CN115307780B - Sea surface temperature prediction method, system and application based on time-space information interaction fusion - Google Patents
Sea surface temperature prediction method, system and application based on time-space information interaction fusion Download PDFInfo
- Publication number
- CN115307780B CN115307780B CN202211194484.9A CN202211194484A CN115307780B CN 115307780 B CN115307780 B CN 115307780B CN 202211194484 A CN202211194484 A CN 202211194484A CN 115307780 B CN115307780 B CN 115307780B
- Authority
- CN
- China
- Prior art keywords
- time
- matrix
- space
- spatial
- surface temperature
- 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.)
- Active
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 59
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000003993 interaction Effects 0.000 title claims description 16
- 239000011159 matrix material Substances 0.000 claims abstract description 194
- 230000002452 interceptive effect Effects 0.000 claims abstract description 36
- 230000002776 aggregation Effects 0.000 claims abstract description 18
- 238000004220 aggregation Methods 0.000 claims abstract description 18
- 238000010276 construction Methods 0.000 claims abstract description 13
- 230000000007 visual effect Effects 0.000 claims abstract description 9
- 230000002123 temporal effect Effects 0.000 claims description 25
- 238000006116 polymerization reaction Methods 0.000 claims description 14
- 230000004931 aggregating effect Effects 0.000 claims description 8
- 238000012935 Averaging Methods 0.000 claims description 6
- 238000013461 design Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 3
- 230000000379 polymerizing effect Effects 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 abstract description 10
- 230000014509 gene expression Effects 0.000 abstract description 3
- ATNHDLDRLWWWCB-AENOIHSZSA-M chlorophyll a Chemical compound C1([C@@H](C(=O)OC)C(=O)C2=C3C)=C2N2C3=CC(C(CC)=C3C)=[N+]4C3=CC3=C(C=C)C(C)=C5N3[Mg-2]42[N+]2=C1[C@@H](CCC(=O)OC\C=C(/C)CCC[C@H](C)CCC[C@H](C)CCCC(C)C)[C@H](C)C2=C5 ATNHDLDRLWWWCB-AENOIHSZSA-M 0.000 description 8
- 229930002868 chlorophyll a Natural products 0.000 description 7
- 238000010586 diagram Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000012549 training Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 239000000126 substance Substances 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 229930002875 chlorophyll Natural products 0.000 description 1
- 235000019804 chlorophyll Nutrition 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01K—MEASURING TEMPERATURE; MEASURING QUANTITY OF HEAT; THERMALLY-SENSITIVE ELEMENTS NOT OTHERWISE PROVIDED FOR
- G01K13/00—Thermometers specially adapted for specific purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2474—Sequence data queries, e.g. querying versioned data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/29—Geographical information databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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
-
- 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)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Computational Mathematics (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Remote Sensing (AREA)
- Algebra (AREA)
- Fuzzy Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the technical field of sea surface temperature prediction, and discloses a sea surface temperature prediction method, a sea surface temperature prediction system and application based on time-space information interactive fusion, wherein the sea surface temperature prediction system comprises an input module, a time-space matrix construction module, a time-space characteristic interactive fusion module, a time-space matrix aggregation module, a sea surface temperature data prediction module and an output module, a time-space matrix of each point in space is constructed through the time-space matrix construction module, and a time visual angle and a space visual angle characteristic expression are extracted from the time-space matrix; by constructing a space-time feature interactive fusion module and utilizing a self-attention mechanism and a mutual-attention mechanism in a Transformer, the mutual guidance relations of the time features at different moments, the mutual guidance relations of the space features in different spaces and the mutual guidance relations of the time features and the space features are fully excavated, so that the interactive fusion of the space-time information of the sea surface temperature data is realized, and the sea surface temperature prediction accuracy is improved.
Description
Technical Field
The invention belongs to the technical field of sea surface temperature prediction, and particularly relates to a sea surface temperature prediction method and system based on space-time information interaction fusion and application.
Background
The sea surface temperature is the water temperature near the surface of the ocean, and the time-space change of the sea surface temperature reflects the dynamic evolution rule of the physical and chemical properties of the ocean, so that the time-space distribution prediction of the sea surface temperature data has important significance for guiding the ocean ecological environment. In the most advanced sea surface temperature data prediction method at present, the influences of time sequence characteristics, space structure characteristics and external factors on marine scientific data can be comprehensively considered, time sequence models such as a long-short term memory neural network model and the like are generally adopted in the aspect of time characteristic data processing, and a convolutional neural network model is generally adopted in the aspect of space characteristic data processing. In the space-time fusion strategy, space fusion is carried out firstly and then time fusion is carried out, although the fusion of time sequence characteristics and space characteristics is realized, the space-time distribution rule and the change trend of the sea surface temperature data can be mined to a certain extent. However, this method has the following problems:
firstly, the existing model structure can not realize the interactive fusion of the sea surface temperature data space-time information. The existing sea surface temperature data prediction model structure mostly adopts a method of performing time characteristic polymerization after spatial characteristic polymerization. Because the spatial dimension is fused before the time dimension, the spatial feature can not be guided by the spatial features at other moments when being fused, and therefore the interactive fusion of the space-time information can not be effectively realized.
Secondly, the characteristic of slow change of the physical process exists in the marine physical field, so that the problem of ultra-long term dependence needs to be considered in time sequence prediction, for example, the marine element characteristics before one month can also influence the prediction of marine scientific data at the current moment. The existing method mostly adopts traditional time sequence prediction models such as a long-term and short-term memory neural network model, the previous time sequence information is lost in the time sequence characteristic transmission process, and although the long-term dependence problem is relieved to a certain extent by the structure of a gate mechanism in the model, the existing time sequence prediction method can not be used for the ultra-long-term dependence phenomenon.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a sea surface temperature prediction method, a sea surface temperature prediction system and application based on space-time information interaction fusion, wherein a space-time matrix of each point in a space is constructed through a space-time matrix construction module, and a time visual angle and a characteristic expression of a space visual angle are extracted from the space-time matrix; by constructing a space-time feature interactive fusion module and utilizing a self-attention mechanism and a mutual-attention mechanism in a Transformer, the mutual guidance relationship of time features at different moments, the mutual guidance relationship of space features in different spaces and the mutual guidance relationship of the time features and the space features are fully mined, the interactive fusion of the space-time information of the sea surface temperature data is realized, and the sea surface temperature prediction accuracy is improved.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
the sea surface temperature prediction method based on the time-space information interactive fusion comprises the following steps:
s1, taking N time dimensions as M N V sea surface temperature data matrixPerforming matrix averaging operation on the V dimension and aggregating the matrixes to obtain a matrix S with the dimension of N x M x N; m, N and V are three dimensions;
s2, extracting a space-time matrix of each space point from the matrix SWherein(ii) a The extraction method comprises extracting characteristic value of point i in t days and characteristic value of D-1 space points in neighborhood thereof as the characteristic value of the point Dimension of being(ii) a Then will bePerforming matrix transformation to obtain a space-time grid with i point and dimension of t x DTo obtain a space-time grid of all time instants;
S3, mixingPerforming feature aggregation to obtain a time feature matrix,With dimensions M x N x t x D, M x N representing the presence of M x N spatial points, and for each spatial point there is a two-dimensional time signature matrix with dimensions t x D, wherein each row represents the time signature of the sea surface temperature, i.e. each spatial pointThe time characteristic of the moment has the dimension of 1 x D;
s4, mixingThe last two dimensions are transposed to obtain a space characteristic matrix,The dimension of (a) is M x N x D x t, M x N represents that M x N space points exist, and a two-dimensional space feature matrix with the dimension of D x t exists for each space point, wherein each row represents the space feature of the sea surface temperature, namely the space feature on each space point, and the dimension is 1 x t;
s5, mixingAndconverting the time characteristic matrix into a two-dimensional time characteristic matrix and a two-dimensional space characteristic matrix of M points by N points, wherein the dimensionalities of the obtained time characteristic matrix and the obtained space characteristic matrix are t f and D f respectively;
s6, interactively fusing the time features and the spatial features obtained in the step S5 through four transform-based attention coding blocks, wherein the transform-based attention coding blocks are respectively based on a time self-attention module and a space self-attention moduleA force module, a time-directed spatial attention module, a spatial-directed temporal attention module, an input temporal feature matrixAnd a spatial feature matrixFinally, four feature matrixes M of t x l, D x l, t x l and D x l of the M x N sample points are obtained respectively 1 、M 2 、M 3 、M 4 ;
S7, combining the four feature matrixes M 1 、M 2 、M 3 、M 4 Carrying out polymerization to obtain a polymerization space-time matrix M with dimension t x D;
s8, subjecting the aggregation space-time matrix M to LSTM operation to obtain the prediction result of the last M-N points(ii) a And performing loss calculation through the real data values y of the M x N space points, and performing back propagation operation to update the model parameters.
Further, in step S6, the time-based self-attention module only receives the time feature matrixAnd set the weight,,Respectively withObtained by multiplying,, :
Finally utilizeWeight of relationship between the intermediate features, calculating M 1 Assigning attention to different features:
the resulting matrix M 1 Is a feature matrix of mutual guidance of temporal features.
Further, in step S6, the spatial self-attention module only receives the spatial feature matrixAnd set the weight,,Respectively withObtained by multiplying,, :
Finally utilizeThe weight of the relation between the intermediate features, calculating M 2 Assigning attention to different features:
the resulting matrix M 2 Is a feature matrix of mutual guidance of spatial features.
Further, in step S6, the temporal guidance spatial attention module first receives the temporal feature matrixAnd set the weightAnd is prepared byObtained by multiplyingAs a time guidance matrix:
then inputting the spatial feature matrixAnd design the weights,Respectively withObtained by multiplying,:
Finally utilizeWeight of relationship between the intermediate features, calculating M 3 Namely, the attention of the time characteristic to the space characteristic is distributed:
the resulting matrix M 3 Is a feature matrix of temporal features guiding spatial features.
Further, in step S6, the spatial guidance temporal attention module first receives a spatial feature matrixAnd set the weightAnd is andobtained by multiplyingAs a spatial guide matrix:
then inputting the time characteristic matrixAnd design the weights,Respectively withObtained by multiplying,:
Finally utilizeThe weight of the relation between the intermediate features, calculating M 4 Namely, the attention of the spatial feature to the temporal feature is distributed:
the resulting matrix M 4 Is a feature matrix of spatial feature guide temporal features.
Further, step S7 obtains an aggregation spatio-temporal matrix M according to the following formula,
wherein,combining the time characteristics after the time information self-fusion with the space characteristics after the time-space information fusion to obtain a multi-view space-time characteristic matrix;combining the spatial features after spatial information self-fusion with the temporal features after spatial-temporal information fusion to obtain a spatial-temporal feature matrix with multiple visual angles; and T represents the transposition of the matrix, and finally cross multiplication operation is carried out to obtain a final space-time fusion characteristic matrix.
Further, in step S8, the prediction result is obtained by using M × N pointsAnd the real data values y of the M x N spatial points are subjected to a loss calculation, as shown in equation (4):
and (3) each time all samples in one batch are predicted and Loss is calculated, obtaining a final Loss value Loss through a formula (5), and performing back propagation operation to update model parameters:
wherein m is the batch size.
The invention also provides a sea surface temperature prediction system based on the time-space information interactive fusion, which is used for realizing the sea surface temperature prediction method based on the time-space information interactive fusion, and comprises an input module, a time-space matrix construction module, a time-space characteristic interactive fusion module, a time-space matrix aggregation module, a sea surface temperature data prediction module and an output module,
the input module is used for inputting sea surface temperature data,obtaining N sea surface temperature data matrixes with time dimension M N VPerforming matrix averaging operation on the V dimension and aggregating the matrices to obtain a matrix S with the dimension of N × M × N;
the space-time matrix construction module is used for constructing a space-time matrix and extracting the space-time matrix of each space point from the matrix SThen will bePerforming matrix transformation to obtain a space-time grid with i point and dimension of t x DTo obtain a space-time grid of all time instants;
The space-time feature interactive fusion module is used for carrying out space-time feature interactive fusion and comprises four attention coding blocks based on a transformer, namely a time self-attention module, a time guidance space attention module, a space guidance time attention module and a space self-attention module,
firstly, the first step is toPerforming feature aggregation to obtain a time feature matrixThen will beThe last two dimensions are transposed to obtain a space characteristic matrix(ii) a Will be provided withAndconverting the two-dimensional time characteristic matrix and the two-dimensional space characteristic matrix into M x N points, and finally obtaining four characteristic matrixes M of the M x N sample points through four transform-based attention coding block processing 1 、M 2 、M 3 、M 4 ;
The space-time matrix aggregation module is used for aggregating four feature matrixes M 1 、M 2 、M 3 、M 4 Polymerizing to obtain a polymerization space-time matrix M;
the sea surface temperature data prediction module is used for carrying out LSTM operation on the polymerization space-time matrix M to obtain the prediction result of the last M points by N points;
And the output module is used for outputting the prediction result.
The invention also provides application of the sea surface temperature prediction system based on the time-space information interaction fusion, which is used for predicting the sea surface temperature, inputting the sea surface temperature data in a period of historical time A days, and outputting the sea surface temperature prediction result in a period of future time B days.
Compared with the prior art, the invention has the advantages that:
(1) And realizing the interactive fusion of the time-space information of the sea surface temperature data. The method constructs a space-time global space through a space-time matrix construction module, extracts a time characteristic matrix and a space characteristic matrix from a time visual angle and a space visual angle, interactively fuses the time characteristic matrix and the space characteristic matrix by utilizing a self-attention mechanism and an interactive attention mechanism in a Transformer, fully excavates the mutual guidance relation of the time characteristic and the space characteristic, realizes the interactive fusion of the space-time information of the sea surface temperature data, and improves the sea surface temperature prediction accuracy.
(2) The problem of ultra-long-term dependence of sea surface temperature time sequence prediction is solved. Through the space-time feature interaction fusion module, the time features at all moments can guide and interact any time feature, the distance and the like in the interaction process are guaranteed, the overall effective ultra-long term dependence relationship is effectively extracted, and the problem that the previous time sequence information is lost in the time sequence feature transmission process in the existing method is solved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a schematic flow diagram of the process of the present invention;
FIG. 2 is a block diagram of the temporal self-attention coding block of the present invention;
FIG. 3 is a block diagram of the spatial self-attention coding block of the present invention;
FIG. 4 is a block diagram of a time-guided spatial interactive attention code block of the present invention;
fig. 5 is a block diagram of a space-directed temporal interactive attention coding block of the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific embodiments.
Example 1
The embodiment designs a sea surface temperature prediction method based on space-time information interactive fusion, and the method replaces a convolution submodule in a traditional ConvLSTM model with a space-time matrix construction module and a space-time characteristic interactive fusion module. Firstly, sea surface temperature data are preprocessed, then a space-time matrix construction module is constructed to obtain a time characteristic matrix and a space characteristic matrix, finally a space-time characteristic interactive fusion module is constructed to interactively fuse the space-time information of the time characteristic matrix and the space characteristic matrix to obtain four characteristic matrices and aggregate the four characteristic matrices, the aggregated space-time matrix is subjected to LSTM operation to obtain a prediction result, and finally the purpose is to predict the sea surface temperature in the same space area at the future n +1 moment. With reference to fig. 1-5, the method comprises the following steps:
s1, taking N sea surface temperature data matrixes with time dimensions of M N VPerforming matrix averaging operation on the V dimension and aggregating the matrices to obtain a matrix S with the dimension of N M N; m, N, V are three dimensions.
S2, extracting a space-time matrix of each space point from the matrix SIn which(ii) a The extraction method comprises extracting characteristic value of point i in t days and characteristic value of D-1 space points in neighborhood thereof as the characteristic value of the point Dimension of being(ii) a Then will bePerforming matrix transformation to obtain a space-time grid with the dimensionality of i points as t x DTo obtain a space-time grid of all time instants。
S3, mixingPerforming feature aggregation to obtain a time feature matrix,With dimensions M x N x t x D, M x N representing the presence of M x N spatial points, and for each spatial point there is a two-dimensional time signature matrix with dimensions t x D, wherein each row represents the time signature of the sea surface temperature, i.e. each spatial pointThe time of day is characterized by a dimension of 1 x d.
S4, mixingThe last two dimensions are transposed to obtain a space characteristic matrix,M × N, which represents the presence of M × N spatial points, and for each spatial point there is a two-dimensional spatial feature matrix with a dimension D × t, wherein each row represents a spatial feature of the sea surface temperature, i.e. on each spatial point, with a dimension of 1 × t.
S5, mixingAndand converting the two-dimensional time characteristic matrix and the two-dimensional space characteristic matrix into M-N points, namely taking each space point as an independent sample, and taking the two-dimensional time characteristic matrix and the two-dimensional space characteristic matrix as the characteristic input of the sample. The invention designs two full-connection layers, and obtains a time characteristic matrix sum of M x N sample pointsThe spatial feature matrix dimensions are t f and D f, respectively.
And S6, interactively fusing the time features and the spatial features obtained in the step S5 through four attention coding blocks based on a transformer.
Wherein, the attention coding block based on the transform is based on a time self-attention module, a space self-attention module, a time guidance space attention module and a space guidance time attention module respectively, and finally four feature matrixes M of t x l, D x l, t x l and D x l of M x N sample points are obtained respectively 1 、M 2 、M 3 、M 4 。
As shown in connection with FIG. 2, the time-based self-attention module accepts only the time feature matrixAnd set the weight,,Respectively withObtained by multiplying,, :
Finally utilizeThe weight of the relation between the intermediate features, calculating M 1 Assigning attention to different features:
the resulting matrix M 1 Is a feature matrix of mutual guidance of temporal features.
The spatial self-attention module, shown in connection with FIG. 3, accepts only the spatial feature matrixAnd sets the weight,,Respectively withObtained by multiplying,, :
Finally utilizeThe weight of the relation between the intermediate features, calculating M 2 Assigning attention to different features:
the resulting matrix M 2 Is a feature matrix of mutual guidance of spatial features.
As shown in FIG. 4, the time feature matrix is first acceptedAnd set the weightAnd is andobtained by multiplyingAs a time guidance matrix:
then inputting the spatial feature matrixAnd designing weights,Respectively withObtained by multiplying,:
Finally utilizeThe weight of the relation between the intermediate features, calculating M 3 Namely, the attention degree of the time characteristic to the space characteristic is distributed:
the resulting matrix M 3 Is a feature matrix of temporal feature guide spatial features.
As shown in connection with FIG. 5, the spatial guidance temporal attention module first receives a spatial feature matrixAnd set weightsAnd is andobtained by multiplyingAs a spatial guide matrix:
then inputting the time characteristic matrixAnd design the weights,Respectively withObtained by multiplying,:
Finally utilizeWeight of relationship between the intermediate features, calculating M 4 Namely, the attention of the spatial feature to the temporal feature is distributed:
the resulting matrix M 4 Is a feature matrix of spatial feature guide temporal features.
S7, combining the four feature matrixes M 1 、M 2 、M 3 、M 4 And (5) carrying out polymerization to obtain a polymerization space-time matrix M with dimensions of t x D.
An aggregate spatio-temporal matrix M is obtained according to the following equation,
wherein,combining the time characteristics after the time information self-fusion with the space characteristics after the time-space information fusion to obtain a multi-view space-time characteristic matrix;self-fusing spatial features and time of spatial informationCombining the time characteristics after the empty information fusion to obtain a space-time characteristic matrix with multiple visual angles; and T represents the transposition of the matrix, and finally cross multiplication operation is carried out to obtain a final space-time fusion characteristic matrix.
S8, subjecting the polymerization space-time matrix M of t x D to LSTM operation to obtain the prediction result of the last M x N points(ii) a The loss calculation is performed by the true data values y of M × N spatial points, as shown in equation (4). And (3) each time all the samples in one batch (the batch size is defined as m) are predicted and Loss is calculated, obtaining a final Loss value Loss through a formula (5), and carrying out back propagation operation to update the model parameters.
Example 2
The embodiment provides a sea surface temperature prediction system based on spatio-temporal information interaction fusion, which is used for realizing the sea surface temperature prediction method based on spatio-temporal information interaction fusion as described in embodiment 1.
The input module is used for inputting sea surface temperature data and acquiring N sea surface temperature data matrixes with time dimensions of M x N x VAnd carrying out matrix averaging operation on the V dimension and aggregating the matrices to obtain a matrix S with the dimension of N M N.
The space-time matrix construction module is used for constructing a space-time matrix and extracting the space-time matrix of each space point from the matrix SThen will bePerforming matrix transformation to obtain a space-time grid with i point and dimension of t x DTo obtain a space-time grid of all time instants。
The space-time feature interactive fusion module is used for carrying out space-time feature interactive fusion and comprises four attention coding blocks based on a transformer, namely a time self-attention module, a time guidance space attention module, a space guidance time attention module and a space self-attention module,
firstly, the first step is toPerforming feature aggregation to obtain a time feature matrixThen will beThe last two dimensions are transposed to obtain a space characteristic matrix(ii) a Will be provided withAndconverting the two-dimensional time characteristic matrix and the two-dimensional space characteristic matrix into M x N points, and finally obtaining four characteristic matrixes M of M x N sample points through four transform-based attention coding block processing 1 、M 2 、M 3 、M 4 ;
The space-time matrix aggregation moduleFor combining four feature matrices M 1 、M 2 、M 3 、M 4 And polymerizing to obtain a polymerization space-time matrix M.
The sea surface temperature data prediction module is used for carrying out LSTM operation on the polymerization space-time matrix M to obtain the prediction result of the last M points by N points。
And the output module is used for outputting the prediction result.
The data processing process of each module can refer to the record in embodiment 1, and is not described herein again.
Example 3
The embodiment provides an application of a sea surface temperature prediction system based on space-time information interactive fusion, which is used for predicting sea surface temperature, inputting sea surface temperature data in a period of historical time A days, and outputting a sea surface temperature prediction result in a period of future time B days.
A specific method of predicting the sea surface temperature is described below by way of example.
Step one, establishing a training data set:
(1) Sea surface temperature data of 42 years from 1981 to 2022 and the like in the world are acquired, sea surface temperature data of a prediction region are extracted from two dimensions of time and space, and sea surface temperature data sets on different space-time dimensions in the region are constructed.
(2) The constructed sea surface temperature space-time data set is divided into a training sample, a verification sample and a test sample according to the proportion of 70%, 20% and 10%, and the data are processed according to a method for predicting B days every A days.
Step two, establishing a sea surface temperature prediction model:
(1) And constructing a deep learning framework by adopting a Pythroch.
(2) A Transformer mechanism-based spatio-temporal information interactive fusion prediction model is constructed, and comprises a spatio-temporal matrix construction module, a spatio-temporal feature interactive fusion module (four Transformer-based attention coding blocks), a spatio-temporal matrix aggregation module and a sea surface temperature data prediction module. The model inputs the sea surface temperature data of the previous A days and outputs the sea surface temperature prediction data of the future B days.
(3) Inputting the training sample obtained in the first step into a model for training, detecting the model prediction effect after each training by using a verification sample, and evaluating the model by using a test sample after the model is trained for 100 times.
Step three, model application:
(1) And setting B-day parameters to be predicted, and inputting the sea surface temperature data of the previous A-day into the trained model to obtain a final B-day sea surface temperature prediction result.
In summary, the invention (1) constructs a spatio-temporal matrix of each point in the space in a spatio-temporal matrix construction module, and extracts feature expressions of a temporal view and a spatial view from the spatio-temporal matrix. (2) The time-space feature interactive fusion module utilizes a self-attention mechanism and a mutual attention mechanism in a transformer to construct a time self-attention coding block which fully excavates the mutual guidance relationship of time features at different moments, a space self-attention coding block which excavates the mutual guidance relationship of space features in different spaces, and a time guidance space interactive attention coding block and a space guidance time interactive attention coding block which fully excavate the mutual guidance relationship of the time features and the space features, and the time-space data ultra-long term dependency relationship is constructed through the four coding blocks, so that the interactive fusion of the time-space information of the sea surface temperature data is realized.
In addition, it should be noted that, in practical application, the method and the system of the present invention can also be applied to the space-time variation of other marine scientific data (for example, time-time data fields such as chlorophyll concentration, wind speed, wind direction, etc.) to obtain the space-time distribution prediction, and reflect the dynamic evolution rule of the physical and chemical properties of the marine physics. For example, the method and the system of the invention are applied to process, obtain the prediction of the concentration of the offshore chlorophyll a, input the offshore chlorophyll a data in a period of historical time period A (the offshore chlorophyll a data with N time dimensions M N V is obtained through the original chlorophyll a remote sensing data, the meteorological remote sensing data and the buoy chlorophyll a monitoring data), and output the chlorophyll a prediction result in a period of time period B in the future (a whole offshore area chlorophyll a prediction graph which is continuous in a long time sequence).
It will be understood that the foregoing description is not intended to limit the invention, and that the invention is not limited to the examples described above, and that various changes, modifications, additions and substitutions which may be made by one of ordinary skill in the art without departing from the spirit of the invention are therefore intended to be included within the scope of the invention.
Claims (9)
1. The sea surface temperature prediction method based on the time-space information interactive fusion is characterized by comprising the following steps of:
s1, taking t time dimensions as M N V sea surface temperature data matrixPerforming matrix averaging operation on the V dimension and aggregating the matrices to obtain a matrix S with the dimension of t M N; m, N and V are three dimensions;
s2, extracting a space-time matrix of each space point from the matrix SWherein(ii) a The extraction method comprises extracting characteristic value of point i in t days and characteristic value of D-1 space points in neighborhood thereof as the characteristic value of the point Dimension of(ii) a Then will bePerforming matrix transformation to obtain a space-time grid with i point and dimension of t x DTo obtain a space-time grid of all time instants;
S3, mixingPerforming feature aggregation to obtain a time feature matrix,With dimensions M x N x t x D, M x N representing the presence of M x N spatial points, and for each spatial point there is a two-dimensional time signature matrix with dimensions t x D, wherein each row represents the time signature of the sea surface temperature, i.e. each spatial pointTime characteristic of time, dimension is 1 × D;
s4, mixingThe last two dimensions are transposed to obtain a space characteristic matrix,With dimensions M x N x D x t, M x N representing the presence of M x N spatial points, and for each spatial point there is a two-dimensional spatial feature matrix with dimensions D x t, wherein each row represents a spatial feature of the sea surface temperature, i.e. each spaceSpatial features on the points, with dimensions 1 × t;
s5, mixingAndconverting the time characteristic matrix into a two-dimensional time characteristic matrix and a two-dimensional space characteristic matrix of M points by N points, wherein the dimensionalities of the obtained time characteristic matrix and the obtained space characteristic matrix are t f and D f respectively;
s6, interactively fusing the time characteristics and the spatial characteristics obtained in the step S5 through four attention coding blocks based on a transformer, wherein the attention coding blocks based on the transformer are respectively based on a time self-attention module, a spatial self-attention module, a time guidance spatial attention module and a spatial guidance temporal attention module, and a time characteristic matrix is inputAnd a spatial feature matrixFinally, four feature matrixes M of t x l, D x l, t x l and D x l of the M x N sample points are obtained respectively 1 、M 2 、M 3 、M 4 ;
S7, combining the four feature matrixes M 1 、M 2 、M 3 、M 4 Carrying out polymerization to obtain a polymerization space-time matrix M with dimension t x D;
2. Sea surface temperature based on spatio-temporal information interaction fusion according to claim 1The degree prediction method is characterized in that, in step S6, the time-based self-attention module only receives the time feature matrixAnd set the weight,,Respectively withObtained by multiplying,,:
Finally utilizeWeight of relationship between the intermediate features, calculating M 1 Assigning attention to different features:
the resulting matrix M 1 Is a feature matrix of mutual guidance of temporal features.
3. The method for sea table temperature prediction based on spatio-temporal information interaction fusion of claim 1, wherein in step S6, the spatial self-attention module only accepts spatial feature matrixAnd set the weight,,Respectively withObtained by multiplying,,:
Finally utilizeWeight of relationship between the intermediate features, calculating M 2 Assigning attention to different features:
the resulting matrix M 2 Is a feature matrix of mutual guidance of spatial features.
4. The method for sea table temperature prediction based on spatio-temporal information interaction fusion of claim 1, wherein in step S6, the time-guided spatial attention module firstly receives the time feature matrixAnd set the weightAnd is andobtained by multiplyingAs a time guidance matrix:
then inputting the spatial feature matrixAnd designing weights,Respectively withObtained by multiplying,:
Finally utilizeThe weight of the relation between the intermediate features, calculating M 3 Namely, the attention degree of the time characteristic to the space characteristic is distributed:
the resulting matrix M 3 Is a feature matrix of temporal feature guide spatial features.
5. The method for sea table temperature prediction based on spatio-temporal information interaction fusion of claim 1, wherein in step S6, the spatial guidance time attention module firstly receives a spatial feature matrixAnd set the weightAnd is prepared byObtained by multiplyingAs a spatial guide matrix:
then inputting the time characteristic matrixAnd design the weights,Respectively withObtained by multiplying,:
Finally utilizeThe weight of the relation between the intermediate features, calculating M 4 Namely, the attention of the spatial feature to the temporal feature is distributed:
the resulting matrix M 4 Is a feature matrix of spatial feature guide temporal features.
6. The sea table temperature prediction method based on spatio-temporal information interaction fusion of claim 1, characterized in that step S7 obtains an aggregation spatio-temporal matrix M according to the following formula,
wherein,combining the time characteristics after the time information self-fusion with the space characteristics after the time-space information fusion to obtain a multi-view space-time characteristic matrix;combining the spatial features of the spatial information after self-fusion with the temporal features of the spatial information after fusion to obtain a spatial-temporal feature matrix with multiple visual angles; and T represents the transposition of the matrix, and finally cross multiplication operation is carried out to obtain a final space-time fusion characteristic matrix.
7. The method for sea table temperature prediction based on spatio-temporal information interaction fusion of claim 1, wherein in step S8, the prediction results of M x N points are obtainedAnd performing loss calculation on the real data values y of the M x N space points, as shown in formula (4):
and (3) each time all samples in one batch are predicted and Loss is calculated, obtaining a final Loss value Loss through a formula (5), and performing back propagation operation to update model parameters:
wherein m is the batch size.
8. The sea surface temperature prediction system based on the time-space information interactive fusion is characterized by comprising an input module, a time-space matrix construction module, a time-space characteristic interactive fusion module, a time-space matrix aggregation module, a sea surface temperature data prediction module and an output module,
the input module is used for inputting sea surface temperature data and acquiring N sea surface temperature data matrixes with time dimensions of M x N x VPerforming matrix averaging operation on the V dimension and aggregating the matrices to obtain a matrix S with the dimension of N × M × N;
the space-time matrix construction module is used for constructing a space-time matrix and extracting the space-time matrix of each space point from the matrix SThen will bePerforming matrix transformation to obtain a space-time grid with the dimensionality of i points as t x DTo obtain a space-time grid at all times;
The space-time feature interactive fusion module is used for carrying out space-time feature interactive fusion and comprises four attention coding blocks based on a transformer, namely a time self-attention module, a time guidance space attention module, a space guidance time attention module and a space self-attention module,
firstly, the first step is toPerforming feature aggregation to obtain a time feature matrixThen will beThe last two dimensions are transposed to obtain a space characteristic matrix(ii) a Will be provided withAndconverting the two-dimensional time characteristic matrix and the two-dimensional space characteristic matrix into M x N points, and finally obtaining four characteristic matrixes M of the M x N sample points through four transform-based attention coding block processing 1 、M 2 、M 3 、M 4 ;
The space-time matrix aggregation module is used for aggregating four feature matrixes M 1 、M 2 、M 3 、M 4 Polymerizing to obtain a polymerization space-time matrix M;
the sea surface temperature data prediction module is used for subjecting the aggregation space-time matrix M to LSTM operation to obtain the prediction result of the last M points by N points;
And the output module is used for outputting the prediction result.
9. The use of the spatiotemporal information interaction fusion-based sea surface temperature prediction system of claim 8, for predicting sea surface temperature, inputting sea surface temperature data over a historical period of time a days, and outputting sea surface temperature prediction results over a future period of time B days.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211194484.9A CN115307780B (en) | 2022-09-29 | 2022-09-29 | Sea surface temperature prediction method, system and application based on time-space information interaction fusion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211194484.9A CN115307780B (en) | 2022-09-29 | 2022-09-29 | Sea surface temperature prediction method, system and application based on time-space information interaction fusion |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115307780A CN115307780A (en) | 2022-11-08 |
CN115307780B true CN115307780B (en) | 2023-01-06 |
Family
ID=83866567
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211194484.9A Active CN115307780B (en) | 2022-09-29 | 2022-09-29 | Sea surface temperature prediction method, system and application based on time-space information interaction fusion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115307780B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115984281B (en) * | 2023-03-21 | 2023-06-20 | 中国海洋大学 | Multi-task complement method of time sequence sea temperature image based on local specificity deepening |
CN116246213B (en) * | 2023-05-08 | 2023-07-28 | 腾讯科技(深圳)有限公司 | Data processing method, device, equipment and medium |
CN116822382B (en) * | 2023-08-30 | 2023-11-21 | 中国海洋大学 | Sea surface temperature prediction method and network based on space-time multiple characteristic diagram convolution |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019216449A1 (en) * | 2018-05-09 | 2019-11-14 | 주식회사 알고리고 | Method and apparatus for time series artificial neural network electric vehicle power demand prediction, using spatio-temporal fusion of power demand data and heterogeneous data |
CN113326981A (en) * | 2021-05-26 | 2021-08-31 | 北京交通大学 | Atmospheric environment pollutant prediction model based on dynamic space-time attention mechanism |
CN113705880A (en) * | 2021-08-25 | 2021-11-26 | 杭州远眺科技有限公司 | Traffic speed prediction method and device based on space-time attention diagram convolutional network |
US11222217B1 (en) * | 2020-08-14 | 2022-01-11 | Tsinghua University | Detection method using fusion network based on attention mechanism, and terminal device |
CN114399073A (en) * | 2021-11-26 | 2022-04-26 | 中国石油大学(华东) | Ocean surface temperature field prediction method based on deep learning |
CN114492992A (en) * | 2022-01-25 | 2022-05-13 | 重庆邮电大学 | Self-adaptive space-time graph neural network traffic flow prediction method and system based on Transformer |
CN114936691A (en) * | 2022-05-06 | 2022-08-23 | 河北工业大学 | Temperature forecasting method integrating relevance weighting and space-time attention |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160297B (en) * | 2019-12-31 | 2022-05-13 | 武汉大学 | Pedestrian re-identification method and device based on residual attention mechanism space-time combined model |
CN111310987B (en) * | 2020-01-23 | 2023-04-07 | 北京百度网讯科技有限公司 | Method and device for predicting free parking space of parking lot, electronic equipment and storage medium |
CN112541302B (en) * | 2020-12-23 | 2024-02-06 | 北京百度网讯科技有限公司 | Air quality prediction model training method, air quality prediction method and device |
-
2022
- 2022-09-29 CN CN202211194484.9A patent/CN115307780B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019216449A1 (en) * | 2018-05-09 | 2019-11-14 | 주식회사 알고리고 | Method and apparatus for time series artificial neural network electric vehicle power demand prediction, using spatio-temporal fusion of power demand data and heterogeneous data |
US11222217B1 (en) * | 2020-08-14 | 2022-01-11 | Tsinghua University | Detection method using fusion network based on attention mechanism, and terminal device |
CN113326981A (en) * | 2021-05-26 | 2021-08-31 | 北京交通大学 | Atmospheric environment pollutant prediction model based on dynamic space-time attention mechanism |
CN113705880A (en) * | 2021-08-25 | 2021-11-26 | 杭州远眺科技有限公司 | Traffic speed prediction method and device based on space-time attention diagram convolutional network |
CN114399073A (en) * | 2021-11-26 | 2022-04-26 | 中国石油大学(华东) | Ocean surface temperature field prediction method based on deep learning |
CN114492992A (en) * | 2022-01-25 | 2022-05-13 | 重庆邮电大学 | Self-adaptive space-time graph neural network traffic flow prediction method and system based on Transformer |
CN114936691A (en) * | 2022-05-06 | 2022-08-23 | 河北工业大学 | Temperature forecasting method integrating relevance weighting and space-time attention |
Non-Patent Citations (5)
Title |
---|
Prediction of gas leakage and dispersion in utility tunnels based on CFD-EnKF coupling model: A 3D full-scale application;JitaoCai 等;《Sustainable Cities and Society Volume 80, May 2022, 103789》;20220531;第80卷;全文 * |
Prediction of Sea Surface Temperature by Combining Interdimensional and Self-Attention with Neural Networks;Xing Guo等;《remote sensing》;20220922;第4737卷(第14期);全文 * |
Study of LSTM Model in Sea Surface Temperature Prediction of the Yellow Sea Cold Water Mass Area;Zikun Chen等;《2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI)》;20200409;全文 * |
基于CESM预测***对全球关键海区海温主要模态后报能力评估;张守文 等;《海洋学报》;20180930;第40卷(第9期);全文 * |
基于时空演变多重特性建模的近海叶绿素浓度时序预测;王成贺 等;《信号处理》;20220630;第38卷(第6期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN115307780A (en) | 2022-11-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN115307780B (en) | Sea surface temperature prediction method, system and application based on time-space information interaction fusion | |
CN111784041B (en) | Wind power prediction method and system based on graph convolution neural network | |
Shiri et al. | Estimation of daily suspended sediment load by using wavelet conjunction models | |
CN112132149B (en) | Semantic segmentation method and device for remote sensing image | |
CN110119854A (en) | Voltage-stablizer water level prediction method based on cost-sensitive LSTM Recognition with Recurrent Neural Network | |
CN108960063A (en) | It is a kind of towards event relation coding video in multiple affair natural language description algorithm | |
CN113627093B (en) | Underwater mechanism trans-scale flow field characteristic prediction method based on improved Unet network | |
CN115222163A (en) | Multi-factor medium-long term real-time forecasting method and system for harbor basin inlet waves and application | |
CN116596150A (en) | Event prediction method of transform Hoxwell process model based on multi-branch self-attention | |
Shao et al. | Ocean reanalysis data‐driven deep learning forecast for sea surface multivariate in the South China Sea | |
Bento et al. | Ocean wave power forecasting using convolutional neural networks | |
CN115604131B (en) | Link flow prediction method, system, electronic device and medium | |
CN110648030A (en) | Method and device for predicting seawater temperature | |
CN113377907B (en) | End-to-end task type dialogue system based on memory mask self-attention network | |
Bosma et al. | Estimating solar and wind power production using computer vision deep learning techniques on weather maps | |
CN115099318A (en) | Training method and application method of wind speed space-time prediction network and electronic equipment | |
CN114661874A (en) | Visual question-answering method based on multi-angle semantic understanding and self-adaptive dual channels | |
CN114202091A (en) | Indian ocean dipole index prediction method | |
CN111626472A (en) | Scene trend judgment index computing system and method based on deep hybrid cloud model | |
CN117173401B (en) | Semi-supervised medical image segmentation method and system based on cross guidance and feature level consistency dual regularization | |
Li et al. | Hierarchical Transformer With Lightweight Attention for Radar-Based Precipitation Nowcasting | |
CN117272002B (en) | Solar radiation amount estimation method and device, electronic equipment and storage medium | |
CN116610770B (en) | Judicial field case pushing method based on big data | |
CN116090509A (en) | Multivariate spatiotemporal data generation method based on separation attention mechanism | |
Jiang et al. | Heterogeneous Dynamic Graph Convolutional Networks: A Deep Learning Framework for Spatiotemporal Flood Forecasting |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |