Sea wave height multi-site joint prediction method based on deep learning model

Abstract

The invention belongs to the technical field of sea wave height prediction, and relates to a sea wave height multi-site joint prediction method based on a deep learning model, which comprises the following steps: acquiring sea wave height observation data of a plurality of sites in the same sea area, and decomposing the data of each site through a seasonal trend decomposition method; constructing a joint model of a fusion graph convolutional neural network and a long short-term memory network for the trend component, and constructing a long short-term memory network model for the season component and the residual component respectively; the extracted spatial-temporal features are transformed through a full connection layer, and prediction sequences of trend components, seasonal components and residual components of multiple time steps in the future are output respectively; and according to the prediction sequences of the trend component, the seasonal component and the residual component in each station, adding to obtain a prediction result of the multi-station sea wave height. According to the method, spatial structure information between observation stations and time evolution characteristics of each station are fully fused, and the accuracy and stability of multi-station sea wave height joint prediction are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of wave height prediction technology, and in particular to a multi-site joint prediction method for wave height based on a deep learning model. Background Technology

[0002] With increasingly scarce land resources, the ocean, with its vast space and abundant resources, is becoming a new focus of global economic development. Countries are continuously increasing their efforts in ocean development, and the marine economy is playing a rising role in their national economies. Wave height, as a crucial parameter reflecting the state of the marine environment, contains rich dynamic information. Accurate wave height prediction not only helps ensure maritime transportation safety and reduce the threat of giant waves to ship navigation, but also provides design and construction references for marine engineering projects such as ports and offshore platforms. In clean energy development, wave height is also a key indicator for assessing wave energy resource potential and formulating development strategies. Furthermore, wave height prediction also has significant application value in coastal disaster prevention and mitigation, and ecological environment monitoring. Therefore, improving the accuracy and efficiency of wave height prediction is of great practical significance for promoting the sustainable development and utilization of marine resources.

[0003] Currently, wave height prediction technology encompasses various methods, including numerical forecasting, machine learning, and deep learning. Traditional numerical forecasting methods are based on hydrodynamic equations, constructing ocean dynamic models and discretizing the sea area into grids to simulate the generation, propagation, and evolution of waves. This method can systematically reflect ocean physical processes and theoretically characterize wave motion, but it requires extremely high accuracy in initial conditions and boundary parameters, has high computational complexity, and demands huge hardware resources. It is also difficult to adapt quickly to complex and ever-changing ocean environments, resulting in poor real-time performance and flexibility.

[0004] With the widespread application of machine learning technology, more and more research is introducing it into the field of marine environmental prediction. Traditional machine learning algorithms, such as support vector machines and random forests, can build predictive models by mining feature relationships in historical data. Under conditions of sufficient data, they can capture nonlinear patterns in wave height changes relatively well and demonstrate a certain generalization ability. However, these methods rely on manually designed features and are difficult to adaptively extract complex spatiotemporal dynamic features from wave data. As a result, in the case of rapidly changing marine environments, the prediction results are easily disturbed, and the accuracy and stability are difficult to guarantee.

[0005] In recent years, the rapid development of deep learning technology has further driven innovation in wave height prediction methods. Convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and their variants, in particular, have demonstrated superior feature extraction and representation capabilities in spatiotemporal sequence modeling, becoming a research hotspot in this field. CNNs, through convolution operations, can automatically extract local spatial features, making them suitable for modeling the spatial distribution patterns of waves. LTM networks, with their gating mechanisms, can effectively capture dependencies in long-term sequences, making them suitable for describing the temporal evolution of wave height. Meanwhile, graph convolutional networks, due to their advantage in handling non-Euclidean structured data, are increasingly being used for multi-site wave height prediction tasks. They can construct topological structures between sites to uncover spatial correlations between different sites. These deep learning models possess powerful automatic feature extraction capabilities, enabling them to autonomously learn the spatiotemporal dynamics of waves from large-scale observational data without the need for manually designed complex features, showing significant advantages in prediction accuracy and efficiency compared to traditional methods.

[0006] While deep learning has demonstrated significant advantages in wave height prediction, existing methods still have several shortcomings in multi-site joint prediction tasks. On the one hand, existing spatial modeling methods suffer from structural limitations when dealing with multi-site observation data. Since observation sites are typically distributed non-Euclideanally, directly applying traditional continuous spatial feature extraction methods makes it difficult to preserve the original spatial topological relationships between sites, affecting the model's ability to model spatial dependencies. On the other hand, most research focuses on single-site wave height prediction, with limited studies on multi-site joint prediction, lacking a unified modeling framework capable of simultaneously modeling spatiotemporal coupling characteristics and adapting to heterogeneous site distributions. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a multi-site joint prediction method for ocean wave height based on a deep learning model. The invention collects time-series data of ocean wave height from multiple observation stations and decomposes the data for each station using a seasonal trend decomposition method to obtain trend components, seasonal components, and residual components. The geographical distances between stations are constructed into an adjacency matrix, which, along with the trend components, is input into a sub-model combining a graph convolutional neural network and a long short-term memory neural network to obtain the predicted value of the trend components. The seasonal components and residual components of each station are then input into the long short-term memory neural network to extract their temporal features and obtain the predicted values ​​of the seasonal components and residual components. Finally, the predicted values ​​of the three components corresponding to each station are summed to obtain the predicted ocean wave height for future times.

[0008] The technical means employed in this invention are as follows:

[0009] A multi-site joint prediction method for ocean wave height based on a deep learning model includes:

[0010] Wave height observation data from multiple stations within the same sea area were acquired, and the data from each station were decomposed using a seasonal trend decomposition method.

[0011] A joint model of fusion graph convolutional neural network and long short-term memory network is constructed for the trend component, and long short-term memory network models are constructed separately for the seasonal component and the residual component.

[0012] The extracted spatiotemporal features are transformed by a fully connected layer, and the predicted sequences of trend components, seasonal components and residual components for multiple future time steps are output respectively.

[0013] The predicted wave heights for multiple stations are obtained by summing the predicted sequences of trend, seasonal, and residual components from each station.

[0014] Furthermore, the acquisition of wave height observation data from multiple stations within the same sea area, and the decomposition of the data from each station using a seasonal trend decomposition method, specifically includes:

[0015] Missing and null values ​​in wave height observation data from multiple stations within the same sea area are removed through data preprocessing. The geographical distance between each pair of stations is calculated, and the reciprocal of the distance is used to construct an adjacency matrix A.

[0016] For each site's data, a seasonal trend decomposition method is applied, using local weighted regression to decompose the data into trend components, seasonal components, and residual components. The trend component represents the long-term trend, the seasonal component represents periodic fluctuations, and the residual component represents random disturbances or anomalous fluctuations. The seasonal trend decomposition method is expressed as follows:

[0017] X i (t)=T i (t)+S i (t)+R i (t), i∈{1,2,...,N}

[0018] Where N represents the number of sites; data X i (t) is decomposed into trend component T i (t), seasonal component S i (t), residual component R i (t); X i (t) represents the wave height observation data value corresponding to the i-th station at the t-th sampling time; T i (t) represents the trend component data value corresponding to the wave height observation data value at the t-th sampling time of the i-th station; S i (t) represents the seasonal component data value corresponding to the wave height observation data value at the t-th sampling time corresponding to the i-th station, Ri (t) represents the residual component data value corresponding to the wave height observation data value at the t-th sampling time of the i-th station.

[0019] Furthermore, the construction of a joint model fusing graph convolutional neural network and long short-term memory network for the trend component specifically includes:

[0020] The trend components corresponding to the sampling time tn to t of each station are constructed into a feature matrix, and together with the adjacency matrix A, they are input into a deep learning model with an L-layer graph convolutional neural network to extract the spatial features of the trend components between stations. The spatial feature vector result obtained after the input data passes through the graph convolutional neural network is expressed by the following formula:

[0021]

[0022] H (0) =P(tn:t-1)

[0023] in, for The degree matrix; W G is the weight matrix; A is the adjacency matrix based on the geographical distance between sites; I is the identity matrix; n is the sliding window length;

[0024] The high-dimensional spatial features output by the multi-layer graph convolutional neural network are mapped to a time series representation of length n through a linear transformation, which is then used for subsequent modeling of the time features by the long short-term memory network.

[0025] B = Linear(H) (L) )

[0026] Among them, H (L) B represents the extracted spatial features; B is the mapped time series with a length of n.

[0027] The time series data with fused spatial features is input into a Long Short-Term Memory (LSTM) network. After passing through input gates, forget gates, and output gates, the dynamic evolution characteristics of the trend components at each station in the time dimension are extracted. The calculation formula for the LTM network is as follows:

[0028] f t =Sigmoid(W f B t +U f h t-1 +b f )

[0029] i t =Sigmoid(W i B t +U i ht-1 +b i )

[0030]

[0031] o t =sigmoid(W o B t +U o h t-1 +b o )

[0032] h t =o t ⊙tanh(c t )

[0033] Where t∈{1,2,...,n} is used to mark the time steps in the sequence; i t f t o t These are the input gate, forget gate, and output gate, respectively, calculated using the Sigmoid activation function. c represents the candidate cell states calculated using the tanh activation function; t The current cell state; W i W f W o W c These are the input weight matrices for the input gate, forget gate, output gate, and candidate cell states, respectively; U i U f U o U c b are the hidden state weight matrices for the input gate, forget gate, output gate, and candidate cell state, respectively; i b f b o b c These are the bias terms for the input gate, forget gate, output gate, and candidate cell state, respectively; ⊙ indicates that elements at the same position are multiplied together.

[0034] The same model, namely the Long Short-Term Memory (LSTM) network model, is constructed for both the seasonal component and the residual component. The seasonal component or residual component corresponding to the sampling time of each station within the range of tn to t is input into the LSM network. Through the input gate, forget gate and output gate in the network, the dynamic evolution characteristics of the seasonal component or residual component of each station in the time dimension are extracted.

[0035] f t =Sigmoid(W f V t +U f h t-1 +b f )

[0036] i t =Sigmoid(W i V t +U i h t-1 +b i )

[0037]

[0038] o t =sigmoid(W o V t +U o h t-1 +b o )

[0039] h t =o t ⊙tanh(c t )

[0040] Among them, V t ∈{S t ,R t}, when V t Take seasonal portion S t When V is obtained, the dynamic evolution characteristics of the seasonal component in the time dimension are obtained. t Take the residual component R t At that time, the dynamic evolution characteristics of the residual components in the time dimension are obtained.

[0041] Furthermore, the step of transforming the extracted spatiotemporal features through a fully connected layer to output predicted sequences of trend components, seasonal components, and residual components for multiple future time steps specifically includes:

[0042] The predicted value of the trend component is:

[0043]

[0044] The predicted value of the seasonal component is:

[0045]

[0046] The predicted value of the residual component is:

[0047]

[0048] Among them, W fc Represents the weights of the fully connected layer; b fc This represents the bias term of the fully connected layer.

[0049] Furthermore, the predicted wave heights at the multiple stations are expressed as follows:

[0050]

[0051] The final prediction of wave height at the multi-site locations is the sum of the predicted values ​​of the trend component, seasonal component, and residual component.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] 1. The multi-site joint prediction method for wave height based on a deep learning model provided by this invention fully exploits spatial correlation features. This invention constructs an adjacency matrix based on geographical distance and combines it with a graph convolutional neural network to model the spatial structure of the trend components of each site. This effectively utilizes the spatial correlation between sites, makes up for the shortcomings of existing single-site prediction methods that cannot capture spatial features, and improves the overall accuracy and stability of multi-site joint prediction.

[0054] 2. The multi-site joint prediction method for ocean wave height based on a deep learning model provided in this invention enhances long-term prediction capabilities. This invention employs a seasonal trend decomposition method to break down the complex ocean wave height time series into trend components, seasonal components, and residual components, modeling each separately. This effectively separates and extracts the long-term evolution trend from the series. Compared to methods that directly model the original series, this invention demonstrates superior prediction accuracy and stronger model stability in medium- and long-term prediction tasks.

[0055] 3. The multi-site joint prediction method for ocean wave height based on a deep learning model provided by this invention supports joint prediction at multiple sites, thus improving efficiency. The deep learning model constructed in this invention can simultaneously receive data input from multiple sites and output the prediction results for all target sites at once, avoiding the redundant calculation process of modeling and predicting each site separately. This significantly improves model inference efficiency and running speed while ensuring prediction accuracy.

[0056] Based on the above reasons, this invention can be widely applied in the field of wave height prediction technology. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart of the multi-site joint prediction method for ocean wave height based on a deep learning model in this invention.

[0059] Figure 2 This is a prediction result diagram of a certain site from May to September, used in an embodiment of the present invention to predict tasks for the next 24 hours.

[0060] Figure 3 This refers to wave height observation data of the station between May and September in this embodiment of the invention. Detailed Implementation

[0061] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0064] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0065] like Figure 1As shown, this invention provides a multi-site joint prediction method for ocean wave height based on a deep learning model, comprising:

[0066] The wave height observation data of multiple stations in the same sea area are obtained, and the data of each station is decomposed by the seasonal trend decomposition method. In a specific implementation, as a preferred embodiment of the present invention, missing values ​​and null values ​​in the wave height observation data of multiple stations in the same sea area are removed by data preprocessing, and the geographical distance between each pair of stations is calculated, and the reciprocal of the distance is used to construct an adjacency matrix A.

[0067] For each site's data, a seasonal trend decomposition method is applied, using local weighted regression to decompose the data into trend components, seasonal components, and residual components. The trend component represents the long-term trend, the seasonal component represents periodic fluctuations, and the residual component represents random disturbances or anomalous fluctuations. The seasonal trend decomposition method is expressed as follows:

[0068] X i (t)=T i (t)+S i (t)+R i (t), i∈{1,2,...,N}

[0069] Where N represents the number of sites; data X i (t) is decomposed into trend component T i (t), seasonal component S i (t), residual component R i (t); X i (t) represents the wave height observation data value corresponding to the i-th station at the t-th sampling time; T i (t) represents the trend component data value corresponding to the wave height observation data value at the t-th sampling time of the i-th station; S i (t) represents the seasonal component data value corresponding to the wave height observation data value at the t-th sampling time corresponding to the i-th station, R i (t) represents the residual component data value corresponding to the wave height observation data value at the t-th sampling time of the i-th station.

[0070] A joint model fusing a graph convolutional neural network and a long short-term memory network is constructed for the trend component, and the same model, the long short-term memory network model, is constructed for the seasonal component and the residual component. In a preferred embodiment, the trend components corresponding to the sampling time tn to t of each station are constructed as feature matrices, and these feature matrices, along with the adjacency matrix A, are input into a deep learning model with an L-layer graph convolutional neural network to extract the spatial features of the trend components between stations. The spatial feature vector obtained after the input data passes through the graph convolutional neural network is represented by the following formula:

[0071]

[0072]

[0073] H (0) =P(tn:t-1)

[0074] in, for The degree matrix; W G is the weight matrix; A is the adjacency matrix based on the geographical distance between sites; I is the identity matrix; n is the sliding window length;

[0075] The high-dimensional spatial features output by the multi-layer graph convolutional neural network are mapped to a time series representation of length n through a linear transformation, which is then used for subsequent modeling of the time features by the long short-term memory network.

[0076] B = Linear(H) (L) )

[0077] Among them, H (L) B represents the extracted spatial features; B is the mapped time series with a length of n.

[0078] The time series data with fused spatial features is input into a Long Short-Term Memory (LSTM) network. After passing through input gates, forget gates, and output gates, the dynamic evolution characteristics of the trend components at each station in the time dimension are extracted. The calculation formula for the LTM network is as follows:

[0079] f t =Sigmoid(W f B t +U f h t-1 +b f )

[0080] i t =Sigmoid(W i B t +U i h t-1 +b i )

[0081]

[0082] o t =sigmoid(W o B t +U o h t-1 +b o )

[0083] h t =ot ⊙tanh(c t )

[0084] Where t∈{1,2,...,n} is used to mark the time steps in the sequence; i t f t o t These are the input gate, forget gate, and output gate, respectively, calculated using the Sigmoid activation function. c represents the candidate cell states calculated using the tanh activation function; t The current cell state; W i W f W o W c These are the input weight matrices for the input gate, forget gate, output gate, and candidate cell states, respectively; U i U f U o U c b are the hidden state weight matrices for the input gate, forget gate, output gate, and candidate cell state, respectively; i b f b o b c These are the bias terms for the input gate, forget gate, output gate, and candidate cell state, respectively; ⊙ indicates that elements at the same position are multiplied together.

[0085] The same model, namely the Long Short-Term Memory (LSTM) network model, is constructed for both the seasonal and residual components. The seasonal or residual components corresponding to the sampling time tn to t of each station are input into the LSM network. Through the input gate, forget gate, and output gate of the network, the dynamic evolution characteristics of the seasonal or residual components of each station in the time dimension are extracted.

[0086] f t =Sigmoid(W f V t +U f h t-1 +b f )

[0087] i t =Sigmoid(W i V t +U i h t-1 +b i )

[0088]

[0089] o t =sigmoid(W o V t+U o h t-1 +b o )

[0090] h t =o t ⊙tanh(c t )

[0091] Among them, V t ∈{S t ,R t}, when V t Take seasonal portion S t When V is obtained, the dynamic evolution characteristics of the seasonal component in the time dimension are obtained. t Take the residual component R t At that time, the dynamic evolution characteristics of the residual components in the time dimension are obtained.

[0092] The extracted spatiotemporal features are transformed by a fully connected layer to output predicted sequences of trend components, seasonal components, and residual components for multiple future time steps. Specifically, in a preferred embodiment of this invention, the predicted value of the trend component is:

[0093]

[0094] The predicted value for seasonal components is:

[0095]

[0096] The predicted values ​​of the residual components are:

[0097]

[0098] Among them, W fc Represents the weights of the fully connected layer; b fc This represents the bias term of the fully connected layer.

[0099] Based on the trend component, seasonal component, and residual component at each station, the predicted wave height at multiple stations is calculated. In a preferred embodiment of this invention, the predicted wave height at multiple stations is expressed as follows:

[0100]

[0101] The final prediction of wave height at multiple sites is the sum of the predicted values ​​of the trend component, seasonal component, and residual component.

[0102] Example

[0103] This embodiment acquires wave height observation data from multiple stations in the same sea area, including observation data from 2020 to 2023. Missing and null values ​​are preprocessed, and the data from 2020 to 2022 is used as the training set, and the data from 2023 is used as the test set.

[0104] according to Figure 1 The flowchart shown demonstrates joint prediction of wave height from multiple stations. For each station, the wave height observation data is seasonally decomposed to obtain three components: trend, seasonal, and residual. By minimizing the model loss function, gradient descent is used to train and learn the parameters of the fusion graph convolutional neural network and long short-term memory network. After model training converges, the final multi-site joint prediction result of wave height is obtained, as shown below. Figure 2 As shown in the figure, this invention predicts the wave height at a certain site from May to September for the next 24 hours. The upper part represents the predicted wave height, and the lower part represents the absolute value of the error between the predicted wave height and the observed wave height data. Figure 3 yes Figure 2 The data includes wave height observations from the stations mentioned in the article between May and September.

[0105] This invention was experimentally compared with the Long Short-Term Memory (LSTM) network model proposed by Hochreiter.etc, the Transformer model based on an attention mechanism proposed by Vaswani.etc, and the Temporal Graph Convolutional Network model proposed by Zhao.etc on the same dataset, as shown in Table 1. A higher coefficient of determination (R²) indicates greater accuracy, as does a smaller mean squared error (MSE) and a smaller root mean square error (RMSE). The model of this invention achieves higher accuracy in terms of R². 2 The present invention outperforms the other three models in terms of mean square error (MAE) and root mean square error (RMSE), therefore it can be considered superior to the other three models.

[0106] Table 1 Comparison of experimental results

[0107]

[0108]

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-site joint prediction method for ocean wave height based on a deep learning model, characterized in that, include: Wave height observation data from multiple stations within the same sea area were acquired, and the data from each station were decomposed using a seasonal trend decomposition method. A joint model of fusion graph convolutional neural network and long short-term memory network is constructed for the trend component, and long short-term memory network models are constructed separately for the seasonal component and the residual component. The extracted spatiotemporal features are transformed by a fully connected layer, and the predicted sequences of trend components, seasonal components and residual components for multiple future time steps are output respectively. The predicted wave heights for multiple stations are obtained by summing the predicted sequences of trend, seasonal, and residual components from each station.

2. The multi-site joint prediction method for ocean wave height based on a deep learning model according to claim 1, characterized in that, The process of acquiring wave height observation data from multiple stations within the same sea area and decomposing the data from each station using a seasonal trend decomposition method specifically includes: Missing and null values ​​in wave height observation data from multiple stations within the same sea area are removed through data preprocessing. The geographical distance between each pair of stations is calculated, and the reciprocal of the distance is used to construct an adjacency matrix A. For each site's data, a seasonal trend decomposition method is applied, using local weighted regression to decompose the data into trend components, seasonal components, and residual components. The trend component represents the long-term trend, the seasonal component represents periodic fluctuations, and the residual component represents random disturbances or anomalous fluctuations. The seasonal trend decomposition method is expressed as follows: X i (t)=T i (t)+S i (t)+R i (t),i∈{1,2,...,N} Where N represents the number of sites; data X i (t) is decomposed into trend component T i (t), seasonal component S i (t), residual component R i (t); X i (t) represents the wave height observation data value corresponding to the i-th station at the t-th sampling time; T i (t) represents the trend component data value corresponding to the wave height observation data value at the t-th sampling time of the i-th station; S i (t) represents the seasonal component data value corresponding to the wave height observation data value at the t-th sampling time corresponding to the i-th station, R i (t) represents the residual component data value corresponding to the wave height observation data value at the t-th sampling time of the i-th station.

3. The multi-site joint prediction method for ocean wave height based on a deep learning model according to claim 1, characterized in that, The construction of a joint model fusing graph convolutional neural network and long short-term memory network for the trend component specifically includes: The trend components corresponding to the sampling time tn to t of each station are constructed into a feature matrix, and together with the adjacency matrix A, they are input into a deep learning model with an L-layer graph convolutional neural network to extract the spatial features of the trend components between stations. The spatial feature vector result obtained after the input data passes through the graph convolutional neural network is expressed by the following formula: H (0) =P(t-n:t-1) in, for The degree matrix; W G is the weight matrix; A is the adjacency matrix based on the geographical distance between sites; I is the identity matrix; n is the sliding window length; The high-dimensional spatial features output by the multi-layer graph convolutional neural network are mapped to a time series representation of length n through a linear transformation, which is then used for subsequent modeling of the time features by the long short-term memory network. B=Linear(H (L) ) Among them, H (L) B represents the extracted spatial features; B is the mapped time series with a length of n. The time series data with fused spatial features is input into a Long Short-Term Memory (LSTM) network. After passing through input gates, forget gates, and output gates, the dynamic evolution characteristics of the trend components at each station in the time dimension are extracted. The calculation formula for the LTM network is as follows: f t =Sigmoid(W f B t +U f h t-1 +b f ) i t =Sigmoid(W i B t +U i h t-1 +b i ) o t =sigmoid(W o B t +U o h t-1 +b o ) h t =o t ⊙tanh(c t ) Where t∈{1,2,...,n} is used to mark the time steps in the sequence; i t f t o t These are the input gate, forget gate, and output gate, respectively, calculated using the Sigmoid activation function. c represents the candidate cell states calculated using the tanh activation function; t The current cell state; W i W f W o W c These are the input weight matrices for the input gate, forget gate, output gate, and candidate cell states, respectively; U i U f U o U c b are the hidden state weight matrices for the input gate, forget gate, output gate, and candidate cell state, respectively; i b f b o b c These are the bias terms for the input gate, forget gate, output gate, and candidate cell state, respectively; ⊙ indicates that elements at the same position are multiplied together. The same model, namely the Long Short-Term Memory (LSTM) network model, is constructed for both the seasonal component and the residual component. The seasonal component or residual component corresponding to the sampling time tn to t of each station is input into the LSM network. After passing through the input gate, forget gate, and output gate of the network, the dynamic evolution characteristics of the seasonal component or residual component of each station in the time dimension are extracted. f t =Sigmoid(W f V t +U f h t-1 +b f ) i t =Sigmoid(W i V t +U i h t-1 +b i ) o t =sigmoid(W o V t +U o h t-1 +b o ) h t =o t ⊙tanh(c t ) Among them, V t ∈{S t ,R t }, when V t Take seasonal portion S t When V is obtained, the dynamic evolution characteristics of the seasonal component in the time dimension are obtained. t Take the residual component R t At that time, the dynamic evolution characteristics of the residual components in the time dimension are obtained.

4. The multi-site joint prediction method for ocean wave height based on a deep learning model according to claim 1, characterized in that, The process involves transforming the extracted spatiotemporal features using a fully connected layer to output predicted sequences for trend components, seasonal components, and residual components at multiple future time steps. Specifically, this includes: The predicted value of the trend component is: The predicted value of the seasonal component is: The predicted value of the residual component is: Among them, W fc Represents the weights of the fully connected layer; b fc This represents the bias term of the fully connected layer.

5. The multi-site joint prediction method for ocean wave height based on a deep learning model according to claim 1, characterized in that, The predicted wave heights at the multiple stations are represented as follows: The final prediction of wave height at the multi-site locations is the sum of the predicted values ​​of the trend component, seasonal component, and residual component.

Images