A radar echo extrapolation method, system and storage medium

Abstract

The application discloses a radar echo extrapolation method and system in the field of electronic communication and a storage medium, and comprises the following steps: extracting radar echo data color pictures to be tested from an obtained real-time radar echo image sequence, inputting the radar echo data color pictures to be tested into a pre-trained Sat-Lstm-RNN network model after pretreatment, obtaining radar echo prediction values, and restoring the radar echo prediction values into radar echoes according to the radar echo prediction values; the training process of the Sat-Lstm-RNN network model comprises the following steps: constructing a Sat-Lstm-RNN network model based on an encoder, a Sat-ConvLstm module and a decoder; training the Sat-Lstm-RNN network model through a training image set, establishing a loss function according to radar echo prediction values output by the Sat-Lstm-RNN network model and real radar echo values, adjusting recurrent neural network parameters by using the loss function, and repeatedly iterating the training process until the Sat-Lstm-RNN network model converges; and the radar echo extrapolation method has higher accuracy and clearer prediction results.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of meteorological disaster short-impending forecast, and particularly relates to a radar echo extrapolation method and system. BACKGROUND

[0002] Radar echo extrapolation is to predict the shape, position, intensity and motion information of future echo according to existing historical radar echo data, so as to realize the tracking and prediction of severe convective weather system, which is of great significance for disaster prevention and reduction, dangerous weather warning and protection of national life and property safety. Radar echo extrapolation can be regarded as a spatiotemporal sequence prediction problem, which predicts the spatiotemporal sequence 1-3 hours in the future by predicting the spatiotemporal sequence before the time node. Traditional radar echo extrapolation often adopts single centroid method, cross correlation method and optical flow method, which are often difficult to predict complex and changeable severe convective weather.

[0003] With the development of deep learning, radar echo extrapolation has a new solution. Compared with the traditional numerical weather prediction method, it can realize faster and more accurate prediction. However, the current radar echo extrapolation method based on deep learning often only considers the sequence of the past one time step, which usually produces fuzzy prediction results, and cannot model the radar echo intensity with multi-modal and skewed distribution, making it difficult to accurately predict rapidly changing severe convective weather. SUMMARY

[0004] The purpose of the present application is to overcome the problems of inaccurate and unclear prediction results of severe convective weather in the prior art, and provide a radar echo extrapolation method, system and storage medium, which realizes a radar echo extrapolation method with higher accuracy and clearer prediction results.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is:

[0006] The present application provides a radar echo extrapolation method in the first aspect, comprising:

[0007] The radar echo data color chart to be tested is extracted from the obtained real-time radar echo image sequence, and after the radar echo data color chart to be tested is preprocessed, it is input into a pre-trained Sat-Lstm-RNN network model to obtain a radar echo prediction value, and the radar echo prediction value is restored into a radar echo;

[0008] The training process of the Sat-Lstm-RNN network model comprises:

[0009] The historical radar echo image sequence is acquired, and radar echo data color images to be trained are extracted from the historical radar echo image sequence at a set time interval; the radar echo data color images to be trained are sequentially subjected to grayscale value mapping and normalization processing to form normalized grayscale data, and a training image set is constructed;

[0010] The Sat-Lstm-RNN network model is constructed based on the encoder, the Sat-ConvLstm module, and the decoder; the Sat-Lstm-RNN network model is trained by using the training image set, a loss function is established according to radar echo predicted values output by the Sat-Lstm-RNN network model and radar echo true values, the recurrent neural network parameters are adjusted by using the loss function, and the training process is repeatedly iterated until the Sat-Lstm-RNN network model converges.

[0011] Preferably, the preprocessing of the radar echo data color image to be tested comprises sequentially performing grayscale value mapping and normalization processing on the radar echo data color image to be tested.

[0012] Preferably, the method for performing grayscale value mapping and normalization processing on the radar echo data color image comprises:

[0013] The radar echo data color image to be trained or the radar echo data color image to be tested with a channel dimension of 3 is mapped to grayscale data with a channel dimension of 1, and the expression formula is:

[0014] GRAY = B·0.114 + G·0.587 + R·0.299

[0015] In the formula, R represents the brightness of a pixel in the red channel of the radar echo data color image, G represents the brightness of the pixel in the green channel of the radar echo data color image, B represents the brightness of the pixel in the blue channel of the radar echo data color image, and GRAY represents the brightness of the pixel in the grayscale channel of the radar echo data color image.

[0016] The grayscale data is subjected to normalization processing to obtain normalized grayscale data, and the expression formula is:

[0017]

[0018] In the formula, Pixel represents the normalized grayscale data.

[0019] Preferably, the method for constructing the Sat-Lstm-RNN network model based on the encoder, the Sat-ConvLstm module, and the decoder comprises:

[0020] The Sat-ConvLstm module is stacked after the encoder and L layers are set; and the output of the Sat-ConvLstm module of the Lth layer is set as the input of the decoder.

[0021] Preferably, the method for training the Sat-Lstm-RNN network model by the training image set comprises:

[0022] performing time and space feature information X on the normalized gray data in the training image set by using an encoder to down-sample dw ; inputting the time and space feature information X dw to the Sat-ConvLstm module with L layers stacked; the Sat-ConvLstm module comprises an attention module and a recurrent neural network module;

[0023] setting the hidden state of the t time step output by the Sat-ConvLstm module of the l layer as hidden state setting the hidden state of the t time step output by the Sat-ConvLstm module of the l-1 layer as hidden state l is in the range of 1 to L; setting hidden state hidden state of the previous τ time steps is a full zero tensor; the hidden state hidden state of the previous τ time steps is recorded as hidden state

[0024] inputting the time and space feature information X dw , the hidden state and the hidden state to the attention module of the l layer Sat-ConvLstm module to obtain motion information H AMI ;

[0025] inputting the motion information H AMI , the time and space feature information X dw and the cell state to the recurrent neural network module of the l layer Sat-ConvLstm module to obtain the cell state and the hidden state

[0026] inputting the cell state and the hidden state to the next layer Sat-ConvLstm module, and repeating iteration until the cell state and the hidden state output by the Sat-ConvLstm module of the L layer are obtained

[0027] performing up-sampling on the cell state and the hidden state by using a decoder to output a radar echo prediction value.

[0028] Preferably, the time and space feature information X dw , the hidden state and the hidden state are input into the attention module of the l-th Sat-ConvLstm module to obtain the motion information H AMI The method comprises the following steps:

[0029] The attention score α j of the hidden state is calculated, and the expression formula is as follows:

[0030]

[0031]

[0032]

[0033] In the formula, ⊙ represents the Hadamard product;

[0034] The long-term motion information H att is calculated based on the attention score α j , and the expression formula is as follows:

[0035]

[0036] The motion information H AMI is calculated using the long-term motion information H att , the time and space feature information X dw , and the expression formula is as follows:

[0037]

[0038] U f = σ (W f * X dw )

[0039] Wherein, σ is a sigmoid function; * represents convolution operation, and W f represents the weight parameter of the time and space feature information X dw .

[0040] Preferably, the motion information H AMI , the time and space feature information X dw and the cell state are input into the recurrent neural network module of the l-th Sat-ConvLstm module to obtain the cell state and the hidden state of the t-th time step, and the expression formula is as follows:

[0041] i t = σ (W xi * Xdw +W hi *H AMI +b i

[0042]

[0043]

[0044] o t =σ(W xo *X dw +W ho *H AMI +b o

[0045]

[0046] In the formula, * represents convolution operation, represents Hadamard product, tanh represents hyperbolic tangent activation function, and represents Sigmoid function. t represents the expression function of the input gate; t represents the expression function of the forget gate, t represents the expression function of the output gate; g , b i , b f , b o respectively represent the set bias parameters, W xi represents the weight matrix corresponding to X dw in the input gate, W hi represents the weight matrix corresponding to H AMI in the input gate, W xf represents the weight matrix corresponding to X dw in the forget gate, W hf represents the weight matrix corresponding to H AMI in the forget gate, W cf represents the weight matrix corresponding to in the forget gate, W xg represents the weight matrix corresponding to X dw in the cell state update, W hg represents the weight matrix corresponding to H AMI in the cell state update, W xo represents the weight matrix corresponding to X dw in the output gate, and W ho represents the weight matrix corresponding to H AMI in the output gate.

[0047] Preferably, the method for establishing a loss function according to the radar echo predicted value output by the Sat-Lstm-RNN network model and the radar echo true value comprises: ​​

[0048]

[0049] In the formula, y t(n,h,w) represents the real value of the radar echo at the t time step; represents the predicted value of the radar echo output by the Sat-Lstm-RNN network model at the t time step; n represents a sequence of radar echoes; w represents the horizontal coordinate of the radar echo; and h represents the vertical coordinate of the radar echo.

[0050] The second aspect provides a radar echo extrapolation system, comprising:

[0051] A prediction module is configured to extract a radar echo data color map to be tested from an obtained real-time radar echo image sequence, input the radar echo data color map to be tested after preprocessing into a pre-trained Sat-Lstm-RNN network model to obtain a radar echo predicted value, and restore the radar echo predicted value into a radar echo.

[0052] An acquisition module is configured to acquire a historical radar echo image sequence, and extract radar echo data color maps to be trained from the historical radar echo image sequence at a set time interval; and sequentially perform grayscale value mapping and normalization processing on the radar echo data color maps to be trained to form normalized grayscale data, and construct a training image set.

[0053] A construction module is configured to construct a Sat-Lstm-RNN network model based on an encoder, a Sat-ConvLstm module, and a decoder.

[0054] A training module is configured to train the Sat-Lstm-RNN network model through the training image set, establish a loss function according to a radar echo predicted value output by the Sat-Lstm-RNN network model and a radar echo real value, adjust recurrent neural network parameters by using the loss function, and repeat an iteration training process until the Sat-Lstm-RNN network model converges.

[0055] The third aspect of the present application provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the radar echo extrapolation method.

[0056] Compared with the prior art, the present application has the following beneficial effects:

[0057] The present application sequentially performs grayscale value mapping and normalization processing on the radar echo data color maps to be trained to form normalized grayscale data, which not only retains more radar echo information, but also does not occupy too many resources, reduces the training time, and improves the training efficiency.

[0058] The application is used for constructing a Sat-Lstm-RNN network model based on an encoder, a Sat-ConvLstm module and a decoder; the Sat-Lstm-RNN network model is trained through a training image set, a loss function is established according to radar echo predicted values and radar echo true values output by the Sat-Lstm-RNN network model, a recurrent neural network parameter is adjusted by using the loss function, and a repeated iteration training process is performed until the Sat-Lstm-RNN network model converges; the Sat-Lstm-RNN network model saves more predicted visual details, and the prediction accuracy and clarity are improved. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is a flowchart of a radar echo extrapolation method provided by an embodiment of the application;

[0060] Figure 2 is a structural diagram of an attention module provided by an embodiment of the application;

[0061] Figure 3 is a structural diagram of a Sat-ConvLstm module provided by an embodiment of the application;

[0062] Figure 4 is a structural diagram of a Sat-Lstm-RNN network model provided by an embodiment of the application. DETAILED DESCRIPTION

[0063] The application will be further described below in combination with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0064] Embodiment one

[0065] As shown in Figures 1 to 4 , a radar echo extrapolation method comprises the following steps:

[0066] Radar echo data color images to be tested are extracted from acquired real-time radar echo image sequences, gray value mapping and normalization processing are performed on the radar echo data color images to be tested to form normalized gray data, the normalized gray data are input to a pre-trained Sat-Lstm-RNN network model to obtain radar echo predicted values, and the radar echo predicted values are restored into radar echoes;

[0067] The training process of the Sat-Lstm-RNN network model comprises the following steps:

[0068] The historical radar echo image sequence is acquired, and radar echo data color images to be trained are extracted from the historical radar echo image sequence at a set time interval; the height of the radar echo data color image is 480, the width is 480, and the channel number C is 1; the radar echo data color image to be trained is sequentially subjected to grayscale value mapping and normalization processing to form normalized grayscale data, thereby constructing a training image set; a total length of a set of training data is set as 20, wherein the input sequence length is 10, and the prediction sequence length is 10;

[0069] The method for performing grayscale value mapping and normalization processing on the radar echo data color image comprises:

[0070] The radar echo data color image to be trained or the radar echo data color image to be tested with a channel dimension of 3 is mapped to grayscale data with a channel dimension of 1, and the expression formula is:

[0071] GRAY = B·0.114 + G·0.587 + R·0.299

[0072] In the formula, R represents the brightness of a pixel in the red channel of the radar echo data color image, G represents the brightness of the pixel in the green channel of the radar echo data color image, B represents the brightness of the pixel in the blue channel of the radar echo data color image, and GRAY represents the brightness of the pixel in the grayscale channel of the radar echo data color image.

[0073] The grayscale data is subjected to normalization processing to obtain normalized grayscale data, and the expression formula is:

[0074]

[0075] In the formula, Pixel represents the normalized grayscale data.

[0076] The method for constructing a Sat-Lstm-RNN network model based on an encoder, a Sat-ConvLstm module and a decoder comprises:

[0077] The Sat-ConvLstm module is stacked after the encoder, and the Lth layer of the Sat-ConvLstm module is set as the input of the decoder. The encoder and the decoder comprise at least 3 convolutional layers, each layer of the encoder is a down-sampling layer, a LayerNorm method is introduced after each down-sampling layer, the size of a convolution kernel is set as 1, the number of convolution kernels is set as 64, the step is set as 2, and the pixel padding is set as 0.

[0078] The Sat-ConvLstm module comprises an attention module and a recurrent neural network module; the recurrent neural network module comprises a forget gate, an input gate and an output gate; the task of the forget gate is to accept a past information requiring long-term memory and to determine which memory information to retain and forget; the task of the input gate is to extract information from a current time sequence and selectively retain in a current memory cell; the task of the output gate is to selectively output the information retained by the forget gate and the input gate; each gate comprises at least one convolution layer, and each convolution layer is followed by a LayerNorm method to accelerate the convergence of the Sat-Lstm-RNN network model.

[0079] In the embodiment, the convolution kernel size F of the hidden layer of the Sat-ConvLstm module is set to 5, the convolution kernel quantity hidden_channel is set to 64, the step S is set to 1, the stacking layer number L of the Sat-ConvLstm module is set to 4, and the batch size is set to 16.

[0080] The method for training the Sat-Lstm-RNN network model by using the training image set comprises the following steps:

[0081] The time and space feature information X is obtained by using an encoder to down-sample normalized gray data in the training image set. dw The side length N of the down-sampled image is

[0082] N=(W-F+2P) / S+1

[0083] Wherein, W is the side length of the input image, F is the convolution kernel size, P is the padding pixel, S is the step, and N is the side length of the down-sampled image.

[0084] The time and space feature information X is input into the Sat-ConvLstm module stacked with L layers. dw

[0085] The hidden state of the Sat-ConvLstm module output by the lth layer at the tth time step is set as hidden state The hidden state of the Sat-ConvLstm module output by the (l-1)th layer at the tth time step is set as hidden state The value range of l is 1 to L; and The hidden state of the Sat-ConvLstm module output by the lth layer at the tth time step is set as hidden state The hidden state of the previous τ time steps is a zero tensor; and the hidden state of the Sat-ConvLstm module output by the lth layer at the tth time step is set as hidden state The hidden state of the previous τ time steps is set as hidden state

[0086] The time and space feature information X is input into the Sat-ConvLstm module stacked with L layers. dw The hidden state of the Sat-ConvLstm module output by the lth layer at the tth time step is set as hidden state The hidden state of the Sat-ConvLstm module output by the lth layer at the tth time step is set as hidden state ​The attention module inputted to the l-th layer Sat-ConvLstm module obtains motion information H AMI The method comprises the following steps:

[0087] The hidden state The attention score α j is calculated, and the expression formula is as follows:

[0088]

[0089]

[0090]

[0091] In the formula, ⊙ represents Hadamard product;

[0092] The long-term motion information H j is calculated based on the attention score α att , and the expression formula is as follows:

[0093]

[0094] The motion information H att is calculated by using the long-term motion information H dw , the time and space feature information X AMI , and the expression formula is as follows:

[0095]

[0096] U f = σ (W f *X dw )

[0097] Wherein, σ is a sigmoid function; * represents convolution operation, and W f represents a weight parameter of the time and space feature information X dw .

[0098] The motion information H AMI , the time and space feature information X dw and the cell state are inputted to the recurrent neural network module of the l-th layer Sat-ConvLstm module to obtain the cell state and the hidden state of the t-th time step, and the expression formula is as follows:

[0099] i t = σ (W xi *X dw +W hi *H AMI +b i )

[0100]

[0101]

[0102] o t = σ(W xo *X dw +W ho *H AMI +b o )

[0103]

[0104] In the formula, * represents convolution operation, represents Hadamard product, tanh represents hyperbolic tangent activation function, and represents Sigmoid function. t represents the expression function of the input gate. t represents the expression function of the forgetting gate. t represents the expression function of the output gate. g , b i , b f , b o respectively represent the set bias parameters. xi represents the weight matrix corresponding to X dw in the input gate. hi represents the weight matrix corresponding to H AMI in the input gate. xf represents the weight matrix corresponding to X dw in the forgetting gate. hf represents the weight matrix corresponding to H AMI in the forgetting gate. cf represents the weight matrix corresponding to in the forgetting gate. xg represents the weight matrix corresponding to X dw in the cell state update. hg represents the weight matrix corresponding to H AMI in the cell state update. xo represents the weight matrix corresponding to X dw in the output gate. ho represents the weight matrix corresponding to H AMI in the output gate.

[0105] The cell state and the hidden state are input to the next layer Sat-ConvLstm module, and the iteration is repeated until the cell state and the hidden state

[0106] Using a decoder on the cell state and the hidden state Upsampling the output radar echo prediction value.

[0107] The method for establishing a loss function according to the radar echo prediction value output by the Sat-Lstm-RNN network model and the true value of the radar echo comprises:

[0108]

[0109] In the formula, y t(n,h,w) represents the true value of the radar echo at the t time step; represents the radar echo prediction value output by the Sat-Lstm-RNN network model at the t time step; n represents the sequence of the radar echo; w represents the horizontal coordinate of the radar echo; and h represents the vertical coordinate of the radar echo.

[0110] Adjust the parameters of the recurrent neural network according to the loss function using an Adam optimizer, and the learning rate lr of the Adam optimizer is:

[0111] lr = lr··(gamma) iter

[0112] The formula is, the initial learning rate lr is 0.01, gamma is 0.9, and iter is the number of current iterations.

[0113] Reading data in the training image set once is equal to 1 iteration, the maximum number of iterations is 20,000 times, and after training 5,000 times, whenever the loss of the model is lower than the previous loss, the training result of this time is saved, and the iteration training process is repeated until the Sat-Lstm-RNN network model converges.

[0114] Embodiment two

[0115] As shown in Figure 1 , a radar echo extrapolation system, the system provided by the embodiment can be applied to the method described in embodiment one, and the radar echo extrapolation system comprises:

[0116] A prediction module is configured to extract radar echo data color maps to be tested from the obtained real-time radar echo image sequence, input the radar echo data color maps to be tested after preprocessing into a pre-trained Sat-Lstm-RNN network model to obtain a radar echo prediction value, and restore the radar echo according to the radar echo prediction value.

[0117] The acquisition module is configured to acquire a historical radar echo image sequence, and extract radar echo data color images to be trained from the historical radar echo image sequence at a set time interval; the radar echo data color images to be trained are sequentially subjected to grayscale value mapping and normalization processing to form normalized grayscale data, and a training image set is constructed;

[0118] The construction module is configured to construct a Sat-Lstm-RNN network model based on the encoder, the Sat-ConvLstm module, and the decoder.

[0119] The training module is configured to train the Sat-Lstm-RNN network model through the training image set, establish a loss function according to radar echo predicted values output by the Sat-Lstm-RNN network model and radar echo true values, adjust recurrent neural network parameters by using the loss function, and repeat the iteration training process until the Sat-Lstm-RNN network model converges.

[0120] Embodiment three

[0121] A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the radar echo extrapolation method according to the embodiment one.

[0122] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0123] The present application is described with reference to flowcharts and / or block diagrams according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks

[0124] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 The flow or flows and / or blocks Figure 1 The flow or flows and / or blocks

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 The flow or flows and / or blocks Figure 1 The flow or flows and / or blocks

[0126] The above description is only preferred embodiments of the present application, it should be pointed out that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A radar echo extrapolation method, characterized in that, include: The radar echo data color image to be tested is extracted from the acquired real-time radar echo image sequence. After preprocessing the radar echo data color image to be tested, it is input into the pre-trained Sat-Lstm-RNN network model to obtain the radar echo prediction value. The radar echo is then restored based on the radar echo prediction value. The training process of the Sat-Lstm-RNN network model includes: Acquire historical radar echo image sequences, and extract radar echo data color images to be trained from the historical radar echo image sequences at set time intervals; perform grayscale value mapping and normalization processing on the radar echo data color images to be trained in sequence to form normalized grayscale data, and construct a training image set. A Sat-Lstm-RNN network model is constructed based on the encoder, Sat-ConvLstm module, and decoder; the Sat-Lstm-RNN network model is trained using a training image set, and the specific process includes: Temporal and spatial feature information generated by downsampling normalized grayscale data in the training image set using an encoder. ; to combine temporal and spatial feature information The input is fed into a Sat-ConvLstm module with L stacked layers; the Sat-ConvLstm module consists of an attention module and a recurrent neural network module. Let the hidden state at time step t of the output of the Sat-ConvLstm module of layer l be the hidden state. Set the hidden state at time step t output by the (l-1)th layer Sat-ConvLstm module as the hidden state. The value of l ranges from 1 to L; (Setting) Hidden state Previous The hidden state at each time step is an all-zero tensor; the hidden state Previous The hidden state at each time step is denoted as the hidden state. ; Time and space feature information Hidden state and hidden state Motion information is obtained by the attention module of the l-th layer Sat-ConvLstm module. ; motion information Temporal and spatial characteristic information and cellular state The recurrent neural network module, whose input is fed into the l-th layer Sat-ConvLstm module, obtains the cell state at time step t. and hidden state ; Cellular state and hidden state The input is fed into the next layer of the Sat-ConvLstm module, and the iteration is repeated until the Sat-ConvLstm module of the Lth layer outputs the cell state. and hidden state ; Using a decoder to analyze cell states and hidden state Upsample and output predicted radar echo values; A loss function is established based on the radar echo prediction value and the actual radar echo value output by the Sat-Lstm-RNN network model. The parameters of the recurrent neural network are adjusted using the loss function, and the training process is repeated iteratively until the Sat-Lstm-RNN network model converges.

2. The radar echo extrapolation method according to claim 1, characterized in that, The preprocessing of the radar echo data color image to be tested includes: performing grayscale value mapping and normalization on the radar echo data color image to be tested in sequence.

3. The radar echo extrapolation method according to claim 2, characterized in that, Methods for grayscale mapping and normalization of radar echo data color images include: Mapping a color image of radar echo data to be trained or tested, with a channel dimension of 3, to grayscale data with a channel dimension of 1, can be expressed by the following formula: ； In the formula, R represents the brightness of a pixel in the red channel of the radar echo data color image, G represents the brightness of a pixel in the green channel of the radar echo data color image, and B represents the brightness of a pixel in the blue channel of the radar echo data color image. This is represented as the brightness of a pixel in the grayscale channel of a color image of radar echo data; Normalized grayscale data is obtained by normalizing grayscale data, expressed by the following formula: ； In the formula, Pixel represents normalized grayscale data.

4. The radar echo extrapolation method according to claim 1, characterized in that, Methods for constructing Sat-Lstm-RNN network models based on encoders, Sat-ConvLstm modules, and decoders include: Stack the Lth layer of Sat-ConvLstm modules after the encoder; set the output of the Lth layer of Sat-ConvLstm modules as the input of the decoder.

5. The radar echo extrapolation method according to claim 1, characterized in that, Time and space feature information Hidden state and hidden state Motion information is obtained by the attention module of the l-th layer Sat-ConvLstm module. The methods include: Calculate hidden state Attention score The formula is as follows: ； ； ； In the formula, Represented as the Hadamard product; Based on attention score Calculate long-term motion information The formula is as follows: ； Utilizing long-term exercise information Temporal and spatial characteristic information Calculate motion information The formula is as follows: ； ； in, It is the sigmoid function; This represents the convolution operation. Represented as temporal and spatial feature information The weight parameters.

6. The radar echo extrapolation method according to claim 1, characterized in that, motion information Temporal and spatial characteristic information and cellular state The recurrent neural network module, whose input is fed into the l-th layer Sat-ConvLstm module, obtains the cell state at time step t. and hidden state The formula is as follows: ； ； ； ； ； In the formula, This represents the convolution operation. Represents the Hadamard product. Represents the hyperbolic tangent activation function. Represents the Sigmoid function; This is represented as the expression function of the input gate; This is represented as the expression function of the forget gate. This is represented as the expression function of the output gate; , , , These represent the set deviation parameters, Indicates the corresponding input gate The weight matrix, Indicates the corresponding input gate The weight matrix, Represents the corresponding part in the forget gate The weight matrix, Represents the corresponding part in the forget gate The weight matrix, Represents the corresponding part in the forget gate The weight matrix, Indicating the corresponding cell state update The weight matrix, Indicates the corresponding cell state update The weight matrix, Indicates the corresponding output gate The weight matrix, Indicates the corresponding output gate The weight matrix.

7. The radar echo extrapolation method according to claim 1, characterized in that, Methods for establishing a loss function based on the radar echo predictions and actual radar echo values ​​output by the Sat-Lstm-RNN network model include: ； In the formula, This is represented as the true value of the radar echo at time step t; denoted as the radar echo prediction output by the Sat-Lstm-RNN network model at time step t; n represents the sequence of radar echoes; w represents the abscissa of the radar echoes; and h represents the ordinate of the radar echoes.

8. A radar echo extrapolation system, characterized in that, include: The prediction module is used to extract the radar echo data color image to be tested from the acquired real-time radar echo image sequence. After preprocessing the radar echo data color image to be tested, it is input into the pre-trained Sat-Lstm-RNN network model to obtain the radar echo prediction value, and the radar echo is restored based on the radar echo prediction value. The acquisition module is used to acquire historical radar echo image sequences and extract radar echo data color images to be trained from the historical radar echo image sequences at set time intervals; the radar echo data color images to be trained are sequentially subjected to grayscale value mapping and normalization processing to form normalized grayscale data, and a training image set is constructed. Building blocks are used to construct Sat-Lstm-RNN network models based on encoders, Sat-ConvLstm modules, and decoders; The training module is used to train the Sat-Lstm-RNN network model using a training image set. It establishes a loss function based on the radar echo prediction value and the true radar echo value output by the Sat-Lstm-RNN network model, adjusts the parameters of the recurrent neural network using the loss function, and repeats the training process iteratively until the Sat-Lstm-RNN network model converges. The training module trains the Sat-Lstm-RNN network model using a training image set, specifically including: Temporal and spatial feature information generated by downsampling normalized grayscale data in the training image set using an encoder. ; to combine temporal and spatial feature information The input is fed into a Sat-ConvLstm module with L stacked layers; the Sat-ConvLstm module consists of an attention module and a recurrent neural network module. Let the hidden state at time step t of the output of the Sat-ConvLstm module of layer l be the hidden state. Set the hidden state at time step t output by the (l-1)th layer Sat-ConvLstm module as the hidden state. The value of l ranges from 1 to L; (Setting) Hidden state Previous The hidden state at each time step is an all-zero tensor; the hidden state Previous The hidden state at each time step is denoted as the hidden state. ; Time and space feature information Hidden state and hidden state Motion information is obtained by the attention module of the l-th layer Sat-ConvLstm module. ; motion information Temporal and spatial characteristic information and cellular state The recurrent neural network module, whose input is fed into the l-th layer Sat-ConvLstm module, obtains the cell state at time step t. and hidden state ; Cellular state and hidden state The input is fed into the next layer of the Sat-ConvLstm module, and the iteration is repeated until the Sat-ConvLstm module of the Lth layer outputs the cell state. and hidden state ; Using a decoder to analyze cell states and hidden state Upsample and output radar echo prediction values.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the radar echo extrapolation method according to any one of claims 1 to 7.

Images