CN115267945A - Thunder and lightning early warning method and system based on graph neural network - Google Patents

Thunder and lightning early warning method and system based on graph neural network Download PDF

Info

Publication number
CN115267945A
CN115267945A CN202210751766.8A CN202210751766A CN115267945A CN 115267945 A CN115267945 A CN 115267945A CN 202210751766 A CN202210751766 A CN 202210751766A CN 115267945 A CN115267945 A CN 115267945A
Authority
CN
China
Prior art keywords
lightning
data
neural network
thunder
group
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
CN202210751766.8A
Other languages
Chinese (zh)
Inventor
包雅孟
童充
徐建波
洪晨威
陈海文
周海阔
陈振伟
袁婧
苏俊霖
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.)
Suzhou Power Supply Co of State Grid Jiangsu Electric Power Co Ltd
Original Assignee
Suzhou Power Supply Co of State Grid Jiangsu Electric Power Co Ltd
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 Suzhou Power Supply Co of State Grid Jiangsu Electric Power Co Ltd filed Critical Suzhou Power Supply Co of State Grid Jiangsu Electric Power Co Ltd
Priority to CN202210751766.8A priority Critical patent/CN115267945A/en
Publication of CN115267945A publication Critical patent/CN115267945A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/10—Devices for predicting weather conditions
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/02—Instruments for indicating weather conditions by measuring two or more variables, e.g. humidity, pressure, temperature, cloud cover or wind speed
    • 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
    • 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
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Environmental & Geological Engineering (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Biophysics (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Atmospheric Sciences (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Ecology (AREA)
  • Environmental Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

A thunder and lightning early warning method based on a graph neural network comprises the following steps: selecting a plurality of observation points, determining a moment to be predicted for any observation point, and acquiring a group of thunder and lightning activity information of the moment to be predicted, a group of meteorological data and a group of thunder and lightning data of a corresponding time period before the moment to be predicted; carrying out sample training on a plurality of groups of lightning activity information, a plurality of groups of meteorological data and a plurality of groups of lightning data which are formed by a plurality of observation points, and constructing a corresponding graph neural network model for predicting lightning activity; and acquiring a group of meteorological data and a group of lightning data at the current moment, inputting the meteorological data and the lightning data into the graph neural network model, and acquiring lightning activity information corresponding to the current moment. Compared with the prior art, the model considers the combination of the neural network from two dimensions, models the relation among different data types from the perspective of variables, and mines the potential relation among the variables. From the angle of time, the data vectors at different moments are modeled, and the mutual influence among the data vectors at different moments is mined.

Description

Thunder and lightning early warning method and system based on graph neural network
Technical Field
The invention belongs to the field of lightning approach early warning, and particularly relates to a lightning early warning method and system based on a graph neural network.
Background
The thunder disaster is a natural disaster, greatly threatens the daily life of human beings, and causes a great amount of casualties and economic losses. In order to effectively reduce the influence of lightning disasters on the development of the economic society and avoid the occurrence of major casualties and economic loss accidents, the research and the application of lightning early warning technology are more and more important.
The early warning of the lightning approach is realized, and the weather forecast generally refers to the forecast of 0-2 hours as the early warning of the approach. The main current technology for the lightning approach early warning at present is a method for obtaining a lightning early warning result by judging and extrapolating the change trend of radar data and lightning positioning data to obtain possible positions and values of radar and lightning data at a future moment, combining the relation between the predicted values of the radar and the lightning and the threshold value of lightning, or regressing according to a historical process to obtain a linear or multivariate multiple equation of the occurrence of the radar, the lightning and the lightning. For example, in the early warning of lightning activity based on an electric field instrument, whether to issue a warning of the impending lightning is determined by checking the variation characteristics of the electric field at a fixed location, the time of which is usually within 30min, and the early warning is mainly aimed at a small area of a specific target (such as a golf course, a chemical industry area). Proximity warnings of greater range, longer time aging require the assistance of other observations.
The methods have high accuracy, but the parameters such as threshold values, weights and the like need to be adjusted manually aiming at different regions and different seasons. And because the generation mechanism of the lightning is complex and has certain nonlinear characteristics, the process of quantitatively analyzing the prediction factors is very complicated and has no general applicability, and the applicability is limited.
In order to solve the nonlinear problem and establish a model with adaptability and fault tolerance, some attempts have been made to perform weather feature-to-time mapping by using a machine learning method to warn thunder and lightning. Lugwavetao et al [ lugwavetao, zhangyi army, bengal, yawn, nux and Wen, nux, maming, zheng dong, wangfei, 2009: the lightning approach early warning method and system are developed [ J ] weather, 35 (5): 10 ] a decision tree and area identification method is integrated, and a lightning approach prediction system is established. Zhou Ming Wei, etc. [ Zhou Ming Wei, shao gan, zhang Qilin, etc. [ thunb storm potential forecast initial exploration date based on support vector machine ] atmospheric science report, 2018,41 (4): 569-576] adopts the support vector machine to establish a thunder and lightning potential early warning model for the NCEP data, applies the characteristic that the correlation coefficient is greater than 0.3, and compares with other classification models. Chen Yong Wei and the like [ Chen Yong Wei, zheng, wang Han 22531, and the like, 2013. Thunder and lightning potential forecast based on a BP neural network model [ J ]. Drought weather, 31 (3): 595-601] extracts convection parameters as input factors of a neural network, and predicts the potential of thunder and lightning activities. Tianhao et al [ Tianhao, chapter connotation, von Wanxing et al ] the lightning prediction method based on BP neural network and atmospheric electric field characteristics [ J ] the electric porcelain lightning arrester, 2018 (6): 27-33] uses the BP neural network to extract the atmospheric electric field characteristics taking 30 minutes as a time slice, and warns about future lightning occurrence events.
The traditional ground lightning early warning method based on the atmospheric electric field comprises a threshold early warning method, an electric field grading early warning method, a dynamic threshold setting method and the like. However, the thresholds need to be set manually, how to optimally select the thresholds is a difficulty, and the value ranges of the thresholds are different for different regions.
In addition, the meteorological signals collected by the sensors are regularly distributed, for example, the temperature sensors are unevenly distributed on the ocean or land and are not grid positions of a fixed structure. Compared with the traditional convolutional neural network which can only simulate standard grid data, the graph neural network can process data of a non-Euclidean domain and directly carry out feature learning and prediction on a graph structure.
Disclosure of Invention
In order to solve the defects in the prior art, the invention aims to solve the problems and further provides a lightning early warning method and system based on a graph neural network.
The invention adopts the following technical scheme.
A thunder and lightning early warning method based on a graph neural network comprises the following steps:
the method comprises the following steps of S1, selecting a plurality of observation points, determining a moment to be predicted for any observation point, and acquiring a group of thunder and lightning activity information of the moment to be predicted, a group of meteorological data and a group of thunder and lightning data of a corresponding time period before the moment to be predicted;
s2, carrying out sample training on a plurality of groups of thunder and lightning activity information, a plurality of groups of meteorological data and a plurality of groups of thunder and lightning data which are formed by a plurality of observation points at a plurality of moments to be predicted, and constructing a corresponding graph neural network model for forecasting the thunder and lightning activity;
and S3, acquiring a group of meteorological data and a group of lightning data at the current moment, inputting the meteorological data and the lightning data into the graph neural network model, and acquiring lightning activity information corresponding to the current moment.
Further, in step S1, the set of lightning activity information, the set of meteorological data and the set of lightning data are
Figure BDA0003721220000000021
Wherein Y represents the lightning level in the lightning activity information, and {0,1,2,3} represents no lightning, small lightning, medium lightning, and extra lightning, respectively; t is t1,…,tTRepresenting a corresponding time period before the moment to be predicted, wherein the variable set M comprises all data types of meteorological data and lightning data; establishing a special graph according to s, wherein one node in the special graph is a variable
Figure BDA0003721220000000031
Wherein k =1,2, \8230, T, m = alpha, beta, gamma \8230; redefining a node to represent as
Figure BDA0003721220000000032
Further, the meteorological data comprise temperature, humidity, wind speed and direction, pressure and precipitation data; the lightning data comprises electric field intensity, satellite observation cloud top brightness temperature and radar echo physical quantity.
Further, step S2 specifically includes:
step S21, mapping the nodes to the same characteristic space according to the types of the nodes through a preset conversion matrix; wherein the node
Figure BDA0003721220000000033
The type of (b) is m;
step S22, according to the mapped nodes, calculating variable embedding representation z of each typeα；
Step S23, embedding representation z in all variablesαCarrying out weighted fusion to obtain a final graph embedding representation R;
step S24, according to the graph embedding expression R, the lightning activity information is predicted through an MLP layer
Figure BDA0003721220000000034
Figure BDA0003721220000000035
Wherein p is0,p1,p2,p3Respectively representing the probability of no lightning, small lightning, medium lightning and extra large lightning;
and S25, constructing a loss function, and minimizing the loss function based on a gradient descent method, thereby training parameters in the neural network model of the learning diagram.
Further, step S21 specifically includes:
according to a predetermined transformation matrix UαThus will not beThe same type of nodes are mapped to the same feature space:
Figure BDA0003721220000000036
wherein the content of the first and second substances,
Figure BDA0003721220000000037
is a mapped feature representation of the node,
Figure BDA0003721220000000038
m，nαall represent dimensions;
Figure BDA0003721220000000039
is a randomly initialized weight learning matrix.
Further, step S22 specifically includes:
step S221, for any node viCompute its neighborhood-embedded representation for any type of beta
Figure BDA00037212200000000310
Figure BDA00037212200000000311
Wherein, the first and the second end of the pipe are connected with each other,
Figure BDA00037212200000000312
for a node v in a heterogeneous graphiIs a set of nodes of type beta in the neighborhood of (a),
Figure BDA00037212200000000313
denotes the adjacency matrix after joining self-join, hjRepresenting a node vjThe mapped feature indicates i, j =1,2, \8230;, Nv。
Step S222, calculating a node v according to the neighborhood embedding representationiAttention score for type beta
Figure BDA0003721220000000041
Figure BDA0003721220000000042
Where | represents a join operation,
Figure BDA0003721220000000043
is an attention vector of the beta type, all beta type nodes share
Figure BDA0003721220000000044
σ (-) represents the activation function;
step S223, according to the attention scores, calculating the node v of the type beta with the time difference of K standard time scalesjTo node viDegree of importance of
Figure BDA0003721220000000045
Figure BDA0003721220000000046
Figure BDA0003721220000000047
Wherein the content of the first and second substances,
Figure BDA0003721220000000048
is the attention vector, t is the node viThe time of the above-mentioned (c) is,
Figure BDA0003721220000000049
node v of type α time tiThe mapped feature representation;
step S224, by
Figure BDA00037212200000000410
Computing variable embedded representation zα；
Figure BDA00037212200000000411
Figure BDA00037212200000000412
I denotes the splicing operation, zα∈RTm×1Tm represents a dimension.
Further, step S23 specifically includes:
according to the self-attention model, embedding expressions for variables of various types, and mapping the embedded expressions to three different spaces to obtain three vectors:
Figure BDA00037212200000000413
and
Figure BDA00037212200000000414
dk,dvare all preset hyper-parameters, qα,kα,vαRespectively represent Query, key, value in the self-attention model.
qα＝WQ·zα
kα＝WK·zα
vα＝WV·zα
Figure BDA0003721220000000051
Figure BDA0003721220000000052
WQ，WK，WVIs three learnable parameter matrixes, the importance mark w of each input to be learntαAs weighting coefficients, weighting and polymerizing to obtain a final graph embedding R; wherein the content of the first and second substances,
Figure BDA0003721220000000053
dvthe dimensions are represented.
Further, step S24 specifically includes:
Figure BDA0003721220000000054
wherein the content of the first and second substances,
Figure BDA0003721220000000055
is a learnable parameter, b ∈ R4×1Represents the bias, σ (·) represents the sigmod function; y belongs to R4×1＝{p0,p1,p2,p3And (4) respectively representing probability estimation of no lightning, small lightning, medium lightning and extra large lightning.
Further, step S25 specifically includes:
the model parameters include: "U" in step S221α", in step S222
Figure BDA0003721220000000056
In step S223
Figure BDA0003721220000000057
"W" in step S23Q，WK，WV", W, b" in step S24.
Model parameters were optimized by computing a minimization loss function:
Figure BDA0003721220000000058
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003721220000000059
wherein c =0,1,2,3.
A lightning early warning system based on a graph neural network comprises: the system comprises a sample acquisition module and a graph neural network module;
the sample acquisition module is used for acquiring a group of thunder and lightning activity information at the moment to be predicted, a group of meteorological data and a group of thunder and lightning data in a corresponding time period before the moment to be predicted, and a group of meteorological data and a group of thunder and lightning data at the current moment;
the graph neural network module is used for constructing a corresponding graph neural network model for predicting the lightning activity, and obtaining the lightning activity information corresponding to the current moment according to a group of meteorological data and a group of lightning data of the current moment.
Compared with the prior art, the invention has the advantages that:
the invention provides a thunder and lightning early warning and forecasting method based on a graph neural network. Then, a corresponding graph neural network is determined based on the graph structure, and sample data, namely a multivariate time sequence of observation points, is input into the graph neural network, so that the lightning activity condition corresponding to the sample is predicted.
Compared with the prior art, the model considers the combination of the neural network from two dimensions, (1) from the perspective of variables, the relation among different data types is modeled, and the potential relation among the variables is mined. (2) From the angle of time, the data vectors at different moments are modeled, and the mutual influence among the data vectors at different moments is mined. Moreover, the model can be trained and learned end to end without manually setting parameter thresholds, and the universality of the model is improved.
In addition, the work of the invention provides a new research idea for applying the graph neural network to the lightning activity prediction.
Drawings
FIG. 1 is a flow chart of a lightning early warning method based on a graph neural network.
FIG. 2 is a schematic diagram of a lightning early warning model based on a graph neural network.
Detailed Description
The present application is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present application is not limited thereby.
As shown in fig. 1, it is a flow chart of the lightning early warning method based on graph neural network of the present invention, and the method includes:
the method comprises the following steps that S1, a plurality of observation points are selected, a moment to be predicted is determined for any observation point, a group of thunder and lightning activity information of the moment to be predicted is obtained, and a group of meteorological data and a group of thunder and lightning data of a corresponding time period before the moment to be predicted are obtained;
in this embodiment, the meteorological data includes data types such as temperature, humidity, wind speed and direction, pressure, precipitation data, and the like; the lightning data comprises data types such as electric field intensity, satellite observation cloud top brightness temperature, radar echo physical quantity and the like; the thunder and lightning activity information mainly refers to the thunder and lightning level.
According to the method and the device, multivariate time sequence data (namely meteorological data and lightning data) are modeled, and model training is performed by utilizing the sample set, so that the relation prediction between the multivariate time sequence sample data and lightning activities is realized.
For convenience in description, a plurality of groups of lightning activity information, a plurality of groups of meteorological data and a plurality of groups of lightning data which are formed by a plurality of observation points are respectively called as a lightning activity information sample set, a meteorological data sample set and a lightning data sample set; the lightning activity information sample set, the meteorological data sample set, and the lightning data sample set may be collectively represented as a sample set S.
A set of lightning activity information, a set of meteorological data, and a set of lightning data may be represented collectively as
Figure BDA0003721220000000071
Figure BDA0003721220000000072
Y belongs to {0,1,2,3}. Obviously, the sample S is one element in the sample set S. Wherein Y represents lightning level, {0,1,2,3} divides lightning activity from weak to strong into 4 levels representing no lightning and small sizeThunder, medium-grade thunder and extra-large thunder; t is t1,…,tTAnd the variable set M comprises all data types of meteorological data and thunder and lightning data, and alpha and beta represent specific data types. For example: α may represent temperature, β may represent humidity, and γ may represent radar echo physical quantity; further, it is possible to prevent the occurrence of,
Figure BDA0003721220000000073
is the temperature data of the corresponding time period before the moment to be predicted,
Figure BDA0003721220000000074
for the humidity data of the corresponding period before the time to be predicted,
Figure BDA0003721220000000075
is humidity data of a corresponding time period before the moment to be predicted.
S2, carrying out sample training through the thunder and lightning activity information sample set, the meteorological data sample set and the thunder and lightning data sample set, and constructing a corresponding graph neural network model for forecasting the thunder and lightning activity;
the schematic diagram of the graph neural network model for lightning activity prediction in step S2 is shown in fig. 2, and specifically includes steps S21-S25:
step S21, firstly, mapping different types of node characteristics to the same characteristic space through a type-specific conversion matrix;
before step S2 is executed, in order to better capture the interaction relationship between the variables in the multivariate time series data, and to consider the influence of the observed values of the variables at different times. And taking the observed quantities of all the variables at different moments as nodes of a graph structure, wherein the relation between the observed quantities represents the edge of the network. Thereby determining a directional weighted differential map G = (V, E). Comprising a set of nodes V and a set of edges E and the corresponding adjacency matrix a. In addition, M = { α, β, γ, \8230; } represents a node type set.
It is understood that the number of nodes included in the heterogeneous graph G is Nv＝NMX T. Wherein, NMFor a set of variables MThe number of elements in (c).
It will be appreciated that each node in the set of nodes V is denoted ViWherein i =1,2, \ 8230, Nv. For convenience of representation, the observed quantity of each node can be represented as
Figure BDA0003721220000000076
At the same time, in some scenarios, will
Figure BDA0003721220000000081
Simplify the record to
Figure BDA0003721220000000082
Thus, α may be referred to as node viType of (1), will tk(or t) is called node viThe time of (c).
Thus, in some embodiments, step S21 may specifically include:
since the heterogeneous graph includes different types of nodes having different dimensions of feature spaces, for each type of node (e.g., a type of a node), a predetermined transformation matrix U is usedαThereby mapping different types of nodes to the same feature space.
Figure BDA0003721220000000083
Wherein the content of the first and second substances,
Figure BDA0003721220000000084
is node viIs used to represent the feature of the image,
Figure BDA0003721220000000085
its dimension nαIs a function of the node type alpha.
Figure BDA0003721220000000086
The dimension m is set in advance by all nodes.
Figure BDA0003721220000000087
Is a weight learning matrix, and is initialized randomly.
Step S22, then enter the hierarchical attention module to obtain the variable embedding representation zα。
Step S22 specifically includes:
step S221, for any node viCompute its neighborhood-embedded representation for any type of beta
Figure BDA0003721220000000088
Figure BDA0003721220000000089
Wherein, the first and the second end of the pipe are connected with each other,
Figure BDA00037212200000000810
for a node v in a heterogeneous graphiOf a neighborhood of (b) is a set of nodes of type beta,
Figure BDA00037212200000000811
denotes the adjacency matrix after joining self-join, hjRepresenting a node vjAnd (4) representing the mapped features.
Step S222, calculating a node v according to the neighborhood embedding representationiAttention score for type beta
Figure BDA00037212200000000812
Figure BDA00037212200000000813
Where | represents a join operation,
Figure BDA00037212200000000814
is an attention vector of type β (learnable parameters, random initialization), shared by all β -type nodes
Figure BDA00037212200000000815
σ (-) represents the activation function, which can be Leaky ReLU.
Step S223, according to the attention scores, calculating the node v of type beta with the time difference of K standard time scalesjTo node viDegree of importance of
Figure BDA00037212200000000816
Figure BDA00037212200000000817
Figure BDA00037212200000000818
Wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003721220000000091
is the attention vector (learnable parameter), t is the node viThe time of (a) is,
Figure BDA0003721220000000092
node v of type α time tiAnd (4) representing the mapped features.
It should be noted that "the time is different by K standard time scales" means: suppose vjCan be expressed as
Figure BDA0003721220000000093
And v isiCan be expressed as
Figure BDA0003721220000000094
(suppose viIs of the alpha type), ki, kj =1,2, \ 8230;, T. Then, | ki-kj | = K.
It should be noted that, in the following description,
Figure BDA0003721220000000095
is asymmetric, i.e. node viTo nodevjImportance of and node vjTo node viThe importance of (c) is different.
Step S224, by
Figure BDA0003721220000000096
Computing variable embedding representation zα；
By passing
Figure BDA0003721220000000097
To node viAll types of neighbor nodes with all historical time scales are subjected to weighted aggregation operation to obtain a new embedded representation of each node (the type is a node v with alpha time being t)iIts embedded representation can be noted as ziOr
Figure BDA0003721220000000098
). And embedding and splicing the nodes of the same type at different time to obtain variable embedding expression, wherein the calculation formula is as follows:
Figure BDA0003721220000000099
Figure BDA00037212200000000910
i denotes the splicing operation, zα∈RTm×1。
Step S23, embedding representation z in all variablesαPerforming weighted fusion to obtain a final graph embedding representation R;
different embedded representations of different types of variables are considered, having different effects on the final prediction result. Here we consider polymerizing them using a self-attention mechanism. According to the self-attention model, for each type of variable embedding representation, mapping the variable embedding representation to three different spaces, obtaining three vectors:
Figure BDA00037212200000000911
and
Figure BDA00037212200000000912
dk,dvare all preset hyper-parameters, qα,kα,vαRespectively represents Query, key and Value in the attention model.
qα＝WQ·zα
kα＝WK·zα
vα＝WV·zα
Figure BDA00037212200000000913
Figure BDA0003721220000000101
WQ，WK，WVThree learnable parameter matrices, randomly initialized. Importance score w of each input to be learnedαAnd as a weighting coefficient, weighting and aggregating to obtain the final graph embedding R. Wherein the content of the first and second substances,
Figure BDA0003721220000000102
step S24, forecasting thunder and lightning activity information through an MLP layer
Figure BDA0003721220000000103
The lightning activity information is subjected to prediction training, and in general, a multi-layer perceptron MLP or a softmax function can be used for predicting sample labels.
Figure BDA0003721220000000104
Wherein the content of the first and second substances,
Figure BDA0003721220000000105
is a learnable parameter, b ∈ R4×1Denotes bias, σ (·) denotes sigmod function. Y is formed by the element R4×1＝{p0,p1,p2,p3And (5) respectively representing probability estimation of no lightning, small lightning, medium lightning and extra lightning. It should be noted that, in the following description,
Figure BDA0003721220000000106
and Y are both vectors { p }0,p1,p2,p3And satisfy p0+p1+p2+p3＝1
It can be understood that, the steps S21 to S24 model any input sample S, mine the relationship between nodes of different data types and time from the perspective of data type and time, and output the final graph embedding R.
And S25, constructing a loss function, and minimizing the loss function based on a gradient descent method, thereby training the learning model parameters. The model learnable training parameters include: "U" in step S221α", in step S222
Figure BDA0003721220000000107
In step S223
Figure BDA0003721220000000108
"W" in step S23Q，WK，WV", W, b" in step S24. For initialization of these parameters, there are generally standard normal distribution initialization, uniform distribution initialization, xavier initialization, and the like. In this embodiment, all training parameters of the model are initialized randomly using Kaiming initialization.
For the design of the loss function, the cross entropy between the prediction label and the real label is adopted for calculation, and the model is trained by minimizing the loss function:
Figure BDA0003721220000000109
wherein the content of the first and second substances,
Figure BDA00037212200000001010
wherein c =0,1,2,3. Under the guidance of the labeled data, the model parameters can be optimized through a back propagation algorithm and learned node embedding.
And S3, acquiring a group of meteorological data and a group of lightning data at the current moment, inputting the meteorological data and the lightning data into the graph neural network model, and acquiring lightning activity information corresponding to the current moment. And finishing early warning of thunder and lightning according to the thunder and lightning grade in the thunder and lightning activity information.
The present applicant has described and illustrated embodiments of the present invention in detail with reference to the accompanying drawings, but it should be understood by those skilled in the art that the above embodiments are only preferred embodiments of the present invention, and the detailed description is only for the purpose of helping the reader to better understand the spirit of the present invention, and not for the purpose of limiting the scope of the present invention, and on the contrary, any modifications or modifications based on the spirit of the present invention should fall within the scope of the present invention.

Claims (10)

1. A thunder and lightning early warning method based on a graph neural network is characterized by comprising the following steps:
the method comprises the following steps that S1, a plurality of observation points are selected, a moment to be predicted is determined for any observation point, a group of thunder and lightning activity information of the moment to be predicted is obtained, and a group of meteorological data and a group of thunder and lightning data of a corresponding time period before the moment to be predicted are obtained;
s2, carrying out sample training on a plurality of groups of thunder and lightning activity information, a plurality of groups of meteorological data and a plurality of groups of thunder and lightning data which are formed by a plurality of observation points at a plurality of moments to be predicted, and constructing a corresponding graph neural network model for thunder and lightning activity prediction;
and S3, acquiring a group of meteorological data and a group of lightning data at the current moment, inputting the meteorological data and the lightning data into the graph neural network model, and acquiring lightning activity information corresponding to the current moment.
2. According to claimThe method for warning lightning based on graph neural network as claimed in claim 1, wherein in step S1, the set of lightning activity information, the set of meteorological data and the set of lightning data are
Figure FDA0003721219990000011
Figure FDA0003721219990000012
Wherein Y represents the lightning level in the lightning activity information, and {0,1,2,3} represents no lightning, small lightning, medium lightning, and extra large lightning, respectively; t is t1,…,tTRepresenting a corresponding time period before the moment to be predicted, wherein the variable set M comprises all data types of meteorological data and lightning data; establishing an abnormal graph according to s, wherein one node in the abnormal graph is a variable
Figure FDA0003721219990000013
Wherein k =1,2, \8230, T, m = α, β, γ \8230; redefining a node to represent as
Figure FDA0003721219990000014
3. The lightning early warning method based on the graph neural network as claimed in claim 2, wherein the meteorological data comprises temperature, humidity, wind speed and direction, pressure and precipitation data; the lightning data comprises electric field intensity, satellite observation cloud top brightness temperature and radar echo physical quantity.
4. The lightning early warning method based on the graph neural network as claimed in claim 2, wherein the step S2 specifically comprises:
step S21, mapping the nodes to the same characteristic space according to the types of the nodes through a preset conversion matrix; wherein the node
Figure FDA0003721219990000015
The type of m is;
step S22, according to the mapped nodes, calculating variable embedding representation z of each typeα；
Step S23, embedding representation z in all variablesαPerforming weighted fusion to obtain a final graph embedding representation R;
step S24, according to the graph embedded representation R, the lightning activity information is predicted through an MLP layer
Figure FDA0003721219990000016
Figure FDA0003721219990000017
Wherein p is0,p1,p2,p3Respectively representing the probability of no lightning, small lightning, medium lightning and extra large lightning;
and S25, constructing a loss function, and minimizing the loss function based on a gradient descent method, thereby training parameters in the neural network model of the learning diagram.
5. The lightning early warning method based on the graph neural network as claimed in claim 4, wherein the step S21 specifically comprises:
according to a predetermined transformation matrix UαThus, different types of nodes are mapped to the same feature space:
Figure FDA0003721219990000021
wherein the content of the first and second substances,
Figure FDA0003721219990000022
is a node v of type alpha time tiThe characteristics after the mapping are represented by the characters,
Figure FDA0003721219990000023
Figure FDA0003721219990000024
m，nαall represent dimensions;
Figure FDA0003721219990000025
is a randomly initialized weight learning matrix.
6. The lightning early warning method based on the graph neural network as claimed in claim 4, wherein the step S22 specifically comprises:
step S221, for any node viComputing its neighborhood-embedded representation for any type of beta
Figure FDA0003721219990000026
Figure FDA0003721219990000027
Wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003721219990000028
for a node v in a heterogeneous graphiIs a set of nodes of type beta in the neighborhood of (a),
Figure FDA0003721219990000029
denotes the adjacency matrix after addition of self-joins, hjRepresenting a node vjMapped feature representation, i, j =1,2, \ 8230;, Nv；
Step S222, calculating a node v according to the neighborhood embedding representationiAttention score for type beta
Figure FDA00037212199900000210
Figure FDA00037212199900000211
Wherein, | | represents the splicing operation,
Figure FDA00037212199900000212
is an attention vector of the beta type, all beta type nodes share
Figure FDA00037212199900000213
σ (-) represents the activation function;
step S223, according to the attention scores, calculating the node v of type beta with the time difference of K standard time scalesjTo node viDegree of importance of
Figure FDA00037212199900000214
Figure FDA00037212199900000215
Figure FDA00037212199900000216
Wherein the content of the first and second substances,
Figure FDA00037212199900000217
is the attention vector, t is the node viThe time of (a) is,
Figure FDA00037212199900000218
node v of type α time tiThe mapped feature representation;
step S224, by
Figure FDA0003721219990000031
Computing variable embedded representation zα；
Figure FDA0003721219990000032
Figure FDA0003721219990000033
I denotes the splicing operation, zα∈RTm×1Tm represents a dimension.
7. The lightning early warning method based on the graph neural network as claimed in claim 4, wherein the step S23 specifically comprises:
according to the self-attention model, embedding expressions for variables of various types, and mapping the embedded expressions to three different spaces to obtain three vectors:
Figure FDA0003721219990000034
and
Figure FDA0003721219990000035
dk,dvare all preset hyper-parameters, qα,kα,vαRespectively represent Query, key, value in the self-attention model.
qα＝WQ·zα
kα＝WK·zα
vα＝WV·zα
Figure FDA0003721219990000036
Figure FDA0003721219990000037
WQ，WK，WVIs three learnable parameter matrixes, the importance scores w of the learnt inputsαAs a weighting coefficient, weighting and aggregating to obtain a final graph embedding R; wherein the content of the first and second substances,
Figure FDA0003721219990000038
dvthe dimensions are represented.
8. The lightning early warning method based on the graph neural network as claimed in claim 4, wherein the step S24 specifically comprises:
Figure FDA0003721219990000039
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA00037212199900000310
is a learnable parameter, b ∈ R4×1Represents the bias, σ (·) represents the sigmod function; y belongs to R4×1＝{p0,p1,p2,p3And (5) respectively representing probability estimation of no lightning, small lightning, medium lightning and extra lightning.
9. The lightning early warning method based on the graph neural network as claimed in claim 4, wherein the step S25 specifically comprises:
the model parameters include: "U" in step S221α", in step S222
Figure FDA0003721219990000041
In step S223
Figure FDA0003721219990000042
"W" in step S23Q，WK，WV"W, b" in step S24;
model parameters were optimized by computing a minimization loss function:
Figure FDA0003721219990000043
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003721219990000044
wherein c =0,1,2,3.
10. A thunder and lightning early warning system based on graph neural network, characterized by comprising: the system comprises a sample acquisition module and a graph neural network module;
the sample acquisition module is used for acquiring a group of thunder and lightning activity information at the moment to be predicted, a group of meteorological data and a group of thunder and lightning data in a corresponding time period before the moment to be predicted, and a group of meteorological data and a group of thunder and lightning data at the current moment;
the graph neural network module is used for constructing a corresponding graph neural network model for predicting the lightning activity, and obtaining the corresponding lightning activity information at the current moment according to a group of meteorological data and a group of lightning data at the current moment.
CN202210751766.8A 2022-06-29 2022-06-29 Thunder and lightning early warning method and system based on graph neural network Pending CN115267945A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210751766.8A CN115267945A (en) 2022-06-29 2022-06-29 Thunder and lightning early warning method and system based on graph neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210751766.8A CN115267945A (en) 2022-06-29 2022-06-29 Thunder and lightning early warning method and system based on graph neural network

Publications (1)

Publication Number Publication Date
CN115267945A true CN115267945A (en) 2022-11-01

Family

ID=83762981

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210751766.8A Pending CN115267945A (en) 2022-06-29 2022-06-29 Thunder and lightning early warning method and system based on graph neural network

Country Status (1)

Country Link
CN (1) CN115267945A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115859222A (en) * 2022-12-28 2023-03-28 四川省气象灾害防御技术中心 Lightning automatic prediction method based on multi-gradient feature fusion 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
CN116106856A (en) * 2023-04-13 2023-05-12 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) Identification model establishment method and identification method for thunderstorm strong wind and computing equipment
CN117688975A (en) * 2024-02-02 2024-03-12 南京信息工程大学 Meteorological event prediction method and system based on evolution rule mining

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115859222A (en) * 2022-12-28 2023-03-28 四川省气象灾害防御技术中心 Lightning automatic prediction method based on multi-gradient feature fusion 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
CN116106856A (en) * 2023-04-13 2023-05-12 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) Identification model establishment method and identification method for thunderstorm strong wind and computing equipment
CN116106856B (en) * 2023-04-13 2023-08-18 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) Identification model establishment method and identification method for thunderstorm strong wind and computing equipment
CN117688975A (en) * 2024-02-02 2024-03-12 南京信息工程大学 Meteorological event prediction method and system based on evolution rule mining
CN117688975B (en) * 2024-02-02 2024-05-14 南京信息工程大学 Meteorological event prediction method and system based on evolution rule mining

Similar Documents

Publication Publication Date Title
Alemany et al. Predicting hurricane trajectories using a recurrent neural network
CN115267945A (en) Thunder and lightning early warning method and system based on graph neural network
Lei et al. Mutual information based anomaly detection of monitoring data with attention mechanism and residual learning
Moradi Kordmahalleh et al. A sparse recurrent neural network for trajectory prediction of atlantic hurricanes
CN111310968A (en) LSTM neural network circulation hydrological forecasting method based on mutual information
CN113496104A (en) Rainfall forecast correction method and system based on deep learning
Saxena et al. A review study of weather forecasting using artificial neural network approach
CN105574669B (en) Power transmission and transformation equipment state abnormity detection method based on time-space joint data clustering analysis
CN110738355A (en) urban waterlogging prediction method based on neural network
Biswas et al. Weather prediction by recurrent neural network dynamics
CN112180471B (en) Weather forecasting method, device, equipment and storage medium
Naik et al. Weather classification and forecasting using back propagation feed-forward neural network
CN112200237A (en) Abnormal diagnosis method for time sequence monitoring data of structural health monitoring system
Li et al. Leveraging LSTM for rapid intensifications prediction of tropical cyclones
CN117556197A (en) Typhoon vortex initialization method based on artificial intelligence
CN116415730A (en) Fusion self-attention mechanism time-space deep learning model for predicting water level
CN114882373A (en) Multi-feature fusion sandstorm prediction method based on deep neural network
Kajbaf et al. Temporal downscaling of precipitation from climate model projections using machine learning
CN115204032A (en) ENSO prediction method and device based on multi-channel intelligent model
CN114139590A (en) Method for estimating ocean temperature
Pai et al. Long range forecast on south west monsoon rainfall using artificial neural networks based on clustering approach
Mar et al. Optimum neural network architecture for precipitation prediction of Myanmar
CN117233869A (en) Site short-term wind speed prediction method based on GRU-BiTCN
CN110826810B (en) Regional rainfall prediction method combining spatial reasoning and machine learning
CN116822716A (en) Typhoon prediction method, system, equipment and medium based on space-time attention

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