A seawater temperature prediction method based on deep learning

Abstract

The present invention provides a seawater temperature prediction method based on deep learning. The method selects and calculates the correlation between each environmental parameter and the seawater surface temperature value, screens out characteristic parameters that affect the seawater surface temperature value according to the correlation, and uses them as input layer neurons of a seawater temperature prediction neural network model; corresponds the characteristic parameter values ​​in a historical data set with the seawater surface temperature value at the next moment, and constructs a training set and a test set; after training the neural network model to obtain network parameters, the characteristic parameters at the current moment are input to obtain the seawater surface temperature value at the next moment; the method improves the accuracy of the prediction results by objectively selecting characteristic parameters and optimizing the convolutional neural network structure.

Description

Technical Field

[0001] The present invention relates to the technical field of seawater temperature prediction, and in particular to a seawater temperature prediction method based on deep learning. Background Art

[0002] Seawater temperature prediction is a crucial task in marine environmental monitoring and marine climate research. Accurately predicting sea surface temperature is crucial for marine ecosystem management, marine resource development, and marine disaster prevention. Traditional seawater temperature prediction methods are primarily based on statistical models, relying on mathematical methods and empirical formulas. However, these methods often fail to capture complex nonlinear relationships and spatiotemporal dynamics, resulting in low prediction accuracy.

[0003] Deep learning is a machine learning method that uses multi-layer neural networks for pattern recognition and prediction. In seawater temperature prediction, deep learning can achieve accurate predictions by learning from a large amount of historical data, extracting features, and establishing complex nonlinear relationship models. However, there are still some problems with the current seawater temperature prediction methods based on deep learning. Traditional deep learning methods often ignore the correlation between seawater temperature and environmental parameters, resulting in inaccurate prediction results. The quality and quantity of training data will also affect the prediction results. In the published patent, CN110648030A provides a method for predicting seawater temperature. Based on a trained neural network model, the temperature prediction value of the next moment of the last moment in the time series is obtained according to the current seawater temperature and salinity data. This method artificially selects a single feature parameter and only considers the impact of seawater salt concentration data on temperature. It does not truly reflect the comprehensive impact of environmental factors on temperature, making the prediction results very random. In addition, the model structure is simple, and the intrinsic relationship of the data cannot be fully reflected by the model.

[0004] Therefore, it is necessary to propose a seawater temperature prediction method based on deep learning, which can accurately capture the correlation between seawater temperature and environmental parameters, and effectively obtain predicted values ​​based on a reasonable model structure to improve the accuracy and reliability of model prediction. Summary of the Invention

[0005] The present invention provides a seawater temperature prediction method based on deep learning, which adopts the following technical solutions:

[0006] A method for predicting seawater temperature based on deep learning, the method comprising the following steps:

[0007] Step 1: Select environmental parameters that affect the sea surface temperature value, and calculate the correlation between each environmental parameter and the sea surface temperature value based on historical observation values ​​of the environmental parameters.

[0008] Step 2: Based on the correlation, characteristic parameters that affect the sea surface temperature value are screened out, and the characteristic parameters are used as input layer neurons of the sea surface temperature prediction neural network model.

[0009] Step 3: Match the characteristic parameter values ​​in the historical data set with the sea surface temperature values ​​at the next moment to construct a training set and a test set.

[0010] Step 4: Input the training set and test set into the neural network model to obtain the network parameters. According to the trained model, input the characteristic parameters at the current moment to obtain the sea surface temperature value at the next moment.

[0011] Furthermore, calculating the correlation between each environmental parameter and the sea surface temperature value based on the historical observation value of the environmental parameter includes:

[0012] Calculate the correlation between environmental parameters and sea surface temperature. , sea surface temperature , the discrete distribution vector between the two at time n is expressed as , n+1 time is expressed as , then the correlation between the environmental parameters and the sea surface temperature value is The calculation formula is:

[0013] ;

[0014] Where, is the probability distribution function; let the environmental parameters and sea surface temperature The joint distribution function of , the marginal distribution function is expressed as , then there exists a function So that: .

[0015] At this time, the probability distribution function Use function Expressed as:

[0016] ;

[0017] Correlation between environmental parameters and sea surface temperature values Through the function Expressed as:

[0018] ;

[0019] The value range is [-1,1]. When it is a negative number, there is a negative correlation between the environmental parameters and the sea surface temperature, and the closer it is to -1, the greater the negative correlation. When it is a positive number, there is a positive correlation between the environmental parameters and the sea surface temperature. The closer it is to 1, the greater the positive correlation.

[0020] Furthermore, based on the correlation, characteristic parameters that affect the sea surface temperature are screened out, specifically including: assuming that the correlation sequence between each environmental parameter and the sea surface temperature is: , is the number of environmental parameters, For the The correlation between the environmental parameters and the sea surface temperature, ;

[0021] like , then this environmental parameter is listed as a characteristic parameter.

[0022] Furthermore, the neural network model is composed of an input layer, a one-dimensional convolutional layer, a pooling layer, a bidirectional gated recurrent layer, a Dropout layer, a fully connected layer and an output layer connected in sequence.

[0023] Furthermore, the one-dimensional convolutional layer performs feature extraction on the input sample data.

[0024] Furthermore, the pooling layer simplifies the data dimension after the one-dimensional convolution and passes the data into the bidirectional gated recurrent layer.

[0025] Furthermore, the bidirectional gated recurrent layer utilizes a bidirectional GRU to extract sequence features of feature parameters and performs a nonlinear transformation on the sea surface temperature data to determine the intrinsic logical relationship between input and output data.

[0026] Further, through After the secondary feature extraction is performed on the data in the layer, some nodes of the data are randomly discarded in the Dropout layer to prevent overfitting and the processed data are passed to the fully connected layer. The fully connected layer maps the output features to the label space of the sample through nonlinear combination, and the output layer performs linear regression operation to output the predicted result of the sea surface temperature value.

[0027] Furthermore, the neural network model is obtained by adding a bidirectional gated recurrent layer and a Dropout layer on the basis of a convolutional neural network;

[0028] In the bidirectional gated recurrent layer, it is assumed that time, use represents the state of the input layer, represents the output layer state, represents the hidden layer state during the bidirectional GRU layer transmission process, then:

[0029] ;

[0030] ;

[0031] ;

[0032] Where, Represents the feature state learned by the forward layer during the one-way transmission process, It is the feature state learned by the backward layer during the one-way transmission process; and It is the weight of the forward layer, the backward layer and the output layer in sequence; It represents the bias vector added to the output layer.

[0033] Furthermore, the prediction results output by the neural network model are compared with the measured data in the test set, and the goodness of fit index is used. To describe the predictive performance of the model, ,when When , the model training stop condition is met. At this time, the model parameters are the optimal parameters of the neural network model structure. If they are not met, the back propagation algorithm is used to adjust the model deviation and continue the cycle.

[0034] Goodness of fit index The calculation formula is:

[0035] ;

[0036] Where, Representative The actual value of the sea surface temperature at the next moment corresponding to the sample, For the The predicted value of the sea surface temperature at the next moment corresponding to the samples, is the average sea surface temperature, is the total number of samples.

[0037] Furthermore, before step three, the process also includes: preprocessing the historical data set, and filling in the missing data in the training set and the test set by using a linear interpolation function.

[0038] Use a straight line to connect the parameter coordinates of adjacent moments to form a linear interpolation function, denoted as If the characteristic value and sea surface temperature value corresponding to time n in the historical data set are missing, the vertical coordinates of the parameters at time n-1 in the training set and the test set will be and the horizontal parameter coordinates at time n Missing.

[0039] The coordinate values ​​of the training set and the test set at time n-2 are , the coordinate value at time n+1 is , then the linear interpolation function is expressed as:

[0040] ;

[0041] The horizontal coordinate parameters of the training set and the test set at time n-1 and the vertical coordinate parameters at time n Substitute the calculation formula respectively to obtain the coordinate value corresponding to time n in the missing historical data set .

[0042] Furthermore, if the horizontal coordinate values ​​at time n-2 in the training set and the test set are and the horizontal coordinate value at time n+1 If they are equal, then the eigenvalue corresponding to time n in the historical data set is

[0043] ;

[0044] The sea surface temperature value corresponding to time n in the historical data set

[0045] ;

[0046] Furthermore, the characteristic parameters include at least: seawater salinity, carbon dioxide flux at the seawater-air interface, air temperature, and seawater flow rate.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The neural network model established by the present invention can maintain the prediction accuracy of sea surface temperature above 98.5% through continuous combination optimization. This method does not rely on any empirical formulas or assumptions about calculation premises. The obtained prediction results are closer to the actual values ​​and can play a good guiding role in the prediction of sea surface temperature.

[0049] 2. The present invention selects characteristic parameters by calculating the correlation between environmental parameters and sea surface temperature values, avoiding the uncertainty caused by human subjective factors in selecting characteristic parameters. When filling missing data, a method combining linear interpolation function with average value is adopted to improve the data fitting degree. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a seawater temperature prediction method based on deep learning of the present invention.

[0051] Figure 2Schematic diagram of the neural network structure of the present invention. DETAILED DESCRIPTION

[0052] The embodiments of the present invention will be described in detail below with reference to examples. However, those skilled in the art will understand that the following examples are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention.

[0053] like Figure 1 FIG. 1 is a flow chart of a method for predicting seawater temperature based on deep learning according to the present invention, and the method includes the following steps:

[0054] Step 1: Select environmental parameters that affect the sea surface temperature value, and calculate the correlation between each environmental parameter and the sea surface temperature value based on historical observation values ​​of the environmental parameters.

[0055] It should be noted that there are many environmental parameters that affect the sea surface temperature value, including: the geographical coordinates of the sample point, ocean currents, seasons, seawater salinity, seawater emissivity, seawater-air interface carbon dioxide flux, seawater oxygen content, air temperature, air humidity, and seawater flow rate, etc. When determining the correlation between a certain environmental parameter and the sea surface temperature value, the correlation between the two features can be characterized by associating the joint distribution function of the environmental parameter and the sea surface temperature value with their corresponding marginal distribution function. The correlation can more accurately characterize the asymmetric and nonlinear relationship between the features and calculate the correlation between the environmental parameter and the sea surface temperature value.

[0056] Specifically, for environmental parameters , sea surface temperature , the discrete distribution vector between the two at time n is expressed as , n+1 time is expressed as , then the correlation between the environmental parameters and the sea surface temperature value is The calculation formula is:

[0057] ;

[0058] Where, is the probability distribution function; let the environmental parameters and sea surface temperature The joint distribution function of , the marginal distribution function is expressed as , then there exists a function So that: .

[0059] At this time, the probability distribution function Use function Expressed as:

[0060] ;

[0061] Correlation between environmental parameters and sea surface temperature values Through the function Expressed as:

[0062] ;

[0063] It should be noted that The value range is [-1,1]. When it is a negative number, there is a negative correlation between the environmental parameters and the sea surface temperature, and the closer it is to -1, the greater the negative correlation. When it is a positive number, there is a positive correlation between the environmental parameters and the sea surface temperature. The closer it is to 1, the greater the positive correlation.

[0064] Step 2: Based on the correlation, characteristic parameters that affect the sea surface temperature value are screened out, and the characteristic parameters are used as input layer neurons of the sea surface temperature prediction neural network model.

[0065] It should be noted that, based on the correlation, the characteristic parameters that affect the sea surface temperature are screened out, specifically including: assuming that the correlation sequence between each environmental parameter and the sea surface temperature is: , is the number of environmental parameters, For the The correlation between the environmental parameters and the sea surface temperature, ;like , then this environmental parameter is listed as a characteristic parameter.

[0066] In addition, it should be noted that before performing step three, it also includes: preprocessing the historical data set and filling the missing data in the training set and the test set through a linear interpolation function.

[0067] The specific method is: use a straight line to connect the parameter coordinates of adjacent moments to form a linear interpolation function, which is recorded as If the characteristic value and sea surface temperature value corresponding to time n in the historical data set are missing, the vertical coordinates of the parameters at time n-1 in the training set and the test set will be and the horizontal parameter coordinates at time n Missing.

[0068] The coordinate values ​​of the training set and the test set at time n-2 are , the coordinate value at time n+1 is , then the linear interpolation function is expressed as:

[0069] ;

[0070] The horizontal coordinate parameters of the training set and the test set at time n-1 and the vertical coordinate parameters at time n Substitute the calculation formula respectively to obtain the coordinate value corresponding to time n in the missing historical data set .

[0071] If the horizontal coordinate value of the training set and the test set at time n-2 is and the horizontal coordinate value at time n+1 If they are equal, then the eigenvalue corresponding to time n in the historical data set is

[0072] ;

[0073] The sea surface temperature value corresponding to time n in the historical data set

[0074] ;

[0075] Step 3: Match the characteristic parameter values ​​in the historical data set with the sea surface temperature values ​​at the next moment to construct a training set and a test set.

[0076] The present invention adopts the method in step 2, and the characteristic parameters screened are: seawater salinity, seawater-air interface carbon dioxide flux, air temperature, and seawater flow rate.

[0077] Step 4: Input the training set and test set into the neural network model to obtain the network parameters. According to the trained model, input the characteristic parameters at the current moment to obtain the sea surface temperature value at the next moment.

[0078] It should be noted that there are many neural network models involved in the field of deep learning, and one of the research hotspots is convolutional neural networks. Compared with traditional neural networks, CNN has a stronger ability to extract features from large amounts of data. Its basic structure is divided into five layers: input layer, convolution layer, pooling layer, fully connected layer and output layer.

[0079] The main function of the input layer of the convolutional neural network is to receive preprocessed data and transmit the data to the neural network. It can receive three-dimensional, two-dimensional data or one-dimensional time series data.

[0080] For all categories of data, CNN uses different levels for processing; the function of the convolution layer is to locally weight the input information to achieve the effect of information integration; due to the selection of different information weight combinations, the output results can reveal various characteristics of different information combinations. The convolution operation process is to use the receptive field of the convolution kernel to periodically slide the input data matrix and perform related mathematical operations to obtain a new feature map; the pooling layer is used to filter the feature data output by the convolution layer, reduce the feature size, and increase the operation speed; the pooling layer generates new values ​​by applying certain algorithms to represent a certain area in the feature map; maximum pooling and mean pooling are the two most commonly used pooling methods.

[0081] The bidirectional gated recurrent neural network is optimized on this basis. The neural network model used in the present invention is composed of an input layer, a one-dimensional convolutional layer, a pooling layer, a bidirectional gated recurrent layer (composed of two GRU layers), a Dropout layer, a fully connected layer, and an output layer connected in sequence. Figure 2 As shown, it is a schematic diagram of the neural network structure of the present invention.

[0082] The number of hidden layer neural units of GRU is set to 4, 8, 16, and 32 respectively. When the number of hidden layer neural units is 16, the fitting index of sea surface temperature is The maximum, as the number of hidden neural units increases, the average error also gradually increases, the fitting index Start to decrease.

[0083] Therefore, in this embodiment, the number of hidden neural units in the GRU network is selected as 16, and its model parameters are shown in the following table.

[0084] Model parameters Value Number of convolutional layers 1 Number of pooling layers 1 The number of convolution kernels in the convolution layer 8 Convolution kernel size in the convolution layer 3 Pooling kernel size in the pooling layer 3 Number of neurons in the bidirectional gated recurrent layer 16 Batch size 128 Optimizer Adam Convolutional layer activation function Relu Number of iterations 150 Dropout 0.05

[0085] The one-dimensional convolution layer extracts features from the input sample data. Figure 2 The neurons in the input layer are x1, x2, ...; the pooling layer simplifies the data dimension after one-dimensional convolution and passes the data into the bidirectional gated recurrent layer; the bidirectional gated recurrent layer uses a bidirectional GRU to extract the sequence characteristics of the feature parameters and performs a nonlinear transformation on the sea surface temperature data to determine the inherent logical relationship between the input and output data.

[0086] through After the secondary feature extraction is performed on the data in the layer, some nodes of the data are randomly discarded in the Dropout layer to prevent overfitting and the processed data are passed to the fully connected layer. The fully connected layer maps the output features to the label space of the sample through nonlinear combination, and the output layer performs linear regression operation to output the predicted result of the sea surface temperature value.

[0087] It should be noted that the neural network model is obtained by adding a bidirectional gated loop layer and a Dropout layer on the basis of the convolutional neural network; in the bidirectional gated loop layer, it is assumed that time, use represents the state of the input layer, represents the output layer state, represents the hidden layer state during the bidirectional GRU layer transmission process, then:

[0088] ;

[0089] ;

[0090] ;

[0091] Where, Represents the feature state learned by the forward layer during the one-way transmission process, It is the feature state learned by the backward layer during the one-way transmission process; and It is the weight of the forward layer, the backward layer and the output layer in sequence; It represents the bias vector added to the output layer.

[0092] The prediction results output by the neural network model are compared with the measured data in the test set, and the goodness of fit index is used. To describe the predictive performance of the model, ,when When , the model training stop condition is met. At this time, the model parameters are the optimal parameters of the neural network model structure. If not, the back propagation algorithm is used to adjust the model deviation and continue the cycle.

[0093] Goodness of fit index The calculation formula is:

[0094] ;

[0095] Where, Representative The actual value of the sea surface temperature at the next moment corresponding to the sample, For the The predicted value of the sea surface temperature at the next moment corresponding to the samples, is the average sea surface temperature, is the total number of samples.

[0096] It should be noted that in the bidirectional gated recurrent layer of the neural network, the number of hidden layer neural units will have a certain impact on the network complexity and learning ability of the model. If the number of neural units is too small, the ability of the hidden layer to receive and process logging information may decrease, resulting in poor performance of the neural network or even inability to train; if the number of neural units is too large, the system error may be reduced, but the training time of the network will be extended, and overfitting is prone to occur.

[0097] The present invention sets the number of hidden layer neural units of GRU to 4, 8, 16, and 32 respectively. When the number of hidden layer neural units is 16, the fitting index of sea surface temperature is The maximum, as the number of hidden neural units increases, the average error also gradually increases, the fitting index It starts to decrease, so the number of hidden neural units in the GRU network is selected as 16.

[0098] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A seawater temperature prediction method based on deep learning, characterized in that: The method comprises the following steps: Step 1: Select environmental parameters that affect the sea surface temperature value, and calculate the correlation between each environmental parameter and the sea surface temperature value based on historical observation values ​​of the environmental parameters; For the environmental parameter x and the sea surface temperature y, the discrete distribution vector between them at time n is expressed as (x n ,y n ), n+1 time is expressed as (x n+1 ,y n+1 ), then the correlation τ between the environmental parameters and the sea surface temperature is calculated as: τ=P[(x n -x n+1 )(and n -and n+1 )>0]-P[(x n -x n+1 )(and n -and n+1 )<0] =2P[(x n -x n+1 )(y n -y n+1 )>0]-1; Where P is the probability distribution function; let the joint distribution function of the environmental parameter x and the sea surface temperature y be H(x,y), and the marginal distribution function be F(x,y), then there exists a function C such that: H(x,y) = C(F(x,y)); At this time, the probability distribution function P is expressed by the function C as: The correlation τ between environmental parameters and sea surface temperature is expressed by the function C: Step 2: Based on the correlation, characteristic parameters that affect the sea surface temperature are screened out, and the characteristic parameters are used as input layer neurons of the sea surface temperature prediction neural network model; Assume that the correlation sequence between each environmental parameter and the sea surface temperature is: τ1, τ2,…τ l ,…τ m , m is the number of environmental parameters, τ l is the correlation between the lth environmental parameter and the sea surface temperature, 1≤l≤m; If |τ l |≥0.5, then this environmental parameter is listed as a characteristic parameter; the characteristic parameters include at least: seawater salinity, seawater-air interface carbon dioxide flux, air temperature, and seawater flow rate; Step 3: Correspond the characteristic parameter values ​​in the historical data set with the sea surface temperature value at the next moment to construct the training set and test set; Step 4: Input the training set and test set into the neural network model to obtain the network parameters. According to the trained model, input the characteristic parameters at the current moment to obtain the sea surface temperature value at the next moment.

2. The deep learning seawater temperature prediction method according to claim 1, characterized in that: The correlation τ between environmental parameters and sea surface temperature values ​​ranges from [-1,1]; When the correlation τ is negative, there is a negative correlation between the environmental parameters and the sea surface temperature, and the closer it is to -1, the greater the negative correlation; When the correlation τ is positive, there is a positive correlation between the environmental parameters and the sea surface temperature. The closer it is to 1, the greater the positive correlation.

3. The deep learning seawater temperature prediction method according to claim 2, characterized in that: The neural network model consists of an input layer, a one-dimensional convolutional layer, a pooling layer, a bidirectional gated recurrent layer, a Dropout layer, a fully connected layer, and an output layer connected in sequence; The one-dimensional convolution layer performs feature extraction on the input sample data; The pooling layer simplifies the data dimension after one-dimensional convolution and passes the data into the bidirectional gated recurrent layer; The bidirectional gated recurrent layer uses a bidirectional GRU to extract sequence features of feature parameters and performs a nonlinear transformation on the sea surface temperature data to determine the inherent logical relationship between input and output data; After the BiGRU layer performs secondary feature extraction on the data, some nodes of the data are randomly discarded in the Dropout layer to prevent overfitting and the processed data is passed to the fully connected layer. The fully connected layer maps the output features to the label space of the sample through nonlinear combination, and the output layer performs linear regression operation to output the predicted result of the sea surface temperature value.

4. The method for predicting seawater temperature based on deep learning according to claim 3, characterized in that: The neural network model is obtained by adding a bidirectional gated recurrent layer and a Dropout layer on the basis of a convolutional neural network; In the bidirectional gated recurrent layer, it is assumed that at the time t of training, x t Represents the state of the input layer, o t represents the output layer state, h t Represents the hidden layer state during the bidirectional GRU layer transmission process, then: Where, Represents the feature state learned by the forward layer during the one-way transmission process, It is the feature state learned by the backward layer during the one-way transmission process; and is the weight of the forward layer, the backward layer and the output layer in sequence; b o It represents the bias vector added to the output layer.

5. The method for predicting seawater temperature based on deep learning according to claim 4, characterized in that: Compare the prediction results output by the neural network model with the measured data in the test set, and use the goodness-of-fit index R 2 to describe the prediction performance of the model, where 0 < R ≤ 1. When R 2 ≥ 0.985, the model training stop condition is satisfied. At this time, the parameters of the model are the optimal parameters of the neural network model structure. If not satisfied, use the backpropagation algorithm to adjust the bias of the model and continue to loop; Goodness of fit index R 2 The calculation formula is: Where, f i represents the actual value of the sea surface temperature at the next moment corresponding to the i-th sample, is the predicted value of the sea surface temperature at the next moment corresponding to the i-th sample, is the average sea surface temperature, and m is the total number of samples.

6. The method for predicting seawater temperature based on deep learning according to claim 1, characterized in that: Before step 3, the method further includes: preprocessing the historical data set and filling in the missing data in the training set and the test set using a linear interpolation function; Use a straight line to connect the parameter coordinates of adjacent moments to form a linear interpolation function, recorded as K(x) = y. If the eigenvalue and sea surface temperature value corresponding to moment n in the historical data set are missing, the parameter ordinate y at moment n-1 in the training set and test set will be n and the horizontal parameter coordinate x at time n n missing; Note that the coordinate values ​​of the training set and the test set at time n-2 are (x n-2 ,y n-1 ), the coordinate value at time n+1 is (x n+1 ,y n+2 ), the linear interpolation function is expressed as: The horizontal coordinate parameter x at time n-1 in the training set and the test set n-1 and the ordinate parameter y at time n n+1 Substitute them into the calculation formula respectively to obtain the coordinate value (x n ,y n ).

7. The method for predicting seawater temperature based on deep learning according to claim 6, characterized in that: If the horizontal coordinate value x at time n-2 in the training set and test set n-2 and the horizontal coordinate value x at time n+1 n+1 If they are equal, then the eigenvalue corresponding to time n in the historical data set is The sea surface temperature value corresponding to time n in the historical data set

Images