CN115310724A - Precipitation prediction method based on Unet and DCN _ LSTM - Google Patents

Precipitation prediction method based on Unet and DCN _ LSTM Download PDF

Info

Publication number
CN115310724A
CN115310724A CN202211233428.1A CN202211233428A CN115310724A CN 115310724 A CN115310724 A CN 115310724A CN 202211233428 A CN202211233428 A CN 202211233428A CN 115310724 A CN115310724 A CN 115310724A
Authority
CN
China
Prior art keywords
representing
gate
dcn
lstm
input
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211233428.1A
Other languages
Chinese (zh)
Inventor
秦华旺
包顺
戴跃伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN202211233428.1A priority Critical patent/CN115310724A/en
Publication of CN115310724A publication Critical patent/CN115310724A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Abstract

The invention discloses a rainfall prediction method based on Unet and DCN _ LSTM, which relates to the technical field of weather forecast, can reduce training time, improve the timeliness of rainfall prediction, effectively combine two models, improve the precision of rainfall prediction, effectively capture space-time correlation, input the offset of hidden states and memory cells by using deformable convolution learning, and adjust the position of a convolution kernel by inputting, so that the position of the convolution kernel is not fixed any more, the characteristics of a rainfall region can be effectively extracted, a Bayesian algorithm is used, the problem of fussy manual parameter adjustment can be solved, an optimal hyper-parameter combination can be learned by the Bayesian algorithm, the precision is higher by using a Unet and DCN _ LSTM mixed model than using a single model, and the effect is better.

Description

Precipitation prediction method based on Unet and DCN _ LSTM
Technical Field
The invention relates to the technical field of weather forecast, in particular to a rainfall prediction method based on Unet and DCN _ LSTM.
Background
The rainstorm incident can influence people's normal life, cause great life loss and loss of property, therefore, accurate rainfall prediction plays vital role to people's life and trip, the forecast in advance of rainfall can be to the public, reduce the mechanism of calamity risk, government department and infrastructure's managers provide early warning, after the rainstorm early warning is released, above-mentioned relevant personnel, mechanism and department take corresponding action according to predetermined standard operating program, in order to save life and protect the property, so the prediction has huge influence to aviation service, public safety and each field of people, the rainfall prediction is a major problem always, accurate prediction rainfall is not only vital to people's trip and to the society, also can avoid the appearance of heavy disasters such as rainstorm and debris flow simultaneously.
The traditional method for forecasting the precipitation is mainly based on a numerical prediction mode (NWP), and means that an ultra-large computer is adopted as a numerical calculation tool through an atmospheric condition, and the motion state and the weather condition of a future period of time are forecasted through a fluid mechanics and thermodynamic equation set.
The existing deep learning rainfall prediction methods are many, but a single model is adopted, two models are not combined, unet is used independently for rainfall prediction, the extracted time correlation is too small, the prediction is not suitable for long-time prediction, for example, one-hour prediction is realized, the LSTM model is used independently for prediction, the calculated amount is complex, the training time is long, the model parameters are large, and the computer memory is consumed; therefore, the rainfall prediction method based on the Unet and the DCN _ LSTM is provided, compared with the existing method, the time-space correlation can be extracted effectively, the global characteristics of rainfall can be extracted better as the characteristic diagram becomes smaller, meanwhile, the model parameters become smaller, the training time becomes shorter, and the precision of rainfall prediction can be improved.
Disclosure of Invention
In order to solve the technical problems, the invention provides a precipitation prediction method based on Unet and DCN _ LSTM, which comprises the following steps
S1, acquiring meteorological radar data and preprocessing the meteorological radar data;
s2, constructing a Unet and DCN _ LSTM mixed model;
s3, adding a Bayesian algorithm to the Unet and DCN _ LSTM mixed model, performing hyper-parameter optimization, and searching for an optimal parameter combination;
s4, testing the Unet and DCN _ LSTM mixed model;
and S5, converting the result of the prediction of the test set into radar reflectivity through a pixel value, and then obtaining rainfall according to the relation between the radar reflectivity and the rainfall.
The technical scheme of the invention is further defined as follows:
further, in step S1, the method for preprocessing the weather radar data includes the following steps
S1.1, removing abnormal values and repeated values of data, and performing bilinear interpolation on missing values of the data;
s1.2, screening the data sets to ensure that each echo sequence has 20% precipitation coverage rate;
s1.3, carrying out normalization processing on the data, wherein a specific formula is as follows,
Figure 971048DEST_PATH_IMAGE001
wherein X * Representing normalized radar echo intensity values, X max Indicating the maximum value of the radar echo intensity, X min Representing the minimum value of the radar echo intensity, and X represents the radar echo intensity value;
s1.4, dividing the data set by adopting a proportion of 8.
In the foregoing precipitation prediction method based on the Unet and the DCN _ LSTM, in step S2, the method for constructing the Unet and DCN _ LSTM hybrid model includes the following steps
S2.1, a hybrid model encoder part inputs training set data into a model, a characteristic diagram is changed into a half of an original size through two 3 x 3 convolutional layers and then a maximum pooling layer, and the number of channels is doubled through the two 3 x 3 convolutional layers;
s2.2, adopting DCN _ LSTM in the middle of the hybrid model for extracting the time characteristics and the space characteristics of the radar echo sequence, wherein the DCN _ LSTM consists of a plurality of DCN _ LSTM circulating units and is used for decomposing the characteristic diagram output by the encoder and sequentially inputting the characteristic diagram into the DCN _ LSTM circulating units for training;
s2.3, the mixed model decoder part splices the radar echo sequence output by the DCN _ LSTM according to channels, then passes through two 3 x 3 convolutional layers, is subjected to up-sampling, is connected with a characteristic diagram output by the encoder in a skipping mode, passes through the two 3 x 3 convolutional layers and one 1 x 1 convolutional layer, and finally outputs a predicted radar echo sequence.
In the foregoing, in step S2.2, the DCN _ LSTM learns the offset of the input X to the hidden state H and the memory cell C by using the feasible variable convolution, so as to update the hidden state H and the memory cell C, and slides on the input picture, the feasible variable convolution takes the obtained feature map as an input, and a convolution layer is applied to the feature map, so as to obtain the deformation offset of the feasible variable convolution, wherein the offset layer is 2N, and the translation is performed on a plane, and both directions X and y need to be changed;
the specific formula of the feasible convolution is as follows,
Figure 100678DEST_PATH_IMAGE002
wherein, R represents a convolution kernel of 3 × 3, (-1, -1), (-1, 0), (0, 1), (1, 1) represents points in the convolution kernel, and coordinates are integers;
Figure 383892DEST_PATH_IMAGE003
wherein the content of the first and second substances,
Figure 206354DEST_PATH_IMAGE004
representing a feature matrix resulting from the variable convolution,
Figure 207808DEST_PATH_IMAGE005
represents the learning amount of each point obtained through neural network learning within a convolution kernel of size 3 x 3,
Figure 875550DEST_PATH_IMAGE006
represents the center point, i.e., (0, 0) point,
Figure 747691DEST_PATH_IMAGE007
representing points defined in the R range, with more offset matrices learned by convolution than standard convolution feasible convolutions
Figure 741055DEST_PATH_IMAGE008
。
In the aforementioned rainfall prediction method based on Unet and DCN _ LSTM, the DCN _ LSTM model includes a plurality of DCN _ LSTM circulation units, feature information is screened and transmitted through a gating mechanism, and forgetting gates, input gates, modulation gates, output gates, time memory cells and hidden states of convLSTM are reserved, wherein f is respectively t、 i t、 g t、 o t、 C t And H t (ii) a Also includes space cells M t For extracting and transferring spatial structure features vertically between different layers, and adding feasible convolution to learn the offset of input X to hidden state H and memory cell C, the concrete formula is as follows,
Figure 229805DEST_PATH_IMAGE009
where DCN represents a variable convolutional network, X t Representing the picture entered, the lower coordinate t representing the moment of entry,
Figure 701237DEST_PATH_IMAGE010
and
Figure 959043DEST_PATH_IMAGE011
respectively representing the hidden state and the memory cells, the lower coordinate t-1 representing the last moment, the upper coordinate 1 representing the first layer,
Figure 857729DEST_PATH_IMAGE010
and
Figure 568196DEST_PATH_IMAGE011
respectively representing the new hidden state and the memory cell obtained after the update.
One of the foregoing precipitation prediction methods based on Unet and DCN _ LSTM, the DCN _ LSTM model, the input gate, update gate and forget gate for updating memory cells, is as follows,
Figure 843320DEST_PATH_IMAGE012
wherein i t An input gate representing a refresh memory cell; sigma represents an activation function sigmoid; w is a group of xi A parameter matrix representing the training of input X, input gate i; w is a group of hi A parameter matrix representing the training of the input gate i for the hidden state H; x t An input representing time t;
Figure 221212DEST_PATH_IMAGE013
representing the hidden state of the l-th layer at the time t-1; b i Represents the offset to input gate i;
g t a refresh gate indicating a refresh of the memory cell; tanh represents an activation function tanh; w xg Representing a parameter matrix trained for updating gate g for input X; w hg Representing the parameter matrix of the gate g training updated for the hidden state H; b g Represents the bias to the refresh gate g;
f t a forgetting gate that indicates a renewed memory cell; w xf A parameter matrix representing the training of the forgetting gate f for the input X; w hf A parameter matrix representing the training of the hidden state H and the forgetting gate f; b f Represents a bias to the forgetting gate f; * Represents a convolution;
an input gate for updating the cells in space, an update gate and a forget gate, the formula is as follows,
Figure 290799DEST_PATH_IMAGE014
wherein the content of the first and second substances,
Figure 754141DEST_PATH_IMAGE015
an input gate representing an update space cell; sigma represents an activation function sigmoid;
Figure 301797DEST_PATH_IMAGE016
a parameter matrix representing the training of input X, input gate i; w mi A parameter matrix representing the training of the input gate i for the space cell M; x t An input representing time t;
Figure 534195DEST_PATH_IMAGE017
represents the space cell of the l-1 st layer at the time point t;
Figure 774684DEST_PATH_IMAGE018
represents the offset to input gate i;
Figure 725322DEST_PATH_IMAGE019
an update gate representing an updated spatial cell; tanh represents an activation function tanh;
Figure 342248DEST_PATH_IMAGE020
representing a parameter matrix trained for updating gate g for input X; w mg Representing a parameter matrix for updating gate g training on the space cell M;
Figure 429153DEST_PATH_IMAGE021
represents the bias to the refresh gate g;
Figure 840543DEST_PATH_IMAGE022
a forgetting gate representing an updated spatial cell;
Figure 12898DEST_PATH_IMAGE023
a parameter matrix representing the training of the forgetting gate f for the input X; w mf A parameter matrix representing the training of the spatial cell M and the forgetting gate f;
Figure 167936DEST_PATH_IMAGE024
represents a bias to the forgetting gate f; * Represents a convolution;
the hidden state, i.e. the output, is updated by the memory cells, the space cells and the output gate, the specific formula is as follows,
Figure 640506DEST_PATH_IMAGE025
wherein the content of the first and second substances,
Figure 222797DEST_PATH_IMAGE026
memory cells representing the first layer at time t; i.e. i t Representing an input gate; g t Represents an update gate; f. of t Indicating a forgetting gate;
Figure 882448DEST_PATH_IMAGE027
memory cells in layer I at time t-1;
Figure 841177DEST_PATH_IMAGE028
represents the space cell of the l layer at the t time;
Figure 902674DEST_PATH_IMAGE029
an input gate representing a spatial cell;
Figure 921445DEST_PATH_IMAGE030
an update gate representing a spatial cell;
Figure 333972DEST_PATH_IMAGE031
a forgetting gate representing a spatial cell;
Figure 830813DEST_PATH_IMAGE032
represents the space cell of l-1 layer at the t time; o t An output gate is shown; sigma represents an activation function sigmoid;
W xo a parameter training matrix representing the output gate o for input X; w ho A parameter training matrix representing the output gate o for the hidden state H; w co A matrix of training parameters representing the input gate o for memory cell C; w is a group of mo Representing the training parameter matrix of the input gate o for the spatial cell M; x t An input representing time t;
Figure 746816DEST_PATH_IMAGE033
representing the hidden state of the l-th layer at the time of t-1;
Figure 936489DEST_PATH_IMAGE034
memory cells representing the first layer at time t;
Figure 570733DEST_PATH_IMAGE035
representing the space cells of the l layer at the t time; b is a mixture of o Represents the offset to the output gate; w 1x1 A convolution kernel representing a size of 1 × 1;
Figure 871264DEST_PATH_IMAGE036
representing the new hidden state obtained by updating; degree represents the multiplication of the matrices in bits; * Representing a convolution.
In the foregoing precipitation prediction method based on Unet and DCN _ LSTM, in step S3, the number of neurons, the batch size and the learning rate of the hidden layer are optimized through a Bayesian algorithm, which includes the following steps
S3.1, assuming a set of hyper-parametric combinations is
Figure 641774DEST_PATH_IMAGE037
Wherein
Figure 2348DEST_PATH_IMAGE038
Respectively representing the parameter combinations of the number of neurons of the hidden layer, the batch size and the learning rate, and assuming that a loss function and a set hyper-parameter have a mapping relation;
assuming the function f x → R, it is necessary to determine the value of x ∈ R
Figure 389467DEST_PATH_IMAGE039
S3.2, obtaining the random initialization point of the hyper-parameter in the parameter range according to the determined and optimized hyper-parameter
Figure 228110DEST_PATH_IMAGE037
Inputting experimental data to train the model, the response value of the loss function is
Figure 118706DEST_PATH_IMAGE040
Establishing a Gaussian regression process;
known data set
Figure 650181DEST_PATH_IMAGE041
Suppose f is obeyed
Figure 993438DEST_PATH_IMAGE042
So that the prediction also follows a normal distribution,
Figure 901351DEST_PATH_IMAGE044
Figure 912032DEST_PATH_IMAGE045
Figure 614409DEST_PATH_IMAGE046
Figure 444962DEST_PATH_IMAGE047
wherein K is a constant, K (x) and
Figure 625407DEST_PATH_IMAGE048
a covariance matrix is represented by a value of the covariance matrix,
Figure 490595DEST_PATH_IMAGE049
represents the variance of sample n;
find out
Figure 363873DEST_PATH_IMAGE050
And
Figure 681722DEST_PATH_IMAGE051
；
s3.3, selecting the next hyper-parameter combination sampling point from the Gaussian regression model based on the sampling function PI
Figure 665859DEST_PATH_IMAGE052
And the sampling function PI is as follows,
Figure 385553DEST_PATH_IMAGE053
where Φ () represents a normal distribution cumulative density function,
Figure 429732DEST_PATH_IMAGE050
and
Figure 234877DEST_PATH_IMAGE051
respectively representing the mean and variance of the objective function value,
Figure 22705DEST_PATH_IMAGE054
the value of the optimum objective function is represented,
Figure 862485DEST_PATH_IMAGE055
representing a parameter;
s3.4, bringing the selected first group of hyper-parameter combinations into model training, outputting real values of ground observation and predicted mean square error values of radar echo sequences, and if the mean square error values are smaller than a preset threshold value, stopping updating and outputting optimal hyper-parameter combinations; if the mean square error value is not less than the preset threshold value, the hyper-parameter is updated to
Figure 806127DEST_PATH_IMAGE052
And repeating the step S3.2 to the step S3.4 until a hyperparameter combination with the mean square error value smaller than a preset threshold value is found.
In the aforementioned method for predicting precipitation based on the Unet and DCN _ LSTM, in step 3.4, the preset threshold is set to 0.0001.
The method for predicting precipitation based on Unet and DCN _ LSTM in step S4 comprises the following steps
S4.1, loading the weight of model training, testing and storing the weight in a picture format;
s4.2, adopting mean square error, structural similarity and critical success index as evaluation indexes of the test set;
the mean square error is used for evaluating the difference of pixel points of two pictures, and a specific formula is as follows,
Figure 98568DEST_PATH_IMAGE056
where n denotes the total number of samples, i denotes the sample number of the sample points, Y denotes the real tag of the real radar echo pattern,
Figure 690086DEST_PATH_IMAGE057
representing a predicted radar echo map;
the structural similarity is used for measuring the similarity of two pictures, and a specific formula is as follows,
Figure 118794DEST_PATH_IMAGE058
wherein u is x And u y Denotes the mean, σ, of x and y, respectively x And σ y Denotes the variance, σ, for x and y, respectively xy Represents the covariance of the two pictures x and y, C 1 And C 2 Represents a constant;
the specific formula for the critical success index is as follows:
Figure 504776DEST_PATH_IMAGE059
wherein, TP indicates that the true category is positive and the prediction result is also positive, FP indicates that the true category is negative and the prediction result is negative, and FN indicates that the true category is positive and the prediction result is negative.
In the aforementioned precipitation prediction method based on the Unet and DCN _ LSTM, in step S5, the result of prediction of the test set is converted into radar reflectivity through pixel values, and the specific formula is as follows,
Figure 284513DEST_PATH_IMAGE060
wherein, radar _ value represents the value of radar reflectivity of each pixel point converted by a formula, and pixel _ value represents the value of each pixel point;
then the rainfall is obtained according to the relation between the radar reflectivity and the rainfall, the concrete formula is as follows,
Figure 414143DEST_PATH_IMAGE061
wherein Z represents radar reflectivity, R represents rainfall, and a, b represent coefficients.
The beneficial effects of the invention are:
according to the rainfall prediction method, training time can be shortened, timeliness of rainfall prediction is improved, the two models are effectively combined, precision of rainfall prediction is improved, space-time correlation can be effectively captured, the deformable convolution is used for learning offset of input to hidden states and memory cells, the position of a convolution kernel can be adjusted through input, the position of the convolution kernel is enabled to be no longer fixed, characteristics of a rainfall area can be effectively extracted, a Bayesian algorithm is used, complexity of manual parameter adjustment can be solved, an optimal hyper-parameter combination can be learned through the Bayesian algorithm, through multi-index evaluation, the accuracy of prediction is higher by using a Unet and DCN _ LSTM mixed model than that of prediction by using a single model, and the effect is better.
Drawings
FIG. 1 is a schematic overall flow diagram of the present invention;
FIG. 2 is a schematic diagram of the hybrid model of Unet and DCN _ LSTM in the present invention;
FIG. 3 is a schematic flow chart of the Bayesian algorithm for hyperparametric optimization in the present invention.
Detailed Description
The present embodiment provides a precipitation prediction method based on Unet and DCN _ LSTM, as shown in fig. 1 to 3, which includes the following steps
S1, acquiring meteorological radar data, wherein the used data is a radar echo sequence dataset, 10 frames are input and predicted, the interval of each frame is 6 minutes, namely, the rainfall of one hour in the future is predicted by historical data of the previous hour, and the meteorological radar data is preprocessed, and the method specifically comprises the following steps:
s1.1, removing abnormal values and repeated values of data, and performing bilinear interpolation on missing values of the data;
s1.2, screening the data set to ensure that each echo sequence has 20% precipitation coverage rate, and predicting that if the data are not screened during rainfall, a plurality of sequences are likely to have no precipitation coverage, so that the training model effect is not ideal;
s1.3, carrying out normalization processing on the data, wherein a specific formula is as follows,
Figure 962936DEST_PATH_IMAGE001
wherein, X * Representing normalized radar echo intensity values, X max Representing the maximum value, X, of the radar echo intensity min Representing the minimum value of the radar echo intensity, and X represents the radar echo intensity value;
s1.4, dividing the data set by adopting a proportion of 8.
S2, constructing a Unet and DCN _ LSTM hybrid model, wherein the Unet and DCN _ LSTM hybrid model mainly comprise a first half part and a second half part of the Unet and respectively form an encoder part and a decoder part of the model, and a convolutional layer is originally arranged in the middle of the Unet, so that the DCN _ LSTM model is replaced in the method, the time and space correlation can be effectively extracted, and the method comprises the following specific steps:
s2.1, a mixed model encoder part, wherein the encoder part adopts the first half part of Unet, training set data is input into the model, and the input is [8,10,200,200], wherein 8 represents batch size number, 10 represents input seq sequence, and 200 respectively represent the length and width of an input picture; after passing through two 3 × 3 convolutional layers, adding a correction linear unit after each layer to enable the model to become nonlinear and batch normalization, then passing through a 2 × 2 maximum pooling layer, doubling the number of channels after pooling, changing the size of an input radar echo picture into half of the original size, enabling the picture length to be 100 and the width to be 100, and then passing through the two 3 × 3 convolutional layers, adding a correction linear unit after each layer to enable the model to become nonlinear and batch normalization, and doubling the number of channels to be 128;
s2.2, adopting DCN _ LSTM in the middle of the hybrid model for extracting the time characteristics and the space characteristics of a radar echo sequence, wherein the DCN _ LSTM is composed of a plurality of DCN _ LSTM circulating units, performing dimension conversion on [8,128,100 and 100] output by an encoder into [8,128,1,100 and 100], and sequentially inputting the converted data into the DCN _ LSTM circulating units for training;
the DCN _ LSTM updates the hidden state H and the memory cell C by using the offset of the feasible variable convolution learning input X to the hidden state H and the memory cell C, slides on an input picture, uses the obtained feature map as an input by the feasible variable convolution, and applies a convolution layer to the feature map so as to obtain the deformation offset of the feasible variable convolution, wherein the offset layer is 2N, and the translation is carried out on a plane, and the two directions of X and y need to be changed;
the specific formula for the feasible convolution is as follows,
Figure 785398DEST_PATH_IMAGE002
wherein, R represents a convolution kernel of 3 × 3, (-1, -1), (-1, 0), (0, 1), (1, 1) represents points in the convolution kernel, and coordinates are integers;
Figure 786852DEST_PATH_IMAGE003
wherein, the first and the second end of the pipe are connected with each other,
Figure 454594DEST_PATH_IMAGE004
representing a feature matrix resulting from the feasible convolution,
Figure 857894DEST_PATH_IMAGE005
is represented in a size ofThe learning amount of each point obtained by neural network learning in the convolution kernel of 3 x 3,
Figure 851257DEST_PATH_IMAGE006
represents the center point, i.e., (0, 0) point,
Figure 74428DEST_PATH_IMAGE007
representing points defined in the R range, with more offset matrices learned by convolution than standard convolution feasible convolutions
Figure 811440DEST_PATH_IMAGE008
；
The DCN _ LSTM model comprises a plurality of DCN _ LSTM circulation units, characteristic information is screened and transmitted through a gate control mechanism, and forgetting gates, input gates, modulation gates, output gates, time memory cells and hidden states of convLSTM are reserved, wherein the forgetting gates, the input gates, the modulation gates, the output gates, the time memory cells and the hidden states are respectively f t、 i t、 g t、 o t、 C t And H t (ii) a Also includes space cells M t For extracting and transmitting spatial structure features vertically between different layers, and adding feasible convolution to learn the offset of input X to hidden state H and memory cell C, the concrete formula is as follows,
Figure 334825DEST_PATH_IMAGE009
where DCN represents a variable convolutional network, X t Representing the picture entered, the lower coordinate t representing the moment of entry,
Figure 233511DEST_PATH_IMAGE010
and
Figure 209558DEST_PATH_IMAGE011
respectively representing the hidden state and the memory cells, the lower coordinate t-1 representing the previous moment, the upper coordinate 1 representing the first layer,
Figure 484681DEST_PATH_IMAGE010
and
Figure 596994DEST_PATH_IMAGE011
respectively representing the new hidden state and the memory cell obtained after updating;
in the DCN _ LSTM model, the input gate, the update gate, and the forget gate for updating the memory cells are defined as follows,
Figure 932160DEST_PATH_IMAGE012
wherein i t An input gate representing a refresh memory cell; sigma represents an activation function sigmoid; w xi A parameter matrix representing the training of input X, input Gate i; w hi A parameter matrix representing the training of the input gate i for the hidden state H; x t An input representing time t;
Figure 129923DEST_PATH_IMAGE013
representing the hidden state of the l-th layer at the time t-1; b i Represents the offset to input gate i;
g t a refresh gate indicating a refresh of the memory cell; tanh represents an activation function tanh; w xg Representing a parameter matrix trained for updating gate g for input X; w hg Representing that the parameter matrix of the training of the gate g is updated for the hidden state H; b is a mixture of g Represents the bias to the refresh gate g;
f t a forgetting gate representing a renewed memory cell; w xf A parameter matrix representing the training of the forgetting gate f for the input X; w is a group of hf Representing a parameter matrix for training a hidden state H and a forgetting gate f; b is a mixture of f Represents a bias to the forgetting gate f; * Represents a convolution;
an input gate for updating the cells in space, an update gate and a forget gate, the formula is as follows,
Figure 474317DEST_PATH_IMAGE014
wherein the content of the first and second substances,
Figure 706715DEST_PATH_IMAGE015
an input gate representing the update space cells; σ represents an activation function sigmoid;
Figure 947204DEST_PATH_IMAGE016
a parameter matrix representing the training of input X, input Gate i; w mi A parameter matrix representing the training of the input gate i for the space cell M; x t An input representing time t;
Figure 897842DEST_PATH_IMAGE017
represents the space cell of the l-1 st layer at the time point t;
Figure 514768DEST_PATH_IMAGE018
represents the offset to input gate i;
Figure 601673DEST_PATH_IMAGE019
an update gate representing an updated spatial cell; tanh represents an activation function tanh;
Figure 13063DEST_PATH_IMAGE020
representing a parameter matrix trained for updating gate g for input X; w mg Representing a parameter matrix for updating gate g training on the space cell M;
Figure 185418DEST_PATH_IMAGE021
represents the bias to the refresh gate g;
Figure 871614DEST_PATH_IMAGE022
a forgetting gate representing an updated spatial cell;
Figure 813025DEST_PATH_IMAGE023
a parameter matrix representing the training of the forgetting gate f for the input X; w mf A parameter matrix representing the training of the spatial cell M and the forgetting gate f;
Figure 660896DEST_PATH_IMAGE024
indicating to left behind door fBiasing; * Represents a convolution;
the hidden state, i.e. the output, is updated by the memory cells, the space cells and the output gate, the specific formula is as follows,
Figure 320547DEST_PATH_IMAGE025
wherein the content of the first and second substances,
Figure 279276DEST_PATH_IMAGE026
memory cells representing the first layer at time t; i.e. i t Representing an input gate; g is a radical of formula t Represents an update gate; f. of t Indicating a forgetting gate;
Figure 340773DEST_PATH_IMAGE027
memory cells in the first layer at time t-1;
Figure 359544DEST_PATH_IMAGE028
representing the space cells of the l layer at the t time;
Figure 240913DEST_PATH_IMAGE029
an input gate representing a spatial cell;
Figure 3332DEST_PATH_IMAGE030
an update gate representing a spatial cell;
Figure 184915DEST_PATH_IMAGE031
a forgetting gate representing a spatial cell;
Figure 374588DEST_PATH_IMAGE032
represents the space cell of l-1 layer at the t time; o t An output gate is shown; σ represents an activation function sigmoid;
W xo a parameter training matrix representing the output gate o for input X; w ho A parameter training matrix representing the output gate o for the hidden state H; w co Representing the training parameter matrix of input gate o for memory cell C; w mo Representing training on spatial cell M, input Gate oA parameter matrix; x t An input representing time t;
Figure 8832DEST_PATH_IMAGE033
representing the hidden state of the l-th layer at the time t-1;
Figure 309363DEST_PATH_IMAGE034
memory cells representing the first layer at time t;
Figure 79873DEST_PATH_IMAGE035
representing the space cells of the l layer at the t time; b is a mixture of o Represents the offset to the output gate; w 1x1 A convolution kernel representing a size of 1 × 1;
Figure 440447DEST_PATH_IMAGE036
representing the new hidden state obtained by updating; degree represents a multiplication of the matrices in bits; * Representing a convolution.
S2.3, a hybrid model decoder part,
the method comprises the steps that the method is composed of the second half part of Unet, the DCN _ LSTM output is used for splicing according to channels, feature fusion is carried out on a feature diagram which is subjected to skip connection and original input, and the number of the channels is changed into 256; capturing the context of the original image through two 3 x 3 convolutional layers, and adding a correction linear unit after each layer so that the model becomes nonlinear and batch normalization; then, through upsampling, bilinear interpolation is used in the upsampling, dimensionality is recovered, and meanwhile, through skipping connection, the number of characteristic channels is changed into 128; through two 3 x 3 convolutional layers and one 1 x 1 convolutional layer. And finally, outputting the predicted radar echo sequence.
S3, adding a Bayesian algorithm to the Unet and DCN _ LSTM mixed model, carrying out hyper-parameter optimization, searching for an optimal parameter combination, and optimizing the number of neurons, batch size and learning rate of a hidden layer through the Bayesian algorithm, wherein the specific steps are as follows:
s3.1, assuming a set of hyper-parametric combinations is
Figure 296408DEST_PATH_IMAGE037
In which
Figure 666209DEST_PATH_IMAGE038
Respectively representing the parameter combinations of the number of neurons of the hidden layer, the batch size and the learning rate, and assuming that a loss function and a set hyper-parameter have a mapping relation;
assuming the function f x → R, it is necessary to determine the value of x ∈ R
Figure 556805DEST_PATH_IMAGE039
S3.2, obtaining the random initialization point of the hyper-parameter in the parameter range according to the determined and optimized hyper-parameter
Figure 88280DEST_PATH_IMAGE037
Inputting experimental data to train the model, the response value of the loss function is
Figure 431537DEST_PATH_IMAGE040
Establishing a Gaussian regression process;
known data set
Figure 73871DEST_PATH_IMAGE041
Suppose f is obeyed
Figure 818973DEST_PATH_IMAGE042
So the prediction also follows a normal distribution,
Figure 521350DEST_PATH_IMAGE044
Figure 351902DEST_PATH_IMAGE045
Figure 797927DEST_PATH_IMAGE046
Figure 928694DEST_PATH_IMAGE047
wherein K is a constant, K (x) and
Figure 801972DEST_PATH_IMAGE048
a covariance matrix is represented by a matrix of covariance,
Figure 854242DEST_PATH_IMAGE049
represents the variance of sample n;
find out
Figure 103958DEST_PATH_IMAGE050
And
Figure 823652DEST_PATH_IMAGE051
；
s3.3, selecting the next hyper-parameter combination sampling point from the Gaussian regression model based on the sampling function PI
Figure 133411DEST_PATH_IMAGE052
The sampling function PI is, as follows,
Figure 672976DEST_PATH_IMAGE053
where Φ () represents a normal distribution cumulative density function,
Figure 460804DEST_PATH_IMAGE050
and
Figure 35005DEST_PATH_IMAGE051
respectively representing the mean and variance of the objective function value,
Figure 250085DEST_PATH_IMAGE054
the value of the optimum objective function is represented,
Figure 542526DEST_PATH_IMAGE055
representing a parameter;
and S3.4, substituting the selected first group of hyper-parameter combinations into model training, outputting a real value of ground observation and a mean square error value of a predicted radar echo sequence, and setting a preset threshold value to be 0.0001, if the mean square error value is smaller than a preset threshold value, stopping updating and outputting the optimal hyper-parameter combination; if the mean square error value is not less than the preset threshold value, the hyper-parameter is updated to
Figure 134045DEST_PATH_IMAGE052
And repeating the step S3.2 to the step S3.4 until the hyper-parameter combination with the mean square error value smaller than the preset threshold value is found.
After a model is built, inputting a training set into the model for training, using an optimized super-parameter, setting a maximum training epoch by using a loss function according to a Mean Square Error (MSE) of an output radar echo diagram and a ground real radar echo diagram, enabling a loss value loss to reach a minimum value through back propagation, wherein the minimum value means that the training loss continuously decreases until the training loss does not decrease, and storing a best training weight;
the loss function uses the sum of the mean error (MSE) and the Mean Absolute Error (MAE), and the equation is as follows:
Figure 828331DEST_PATH_IMAGE062
Figure 214313DEST_PATH_IMAGE063
where n denotes the total number of samples, i denotes the sample number of the sample points, Y denotes the real tag of the real radar echo pattern,
Figure 728471DEST_PATH_IMAGE057
representing a predicted radar echo map;
the total loss value is loss = loss1+ loss2.
S4, testing the Unet and DCN _ LSTM mixed model, and evaluating the model by adopting a plurality of indexes, wherein the specific steps are as follows:
s4.1, loading the weight of model training, testing and storing the weight in a picture format;
s4.2, adopting Mean Square Error (MSE), structural Similarity (SSIM) and Critical Success Index (CSI) as evaluation indexes of the test set;
the mean square error is used for evaluating the difference of pixel points of two pictures, and a specific formula is as follows,
Figure 858101DEST_PATH_IMAGE056
where n denotes the total number of samples, i denotes the number of sample points, Y denotes the real tag of the real radar echo pattern,
Figure 672473DEST_PATH_IMAGE057
representing a predicted radar echo map;
the structural similarity is used for measuring the similarity of two pictures, and the specific formula is as follows,
Figure 229357DEST_PATH_IMAGE058
wherein u is x And u y Denotes the mean, σ, of x and y, respectively x And σ y Denotes the variance, σ, for x and y, respectively xy Represents the covariance, C, of the two pictures x and y 1 And C 2 Represents a constant;
the specific formula for the critical success index is as follows:
Figure 230811DEST_PATH_IMAGE059
wherein, TP indicates that the true category is positive and the prediction result is also positive, FP indicates that the true category is negative and the prediction result is negative, and FN indicates that the true category is positive and the prediction result is negative.
S5, converting the result of the prediction of the test set into radar reflectivity through a pixel value, wherein the specific formula is as follows,
Figure 158272DEST_PATH_IMAGE060
wherein, radar _ value represents the value of radar reflectivity of each pixel point converted by a formula, and pixel _ value represents the value of each pixel point;
then the rainfall is obtained according to the relation between the radar reflectivity and the rainfall, the concrete formula is as follows,
Figure 561572DEST_PATH_IMAGE061
wherein Z represents radar reflectivity, R represents rainfall, and A and b represent coefficients.
In addition to the above embodiments, the present invention may have other embodiments. All technical solutions formed by adopting equivalent substitutions or equivalent transformations fall within the protection scope of the claims of the present invention.

Claims (10)

1. A precipitation prediction method based on Unet and DCN _ LSTM is characterized in that: comprises the following steps
S1, acquiring meteorological radar data and preprocessing the meteorological radar data;
s2, constructing a hybrid model of Unet and DCN _ LSTM;
s3, adding a Bayesian algorithm to the Unet and DCN _ LSTM mixed model, performing hyper-parameter optimization, and searching for an optimal parameter combination;
s4, testing the Unet and DCN _ LSTM mixed model;
and S5, converting the result of the prediction of the test set into radar reflectivity through a pixel value, and then obtaining rainfall according to the relation between the radar reflectivity and the rainfall.
2. The method of claim 1, wherein the method of predicting precipitation based on Unet and DCN _ LSTM comprises: in the step S1, the method for preprocessing the meteorological radar data comprises the following steps
S1.1, removing abnormal values and repeated values of data, and performing bilinear interpolation on missing values of the data;
s1.2, screening the data set to ensure that each echo sequence has 20% precipitation coverage rate;
s1.3, carrying out normalization processing on the data, wherein a specific formula is as follows,
Figure 521328DEST_PATH_IMAGE001
wherein, X * Representing normalized radar echo intensity values, X max Indicating the maximum value of the radar echo intensity, X min Representing the minimum value of the radar echo intensity, and X represents the radar echo intensity value;
s1.4, dividing the data set by adopting a proportion of 8.
3. The method of claim 1, wherein the method of predicting precipitation based on Unet and DCN _ LSTM comprises: in the step S2, the method for constructing the hybrid model of Unet and DCN _ LSTM comprises the following steps
S2.1, inputting training set data into a model by a hybrid model encoder part, changing a characteristic diagram into a half of the original size through two 3 x 3 convolutional layers and a maximum pooling layer, and doubling the number of channels through the two 3 x 3 convolutional layers;
s2.2, adopting DCN _ LSTM in the middle of the hybrid model for extracting the time characteristics and the space characteristics of the radar echo sequence, wherein the DCN _ LSTM consists of a plurality of DCN _ LSTM circulating units and is used for decomposing the characteristic diagram output by the encoder and sequentially inputting the characteristic diagram into the DCN _ LSTM circulating units for training;
s2.3, the mixed model decoder part splices the radar echo sequence output by the DCN _ LSTM according to channels, then passes through two 3 x 3 convolutional layers, is subjected to up-sampling, is connected with a characteristic diagram output by the encoder in a skipping mode, passes through the two 3 x 3 convolutional layers and one 1 x 1 convolutional layer, and finally outputs a predicted radar echo sequence.
4. The method of claim 3, wherein the method of predicting precipitation based on Unet and DCN _ LSTM comprises: in the step S2.2, the DCN _ LSTM learns the offset of the input X to the hidden state H and the memory cell C by using the variable convolution, so as to update the hidden state H and the memory cell C, slides on the input image, and the variable convolution takes the obtained feature map as an input, and applies a convolution layer to the feature map, so as to obtain the variable convolution offset, wherein the offset layer is 2N, and is translated on a plane, and two directions X and y need to be changed;
the specific formula for the feasible convolution is as follows,
Figure 641731DEST_PATH_IMAGE002
wherein, R represents a convolution kernel of 3 × 3, (-1, -1), (-1, 0), (0, 1), (1, 1) represents points in the convolution kernel, and coordinates are integers;
Figure 890310DEST_PATH_IMAGE003
wherein, the first and the second end of the pipe are connected with each other,
Figure 551098DEST_PATH_IMAGE004
representing a feature matrix resulting from the feasible convolution,
Figure 568733DEST_PATH_IMAGE005
represents the amount of learning of each point obtained by neural network learning within a convolution kernel of size 3 x 3,
Figure 656774DEST_PATH_IMAGE006
represents the center point, i.e., (0, 0) point,
Figure 127070DEST_PATH_IMAGE007
representing points defined in the R range, with more offset matrices learned by convolution than standard convolution feasible convolutions
Figure 591549DEST_PATH_IMAGE008
。
5. The method of claim 4 for prediction of precipitation based on Unet and DCN _ LSTM, wherein: the DCN _ LSTM model comprises a plurality of DCN _ LSTM circulation units, the characteristic information is screened and transmitted through a gating mechanism, and a forgetting gate, an input gate, a modulation gate, an output gate, a time memory cell and a hidden state of convLSTM are reserved and are respectively f t、 i t、 g t、 o t、 C t And H t (ii) a Also includes space cells M t For extracting and transferring spatial structure features vertically between different layers, and adding feasible convolution to learn the offset of input X to hidden state H and memory cell C, the concrete formula is as follows,
Figure 463690DEST_PATH_IMAGE009
where DCN represents a variable convolutional network, X t Representing the picture entered, the lower coordinate t representing the moment of entry,
Figure 457054DEST_PATH_IMAGE010
and
Figure 476963DEST_PATH_IMAGE011
respectively representing the hidden state and the memory cells, the lower coordinate t-1 representing the previous moment, the upper coordinate 1 representing the first layer,
Figure 417237DEST_PATH_IMAGE010
and
Figure 206201DEST_PATH_IMAGE011
respectively representing the new hidden state and the memory cell obtained after the updating.
6. The method of claim 4 for predicting precipitation based on Unet and DCN _ LSTM, wherein: in the DCN _ LSTM model, an input gate, an update gate and a forgetting gate for updating memory cells are defined as follows,
Figure 839308DEST_PATH_IMAGE012
wherein i t An input gate representing a refresh memory cell; σ represents an activation function sigmoid; w xi A parameter matrix representing the training of input X, input gate i; w hi A parameter matrix representing the training of the input gate i for the hidden state H; x t An input representing time t;
Figure 346513DEST_PATH_IMAGE013
representing the hidden state of the l-th layer at the time t-1; b i Represents the offset to input gate i;
g t a refresh gate indicating a refresh of the memory cell; tanh represents an activation function tanh; w xg Representing a parameter matrix trained for updating gate g for input X; w hg Representing that the parameter matrix of the training of the gate g is updated for the hidden state H; b g Represents the bias to the refresh gate g;
f t a forgetting gate that indicates a renewed memory cell; w xf A parameter matrix representing the training of the forgetting gate f for the input X; w hf Representing a parameter matrix for training a hidden state H and a forgetting gate f; b is a mixture of f Represents a bias to the forgetting gate f; * Represents a convolution;
an input gate for updating the cells in space, an update gate and a forgetting gate, the formula is as follows,
Figure 824899DEST_PATH_IMAGE014
wherein the content of the first and second substances,
Figure 468370DEST_PATH_IMAGE015
an input gate representing the update space cells; σ represents an activation function sigmoid;
Figure 69115DEST_PATH_IMAGE016
a parameter matrix representing the training of input X, input Gate i; w mi A parameter matrix representing the training of the input gate i for the space cell M; x t An input representing time t;
Figure 1299DEST_PATH_IMAGE017
represents the space cell of the l-1 st layer at the time point t;
Figure 345693DEST_PATH_IMAGE018
represents the offset to input gate i;
Figure 46933DEST_PATH_IMAGE019
an update gate representing an updated spatial cell; tanh represents an activation function tanh;
Figure 818580DEST_PATH_IMAGE020
representing a parameter matrix trained for updating gate g for input X; w mg Representing a parameter matrix for updating gate g training on the space cell M;
Figure 34797DEST_PATH_IMAGE021
represents the bias to the refresh gate g;
Figure 120565DEST_PATH_IMAGE022
a forgetting gate representing an updated spatial cell;
Figure 738628DEST_PATH_IMAGE023
a parameter matrix representing the training of the forgetting gate f for the input X; w mf A parameter matrix representing the training of the spatial cell M and the forgetting gate f;
Figure 884439DEST_PATH_IMAGE024
represents a bias to the forgetting gate f; * Represents a convolution;
the hidden state, i.e. the output, is updated by the memory cells, the space cells and the output gate, the specific formula is as follows,
Figure 587952DEST_PATH_IMAGE025
wherein the content of the first and second substances,
Figure 211832DEST_PATH_IMAGE026
memory cells representing the first layer at time t; i.e. i t Representing an input gate; g is a radical of formula t Represents an update gate; f. of t Indicating a forgotten door;
Figure 949981DEST_PATH_IMAGE027
memory cells in layer I at time t-1;
Figure 63430DEST_PATH_IMAGE028
represents the space cell of the l layer at the t time;
Figure 191923DEST_PATH_IMAGE029
an input gate representing a spatial cell;
Figure 681810DEST_PATH_IMAGE030
an update gate representing a spatial cell;
Figure 946570DEST_PATH_IMAGE031
a forgetting gate representing a spatial cell;
Figure 496500DEST_PATH_IMAGE032
represents the space cell of l-1 layer at the t-th time; o t An output gate is shown; σ represents an activation function sigmoid;
W xo a parameter training matrix representing the output gate o for input X; w is a group of ho A parameter training matrix representing the output gate o for the hidden state H; w co A matrix of training parameters representing the input gate o for memory cell C; w is a group of mo Representing training parameters for spatial cell M, input Gate oA number matrix; x t An input representing time t;
Figure 909026DEST_PATH_IMAGE033
representing the hidden state of the l-th layer at the time t-1;
Figure 140288DEST_PATH_IMAGE034
memory cells representing the first layer at time t;
Figure 587449DEST_PATH_IMAGE035
representing the space cells of the l layer at the t time; b is a mixture of o Indicating an offset to the output gate; w is a group of 1x1 A convolution kernel representing a size of 1 × 1;
Figure 980385DEST_PATH_IMAGE036
representing the new hidden state obtained by updating; degree represents a multiplication of the matrices in bits; * Representing a convolution.
7. The method of claim 1, wherein the method of predicting precipitation based on Unet and DCN _ LSTM comprises: in the step S3, the number of the neurons, the batch size and the learning rate of the hidden layer are optimized through a Bayesian algorithm, and the method comprises the following steps
S3.1, assuming a set of hyper-parametric combinations is
Figure 145787DEST_PATH_IMAGE037
Wherein
Figure 711897DEST_PATH_IMAGE038
Respectively representing the parameter combinations of the number of neurons of the hidden layer, the batch size and the learning rate, and assuming that a loss function and a set hyper-parameter have a mapping relation;
assuming the function f x → R, it is necessary to determine the value of x ∈ R
Figure 210969DEST_PATH_IMAGE039
S3.2, obtaining the random initialization point of the hyper-parameter in the parameter range according to the determined and optimized hyper-parameter
Figure 837122DEST_PATH_IMAGE037
Inputting experimental data to train the model, the response value of the loss function is
Figure 427503DEST_PATH_IMAGE040
Establishing a Gaussian regression process;
known data set
Figure 62884DEST_PATH_IMAGE041
Suppose f is obeyed
Figure 156742DEST_PATH_IMAGE042
So that the prediction also follows a normal distribution,
Figure 219376DEST_PATH_IMAGE044
Figure 828212DEST_PATH_IMAGE045
Figure 204967DEST_PATH_IMAGE046
Figure 481227DEST_PATH_IMAGE047
wherein K is a constant, K (x) and
Figure 652445DEST_PATH_IMAGE048
a covariance matrix is represented by a value of the covariance matrix,
Figure 14157DEST_PATH_IMAGE049
represents the variance of sample n;
find out
Figure 929023DEST_PATH_IMAGE050
And
Figure 794211DEST_PATH_IMAGE051
；
s3.3, selecting the next hyper-parameter combination sampling point from the Gaussian regression model based on the sampling function PI
Figure 198647DEST_PATH_IMAGE052
The sampling function PI is, as follows,
Figure 985338DEST_PATH_IMAGE053
where Φ () represents a normal distribution cumulative density function,
Figure 500633DEST_PATH_IMAGE050
and
Figure 689169DEST_PATH_IMAGE051
respectively representing the mean and variance of the objective function value,
Figure 264506DEST_PATH_IMAGE054
the value of the optimum objective function is represented,
Figure 335231DEST_PATH_IMAGE055
representing a parameter;
s3.4, bringing the selected first group of hyper-parameter combinations into model training, outputting real values of ground observation and predicted mean square error values of radar echo sequences, and if the mean square error values are smaller than a preset threshold value, stopping updating and outputting optimal hyper-parameter combinations; if the mean square error value is not less than the preset threshold value, the hyper-parameter is updated to
Figure 591900DEST_PATH_IMAGE052
And repeating the step S3.2 to the step S3.4 until the hyper-parameter combination with the mean square error value smaller than the preset threshold value is found.
8. The method of claim 7, wherein the method of predicting precipitation based on Unet and DCN _ LSTM comprises: in step 3.4, the preset threshold is set to 0.0001.
9. The method of claim 1 for prediction of precipitation based on Unet and DCN _ LSTM, wherein: in step S4, the method for testing the hybrid model of Unet and DCN _ LSTM includes the following steps
S4.1, loading the weight of model training, testing and storing the weight in a picture format;
s4.2, adopting mean square error, structural similarity and critical success index as evaluation indexes of the test set;
the mean square error is used for evaluating the difference of pixel points of two pictures, and a specific formula is as follows,
Figure 697259DEST_PATH_IMAGE056
where n denotes the total number of samples, i denotes the number of sample points, Y denotes the real tag of the real radar echo pattern,
Figure 381181DEST_PATH_IMAGE057
representing a predicted radar echo map;
the structural similarity is used for measuring the similarity of two pictures, and the specific formula is as follows,
Figure 939201DEST_PATH_IMAGE058
wherein u is x And u y Denotes the mean, σ, of x and y, respectively x And σ y The variance for x and y is represented separately,σ xy represents the covariance of the two pictures x and y, C 1 And C 2 Represents a constant;
the specific formula for the critical success index is as follows:
Figure 796299DEST_PATH_IMAGE059
wherein TP indicates that the true category is a positive prediction result and also a positive prediction result, FP indicates that the true category is a negative prediction result and is a positive prediction result, and FN indicates that the true category is a positive prediction result and is a negative prediction result.
10. The method of claim 1, wherein the method of predicting precipitation based on Unet and DCN _ LSTM comprises: in step S5, the result of the prediction of the test set is converted into radar reflectivity through a pixel value, and the specific formula is as follows,
Figure 693848DEST_PATH_IMAGE060
wherein, radar _ value represents the value of radar reflectivity of each pixel point converted by a formula, and pixel _ value represents the value of each pixel point;
then, the rainfall is obtained according to the relation between the radar reflectivity and the rainfall, the concrete formula is as follows,
Figure 610988DEST_PATH_IMAGE061
wherein Z represents radar reflectivity, R represents rainfall, and A and b represent coefficients.
CN202211233428.1A 2022-10-10 2022-10-10 Precipitation prediction method based on Unet and DCN _ LSTM Pending CN115310724A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211233428.1A CN115310724A (en) 2022-10-10 2022-10-10 Precipitation prediction method based on Unet and DCN _ LSTM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211233428.1A CN115310724A (en) 2022-10-10 2022-10-10 Precipitation prediction method based on Unet and DCN _ LSTM

Publications (1)

Publication Number Publication Date
CN115310724A true CN115310724A (en) 2022-11-08

Family

ID=83867631

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211233428.1A Pending CN115310724A (en) 2022-10-10 2022-10-10 Precipitation prediction method based on Unet and DCN _ LSTM

Country Status (1)

Country Link
CN (1) CN115310724A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116047631A (en) * 2023-03-31 2023-05-02 中科星图维天信(北京)科技有限公司 Precipitation prediction method and device, electronic equipment and storage medium
CN116089884A (en) * 2023-02-09 2023-05-09 安徽省气象台 Method for constructing near-real-time precipitation estimation model and near-real-time precipitation estimation method
CN116307267A (en) * 2023-05-15 2023-06-23 成都信息工程大学 Rainfall prediction method based on convolution
CN116307283A (en) * 2023-05-19 2023-06-23 青岛科技大学 Precipitation prediction system and method based on MIM model and space-time interaction memory
CN116485010A (en) * 2023-03-20 2023-07-25 四川省雅安市气象局 S2S precipitation prediction method based on cyclic neural network
CN116719002A (en) * 2023-08-08 2023-09-08 北京弘象科技有限公司 Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210004727A1 (en) * 2019-06-27 2021-01-07 Mohamad Zaim BIN AWANG PON Hyper-parameter tuning method for machine learning algorithms using pattern recognition and reduced search space approach
CN113156325A (en) * 2021-03-18 2021-07-23 吉林大学 Method for estimating state of health of battery
CN113554148A (en) * 2021-06-07 2021-10-26 南京理工大学 BiLSTM voltage deviation prediction method based on Bayesian optimization

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210004727A1 (en) * 2019-06-27 2021-01-07 Mohamad Zaim BIN AWANG PON Hyper-parameter tuning method for machine learning algorithms using pattern recognition and reduced search space approach
CN113156325A (en) * 2021-03-18 2021-07-23 吉林大学 Method for estimating state of health of battery
CN113554148A (en) * 2021-06-07 2021-10-26 南京理工大学 BiLSTM voltage deviation prediction method based on Bayesian optimization

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
FENG JIANG: "DLU-Net for Pancreatic Cancer Segmentation", 《2020 IEEE INTERNATIONAL CONFERENCE ON BIOINFORMATICS AND BIOMEDICINE (BIBM)》 *
HAORAN CHEN: "FC-ZSM: Spatiotemporal", 《FRONTIERS IN EARTH SCIENCE》 *
何津祥等: "肿瘤治疗与循证医学及相关网址", 《卫生职业教育》 *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116089884A (en) * 2023-02-09 2023-05-09 安徽省气象台 Method for constructing near-real-time precipitation estimation model and near-real-time precipitation estimation method
CN116485010A (en) * 2023-03-20 2023-07-25 四川省雅安市气象局 S2S precipitation prediction method based on cyclic neural network
CN116485010B (en) * 2023-03-20 2024-04-16 四川省雅安市气象局 S2S precipitation prediction method based on cyclic neural network
CN116047631A (en) * 2023-03-31 2023-05-02 中科星图维天信(北京)科技有限公司 Precipitation prediction method and device, electronic equipment and storage medium
CN116307267A (en) * 2023-05-15 2023-06-23 成都信息工程大学 Rainfall prediction method based on convolution
CN116307267B (en) * 2023-05-15 2023-07-25 成都信息工程大学 Rainfall prediction method based on convolution
CN116307283A (en) * 2023-05-19 2023-06-23 青岛科技大学 Precipitation prediction system and method based on MIM model and space-time interaction memory
CN116307283B (en) * 2023-05-19 2023-08-18 青岛科技大学 Precipitation prediction system and method based on MIM model and space-time interaction memory
CN116719002A (en) * 2023-08-08 2023-09-08 北京弘象科技有限公司 Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium
CN116719002B (en) * 2023-08-08 2023-10-27 北京弘象科技有限公司 Quantitative precipitation estimation method, quantitative precipitation estimation device, electronic equipment and computer storage medium

Similar Documents

Publication Publication Date Title
CN115310724A (en) Precipitation prediction method based on Unet and DCN _ LSTM
CN111859800B (en) Space-time estimation and prediction method for PM2.5 concentration distribution
Wang et al. Comparison of convolutional neural networks for landslide susceptibility mapping in Yanshan County, China
CN113011397B (en) Multi-factor cyanobacterial bloom prediction method based on remote sensing image 4D-Fractalnet
CN112415521A (en) CGRU (China-swarm optimization and RU-based radar echo nowcasting) method with strong space-time characteristics
CN114943365A (en) Rainfall estimation model establishing method fusing multi-source data and rainfall estimation method
CN114611608A (en) Sea surface height numerical value prediction deviation correction method based on deep learning model
CN115629160A (en) Air pollutant concentration prediction method and system based on space-time diagram
CN116844041A (en) Cultivated land extraction method based on bidirectional convolution time self-attention mechanism
CN117233869B (en) Site short-term wind speed prediction method based on GRU-BiTCN
CN112766099B (en) Hyperspectral image classification method for extracting context information from local to global
CN113591608A (en) High-resolution remote sensing image impervious surface extraction method based on deep learning
CN117194926A (en) Method and system for predicting hoisting window period of land wind power base
Kaparakis et al. Wf-unet: Weather fusion unet for precipitation nowcasting
CN112529270A (en) Water flow prediction model based on deep learning
CN117011668A (en) Weather radar echo extrapolation method based on time sequence prediction neural network
CN115984132A (en) Short-term prediction method based on CBAIM differential recurrent neural network
CN116148864A (en) Radar echo extrapolation method based on DyConvGRU and Unet prediction refinement structure
Yao et al. A Forecast-Refinement Neural Network Based on DyConvGRU and U-Net for Radar Echo Extrapolation
CN113222206B (en) Traffic state prediction method based on ResLS-C deep learning combination
CN112380985A (en) Real-time detection method for intrusion foreign matters in transformer substation
Zhang et al. MMSTP: Multi-modal Spatiotemporal Feature Fusion Network for Precipitation Prediction
CN117808650B (en) Precipitation prediction method based on Transform-Flownet and R-FPN
CN112926619B (en) High-precision underwater laser target recognition system
Zhang et al. A Multi-task two-stream spatiotemporal convolutional neural network for convective storm nowcasting

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20221108

RJ01 Rejection of invention patent application after publication