CN115204437A - Sea surface temperature forecasting method based on memory map convolution network - Google Patents

Sea surface temperature forecasting method based on memory map convolution network Download PDF

Info

Publication number
CN115204437A
CN115204437A CN202110399839.7A CN202110399839A CN115204437A CN 115204437 A CN115204437 A CN 115204437A CN 202110399839 A CN202110399839 A CN 202110399839A CN 115204437 A CN115204437 A CN 115204437A
Authority
CN
China
Prior art keywords
surface temperature
sea surface
convolution
memory map
memory
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
CN202110399839.7A
Other languages
Chinese (zh)
Inventor
任鹏
张孝宇
赵鹏
赵一丁
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.)
North China Sea Marine Forecasting Center Of State Oceanic Administration
China University of Petroleum East China
Original Assignee
North China Sea Marine Forecasting Center Of State Oceanic Administration
China University of Petroleum East China
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 North China Sea Marine Forecasting Center Of State Oceanic Administration, China University of Petroleum East China filed Critical North China Sea Marine Forecasting Center Of State Oceanic Administration
Priority to CN202110399839.7A priority Critical patent/CN115204437A/en
Publication of CN115204437A publication Critical patent/CN115204437A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Abstract

The invention provides an intelligent sea surface temperature forecasting method, which is characterized in that a deep learning network based on memory map convolution is established through a sea surface temperature training data set constructed by satellite data. The memory map convolution deep learning network consists of a memory layer, a map convolution layer and an output layer. The network trains and modifies the model parameters based on the training data set. And the trained memory map convolution deep learning network is used for forecasting the sea surface temperature. The method fully considers the time and space information of the sea surface temperature, has low dependence on ocean dynamics factors and low dependence on manpower, and can improve the accuracy of the sea surface temperature forecast result in a complex and changeable ocean environment.

Description

Sea surface temperature forecasting method based on memory map convolutional network
Technical Field
The invention relates to the technical field of ocean forecast, in particular to a sea surface temperature forecasting method based on a memory map convolution network.
Background
Sea surface temperature is an important variable for studying global climate and weather changes. The rising sea temperature is an important factor for drought or strong precipitation in the northern hemisphere area. In addition, sea temperature also has an effect on the development of tropical cyclones. Therefore, the forecast of sea surface temperature plays a profound role in weather forecast, environmental protection, fishery production, marine military activities and the like.
Sea surface temperature changes are caused by various factors such as wind, tide, short wave radiation and the like. The sea temperature near the coast varies more complexly due to the complex interaction of terrain, sea and atmosphere. Accurate prediction of offshore surface temperature is an important topic of ocean research. The traditional sea surface temperature prediction method is mainly a numerical prediction method. Numerical methods use the ocean to describe dynamics and thermodynamics to build a predictive model. Numerical mode methods require complex kinetic and thermodynamic equations that tend to underestimate the features associated with the observed sea temperature variability.
Disclosure of Invention
The invention aims to provide a sea surface temperature forecasting method based on a memory map convolution network so as to obtain a more accurate sea surface temperature forecasting result.
In order to achieve the above object, some embodiments of the present invention provide the following technical solutions:
a sea surface temperature forecasting method based on a memory map convolution network comprises the following steps:
s1: constructing a sea surface temperature data set: extracting sea surface temperature data of the region to be predicted with the same position size of M x N from each time data by using the global sea surface temperature field data at T + tau times to construct data, and extracting the sea surface temperature of the region to be predicted from the 1 st to the T th times as historical data
Figure BDA0003020006020000014
Extracting at T +1 th to T + tau th timeSea surface temperature of area to be predicted as future data
Figure BDA0003020006020000013
Constructing a data set;
s2: and constructing a weighted adjacency matrix, and representing the position relation of the sea temperature data by using the weighted adjacency matrix based on the sea surface temperature data set data. For sea surface temperature sequences defined on a regular grid of longitude and latitude coordinates
Figure BDA0003020006020000011
Figure BDA0003020006020000012
Each time t is composed of M x N effective sea surface temperature grid points, and is calculated
Figure BDA0003020006020000015
Each term W in the weighted adjacency matrix W of ij Comprises the following steps:
Figure BDA0003020006020000021
wherein, w mn Representing a weighted adjacency between sea surface temperature points m and n, d mn Representing the distance, w, between the recorded sea surface temperature points m and n on the grid mn Is represented by the distance d between grid points mn And calculating the obtained weight. r and d min Respectively representing the scale parameter and the minimum threshold distance parameter of the formula.
S3: and forecasting future sea surface temperature based on the memory map convolutional network.
In some embodiments of the present invention, a memory map convolution network includes a memory layer, a map convolution layer, and an output layer; the method for calculating the memory map convolutional network output comprises the following steps:
the memory layer is composed of a time convolution unit and a gate control circulation unit. The time convolution unit is:
Figure BDA0003020006020000022
wherein the content of the first and second substances,
Figure BDA0003020006020000023
represents the input of the memory convolution unit,
Figure BDA0003020006020000024
the output of the time convolution unit is represented,
Figure BDA0003020006020000025
and B mem Representing the convolution kernel and the offset, respectively. The height and width of the convolution kernel are set to 1 and K, respectively m Time convolution unit output
Figure BDA0003020006020000026
Divided into two identical channels
Figure BDA0003020006020000027
And
Figure BDA0003020006020000028
and then as an input to the gated loop unit. The gating cycle unit is:
Figure BDA0003020006020000029
wherein the content of the first and second substances,
Figure BDA00030200060200000210
by
Figure BDA00030200060200000211
The obtained product is intercepted and obtained,
Figure BDA00030200060200000212
where e denotes a natural constant and O denotes a Hamdamard product. Obtaining the output of the memory convolution layer after the sea surface temperature sequence passes through the memory convolution layer
Figure BDA00030200060200000213
S4 memory convolution layer output
Figure BDA00030200060200000214
And S2, the constructed weighted adjacency matrix is used for calculating the output of the graph convolution layer. The method of computing graph convolution layer output includes:
for the position information of the sea surface temperature, it is encoded using a graph laplacian matrix, L, which is expressed as:
Figure BDA00030200060200000215
in which I N Denotes an identity matrix, D ii ＝∑ j w ij The graph convolution is approximated by a chebyshev polynomial:
Figure BDA00030200060200000216
wherein the content of the first and second substances,
Figure BDA00030200060200000217
representing the output of the graph convolution, T k () Expressing the Chebyshev polynomial of order K, theta k Representing a polynomial coefficient matrix, λ max Represents the maximum eigenvalue of L;
output of graph volume layer
Figure BDA00030200060200000220
Obtaining the output of the graph volume layer through residual connection
Figure BDA00030200060200000218
Figure BDA00030200060200000219
Wherein ReLU represents a nonlinear activation function;
s5, outputting based on graph volume layer
Figure BDA0003020006020000031
Calculating the output of the second memory layer, wherein the step of calculating the output of the second memory layer comprises the following steps:
the second memory layer consists of a time convolution unit and a gated cyclic unit. The time convolution unit is:
Figure BDA0003020006020000032
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003020006020000033
represents the output of the time convolution unit and,
Figure BDA0003020006020000034
and B mem2 Representing the convolution kernel and the offset, respectively. The height and width of the convolution kernel are set to 1 and K, respectively m2 Inputting data
Figure BDA0003020006020000035
After passing through a time convolution unit, output
Figure BDA0003020006020000036
Divided into two identical channels
Figure BDA0003020006020000037
And
Figure BDA0003020006020000038
and
Figure BDA0003020006020000039
input gated cyclic unit to obtain output
Figure BDA00030200060200000310
Figure BDA00030200060200000311
Wherein, the first and the second end of the pipe are connected with each other,
Figure BDA00030200060200000312
by
Figure BDA00030200060200000313
And (6) intercepting to obtain the product.
S6 outputting through the second memory layer
Figure BDA00030200060200000314
The steps of calculating the output of the output layer are as follows:
the output layer is composed of a convolution layer and a full-connection layer, and the convolution layer is as follows:
Figure BDA00030200060200000315
wherein the content of the first and second substances,
Figure BDA00030200060200000316
the output of the convolutional layer is shown,
Figure BDA00030200060200000317
and B o Respectively representing the convolution kernel and the offset. The height and width of the convolution kernel are respectively set as T and 1, and the output of the full connection layer is obtained based on the output calculation of the convolution layer
Figure BDA00030200060200000318
Figure BDA00030200060200000319
Wherein, a o And b o Representing the weight and bias of the fully connected layer.
And finishing the memory map convolution network model construction process.
S7, training the memory map convolutional network, updating parameters in the memory map convolutional network in the training process, and defining a loss function L of the model as follows:
Figure BDA00030200060200000320
wherein L is reg Represents L 2 A regularization term. And training the memory map convolution network by adopting a loss function.
S8, sea surface temperature prediction is carried out by adopting a memory map convolution network:
extracting sea surface temperature data of T moments before the p moments;
generating a sea surface temperature forecast value at the T +1 moment by adopting a memory map convolution network model, inputting the generated sea surface temperature forecast value at the T +1 moment into the memory map convolution network model as input data, forming the input data with the sea surface temperature data at T-1 moments before the p moment to generate a sea surface temperature forecast value at the T +2 moment, repeating the process, knowing the last moment T + tau of the forecast, and finishing the sea surface temperature forecast.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that:
according to the sea surface temperature method, the sea surface temperature result is forecasted by obtaining the memory map convolution network model through deep learning by utilizing regional sea surface temperature data. The method has low artificial participation degree and low dependence degree on ocean thermodynamic and kinetic influence factors, and can improve the accuracy of the sea surface temperature forecast result in a complex and changeable ocean environment. Compared with the traditional method for manually correcting the forecast result, the method saves manpower, can improve the working efficiency, and has important significance in the aspects of service weather forecast, environmental protection, fishery production, offshore military activities, ocean coastal production and life and the like.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the embodiments or the prior art descriptions will be briefly described 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 inventive exercise.
FIG. 1 is a flow chart of sea surface temperature prediction provided by the present invention;
FIG. 2 is a 1 hour correction of the sea surface temperature using the method of the present invention;
FIG. 3 is a 3 hour correction of sea surface temperature using the method of the present invention;
FIG. 4 is a 7 hour correction of the sea surface temperature using the method of the present invention;
FIG. 5 is a 1 hour corrected result curve for single point sea surface temperature prediction using the method of the present invention;
FIG. 6 is a 3 hour corrected result curve for single point sea surface temperature prediction using the method of the present invention;
FIG. 7 is a 7 hour correction result curve for single point sea surface temperature prediction using the method of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantageous effects to be solved by the present invention more clearly apparent, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The sea surface temperature forecasting method comprises the following steps:
s1: a sea surface temperature data set is constructed.
The construction of the data set relies on global sea surface temperature data from satellite observations. Collecting global sea surface temperature data, and taking I sea surface temperature measured value sequences S i,1:T+τ . Specifically, extracting M x N regions to be predicted from each time data, and selecting S, wherein the I time sequences of observation data are formed by T + tau times, and each time is formed by M x N effective sea surface temperature grid points i,T+τ The memory map convolution network is used for training the memory map convolution network; length and sea of time seriesThe number of the surface temperature grid points can be selected according to requirements and is not limited.
Wherein the sea surface measured value sequence is expressed as: s i,1:T ＝[S i,1 ,S i,2 ,…S i,T+τ ](ii) a Wherein S is i,t Expressed as the sea surface temperature actual measured value at the t-th time in the ith sea surface temperature actual measured value sequence, t ∈ (1, t + τ); i ∈ (1, I).
S2: a weighted adjacency matrix of sea surface temperatures is calculated.
Defining S based on the set sea surface area location i,1:T+τ Each term W in the weighted adjacency matrix W of ij Comprises the following steps:
Figure BDA0003020006020000051
wherein, w mn Representing a weighted adjacency between sea surface temperature points m and n, d mn Representing the distance, w, between the recorded sea surface temperature points m and n on the grid mn Representing the distance d between grid points mn And calculating the obtained weight. r and d min Respectively representing the scale parameter and the minimum threshold distance parameter of the formula.
S3: and constructing a memory map convolution network. The sea surface temperature forecasting model based on memory map convolution is composed of two memory layers, a map convolution layer and an output layer.
S31: the memory layer is composed of a time convolution unit and a gating circulation unit.
Define the input of the first memory layer corresponding to the ith sequence as
Figure BDA0003020006020000052
The output of the first memory layer corresponding to the ith sequence is
Figure BDA0003020006020000053
Specifically, the time convolution unit is:
Figure BDA0003020006020000054
wherein the content of the first and second substances,
Figure BDA0003020006020000055
the output of the time convolution unit is represented,
Figure BDA0003020006020000056
and B mem Representing the convolution kernel and the offset, respectively. The height and width of the convolution kernel are set to 1 and K, respectively mem After input data passes through the time convolution unit, the input data is output
Figure BDA00030200060200000517
Divided into two identical channels
Figure BDA0003020006020000057
And
Figure BDA0003020006020000058
and
Figure BDA0003020006020000059
input gated loop unit to obtain output
Figure BDA00030200060200000510
Figure BDA00030200060200000511
Wherein the content of the first and second substances,
Figure BDA00030200060200000512
by
Figure BDA00030200060200000513
The obtained product is intercepted and obtained,
Figure BDA00030200060200000514
where e denotes a natural constant and O denotes a Hamdamard product.
S32, based on the firstOutput of memory layer
Figure BDA00030200060200000515
And S2, calculating the output of the graph convolution layer by the weighted adjacent matrix W, wherein the step of outputting the graph convolution layer is as follows:
for the position information of the sea surface temperature, using graph Laplace matrix coding, defining a graph Laplace matrix L as:
Figure BDA00030200060200000516
in which I N Representing an identity matrix, D ii ＝∑ j w ij The graph convolution is approximated by a chebyshev polynomial:
Figure BDA0003020006020000061
wherein the content of the first and second substances,
Figure BDA0003020006020000062
representing the output of the graph convolution, T k () Expressing the Chebyshev polynomial of order K, theta k Representing a polynomial coefficient matrix, λ max Represents the maximum eigenvalue of L;
output of graph convolution
Figure BDA0003020006020000063
Obtaining the output of the convolution layer from residual connection
Figure BDA0003020006020000064
Figure BDA0003020006020000065
Wherein ReLU represents a non-linear activation function; output of graph convolution unit
Figure BDA0003020006020000066
S33 outputting based on graph convolution layer
Figure BDA0003020006020000067
Calculating the output of the second memory layer
Figure BDA0003020006020000068
The step of calculating the output of the second memory layer is as follows:
S331:
Figure BDA0003020006020000069
input time convolution layer:
Figure BDA00030200060200000610
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA00030200060200000611
represents the output of the time convolution unit and,
Figure BDA00030200060200000612
and B mem2 Respectively representing the convolution kernel and the offset. The height and width of the convolution kernel are set to 1 and K, respectively m2 ，
S332, inputting data
Figure BDA00030200060200000613
After passing through a time convolution unit, output
Figure BDA00030200060200000614
Divided into two identical channels
Figure BDA00030200060200000615
And
Figure BDA00030200060200000616
and
Figure BDA00030200060200000617
input gated loop unit to obtain output
Figure BDA00030200060200000618
Figure BDA00030200060200000619
Wherein the content of the first and second substances,
Figure BDA00030200060200000620
by
Figure BDA00030200060200000621
And (6) intercepting to obtain the product.
S34 outputting based on the second memory layer
Figure BDA00030200060200000622
Computing output of an output layer
Figure BDA00030200060200000623
The output of the whole memory map convolution network is also realized, and the steps are as follows:
the output layer is composed of a convolution layer and a full-connection layer, and the convolution layer is as follows:
Figure BDA00030200060200000624
wherein
Figure BDA00030200060200000625
And B o Respectively representing the convolution kernel and the offset. The height and width of the convolution kernel are set to T and 1 respectively,
calculating output based on convolution layer to obtain output of full connection layer
Figure BDA00030200060200000626
Figure BDA00030200060200000627
Wherein, a o And b o Representing the weight and bias of the fully connected layer. So far, the whole memory map convolution network construction process is finished.
S4: and predicting the sea surface temperature by adopting a memory map convolution network. During the prediction process, sea surface temperature observation data is adopted. Furthermore, in order to obtain better correction data, as a preferable scheme, the memory map convolution network can be trained first, and the trained memory map convolution network is used for prediction.
S41: and optimizing the training of the memory map convolution network.
And training a deep autoregressive circulating network based on the construction of a training sea surface temperature data set. The training process uses the time t 0 -1 and prediction interval data defining a loss function L of the network as:
Figure BDA0003020006020000071
wherein L is reg Represents L 2 A regularization term. And training the memory map convolution network by adopting a loss function.
The specific training process is as follows.
Figure BDA0003020006020000072
S42: sea surface temperature prediction.
And for the trained model, forecasting the sea surface temperature from T +1 to T + tau according to the sea surface temperature observed value at the previous T moment.
Specifically, for the sea surface temperature moment T +1 to be forecasted, the sea surface temperature data before the T +1 moment is extracted, and the sea surface temperature observation value from the T +1 moment to the T + tau moment is extracted and used for detecting and correcting the result.
Using the deep autoregressive circulating network model trained in the steps S1 to S3 to perform the operation on the time from T +1 to T + tauThe predicted wave height value of (2) is corrected. Obtain the corrected result
Figure BDA0003020006020000073
Specifically, it is known that the sea surface temperature at the future time is corrected by using the learned sea surface temperature relationship at the past time based on the measured sea surface temperature data at a certain time node for a certain period of time.
The method described in the present invention is used to predict sea surface temperature. In the present embodiment, 1 hour, 3 hours, and 7 hours of correction were performed, respectively, using the observation data within one year of the region to be corrected as a data source.
Fig. 2 to 4 are schematic diagrams showing the forecast results and the actual results after 1 hour, 3 hours and 7 hours of the time to be predicted, respectively. Fig. 5-7 are plots of 1 hour, 3 hours, and 7 hour corrections predicted by the method of the present invention for the sea surface temperature at a grid point in the area.
Taking fig. 2 as an example, the left graph in fig. 2 shows a graph of the predicted result of the sea surface temperature of 1 hour obtained by the method of the present invention; the right image is the result of satellite measurement during the forecasting process. As can be seen from the figure, the sea surface temperature value predicted by the memory map convolution network is close to the sea surface actual measurement result.
Taking fig. 3 as an example, the left graph in fig. 3 shows a graph of the predicted sea surface temperature for 3 hours obtained by the method of the present invention; the right image is the result of satellite measurement during the forecasting process. As can be seen from the figure, the sea surface temperature value predicted by the memory map convolution network is close to the sea surface actual measurement result.
Taking fig. 4 as an example, the left graph in fig. 4 shows a 7-hour prediction result of the sea surface temperature obtained by the method of the present invention; the right image is the result of satellite measurement during the forecasting process. As can be seen from the figure, the sea surface temperature value predicted by the memory map convolution network is close to the sea surface actual measurement result.
Taking fig. 5 as an example, a certain grid point in the area takes every hour as a time step, and a graph between the sea surface temperature measured value and the sea surface temperature predicted value is performed for 100 hours; the forecast value of the sea surface temperature is the result obtained by the method of the invention, and the measured value of the sea surface temperature is the result of the satellite measurement in 100 hours of the forecasting process. As can be seen from the figure, the sea surface temperature value predicted by the memory map convolution network is close to the sea surface actual measurement result.
Taking fig. 6 as an example, a certain grid point in the area takes every three hours as a time step, and a graph between the sea surface temperature measured value and the sea surface temperature predicted value is performed for 100 hours; the predicted sea surface temperature value is the result obtained by the method of the present invention, and the measured sea surface temperature value is the result of satellite measurement in 100 hours of the prediction process. As can be seen from the figure, the sea surface temperature value predicted by the memory map convolution network is close to the sea surface actual measurement result.
Taking fig. 7 as an example, a grid point in the area takes every seven hours as a time step, and a graph between the sea surface temperature measured value and the sea surface temperature predicted value is performed for 100 hours; the forecast value of the sea surface temperature is the result obtained by the method of the invention, and the measured value of the sea surface temperature is the result of the satellite measurement in 100 hours of the forecasting process. As can be seen from the figure, the sea surface temperature value predicted by the memory map convolution network is close to the sea surface actual measurement result.
Based on the above, the sea surface temperature forecasting method provided by the invention can effectively forecast the sea surface temperature and improve the precision of sea surface temperature forecasting. Is beneficial to disaster prevention and reduction and offshore construction.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims (2)

1. The sea surface temperature forecasting method based on the memory map convolutional network is characterized by comprising the following steps:
s1: constructing a sea surface temperature data set: extracting phase from data of each time by using global sea surface temperature field data of T + T timesConstructing a data set of the sea surface temperature of the area to be predicted with the same position size of M x N, and extracting the sea surface temperature of the area to be predicted from the 1 st to the T th time as historical data
Figure FDA00030200060100000115
Extracting sea surface temperature of the area to be predicted as future data at the T +1 th to the T + tau th moments
Figure FDA00030200060100000116
Constructing a training data set;
s2: constructing a weighted adjacency matrix: based on the sea surface temperature data set, representing the position relation of the sea temperature data by using a weighted adjacency matrix; sea surface temperature sequence defined on a regular grid of longitude and latitude coordinates
Figure FDA00030200060100000117
Each time t is composed of M x N effective sea surface temperature points, and is calculated
Figure FDA00030200060100000118
Each term W in the weighted adjacency matrix W of ij Comprises the following steps:
Figure FDA0003020006010000011
wherein w mn Representing a weighted adjacency between sea surface temperature points m and n, d mn Representing the distance, w, between the recorded sea surface temperature points m and n on the grid mn Representing the distance d between grid points mn Calculating the obtained weight; r and d min Respectively representing a scale parameter and a minimum threshold distance parameter of a formula;
s3: constructing a memory map convolution model, constructing a memory map convolution depth model based on sea surface temperature data set data, training the memory map convolution depth model, correcting model parameters, and taking the corrected memory map convolution depth model as a sea surface temperature prediction model;
and S4, predicting the sea surface temperature by adopting the trained memory map convolution model.
2. The sea surface temperature forecasting method of claim 1, wherein the memory map convolution model includes a memory layer, a map convolution layer and an output layer; the method for calculating the output of the memory layer comprises the following steps:
the memory layer is composed of a time convolution unit and a gate control circulation unit, wherein the time convolution unit is as follows:
Figure FDA0003020006010000012
wherein the content of the first and second substances,
Figure FDA0003020006010000013
representing the input to the memory convolution unit,
Figure FDA0003020006010000014
the output of the time convolution unit is represented,
Figure FDA0003020006010000015
and B mem Respectively representing a convolution kernel and an offset, the height and width of the convolution kernel being set to 1 and K, respectively mem Time convolution unit output
Figure FDA0003020006010000016
Divided into two identical channels
Figure FDA0003020006010000017
And
Figure FDA0003020006010000018
then as the input of the gating cycle unit, the gating cycle unit is:
Figure FDA0003020006010000019
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA00030200060100000110
by
Figure FDA00030200060100000111
The obtained product is intercepted and obtained,
Figure FDA00030200060100000112
wherein, e represents a natural constant,
Figure FDA00030200060100000113
representing the Hamdamard product;
after the input data passes through the memory convolution layer, the output of the memory convolution layer is obtained
Figure FDA00030200060100000114
CN202110399839.7A 2021-04-14 2021-04-14 Sea surface temperature forecasting method based on memory map convolution network Pending CN115204437A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110399839.7A CN115204437A (en) 2021-04-14 2021-04-14 Sea surface temperature forecasting method based on memory map convolution network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110399839.7A CN115204437A (en) 2021-04-14 2021-04-14 Sea surface temperature forecasting method based on memory map convolution network

Publications (1)

Publication Number Publication Date
CN115204437A true CN115204437A (en) 2022-10-18

Family

ID=83573963

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110399839.7A Pending CN115204437A (en) 2021-04-14 2021-04-14 Sea surface temperature forecasting method based on memory map convolution network

Country Status (1)

Country Link
CN (1) CN115204437A (en)

Similar Documents

Publication Publication Date Title
CN114445634A (en) Sea wave height prediction method and system based on deep learning model
CN112307676B (en) Wave height numerical prediction model result correction method
CN111144666B (en) Ocean thermocline prediction method based on deep space-time residual error network
CN111242351A (en) Tropical cyclone track prediction method based on self-encoder and GRU neural network
CN115951014A (en) CNN-LSTM-BP multi-mode air pollutant prediction method combining meteorological features
CN114004428B (en) Fishing situation prediction method and system based on cellular grid integration mechanism and space-time Transformer
US20230393304A1 (en) El nino extreme weather early warning method and device based on incremental learning
Prasetya et al. Rainfall forecasting for the natural disasters preparation using recurrent neural networks
CN116822382A (en) Sea surface temperature prediction method and network based on space-time multiple characteristic diagram convolution
Grönquist et al. Predicting weather uncertainty with deep convnets
CN116699731B (en) Tropical cyclone path short-term forecasting method, system and storage medium
CN116976227B (en) Storm water increasing forecasting method and system based on LSTM machine learning
CN117556197A (en) Typhoon vortex initialization method based on artificial intelligence
CN116341391B (en) Precipitation prediction method based on STPM-XGBoost model
CN117253344A (en) Seawater acidification early warning and forecasting method, system and electronic equipment
CN111811465A (en) Method for predicting sea wave effective wave height based on multi-sine function decomposition neural network
CN110648030A (en) Method and device for predicting seawater temperature
CN115204437A (en) Sea surface temperature forecasting method based on memory map convolution network
CN116152206A (en) Photovoltaic output power prediction method, terminal equipment and storage medium
CN114169646B (en) Water bloom prediction method, device, electronic equipment and computer readable storage medium
Jonnalagadda et al. Spatial-Temporal Forecast of the probability distribution of Oceanic Nino Index for various lead times.
Wang et al. The prediction method of tropical cyclone intensity change based on deep learning
Li et al. DeepPhysiNet: Bridging Deep Learning and Atmospheric Physics for Accurate and Continuous Weather Modeling
CN117434624B (en) Strong convection weather identification and development prejudgment method based on semantic segmentation
Sabat et al. Investigation on the Impact of Attention Mechanism in Deep Learning Models for Temperature Prediction

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