CN110197307B - Regional sea surface temperature prediction method combined with attention mechanism - Google Patents

Abstract

The invention discloses a regional sea surface temperature prediction method combined with an attention mechanism. The steps include: 1), processing the daily SST data in the region into a matrix, and arranging them sequentially in chronological order to form a matrix sequence, as the input of the CA-ConvLSTM model; 2), processing the SST matrix, and extracting the distribution characteristics of each recording point through a convolution layer; 3), using the attention mechanism to assign attention weights to the obtained matrix characteristics, and then multiplying the attention weights by the corresponding matrix characteristics to obtain weighted characteristics; 4), finally, using the weighted characteristics as Con The input of the vLSTM model, use ConvLSTM to train the prediction model, and finally obtain the prediction result. The invention organizes the SST in the region into a matrix, which is input into the model as a whole, so that the model can extract the time and space correlation of the SST.

Description

A Regional Sea Surface Temperature Prediction Method Combined with Attention Mechanism

technical field

The invention relates to the field of sea surface temperature prediction, in particular to a regional sea surface temperature prediction method combined with an attention mechanism.

Background technique

In recent years, sea surface temperature prediction has attracted increasing attention in various ocean-related fields such as marine meteorology, navigation, marine disaster prevention and mitigation, and marine fisheries. So far, many methods have been proposed to predict Sea Surface Temperature (SST), and good results have been achieved. These methods can be mainly divided into three categories: statistical forecasting methods, numerical forecasting methods and empirical forecasting methods. With the continuous improvement of information collection technology, more and more SST data are collected, stored and processed, and the researchers also analyzed the spatiotemporal characteristics of SST. SST is a typical time series data, so many scholars regard Sea Surface Temperature Prediction (SSTP) as a time series regression problem, and try to use time series prediction methods to predict SST, hoping to obtain better prediction results.

SSTP is not only important in theory, but also has practical applications in many ocean-related fields. The historical temperature data of a single record point on the ocean surface is a typical long-term series data. Therefore, many researchers regard SSTP as a time series regression problem and apply many time series forecasting methods to SSTP. Traditional time series forecasting models such as Autoregressive (AR) model, Moving Average (MA) model, and Autoregressive Moving Average (ARMA) model are all linear models, which cannot meet the needs of practical applications. Therefore, researchers have continuously proposed many nonlinear forecasting methods. Neural network is a nonlinear prediction method with strong nonlinear fitting ability. Currently, there are many different forms of neural network used in SSTP. Aparna et al. proposed an artificial neural network to use the current SST to predict the SST in the northeast Arabian Sea for the next day, with a prediction model error of ≤±0.5°. Long Short Term Memory (LSTM) is a variant of Recurrent Neural Network (RNN), which can solve the problem of RNN "gradient disappearance", and has a strong modeling ability for time series data. It has also been applied to SSTP and achieved remarkable results. Besides neural networks, Support Vector Machines (SVM) are also used for time series forecasting. Lins et al. considered the seasonality of SST changes and the regularity of intra-seasonal changes, obtained the curvature and slope information of SST from buoy data, and used SVM to predict SST.

However, the above methods only consider the temporal correlation of SST, but do not consider the spatial correlation between SSTs, so a large amount of information will be lost when predicting SST; secondly, when predicting SST, the different influences of historical SST on the SST to be predicted are not reflected, making the model not comprehensive enough, and the final prediction accuracy is not high.

Contents of the invention

The object of the present invention is to address the deficiencies in the prior art and provide a regional sea surface temperature prediction method combined with an attention mechanism to solve the above problems.

The technical problem solved by the present invention can adopt following technical scheme to realize:

A regional sea surface temperature prediction method combined with an attention mechanism, comprising the following steps:

1) Process the daily SST data in the area into a matrix, and arrange them sequentially according to time to form a matrix sequence, which is used as the input of the CA-ConvLSTM model;

2), process the SST matrix, and extract the distribution characteristics of each recording point through the convolution layer;

3) Use the attention mechanism to assign attention weights to the obtained matrix features, and then multiply the attention weights by the corresponding matrix features to obtain weighted features;

4) Finally, the weighted feature is used as the input of the ConvLSTM model, and the prediction model is trained using ConvLSTM, and finally the prediction result is obtained.

3. The regional sea surface temperature prediction method combined with attention mechanism according to claim 1, characterized in that: the SST matrix convolution is after the SST data is sorted into a digital matrix sequence with a length of |F|, and the convolution is used to extract the features of the matrix to obtain local features; the convolution operation is completed by the movement of the convolution kernel, and each value of the output matrix is the sum of the product of the value of each 3 * 3 area in the input matrix and the value of the corresponding position of the 3 * 3 convolution kernel.

Further, the daily SST data in the region is processed into a W·H matrix, the SST sequence F=F ₁ , _{F 2} , ..., F _|F|

Further, the matrix feature sequence is obtained through the convolution operation. In the prediction, the SST of k days is used to predict the SST of one day or five days in the future, and the value of k is different; starting from the nth day, the matrix feature sequence of this k day is M=M _n+1 , M _n+2 ,..., M _n+k-1 , M _n+k , and _Mi is the ith (n≤i≤n+k, i∈Z) matrix feature, that is, the matrix of (W-2) (H-2).

Further, the method for automatically learning attention weights of the CA-ConvLSTM model is as follows:

First, the mean value of each matrix feature is obtained through the pooling layer, that is, V=[v _n+1 , v _n+2 , ..., v _n+k-1 , v _n+k ], at this time v _i is the average value of all values in the _Mi matrix; the definition of the attention model Φ is as follows:

D＝tanh(w*V+b) (1)

A＝softmax(W'*D+b') (2)

Where w, b are the weight and bias of the fully connected layer, D is the output of the fully connected layer and the input of softmax, W' and b' are the weight and bias of softmax, A is the output of softmax, that is, the attention weight, which is a vector of length k, expressed as A=[a_n+1, a_n+2,...,a_n+k-1, a_n+k,]; After that, the attention weights are multiplied by the corresponding matrix features in turn to obtain a new feature sequence, that is, the weighted feature sequence, expressed as N=N_n+1, N_n+2,...,N_n+k-1, N_n+k, N_iis the i-th weighted feature, N_i=a_i*M_i, is a matrix of (W-2)·(H-2);

Before the matrix feature sequence M enters the ConvLSTM model, the matrix feature sequence M is automatically weighted through the attention mechanism to obtain the weighted feature sequence N.

Further, the calculation formula of the ConvLSTM is as follows:

Among them, * represents the convolution operation, ○ represents the corresponding multiplication, and σ is the sigmoid activation function.

Compared with the prior art, the beneficial effects of the present invention are as follows:

The invention organizes the SST in the region into a matrix, and inputs it into the model as a whole, so that the model can extract the time and space correlation of the SST. The attention mechanism is used to assign weights to features, reflecting the different degrees of influence of historical SST on the time dimension of SST to be predicted, so that key information is highlighted and irrelevant information is ignored, which is conducive to improving the accuracy of SSTP. Using the strong modeling ability of ConvLSTM for spatiotemporal information, the spatiotemporal information of SST is modeled. In the prediction, not only the temporal correlation of SST, but also the spatial correlation is considered, so that the information contained in the model is more comprehensive.

The CA-ConvLSTM model has some practicality, and it combines temporal and spatial information to improve the accuracy of SSTP. Comparing the prediction results of different methods in the same data set, respectively predicting the SST within one day and within five days, the daily average deviation of SST within one day of CA-ConvLSTM prediction is about 0.18°, and the average daily deviation of SST within five days is about 0.33°. The experimental results show that CA-ConvLSTM has achieved the best results in SSTP, thus verifying the effectiveness of the method. On the other hand, this method can directly perform regional SSTP without modeling the SST of each station, which greatly improves the efficiency of forecasting.

Description of drawings

Fig. 1 is a flow chart of CA-ConvLSTM according to the present invention.

Fig. 2 is a schematic diagram of the learning process of the attention mechanism of the present invention.

Fig. 3 is a schematic diagram of the working principle of the ConvLSTM described in the present invention.

Detailed ways

In order to make the technical means, creative features, goals and effects achieved by the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.

Referring to Fig. 1 to Fig. 3, a regional sea surface temperature prediction method combined with an attention mechanism according to the present invention. The daily SST data in the region is processed into a W·H matrix. The SST sequence F=F _{1 ,} F ₂ ,..., F _|F| In order to fully consider the correlation of SST in time and space, and the influence of historical SST on the time dimension of SST to be predicted, the CA-ConvLSTM model was proposed for the first time, and it was successfully applied to SSTP and achieved good results.

Figure 1 is a flowchart of the CA-ConvLSTM model, where |F| represents the time length of the SST matrix, W represents the width of the region, H represents the height of the region, and k represents the number of days of historical SST used for prediction. The method comprises the steps of:

1) Process the daily SST data in the area into a matrix, and arrange them in chronological order to form a matrix sequence, which is used as the input of the CA-ConvLSTM model.

2) Process the SST matrix, and extract the distribution features of each recording point through the convolutional layer.

3) Use the attention mechanism to assign attention weights to the obtained matrix features, and then multiply the attention weights by the corresponding matrix features to obtain weighted features.

4) Finally, the weighted feature is used as the input of the ConvLSTM model, and the prediction model is trained using ConvLSTM, and finally the prediction result is obtained.

In the above steps, attention weight and ConvLSTM are the two key components of CA-ConvLSTM to predict SST. Reasonable distribution of attention weight and better training of ConvLSTM are the key to ensure that CA-ConvLSTM can obtain excellent performance.

SST matrix convolution:

After sorting the SST data into a sequence of digital matrices with length |F|, convolution is used to perform feature extraction on the matrix to obtain local features. The convolution operation is mainly done by moving the convolution kernel. Each value of the output matrix is the sum of the product of the value of each 3×3 area in the input matrix and the value of the corresponding position of the 3×3 convolution kernel. Table 1 is an information table of parameters related to the convolutional layer, as follows:

Table 1: Convolutional layer related parameter information table

The input is a W·H matrix, and the convolution kernel moves with a step size of 1×1, then the convolution kernel moves (W-2)·(H-2) times, and the output is a matrix of (W-2)·(H-2). This not only preserves the spatial information of SST, but also extracts local features.

Attention Mechanism:

The matrix feature sequence is obtained through the convolution operation. In the prediction, the SST of k days is used to predict the SST of one day or five days in the future, and the value of k is different. Starting from the nth day, the matrix feature sequence of the k day can be expressed as M=M _n+1 , M _n+2 , ..., M _n+k-1 , M _n+k , and M _i is the ith (n≤i≤n+k, i∈Z) matrix feature, that is, the matrix of (W-2)·(H-2).

Figure 2 is a schematic diagram of the attention learning process. In order to allow the model to automatically learn the attention weight, first obtain the mean value of each matrix feature through the pooling layer, that is, V=[v _n+1 , v _n+2 ,...,v _n+k-1 , v _n+k ], at this time v _i is the average value of all values in the _Mi matrix; the definition of the attention model Φ is as follows:

D＝tanh(w*V+b) (1)

A＝softmax(W'*D+b') (2)

Where w, b are the weight and bias of the fully connected layer, D is the output of the fully connected layer and the input of softmax, W' and b' are the weight and bias of softmax, A is the output of softmax, that is, the attention weight, which is a vector of length k, expressed as A=[a_n+1, a_n+2,...,a_n+k-1, a_n+k,]; After that, multiply the attention weight and the corresponding matrix features in turn to get a new feature sequence, that is, the weighted feature sequence, which can be expressed as N=N_n+1, N_n+2,...,N_n+k-1, N_n+k, N_iis the i-th weighted feature, N_i=a_i*M_i, is a (W-2)·(H-2) matrix.

Before the matrix feature sequence M enters the ConvLSTM model, the matrix feature sequence M is automatically weighted through the attention mechanism to obtain the weighted feature sequence N, so that important information in the feature sequence can be assigned a larger attention weight, and irrelevant or unimportant information will be assigned a smaller attention weight, which can provide more comprehensive information for subsequent predictions, and the prediction results will be more accurate.

ConvLSTM:

LSTM has a strong modeling ability for time series data, but if these data are matrices, these data have rich spatial information and strong local features, directly using LSTM to predict SST, not only does not consider spatial correlation, but also cannot extract local features, which will limit SSTP and the prediction accuracy will not be very high. Convolutional Neural Networks (CNN) have powerful feature extraction capabilities, so adding convolution operations on the basis of LSTM can improve the feature extraction capabilities of the model. ConvLSTM was inspired by this. It can not only express the spatial correlation of data, but also reflect the temporal correlation. Considering that SST data has both temporal and spatial correlations in prediction, it is theoretically feasible to model the temporal and spatial correlations of SST with ConvLSTM. The working principle of ConvLSTM is shown in Figure 3.

The calculation formula of ConvLSTM is as follows:

Among them, * represents the convolution operation, ○ represents the corresponding multiplication, and σ is the sigmoid activation function. The model has a strong ability to obtain the spatio-temporal relationship, and the previously obtained weighted feature sequence is used as the input of the ConvLSTM model. Using this method to predict SST, not only the time correlation of SST, but also the spatial correlation of SST is considered. In the actual prediction, the prediction accuracy will be improved to a certain extent. In the actual prediction, when using CA-ConvLSTM to predict the SST of one day, the prediction result is an SST matrix; when predicting the SST of five days, the prediction result is five SST matrices, that is, the SST in the next five consecutive days.

The method of the present invention when actual implementation is as follows:

Experimental environment and data:

In order to verify the performance of CA-ConvLSTM in SSTP, the Keras framework was used to build the model, the Adam optimizer was used for training, and the initial learning rate was 0.001. The experimental environment is Windows10, Intel Core i5, 2.6GHz, 8G RAM, and the algorithm is implemented using python3.

The data used in the experiment is SST data for a period of time. The data is collected near 30N and 130E. The data is arranged in rows and columns of 80*40, including invalid values, represented by -999.0.

Experimental performance indicators:

In order to verify the effectiveness of the CA-ConvLSTM model, the experiment uses Root Mean Square Error (Root Mean Square Error, RMSE) and prediction accuracy (Prediction Accuracy, PACC) to evaluate the performance of different prediction methods. The formulas of RMSE and PACC are expressed as follows:

Where Y_real is the real value and Y_pred is the predicted value. i represents the i-th value arranged in row and column order, and n represents the product of the width and height of the area to be predicted.

When the model is predicting, for the RMSE index, the smaller the RMSE, the better the model performance, and the opposite of PACC, the larger the PACC, the better the model performance. Then when the model performance is the best, it is necessary to satisfy the minimum RMSE and the maximum PACC. When training the model, the appropriate model structure and model parameters can be determined by comparing the values of RMSE and PACC.

Analysis of results:

Due to the lack of SST data used, that is, the SST data of some SST record points within 4749 days are all invalid values, so the SST data needs to be preprocessed. In order to ensure the integrity of the SST data, the 80*40 SST data is cut to 30*30, which removes the invalid value, then the 30*30 SST data after cutting are all valid values, and there are no invalid values. Then, 4749 days of SST data are divided, 75% of the data is used as a training set to train the parameters of the CA-ConvLSTM prediction model, and the remaining 25% of the data is used as a validation set to verify the learning effect of the model.

Determine the model structure: the role of the convolutional layer is to extract local features, and ConvLSTM itself already has such capabilities, so it becomes a question whether it is necessary to add a convolutional layer before A-ConvLSTM. In order to determine the model structure, the 30*30 SST data is used as a data set, and the SST of seven days and fifteen days is used to predict the SST of one day and five days, respectively, and the prediction results of A-ConvLSTM and CA-ConvLSTM are compared. A-ConvLSTM does not add a convolutional layer, and CA-ConvLSTM adds a convolutional layer.

Table 2: Performance comparison of A-ConvLSTM and CA-ConvLSTM

Table 2 shows the results of A-ConvLSTM and CA-ConvLSTM predicting SST within one day and five days under the same parameters and the same data set. For the PACC index, the accuracy of CA-ConvLSTM prediction for one day and five days is 99.33% and 98.78%, respectively, while the PACC index of A-ConvLSTM is 99.16% and 98.74%, which is lower than CA-ConvLSTM; for the RMSE index, the errors of CA-ConvLSTM prediction for one day and five days are 0.2347 and 0.4312, respectively, while the RMSE index of A-ConvLSTM is 0.327 3 and 0.4470, higher than CA-ConvLSTM. Through experimental comparison, it is found that CA-ConvLSTM is better than A-ConvLSTM no matter in RMSE index or PACC index. The experimental results show that adding convolutional layers has a certain effect on improving the accuracy of SSTP, so the structure of the model is determined to be CA-ConvLSTM. The main reason for this situation is that the local features extracted from the SST data are not obvious enough through the convolution operation of ConvLSTM itself, and adding a convolution layer before that improves the feature extraction ability of the model, making the data more obvious in the ConvLSTM model, which is conducive to improving the accuracy of SSTP.

Determine the input size: When using CA-ConvLSTM to predict SST, the size of the data set area is 30*30. Considering that the input size has an impact on the accuracy of SSTP, the edge part is cut from the 30*30 SST to get the 20*20 SST in the center, and then the edge part is cut to get the 10*10 SST in the center, and finally the SST datasets with different sizes are obtained. If the input size is too small, the amount of information contained in the SST data will be relatively small, which is not conducive to prediction; if the input size is too large, the prediction will be interfered by other information, making the prediction accuracy not high. Therefore, three data sets with different sizes are used as the input of the CA-ConvLSTM model, and the appropriate size is selected as the input of the model by comparing the evaluation indicators. Considering that when predicting one-day and five-day SST, the amount of information required is different, and inputs of different sizes will have different effects. For this reason, we still use seven-day and fifteen-day SST to predict one-day and five-day SST, respectively, and compare the RMSE and PACC indicators to determine the appropriate size of the model input when predicting one day and five days, respectively. Since the maximum size of the data set is 30*30, the size larger than 30*30 is not discussed in the experiment.

Table 3: Comparison of evaluation indicators of CA-ConvLSTM predicting SST within one day and five days under different input sizes

Table 3 reflects the comparison of evaluation indicators of CA-ConvLSTM predicting SST within one day and within five days under different input sizes, where k represents the number of days of historical SST used for prediction. When predicting the SST within a day, when 10*10 and 20*20 are used as input, the RMSE indicators are 0.3688 and 0.2426 respectively, both of which are higher than the RMSE indicators when 30*30 is used as the input; the PACC indicators of 10*10 and 20*20 are 98.78% and 99.27%, respectively, which are lower than the PACC indicators of 30*30. Through the above comparison, it can be determined that when predicting the SST of a day, the input size is 3 0*30 can get better results. When predicting the SST within five days, comparing the RMSE index and PACC index predicted by the CA-ConvLSTM model under different sizes, it can be found that the prediction effect is the best when the input size is 30*30. In SSTP, sufficient information support is needed to effectively improve the accuracy of SST area prediction, and the 10*10 and 20*20 areas contain less information, so when training the model, the amount of data is too small, and the model training results will not be very good. Perhaps increasing the input size can achieve better results, but since there is not enough SST data in the area, we will not discuss it here. To sum up, if the input size of the model is determined to be 30*30, then the SST data of this size will be used as the input of the model in subsequent experiments.

Determining the k value: In the previous experiment of determining the model structure and determining the input size, according to the expert experience knowledge, the SST of seven days and fifteen days were used to predict the SST of one day and five days respectively. Considering that the value of k has a greater or lesser impact on the accuracy of SSTP, when using seven-day and fifteen-day SST to predict one-day and five-day SST respectively, the results may not be the best. Therefore, k=4, 7, and 15 are respectively used to predict SST within one day; k=10, 15, and 25 are used to predict SST within five days. By comparing RMSE and PACC indicators through experiments, the appropriate k value is determined for predicting SST within one day and five days.

Table 4: Comparison of evaluation indicators of CA-ConvLSTM predicting SST within one day and five days under different k values

Table 4 reflects the comparison of evaluation indicators of CA-ConvLSTM predicting SST within one day and within five days under different k values. The input size represents the information of the spatial dimension, and the value of k represents the information of the time dimension. The values of both have an impact on the performance of the model. The input size has been determined in the above experiment, so the next step is to determine the size of the k value. When predicting a day, comparing the RMSE and PACC indicators under different k values, when k=7, RMSE is 0.2347, and PACC is 99.33%, which is better than the evaluation indicators when k=4, 15, so k=7 is used to predict SST within a day. When predicting five days, the PACC index of k=15 is slightly higher than the PACC index of k=10 and 25, and the RMSE index of k=15 is also better than k=10 and 25, so when predicting the SST within five days, take k=15. It can be seen that when predicting SST, the amount of information in the spatial and temporal dimensions should be moderate, too much or too little will affect the performance of the model, so the above experiments are needed. In summary, the present invention takes k=7 to predict the SST within one day, and k=15 to predict the SST within five days.

Comparison of prediction methods: After determining the model structure, input size and k value, the prediction model has basically been determined. In order to verify the effectiveness of the CA-ConvLSTM model, the performance of the CA-ConvLSTM model is compared with the current advanced SSTP methods. These methods include: SVR, LSTM and ConvLSTM. Among them, ConvLSTM can perform regional SSTP. The same experiment as above is also carried out on ConvLSTM. It is determined that the suitable input size of ConvLSTM is 30*30, and the SST of seven days and fifteen days is used to predict the SST within one day and five days in the future. Since both SVR and LSTM are single-point predictions, they can only build a model for each recording point separately, so the SST of 900 recording points in the 30*30 area needs to be predicted in the experiment. Among them, the Radial Basis Function (RBF) used by SVR can realize nonlinear mapping and has few parameters. For SVR and LSTM, the seven-day and fifteen-day SST are still used to predict the SST within one day and five days. By comparing the RMSE and PACC indicators of different prediction methods when predicting SST, the difference in prediction performance between CA-ConvLSTM and SVR, LSTM and ConvLSTM models is compared.

Table 5: Performance comparison of different prediction methods

Table 5 is a comparison of the evaluation indicators of CA-ConvLSTM and SVR, LSTM and ConvLSTM models. When predicting one day, the PACC indicators of SVR and LSTM are 98.92%, 98.86%, and the RMSE indicators are 0.3908 and 0.4655; the PACC indicators of ConvLSTM are 99.04%, and the RMSE indicators are 0.3602. Both indicators are better than SVR and LSTM, indicating that ConvLSTM is better than SVR and LSTM, which verifies the importance of spatial correlation to SSTP. Compared with CA-ConvLSTM, the evaluation indicators of ConvLSTM are worse. The experimental results show that when predicting a day, compared with the other three methods, the CA-ConvLSTM model has the best performance. When predicting five days, CA-ConvLSTM also has the best performance compared to SVR, LSTM and ConvLSTM. Since CA-ConvLSTM not only considers the time and space correlation of SST, but also expresses the different influences of historical SST on the predicted SST by assigning different weights, which makes the model closer to reality, contains more comprehensive information, and finally improves the accuracy of SSTP. In summary, compared with SVR, LSTM and ConvLSTM, CA-ConvLSTM has the best performance in predicting SST within one day and within five days, which verifies the effectiveness of the method.

The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above-mentioned embodiments. What are described in the above-mentioned embodiments and the description only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention also has various changes and improvements, and these changes and improvements all fall within the scope of the claimed invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims (1)

1. A regional sea surface temperature forecasting method in conjunction with attention mechanism, is characterized in that: comprise the steps: 1) Process the daily SST data in the area into a matrix, and arrange them sequentially according to time to form a matrix sequence, which is used as the input of the CA-ConvLSTM model; 2), process the SST matrix, and extract the distribution characteristics of each recording point through the convolution layer; 3) Use the attention mechanism to assign attention weights to the obtained matrix features, and then multiply the attention weights by the corresponding matrix features to obtain weighted features; 4) Finally, the weighted feature is used as the input of the ConvLSTM model, and the ConvLSTM is used to train the prediction model, and finally the prediction result is obtained; The SST matrix convolution is to organize the SST data into a digital matrix sequence with a length of |F|, and then use the convolution to extract the features of the matrix to obtain local features; the convolution operation is completed by moving the convolution kernel, and each value of the output matrix is the sum of the product of the value of each 3 × 3 area in the input matrix and the value of the corresponding position of the 3 × 3 convolution kernel; The daily SST data in the area is processed into a W H matrix, the SST sequence F in the area F=F ₁ , F ₂ ,..., F _|F|, |F| represents the time length of the SST sequence, and F _i =W H is the SST of all recording points on the i-th (1≤i≤|F|, i∈Z) day in the area, that is, a W H matrix, and the sequence formed by these matrices is used as the input of the CA-ConvLSTM model; The matrix feature sequence is obtained through the convolution operation. In the prediction, the SST of k days is used to predict the SST of the next day or five days, and the value of k is different; starting from the nth day, the matrix feature sequence of this k day is M=M _n+1 , M _n+2 ,..., M _n+k-1 , M _n+k , M _i is the ith (n≤i≤n+k, i∈Z) matrix feature, that is, the matrix of (W-2) (H-2); The method for automatically learning attention weights of the CA-ConvLSTM model is as follows: First, the mean value of each matrix feature is obtained through the pooling layer, that is, V=[v _n+1 , v _n+2 , ..., v _n+k-1 , v _n+k ], at this time v _i is the average value of all values in the _Mi matrix; the definition of the attention model Φ is as follows: D＝tanh(w*V+b) (1) A＝softmaχ(W’*D+b’) (2) Where w, b are the weight and bias of the fully connected layer, D is the output of the fully connected layer and the input of softmax, W' and b' are the weight and bias of softmax, A is the output of softmax, that is, the attention weight, which is a vector of length k, expressed as A=[a_n+1, a_n+2,...,a_n+k-1, a_n+k,]; After that, the attention weights are multiplied by the corresponding matrix features in turn to obtain a new feature sequence, that is, the weighted feature sequence, expressed as N=N_n+1, N_n+2,...,N_n+k-1, N_n+k, N_iis the i-th weighted feature, N_i=a_i*M_i, is a matrix of (W-2)·(H-2); Before the matrix feature sequence M enters the ConvLSTM model, the matrix feature sequence M is automatically weighted through the attention mechanism to obtain the weighted feature sequence N; The calculation formula of the ConvLSTM is as follows: i _t ＝σ(w _χi *χ _t +w _hi *H _t-1 +wci·Ct-1+bi) (3) f _t ＝σ(w _χf *χ _t +w _hf *H _t-1 +wci·Ct-1+bf) (4) o _t ＝σ(w _χo *χ _t +w _ht *H _t-1 +wct·Ct+bt) (6) H _t ＝o _t ·tanh(C _t ) (7) Where * represents the convolution operation, · represents the corresponding multiplication, and σ is the sigmoid activation function.