as floating point type variables.

It should be noted that, all the fully-connected layers in the coding module are followed by the Relu activation function, and all the convolutional layers are followed by the normalization layer and the Relu activation function.

The input of the one-dimensional data encoder is fusion data of the spatial characteristics of the one-dimensional data and the two-dimensional data, namely the spatial characteristics of the one-dimensional data and the two-dimensional data are spliced and then input to a full connection layer of the one-dimensional data encoder, and after fusion processing of the full connection layer, an LSTM unit is used for extracting the time sequence characteristics after fusion.

The decoding module mainly comprises a generator selector and a plurality of same generators. In the embodiment, the number of generators is set to be 6, the generator selector consists of three fully-connected layers, and the first fully-connected layer and the second fully-connected layer are followed by the Relu activation function. The number of the neurons of the three fully-connected layers is 32, 32,6 respectively, the weights are W _16, W _17and W _18respectively, and the variables are defined as floating point type variables. The offsets are b _16 to b _18, respectively, and are defined as floating point type variables. The generator selector predicts the probability of each generator being selected according to the features extracted by the encoding module, and then selects 6 generators for final typhoon prediction by using a Mongolian card Luo Caiyang method according to the probability sequence. Each generator consists of two full-connection layers and a plurality of same LSTM units, the number of neurons of the two full-connection layers is respectively set to be 32 and 64, the weights are W _19 and W _20 and are defined as floating point type variables, and the offsets are b _19 and b _20 and are defined as floating point type variables; the number of hidden layer neurons of the LSTM unit is set to be 64, the weight value is W _ L _2, the number is defined as a floating point type variable, no bias exists, and the number of the LSTM unit is set to be 4. The full-connection layer of the generator is used for fusing and extracting the features extracted by the coding module and the predicted two-dimensional data, the LSTM unit is used for extracting the time sequence features and predicting typhoon, and Relu activation functions are connected behind all the full-connection layers in the generator.

The output of the environment data encoder and the output of the one-dimensional data encoder are input to the generator selector after being spliced and fused, and the generator selector outputs the probability that each generator is selected. For example, the probability of selecting generator G1 is 0.25, the probability of selecting generator G2 is 0.15, …, and the probability of selecting generator Gk is 0.10.

As shown in fig. 2, the feature splicing module is configured to splice an output of the environment data encoder, an output of the one-dimensional data encoder, and an output feature of the two-dimensional data encoder, and input the spliced output features to each generator.

In this example, 6 selected generators are used to generate the predicted typhoon sequence in turn. Finally, 6 predicted possible typhoon trends are obtained, and the method is more valuable to meteorological experts.

The discrimination network of the embodiment is composed of two full-connection layers, a plurality of LSTM units and another full-connection layer and is used for judging the probability value of the authenticity of the input typhoon sequence. The number of the first two full connection layer neurons is set to be 32, the weight values are W _21 and W _22, the floating point type variables are defined, the bias values are b _21 and b _22, the floating point type variables are defined, and then the Relu activation function is connected. The number of hidden layer neurons of the LSTM unit is set to be 128, the weight value is W _ L _3, the number is defined as a floating point type variable, and no bias exists. The number of LSTM units is set to 12. The number of the second full connection layer neurons is set to be 2, the weight value is W _23, the floating point type variable is defined, the bias value is b _24, and the floating point type variable is defined.

And S2, acquiring a training data set, training the typhoon multi-trend prediction model, and predicting the typhoon by adopting the trained typhoon multi-trend prediction model.

After the typhoon multi-trend prediction model is built, a training data set is obtained to train the typhoon multi-trend prediction model. Specifically, the data set is divided into a training set, a verification set and a test set. The test set contained 82 typhoon data that occurred in western pacific and south-ocean waters of china between 2017 and 2019. The training and validation sets contained typhoon data from 1950-2016. The ratio of 80% to 20% is randomly divided. The training set contains 1303 typhoon data and the verification set contains 318 typhoon data. And carrying out standardized processing on the one-dimensional data, the two-dimensional data and the environmental data in the data set, so that the data is more beneficial to the training of the model.

During training, initializing a generated network and judging network parameters by using random weights, setting 1 iteration of the generated network, judging that 1 iteration of the network is one complete network training, and totally performing 200 complete training. The learning rates of the discrimination network and the generation network were set to 0.0001. The training process is as follows:

the sequence including typhoon one-dimensional data (typhoon longitude and latitude, central wind speed and air pressure) and typhoon two-dimensional data (potential height data of a preset equal pressure surface, such as the potential height data of 500hPa, which represents the height of each coordinate point on the 500hPa equal pressure surface from the ground in a certain area, and different meteorological phenomena can affect the height of the 500hPa equal pressure surface) and typhoon environment data (environment data includes month, typhoon moving speed, typhoon moving direction in the past 24 hours, typhoon intensity change in the past 24 hours, range of subtropical zone high pressure and position of typhoon) are input into the generation network.

And a corresponding encoder in the generation network performs characteristic extraction on the input data of the plurality of modes.

The decoding module decodes the features, selects a proper generator to carry out typhoon prediction and finally gives out a plurality of possible future trends of typhoons.

And calculating the error of the predicted typhoon and the true value of the typhoon to obtain the L2 loss, feeding the error back to the network, adjusting the parameters of the network and optimizing the prediction performance.

And inputting the predicted typhoon and the real typhoon sequence into a discrimination network, and calculating the real probability value of each typhoon sequence.

And calculating cross entropy loss according to the probability value and the corresponding label of each typhoon sequence obtained by calculation, minimizing a loss function by using an Adam algorithm, training the model to be convergent, and obtaining an optimal performance model to predict typhoons.

Through the operation of the steps, the training of the typhoon multi-trend prediction model is completed, and the prediction of various possible trends of typhoon can be realized by adopting the trained typhoon multi-trend prediction model.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, and these are all within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims (2)

1. A multi-generator-based environmental factor guided typhoon multi-trend prediction method is characterized by comprising the following steps:

the method comprises the steps that a typhoon multi-trend prediction model is built, the typhoon multi-trend prediction model comprises a generation network and a judgment network, the generation network comprises an encoding module and a decoding module, the encoding module comprises a two-dimensional data encoder, a one-dimensional data encoder and an environment data encoder, and the decoding module comprises a generator selector and a plurality of identical generators;

the two-dimensional data encoder comprises a 3D U-net convolutional neural network and a full connection layer, the one-dimensional data encoder comprises two full connection layers and a plurality of LSTM units, and the environment data encoder comprises a full connection layer and a convolutional layer;

the generator selector comprises a plurality of fully connected layers, and the generator comprises a fully connected layer and a plurality of same LSTM units;

the discrimination network comprises two full connection layers, a plurality of LSTM units and another full connection layer which are connected in sequence;

the two-dimensional data encoder is used for extracting spatial features of the two-dimensional data and generating the two-dimensional data of a future preset time point;

the input of the one-dimensional data encoder is fusion data of spatial characteristics of the one-dimensional data and the two-dimensional data, and the fusion data is used for extracting time sequence characteristics;

the environment data encoder is used for extracting environment data characteristics;

the output of the environment data encoder and the output of the one-dimensional data encoder are input to a generator selector after splicing and fusion, and the generator selector outputs the probability of each generator being selected;

the output of the environment data encoder, the output of the one-dimensional data encoder and the output characteristics of the two-dimensional data encoder are spliced and input into each generator after splicing;

the generator selector predicts the probability sequence selected by each generator according to the features extracted by the coding module, selects the generator to predict typhoon according to the obtained probability sequence and outputs the predicted typhoon sequence; the discrimination network is used for judging the probability value of the authenticity of the input typhoon sequence;

and acquiring a training data set, training the typhoon multi-trend prediction model, and predicting the typhoon by adopting the trained typhoon multi-trend prediction model.

2. The multi-generator environmental factor-guided typhoon multi-trend prediction method as claimed in claim 1, wherein the one-dimensional data comprises typhoon longitude and latitude and central wind speed air pressure, and the two-dimensional data comprises potential height data of a preset isobaric surface.

Images