Sea surface temperature prediction method

Abstract

The invention relates to a sea surface temperature prediction method, which comprises the following steps of: acquiring sea surface temperature grid data of a target area, and converting the sea surface temperature grid data into a temperature vector time sequence containing relative position information; inputting the temperature vector time sequence into a sea surface temperature prediction model based on causal convolution and an extended long-short-term memory network to obtain a temperature prediction result; the sea surface temperature prediction model comprises a causal convolution layer used for extracting spatial-temporal characteristics of the temperature vector time sequence; and the extended LSTM layer comprises a plurality of xLSTM layers which are mutually stacked and is used for obtaining a temperature prediction result according to the spatial-temporal characteristic analysis. According to the method, the long-term prediction capability can be improved, and the dynamic spatial correlation is enhanced, so that the precision, stability and adaptability of sea surface temperature prediction are improved.

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. Background Art

[0002] Sea surface temperature (SST) is a core parameter in oceanographic and climatological research, profoundly impacting global climate patterns, marine ecosystems, and human activities. Studies have shown that long-term trends in SST contribute significantly to sea level rise through thermal expansion, while its short-term fluctuations are closely associated with extreme weather events such as typhoons. Furthermore, the spatiotemporal evolution of SST anomalies (SSTAnomalies, SSTA) reveals their association with climate phenomena. For example, SSTA fluctuations in the eastern Pacific fishing grounds are significantly affected by the warm and cold phases of ENSO. Therefore, accurately predicting SST and its anomalies not only contributes to understanding the inherent laws of the climate system but also provides a scientific basis for distant-water fishery management, disaster warning, and ecological protection.

[0003] Traditional SST prediction methods primarily rely on numerical models and statistical methods. With the development of deep learning technology, data-driven prediction methods have gradually emerged. By learning from the inherent patterns of the data, they can predict SST changes with high resolution. However, existing deep learning methods still face challenges in simultaneously processing short-term local features and long-term spatiotemporal dependencies. In particular, in regional predictions, they struggle to adaptively learn dynamic spatial correlations that change over time.

[0004] For example, LSTM networks, widely used in time series analysis, perform poorly in regional forecasting due to a lack of modeling of spatial correlations. Furthermore, their inability to model long-term dependencies leads to a significant decrease in forecast accuracy as the forecast time span increases. Furthermore, because LSTMs rely on gating mechanisms (such as input and forget gates) to control the flow and updating of information, these mechanisms lack the speed and sensitivity to rapidly changing short-term patterns. Consequently, they are less able to capture short-term fluctuations and localized patterns in SST data (such as outliers or rapid temperature changes).

[0005] Although graph neural networks (GNNs) can handle spatial dependencies, they rely on static graph structures and cannot adapt to the dynamic spatial correlations in SST data that change over time, resulting in limited prediction accuracy and stability.

[0006] The ConvLSTM model, which combines the features of a convolutional neural network (CNN) and a long-term short-term memory (LSTM), significantly increases the number of parameters and computational cost when processing large-scale gridded data, especially in scenarios with high-resolution SST data (e.g., 40×40 grids). Furthermore, the ConvLSTM model inherits the vanishing gradient problem of LSTM when processing long sequences, which limits the model's ability to effectively utilize early historical information for long-term predictions. Furthermore, the ConvLSTM's convolution operation is based on a fixed kernel and cannot adjust weights based on dynamic changes in SST data. Therefore, it cannot adaptively learn spatial correlations that change over time, such as changes in inter-regional dependencies caused by shifting ocean currents. The convolution operation may also introduce future information, thereby disrupting causality and reducing prediction accuracy. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a sea surface temperature prediction method that can improve long-term prediction capabilities and enhance dynamic spatial correlation, thereby improving the accuracy, stability and adaptability of sea surface temperature prediction.

[0008] The present invention solves the technical problem by providing a method for predicting sea surface temperature, comprising the following steps:

[0009] Obtain the sea surface temperature grid data of the target area within the set time period and convert it into a temperature vector time series containing relative position information;

[0010] The temperature vector time series is input into a sea surface temperature prediction model based on causal convolution and extended long short-term memory network to obtain a temperature prediction result; the sea surface temperature prediction model includes:

[0011] A causal convolutional layer, for extracting spatiotemporal features of the temperature vector time series;

[0012] The extended LSTM layer includes a plurality of xLSTM layers stacked on each other, and is used to obtain a temperature prediction result based on the spatiotemporal feature analysis.

[0013] Furthermore, the plurality of xLSTM layers stacked on each other include an mLSTM layer, an sLSTM layer and an mLSTM layer connected in sequence.

[0014] Furthermore, the sea surface temperature grid data includes a matrix with rows and columns divided by longitude and latitude and with temperature as an element.

[0015] Furthermore, the temperature vector time series is obtained by data expansion in one of the following ways: longitude priority, latitude priority, or serpentine expansion according to the marine meteorological characteristics of the target area.

[0016] Furthermore, the causal convolutional layer is set to cover 4 consecutive time steps.

[0017] Furthermore, the causal convolutional layer adopts SiLU function as the activation function.

[0018] Furthermore, the method further includes the step of normalizing the temperature vector time series.

[0019] Furthermore, when training the sea surface temperature prediction model, the step of dividing the normalized temperature vector time series into an input sequence for training and a target sequence for verification is also included, using a set time length as a window.

[0020] Beneficial effects

[0021] Due to the adoption of the above-mentioned technical solution, the present invention has the following advantages and positive effects compared with the prior art: the present invention introduces an extended long short-term memory network, which enhances the modeling ability of long-term dependencies through exponential gating and optimized memory unit design, ensuring high accuracy in SST prediction over long time spans and overcoming the limitations of traditional LSTM; in addition, by processing the sea surface temperature grid data into a temperature vector time series that meets the input requirements of the causal convolution layer and retains relative position information, causal convolution is used to extract short-term local spatial features in the SST grid data, thereby strengthening the compliance with the causal requirements of the prediction task. Combined with xLSTM to process time dependencies, the spatiotemporal correlation of SST can be effectively modeled, so that the model can adapt to complex changes without relying on fixed spatial assumptions, thus making up for the defects of the prior art in dynamic spatial feature extraction; at the same time, by setting the SiLU function as the activation function of the causal convolution layer, not only can the nonlinear expression ability of the prediction model be enhanced, but its adaptive gating mechanism can also be used to dynamically learn spatial correlations that change over time, thereby improving the accuracy, stability and adaptability of sea surface temperature prediction without increasing the complexity of the model and the amount of computation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of an embodiment of the present invention;

[0023] Figure 2 Schematic diagram of the sea surface temperature prediction model structure according to an embodiment of the present invention;

[0024] Figure 3 This is a flow chart of the sea surface temperature prediction model training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0026] The embodiment of the present invention relates to a method for predicting sea surface temperature, such as Figure 1 As shown, the following steps are included:

[0027] Obtain sea surface temperature grid data of the target area and convert it into a temperature vector time series containing relative position information;

[0028] The temperature vector time series is input into the sea surface temperature prediction model based on causal convolution and extended long short-term memory network to obtain the temperature prediction results. The sea surface temperature prediction model includes:

[0029] Causal convolutional layer, used to extract the first feature of the temperature vector time series;

[0030] The extended LSTM layer includes several xLSTM layers stacked on each other, and is used to obtain a temperature prediction result based on the first feature analysis.

[0031] The original sea surface temperature (SST) grid data is represented as (time, longitude, latitude). Because the causal convolution layer requires one-dimensional input data, the data needs to be flattened into (time, longitude * latitude). The temperature vector obtained after flattening must retain the original relative position information to ensure that the sea surface temperature prediction model can capture the accurate spatiotemporal characteristics of the input temperature vector. In some preferred methods, the appropriate flattening method can be selected based on the marine meteorological characteristics of the target area. For example, for the target area near the equator, the temperature gradient changes more significantly along the latitude direction (i.e., the east-west direction), so it can be expanded in a longitude-first manner, that is, all the longitude points of each latitude are arranged in sequence, and then connected end to end in sequence according to the latitude; for some mid-latitude areas where the temperature, air pressure and ocean currents change drastically, it can be expanded in a latitude-first manner, that is, all the latitude points of each longitude are arranged in sequence, and then connected end to end in sequence according to the longitude; it can also be expanded in a serpentine manner, that is, all the longitude points of each latitude are arranged in sequence, and then connected end to end and end to end in sequence according to the latitude, so that the edge longitude points of the upper and lower latitudes are closer, reducing the jump distance between adjacent latitudes, and more conducive to capturing local features across latitudes.

[0032] The flattened data needs to be preprocessed before being input into the prediction model, including:

[0033] Normalization: The Min-Max Normalization method is used to scale the SST data to the interval [0,1].

[0034] Time series construction: Use the sliding window method to divide the historical SST data into input series and target series. For example, with a window length of 30 days, generate the input series (30 days of historical data) and the target series (forecast data for the next 7 days).

[0035] The sea surface temperature prediction model of this embodiment is constructed based on causal convolution and extended long short-term memory network, such as Figure 2 As shown, it includes the following components:

[0036] Causal Convolution Layer: extracts short-term local features from SST data;

[0037] xLSTM layer (extended long short-term memory network layer): models the long-term dependencies of SST data.

[0038] The input data first passes through the causal convolution layer to extract local patterns in the time dimension. In the causal convolution, padding is used on the left, so the output shape remains unchanged and can be directly input into the xLSTM layer to capture long-term trends and spatial correlations, and then output the prediction results.

[0039] Among them, the causal convolution layer can ensure that the model only uses current and historical data to avoid future information leakage. Its technical implementation is as follows: Causal convolution performs zero padding on the left side (past direction) of the time series to ensure that when the convolution kernel calculates the output of the current time step t, it only covers the data from the start time to t, and does not touch the information from t+1 and later. The calculation formula of causal convolution is:

[0040]

[0041] Where: y t is the output of causal convolution, k is the convolution kernel size, ω i is the weight of the convolution kernel, x t-1 is the input of time step ti, when ti<1, x t-1 Filled with 0.

[0042] Convolution kernel size: set to 4, indicating a short-term window covering 4 time steps.

[0043] Activation function: SiLU (Sigmoid Linear Unit) function is used, which enhances the nonlinear expression ability of the model. The calculation formula of SiLU activation function is:

[0044] SiLU(x)=x·σ(x) (2)

[0045] Here, σ(x) is the Sigmoid function. Simply put, the SiLU function multiplies the input x by the output of the Sigmoid function, forming a nonlinear transformation. The advantages of the SiLU function for prediction tasks lie in its nonlinear expressiveness, smoothness, sparsity, and ability to mitigate vanishing gradients. It also avoids the "neuron death" issue of ReLU and improves model flexibility and generalization through an adaptive gating mechanism.

[0046] The xLSTM layer enhances long-term dependency modeling capabilities by combining sLSTM and mLSTM blocks. It employs an "msm" stacking structure, consisting of an mLSTM layer, an sLSTM layer, and an mLSTM layer. The mLSTM layer, through its matrix memory and multi-head mechanism, effectively captures complex features and multi-dimensional information in the input data. The sLSTM layer excels at capturing long-term trends, utilizing an exponential gating mechanism to enhance its ability to model long-term dependencies. This stacking approach ensures that both the input and output of the xLSTM model undergo multi-dimensional feature extraction. The middle sLSTM layer, through its stable memory mechanism, ensures that the model effectively utilizes long-range historical information.

[0047] The specific implementation of each module is as follows:

[0048] sLSTM blocks (scalar long short-term memory blocks): By retaining scalar memory and scalar updates, combined with memory mixing and a multi-head mechanism, the model's ability to model long-term dependencies is enhanced. Memory mixing is implemented through recurrent connections of hidden states and is suitable for tasks requiring state tracking, such as code evaluation or entity tracking in long narratives. The multi-head mechanism enables the model to capture sequence features from multiple perspectives, improving expressiveness. Memory mixing is performed within each head, while the heads remain independent and non-mixed. Similar to the multi-head attention in the Transformer, this design enables the model to capture sequence features from multiple perspectives while avoiding interference between heads.

[0049] The forward propagation of sLSTM is shown in formulas (3) to (12).

[0050] c t =f t c t-1 +i t z t , cell state (3)

[0051] n t =f t n t-1 +i t, normalized state (4)

[0052] h t =o t c t / n t , hidden state (5)

[0053]

[0054]

[0055] where c t is the cell state at time step t. State n t Used for normalization. It preserves the long-term memory of the network, f t It is the forget gate, i t is the input gate, z t Control the input amount and the previous hidden state h t-1 Added to the cell state. Matrix ω t 、ω i 、ω f and ω o are the input weights that map the input weights to the unit input, input gate, forget gate, and output gate, respectively.

[0056] sLSTM updates the activation function to an exponential function. The exponential function grows faster than the sigmoid function and is more sensitive to input changes. Therefore, it can quickly and effectively adjust the output of the input and forget gates, allowing the model to more quickly capture changes in input information and selectively remember or forget information, thereby improving the model's memory and forgetting capabilities. t , to prevent the overflow caused by the exponential activation function from causing the value to be too large.

[0057] m t =max(log(f t )+m t-1 ,log(i t ), steady state (10)

[0058] i t ′=exp(log(i t )-m t ), stable. Input gate (11)

[0059] f t ′=exp(log(f t )+m t-1 -m t ). Stable. Forget gate (12)

[0060] mLSTM Block (Matrix Long Short-Term Memory Block): To increase the storage capacity of LSTM, the most significant difference between the Matrix Long Short-Term Memory (mLSTM) model and the previous sLSTM is that the state and weight parameters are now in matrix form. The forward propagation of the mLSTM is shown in Equations (13) to (2-26).

[0061]

[0062] n t =f t n t-1 +i t k t , normalized state (14)

[0063]

[0064] q t =W q x t +b q , query vector (16)

[0065]

[0066] v t =W v x t +b v , value vector (17)

[0067]

[0068] Among them C t is a matrix memory, when it is the input gate, f t It's the forget gate, t is the output gate.

[0069] The model training process is as follows Figure 3 As shown in the figure, the mean squared error (MSE) is used as the loss function during training, the Adam optimizer (adaptive moment estimation optimizer) is used, the initial learning rate is set to 0.0001, the batch size is set to 512 samples, and the number of training rounds is 200 rounds.

[0070] The technical effects of this embodiment are further illustrated below through specific data comparison.

[0071] Based on the OISST V2 high-resolution dataset (spatial resolution 0.25°×0.25°, temporal resolution daily, covering 1981-2024) provided by the National Oceanic and Atmospheric Administration (NOAA), the RMSE, MAE, and R of the proposed method (CausalConv-xLSTM) and the existing methods (LSTM and ConvLSTM) on different prediction days are compared. 2 The performance is shown in Tables 1 to 3.

[0072] As can be seen from the following table, this implementation is technically superior to existing technologies. Causal convolution not only improves the accuracy of short-term feature extraction, but also ensures causal consistency; xLSTM enhances long-term prediction capabilities and dynamic correlation modeling, overcoming the limitations of LSTM and ConvLSTM. Experiments show that this implementation has higher RMSE, MAE, and R in 1-, 3-, 5-, and 7-day forecasts. 2 Both are lower than LSTM and ConvLSTM, proving its superiority in SST prediction.

[0073] Table 1 RMSE (℃) statistics of the CausalConv-xLSTM model and different models

[0074] Model 1 day 3 days 5 days 7 days LSTM 0.40 0.43 0.46 0.48 ConvLSTM 0.24 0.30 0.35 0.41 CausalConv-xLSTM 0.14 0.21 0.26 0.31

[0075] Table 2 MAE (℃) statistics of CausalConv-xLSTM model and different models

[0076] Model 1 day 3 days 5 days 7 days LSTM 0.29 0.32 0.34 0.35 ConvLSTM 0.18 0.22 0.26 0.29 CausalConv-xLSTM 0.10 0.15 0.18 0.21

[0077] Table 3 R of CausalConv-xLSTM model and different models 2 statistics

[0078] Model 1 day 3 days 5 days 7 days LSTM 0.898 0.883 0.869 0.856 ConvLSTM 0.963 0.943 0.922 0.896 CausalConv-xLSTM 0.986 0.970 0.954 0.938

[0079] From the above data, it is clear that this implementation outperforms existing technologies in terms of high-precision prediction, robustness, dynamic spatial correlation learning, and causal consistency. These advantages are due to the integration of causal convolution and xLSTM (extended long short-term memory) technology in this implementation, which addresses the shortcomings of traditional methods such as LSTM and ConvLSTM in long-term dependency modeling and short-term feature extraction. This is specifically reflected in the following two aspects:

[0080] 1. High-precision prediction, especially in long-term prediction. The present invention achieves low root mean square error (RMSE) in 1-day, 3-day, 5-day and 7-day predictions. For example, the RMSE for 7-day predictions is only 0.31°C, which is much lower than the 0.48°C of traditional LSTM and 0.41°C of ConvLSTM. This advantage stems from the multi-head mechanism and optimized memory units of xLSTM, which enhance the model's ability to model long-term trends and overcome the problem of insufficient capture of long-term dependencies caused by the vanishing gradient of traditional LSTM. At the same time, causal convolution accurately extracts short-term local features through the padding mechanism and convolution kernel design, further improving the prediction accuracy.

[0081] 2. Adaptive learning of dynamic spatial correlations. Unlike existing techniques that require a predefined static graph structure, this method can adaptively capture dynamic spatial patterns. This feature is achieved by combining causal convolution with xLSTM: causal convolution extracts short-term local features, while xLSTM dynamically models spatial correlations through a multi-head mechanism, allowing the model to adapt to complex changes without relying on fixed spatial assumptions.

[0082] 3. Ensure causal consistency of predictions. This method ensures that the model relies solely on historical data and does not introduce future information, meeting the causal requirements of the prediction task. This advantage is achieved through the padding mechanism of causal convolution, which restricts the convolution kernel to access only current and historical data, avoiding the information leakage issues that may occur in traditional methods.

[0083] 4. Short-term forecasting also performed well. For example, the RMSE for a one-day forecast was 0.14°C, surpassing both the LSTM (0.40°C) and the ConvLSTM (0.24°C). This advantage stems from the causal convolution algorithm's ability to accurately extract rapidly changing local features, overcoming the traditional LSTM's inability to capture short-term dynamic patterns. Combined with the optimization of the xLSTM, this further improved overall performance.

[0084] It can be seen that the present invention optimizes short-term feature extraction and long-term dependency modeling by fusing causal convolution and xLSTM, respectively. It not only surpasses the existing technology in accuracy and robustness, but also demonstrates unique advantages in adaptability and causality, providing a more efficient and accurate solution for SST prediction.

Claims

1. A sea surface temperature prediction method, characterized in that: The following steps are involved: Obtain the sea surface temperature grid data of the target area within the set time period and convert it into a temperature vector time series containing relative position information; Inputting the temperature vector time series into a sea surface temperature prediction model based on causal convolution and extended long short-term memory network to obtain a temperature prediction result; The sea surface temperature prediction model includes: A causal convolutional layer, for extracting spatiotemporal features of the temperature vector time series; The extended LSTM layer includes a plurality of xLSTM layers stacked on each other, and is used to obtain a temperature prediction result based on the spatiotemporal feature analysis.

2. The method according to claim 1, characterized in that The multiple xLSTM layers stacked on each other include an mLSTM layer, an sLSTM layer and an mLSTM layer connected in sequence.

3. The method according to claim 1, characterized in that The sea surface temperature grid data includes a matrix with rows and columns divided by longitude and latitude and with temperature as an element.

4. The method according to claim 1, wherein The temperature vector time series is obtained by performing data expansion in one of the following ways: longitude priority, latitude priority, or serpentine expansion according to the marine meteorological characteristics of the target area.

5. The method according to claim 1, wherein The causal convolutional layer is set to cover 4 consecutive time steps.

6. The method according to claim 1, characterized in that The causal convolutional layer uses the SiLU function as the activation function.

7. The method according to claim 1, characterized in that The method also includes a step of normalizing the temperature vector time series.

8. The method according to claim 7, characterized in that When training the sea surface temperature prediction model, the method further includes dividing the normalized temperature vector time series into an input sequence for training and a target sequence for verification using a set time length as a window.

Images