Sea surface temperature prediction method based on SA-ConvGRU

Abstract

The invention provides a sea surface temperature prediction method based on SA-ConvGRU, and the method comprises the steps: carrying out the region segmentation operation of a sea surface temperature historical data set, and obtaining a segmented data set; performing three-dimensional conversion and data merging on the segmented data set to obtain an initial data set; performing standardization processing on the initial data set to obtain a standard data set; and performing data prediction on the standard data set through the constructed prediction model to obtain a prediction result. According to the method, the spatial feature relevance of the SST is fully considered, and the sea surface temperature prediction is realized through the constructed SA-ConvGRU model.

Description

Technical Field

[0001] The present invention relates to the technical field of sea surface temperature prediction, and in particular to a sea surface temperature prediction method based on SA-ConvGRU. Background Art

[0002] At present, sea surface temperature prediction methods can be roughly divided into two categories: numerical methods and data-driven methods. Numerical methods usually use dynamic and thermodynamic equations to describe sea surface temperature changes, which are usually complex and require a lot of calculations and time, making it difficult to solve the equations that characterize sea temperature changes. Data-driven methods mainly predict future sea surface temperature values ​​from a data perspective. Among data-driven methods, the use of classical machine learning methods to predict sea surface temperature has become a basic method to capture the law of sea surface temperature changes. In recent years, with the development of deep learning, some deep learning methods have gradually been applied to the field of sea surface temperature (SST) prediction. The prediction of sea surface temperature through traditional recurrent neural networks will encounter gradient vanishing or gradient explosion problems, and cannot be applied to the field of long-term sequence prediction.

[0003] From the existing deep learning methods, most research results tend to only consider time series dependency. However, in the process of sea surface temperature prediction, the temperature of adjacent observation points at the same time is actually similar, which means that it is necessary to study the spatial feature correlation of sea surface temperature data. However, the extraction of spatial features is not deep enough, and it is impossible to obtain the accurate change trend of SST in space, resulting in low prediction accuracy. Therefore, it is necessary to design a sea surface temperature prediction method based on SA-ConvGRU. Summary of the invention

[0004] The purpose of the present invention is to provide a sea surface temperature prediction method based on SA-ConvGRU, so as to fully consider the spatial feature correlation of SST and realize the prediction of sea surface temperature through the constructed SA-ConvGRU model.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A sea surface temperature prediction method based on SA-ConvGRU, comprising the following steps:

[0007] Performing a regional segmentation operation on the sea surface temperature historical data set to obtain a segmented data set;

[0008] Perform three-dimensional transformation and data merging on the segmented data set to obtain an initial data set;

[0009] Standardize the initial data set to obtain a standard data set;

[0010] The constructed prediction model is used to predict the standard data set and obtain the prediction results; the prediction model has multiple SA-ConvGRU network models built in; the SA-ConvGRU network model includes a ConvGRU model and a feature aggregation mechanism.

[0011] Optionally, the sea surface temperature historical data set is subjected to a regional segmentation operation to obtain a segmented data set, and the specific steps are: the sea surface temperature historical data set is subjected to a regional non-overlapping segmentation operation according to a preset segmentation threshold to obtain a plurality of non-repeating small-range regional data sets.

[0012] Optionally, the segmented data set is three-dimensionally transformed and merged to obtain an initial data set. The specific steps are: using the single-region 2D sea surface temperature data tensor in the segmented data set as the plane basis, and using the historical time length as the vertical height to construct a single-region 3D tensor sea surface temperature data set, and merging the single-region 3D tensor sea surface temperature data set to obtain the initial data set.

[0013] Optionally, the calculation formula for the normalization process is: in, is the standardized data, x is the actual regional observation value, μ is the regional mean, and σ is the regional standard deviation.

[0014] Optionally, the prediction model includes an encoder and a decoder; both the encoder and the decoder have a built-in SA-ConvGRU network model; the structural formula of the SA-ConvGRU network model includes: Z t =σ(W xz x t +W hz H t-1 );R t =σ(W xr x t +W hr H t-1 ); Among them, Y is the self-attention module, To input information, is the previous state information, x t is the input, H t-1 is the input vector of the t-1th time step, Z t is the update gate, σ is the sigmoid activation function, R t is the reset gate, H' t For the candidate gate, is the Hadamard product, f is the nonlinear activation function, x is the convolution operation, H and W are the height and width of the input tensor respectively, H t is the input vector for the tth time step.

[0015] Optionally, the ConvGRU model includes a convolutional neural network and a gated recurrent unit; the structural formula of the ConvGRU model includes: Z t =σ(W xz x t +W hz H t-1 );R t =σ(W xr x t +W hr H t-1 ); Among them, W xz , W xr and W xh Both are x t The associated weight matrix, W hz , W hr and W hh are all weight matrices related to the hidden state at the previous moment.

[0016] Optionally, the calculation formula of the feature aggregation mechanism includes: i' t =σ(W iz *Z+W ih *H t +b i );g' t =tanh(W gz *Z+W gh *H t +b g );o' t =σ(W oz *Z+W oh *H t +b o ), where i' t is the state at time t, Z is the aggregate feature of the final output, and W iz , W gz and W oz are weight matrices related to Z, W ih , W gh and W oh are weight matrices associated with the input vector at the tth time step, b i is the bias of the new input gate, g' t is the state of the memory unit at time t, b g is the offset of the memory unit, o' t is the state of the new output gate at time t, b o is the bias of the new output gate.

[0017] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the sea surface temperature prediction method based on SA-ConvGRU provided by the present invention comprises: performing a regional segmentation operation on the sea surface temperature historical data set to obtain a segmented data set; performing three-dimensional transformation and data merging on the segmented data set to obtain an initial data set; performing standardization processing on the initial data set to obtain a standard data set; performing data prediction on the standard data set through the constructed prediction model to obtain a prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0019] Figure 1 This is a flow chart of sea surface temperature prediction according to an embodiment of the present invention;

[0020] Figure 2 It is a SA-ConvGRU network model structure diagram of an embodiment of the present invention;

[0021] Figure 3 This is a structural diagram of the SAM module of an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 As shown, the embodiment of the present invention provides a sea surface temperature prediction method based on SA-ConvGRU, comprising the following steps:

[0025] Step 100: performing a region segmentation operation on the sea surface temperature historical data set to obtain a segmented data set;

[0026] Specifically, the sea surface temperature historical data set of this embodiment is the OISST (optimum interpolation sea surface temperature) data set of the global sea surface temperature from 1982 to 2022 of the National Oceanic and Atmospheric Administration (NOAA) of the United States. The OISST data set is divided into equal regions at a resolution of 0.25 degrees to obtain 24 regions, each of which is 60*60 in size.

[0027] Step 200: performing three-dimensional transformation and data merging on the segmented data set to obtain an initial data set;

[0028] Specifically, the single-region 2D sea surface temperature data tensor in the segmented dataset is used as the plane basis, and the historical time length that needs to be input is used as the longitudinal height to construct a single-region 3D tensor sea surface temperature dataset, thereby converting the SST time series data into SST space-time series data.

[0029] It should be noted that the standard space-time sequence generally adopts X 1:T =[X 1 ,X 2 ,...,X T ] to represent, where each X i can be described by a 3D tensor [w,h,feature], where feature represents the number of features of each tensor. w and h represent the monitoring width and monitoring height on this feature variable, respectively, which can be used to obtain the size of a single region after region segmentation. The length of the sequence T corresponds to the size of the historical time used, so the spatiotemporal sequence X 1:T It can be described by a 4D tensor [T, 60, 60, feature]. Each single region contains the SST information of the region for 41 years.

[0030] More specifically, in this embodiment, the SST time series data set of a single region is represented by F = {F 1 ,F 2 ,...F k}, where k is the number of days of historical SST data used, F k represents the matrix of all SST data values ​​detected in the area on a single day, F k =w×h, w represents the monitoring width of the area, h represents the monitoring height of the area, both are 60. Then these matrix data are preprocessed and converted into the SST time series of single point detection X={X 1 ,X 2 ,...X n}, where |X i |=k,(1≤i≤n,i∈Z).

[0031] Step 300: Standardize the initial data set to obtain a standard data set;

[0032] Specifically, after the data is segmented and merged into different regions, the data in the same region needs to be standardized so that the data differences in the same batch of data can be reduced to an appropriate range, thereby greatly improving the training speed of the neural network and reducing the usage overhead of computer hardware resources.

[0033] More specifically, this embodiment uses z-score standardization for standardization processing, and the calculation formula is:

[0034]

[0035] in, is the standardized data, x is the actual regional observation value, μ is the regional mean, and σ is the regional standard deviation.

[0036] Step 400: Perform data prediction on the standard data set using the constructed prediction model to obtain a prediction result;

[0037] Specifically, the prediction model adopts a codec structure, including an encoder and a decoder, and builds different numbers of SA-ConvGRU network models in the encoder and decoder structures. The encoder part is used to extract spatiotemporal feature information in the data; the decoder part is a prediction network, which is used to analyze and process the effective features output from the encoder to achieve target prediction.

[0038] It should be noted that the encoding and decoding structure is highly practical and generalizable. Through the multi-layer network layers built in the encoder, the spatiotemporal feature information is regulated by multi-layer features to achieve multi-layer attention coordination, multi-dimensional expansion of the local spatial receptive field, and enhanced global spatiotemporal feature correlation dependency, achieving better feature aggregation processing effects, thereby ensuring the effective transmission of important spatiotemporal features. After the spatiotemporal features output by the encoder are adjusted and processed by the multi-layer attention mechanism, they are sent to the decoder for corresponding decoding, and finally the predicted target is output.

[0039] Furthermore, the structure of the SA-ConvGRU network model is as follows Figure 2 As shown. Among them, the structural formula of the SA-ConvGRU network model is as follows:

[0040]

[0041] Z t =σ(W xz x t +W hz H t-1 );

[0042] Rt =σ(W xr x t +W hr H t-1 );

[0043]

[0044] Among them, Y is the self-attention module, To input information, is the previous state information, x t is the input, H t-1 is the input vector of the t-1th time step, Z t is the update gate, σ is the sigmoid activation function, R t is the reset gate, H' t For the candidate gate, is the Hadamard product, f is the nonlinear activation function, x is the convolution operation, H and W are the height and width of the input tensor respectively, H t is the input vector of the tth time step; W xz is the same as the current input x t The associated weight matrix is ​​responsible for controlling the effect of the current input in the update gate; W hz is the weight matrix associated with the hidden state at the previous moment, responsible for controlling the role of historical information in the update gate; W r is with x t The associated weight matrix determines the role of the current input in calculating the reset gate; W hr is the weight matrix associated with the hidden state at the previous moment, which determines the influence of the hidden state at the previous moment on the calculation of the reset gate; W xh is with x t The relevant weight matrix is ​​responsible for introducing the current input information when calculating the candidate hidden state; W hh It is a weight matrix related to the hidden state at the previous moment, which controls the role of the hidden state at the previous moment.

[0045] It should be noted that in each time step, the attention module SAM selectively performs a weighted summation of the input features of each position and all position information.

[0046] Specifically, the SA-ConvGRU network model includes the ConvGRU model and the feature aggregation mechanism. The ConvGRU model combines the convolutional neural network and the gated recurrent unit (GRU), takes into account the heterogeneity of spatiotemporal data, and realizes the simultaneous processing of SST spatiotemporal features. The structural formula of the ConvGRU model is as follows:

[0047] Z t =σ(Wxz x t +W hz H t-1 );

[0048] R t =σ(W xr x t +W hr H t-1 );

[0049]

[0050] It should be noted that the update gate adds the information of the convolutional neural network and the gated recurrent unit and puts it into the sigmoid activation function, and compresses the activation result to between 0 and 1. The reset gate combines the new input information with the previous memory, and determines the amount of information passed from the past to the future and the amount of information passed from the previous time step to the current time step.

[0051] Specifically, this embodiment embeds SAM into a standard ConvGRU to construct a SA-ConvGRU model, which memorizes features with long-term dependencies in spatial and temporal domains through self-attention memory.

[0052] It should be noted that the introduced self-attention module SAM can improve the spatial feature utilization of the three-dimensional convolutional network, show the importance of different environmental variable factors to the results, and at the same time adjust the importance of information at different historical moments in the variables, thereby enhancing the model's feature extraction ability and prediction accuracy.

[0053] Furthermore, the SAM structure is Figure 3 SAM is a feature aggregation mechanism that uses self-attention to fuse current features and previously memorized features by calculating pairwise similarity scores. t-1 and M t are the self-attention memory states at time t-1 and t, respectively, H t is the input at time t, is the output at time t, Represents matrix multiplication; Adding matrix elements, Q h , K h and V h H t The three feature spaces of the mapping are query, key and value, K m and V m M t-1 The corresponding mapping, Z h is the standard self-attention network aggregation feature, Z mis the aggregate feature corresponding to the newly added memory unit, and Z is the aggregate feature of the final output.

[0054] It should be noted that this structure adds a memory module to the SA mechanism. Different from the SA mechanism, SAM uses the input information H t The generated Q h With K h and K m Calculate the similarity and get A h and A m , and finally get Z h and Z m , and combine the two into Z to get the output result Generate the corresponding gate control unit. The calculation formula of this process is:

[0055] i' t =σ(W iz *Z+W ih *H t +b i );

[0056] g' t =tanh(W gz *Z+W gh *H t +b g );

[0057] o' t =σ(W oz *Z+W oh *H t +b o );

[0058] Among them, i' t is the state of the new input gate in SAM at time t; W iz is the weight matrix associated with Z, W ih is with H t The related weight matrix, together they constitute the weight matrix W of the new input gate i ; b i is the bias of the new input gate; g' t is the state of the memory unit in SAM at time t; W gz is the weight matrix associated with Z, W gh is with H t The related weight matrix, together they constitute the weight matrix W of the memory unit g ; b g is the bias of the memory unit; o' t is the state of the new output gate in SAM at time t; W oz is the weight matrix associated with Z, Woh is with H t The related weight matrix, together they constitute the weight matrix W of the new output gate o ; b o is the bias of the new output gate.

[0059] According to i' t 、g' t 、o' t The memory state and output results of self-attention are used to update the structure and obtain M t and The calculation formula for this process is:

[0060]

[0061] It is understandable that this embodiment introduces a separate memory unit M inside SAM. The memory unit M can extract and store the general features of the local spatial receptive field at the historical and current time steps in the spatial domain; the memory unit M can capture the long-range time feature dependencies through the gating mechanism in the time domain. The function of strengthening the dependency between long-term temporal features while capturing the long-range spatial information features is realized through the SAM module and the memory unit M.

[0062] Specifically, this embodiment uses mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) as evaluation indicators to evaluate the prediction results of the SA-ConvGRU model. The calculation formulas of each indicator are as follows:

[0063]

[0064] Among them, y i and They represent the true value and predicted result of SST on the i-th day respectively.

[0065] The beneficial effects of the present invention are as follows:

[0066] 1) The influence of the spatial characteristic correlation of SST on SST data is fully considered, and the time series is converted into spatiotemporal series through regional segmentation and three-dimensional transformation;

[0067] 2) Using the encoder-decoder structure to build a network model improves the practicality and generalization of the model;

[0068] 3) The constructed SA-ConvGRU network model introduces the feature aggregation mechanism SAM, which expands the local spatial receptive field and enhances the correlation dependency of global spatiotemporal features.

[0069] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0070] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A sea surface temperature prediction method based on SA-ConvGRU, characterized in that: The steps include: Performing a regional segmentation operation on the sea surface temperature historical data set to obtain a segmented data set; Performing three-dimensional transformation and data merging on the segmented data set to obtain an initial data set; Standardizing the initial data set to obtain a standard data set; The constructed prediction model is used to perform data prediction on the standard data set to obtain a prediction result; the prediction model has multiple SA-ConvGRU network models built in; the SA-ConvGRU network model includes a ConvGRU model and a feature aggregation mechanism.

2. The sea surface temperature prediction method based on SA-ConvGRU according to claim 1, characterized in that: The sea surface temperature historical data set is subjected to a regional segmentation operation to obtain a segmented data set. The specific steps are: the sea surface temperature historical data set is subjected to a regional non-overlapping segmentation operation according to a preset segmentation threshold to obtain a plurality of non-repeating small-range regional data sets.

3. The sea surface temperature prediction method based on SA-ConvGRU according to claim 1, characterized in that: The segmented data set is subjected to three-dimensional transformation and data merging to obtain an initial data set, and the specific steps are: using the single-region 2D sea surface temperature data tensor in the segmented data set as a plane basis, and using the historical time length as a longitudinal height to construct a single-region 3D tensor sea surface temperature data set, and data merging the single-region 3D tensor sea surface temperature data set to obtain the initial data set.

4. The sea surface temperature prediction method based on SA-ConvGRU according to claim 1, characterized in that: The calculation formula for the standardization process is: in, is the standardized data, x is the actual regional observation value, μ is the regional mean, and σ is the regional standard deviation.

5. The sea surface temperature prediction method based on SA-ConvGRU according to claim 1, characterized in that: The prediction model includes an encoder and a decoder; the encoder and the decoder are both equipped with the SA-ConvGRU network model; the structural formula of the SA-ConvGRU network model includes: Z t =σ(W xz x t +W hz H t-1 );R t =σ(W xr x t +W hr H t-1 ); Among them, Y is the self-attention module, To input information, is the previous state information, x t is the input, H t-1 is the input vector of the t-1th time step, Z t is the update gate, σ is the sigmoid activation function, R t H is the reset gate. t 'For candidate gates, is the Hadamard product, f is the nonlinear activation function, x is the convolution operation, H and W are the height and width of the input tensor respectively, H t is the input vector for the tth time step.

6. The sea surface temperature prediction method based on SA-ConvGRU according to claim 5, characterized in that: The ConvGRU model includes a convolutional neural network and a gated recurrent unit; the structural formula of the ConvGRU model includes: Z t =σ(W xz x t +W hz H t-1 );R t =σ(W xr x t +W hr H t-1 ); Among them, W xz , W xr and W xh Both are x t The associated weight matrix, W hz , W hr and W hh are all weight matrices related to the hidden state at the previous moment.

7. The sea surface temperature prediction method based on SA-ConvGRU according to claim 1, characterized in that: The calculation formula of the feature aggregation mechanism includes: i′ t =σ(W iz *Z+W ih *H t +b i ); g′ t =tanh(W gz *Z+W gh *H t +b g );o′ t =σ(W oz *Z+W oh *H t +b o ), where i′ t is the state at time t, Z is the aggregate feature of the final output, and W iz , W gz and W oz are weight matrices related to Z, W ih , W gh and W oh are weight matrices associated with the input vector at the tth time step, b i is the bias of the new input gate, g′ t is the state of the memory unit at time t, b g is the offset of the memory unit, o′ t is the state of the new output gate at time t, b o is the bias of the new output gate.

Images