Marine drifting buoy trajectory prediction method based on FGGNN model

Abstract

The invention provides an ocean drifting buoy trajectory prediction method based on an FGGNN model, and relates to the technical field of ocean drifting buoy trajectory prediction. The action mechanism between the marine environment elements and the positions of the marine drifting buoys is simulated through graph aggregation operation, the characteristic relation between the tracks of the marine drifting buoys and the marine environment elements is effectively extracted, the reasoning efficiency of the gating circulation unit is effectively improved by means of the improved gating circulation unit, and the reliability of the gating circulation unit is improved. A trajectory prediction model FGGNN based on a graph structure neural network is utilized, an encoder-decoder architecture is adopted, trajectory time sequence data are received by using a graph structure, time sequence reasoning is realized by using an improved gating circulation unit, and a complex dependency relationship between features and between time sequences in buoy trajectory data is captured, so that accurate prediction of a buoy trajectory is realized.

Description

Technical Field

[0001] The invention relates to the technical field of ocean drifting buoy trajectory prediction, and in particular to an ocean drifting buoy trajectory prediction method based on a FGGNN model. Background Art

[0002] Currently, the methods for predicting the trajectory of drifting buoys are mainly divided into physical model-based methods and data-driven methods. Physical model-based methods usually simulate the motion trajectory of the buoy by establishing fluid dynamics equations and combining them with marine environmental data. However, such methods are limited by the accuracy of marine environmental data acquisition and the ability of physical models to depict a variety of marine environmental elements. It is difficult to consider complex factors such as wind, current, and tide at the same time. There is still a certain deviation from the real marine environment, which limits the accuracy of the prediction results. In addition, physical models usually require parameter adjustments for different regions, which limits their applicability.

[0003] With the development of deep learning technology, researchers have gradually applied data-driven neural network models to trajectory prediction tasks. Since buoy trajectory data exhibit typical time series characteristics, recurrent neural networks (RNNs) and their improved models, such as long short-term memory networks (LSTMs) and gated recurrent units (GRUs), are widely used in time series data prediction tasks. These models can effectively capture the temporal dependency of trajectory data, but they still have certain limitations. For example, RNNs are prone to gradient vanishing or gradient exploding problems when processing long sequence data, making it difficult for the model to capture long-term dependency information; although LSTM alleviates this problem to a certain extent, it has a high computational complexity and is difficult to model complex relationships between feature dimensions. In addition, most traditional neural network models are based on fully connected operations, which cannot effectively model the non-Euclidean relationship between buoy trajectories and marine environmental elements, limiting the model's predictive ability.

[0004] In recent years, graph neural networks (GNNs) have gradually been applied to trajectory prediction tasks due to their advantages in modeling complex relational networks. GNNs can effectively use graph structures to represent the dependencies between different features, allowing the model to learn spatial and temporal information at the same time. However, existing GNNs are mainly used to model spatial relationships between nodes, and rarely consider the interactions between feature dimensions, resulting in limitations in the model when processing multimodal marine environmental data. Therefore, for the task of trajectory prediction of drifting buoys in the ocean, it is necessary to design a neural network model that can simultaneously model the interactions between time series features and feature dimensions to improve the accuracy of trajectory prediction. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a method for predicting the trajectory of an ocean drifting buoy based on a FGGNN model, comprising the following steps: Step S1, obtaining time series data from an ocean drifting buoy, and obtaining features in the time series data through interpolation calculation; Step S2, analyzing the interaction relationship between features in the time series data, generating a directed adjacency matrix of the feature graph, and performing normalization processing; Step S3, using the weighted adjacency matrix to perform a convolution operation on the feature graph to obtain time series data; Step S4, extracting the time features of the time series data using an improved gated recurrent unit; Step S5: construct an FGGNN model to predict the trajectory of the ocean drifting buoy based on the time characteristics obtained in step S4.

[0006] In a preferred embodiment, step S2 comprises: Step S201, analyzing the interaction relationship between features in the time series data, and generating a directed adjacency matrix through an edge set; Step S202, adding self-loops to the directed adjacency matrix so that each node is at least connected to itself; ; in, is the identity matrix, is the original directed adjacency matrix, is the adjacency matrix after adding self-loops; Step S203: Calculate the degree matrix D to obtain the degree-normalized directed adjacency matrix : .

[0007] In a preferred embodiment, step S3 comprises: Step S301: Adjacency matrix weights The normalized adjacency matrix Perform weighted operation to obtain the weighted adjacency matrix : ; in, represents element-wise product; Step S302: Based on the weighted adjacency matrix, a convolution operation is performed on the structure of the feature graph, and feature data after convolution of each node graph is output; ; in Represents the i-th feature data of the feature map after convolution, is the jth feature of the input data X, is an element in the weighted adjacency matrix, representing the connection weight between node i and node j; Step S303: Overall scale the feature data after graph convolution to obtain the intermediate representation T2 of the time series data: ; in, is a learnable parameter, and ReLU is the activation function; Step S304: perform batch normalization on the intermediate representation of the time series data to obtain the time series data after batch normalization. : ; In the formula, BatchNorm means batch normalization.

[0008] In a preferred embodiment, step S4 comprises: Step S401: Use graph convolution operation to replace the fully connected operation in the gated recurrent unit to form an improved gated recurrent unit, and extract the relationship between time series data features while learning the time relationship; Step S402: In the improved gated recurrent unit, at each moment, graph convolution is performed on the hidden state of the previous moment to obtain the influence of the state of the previous moment on the state of the current moment, thereby forming time series information; Step S403: Use the improved gated recurrent unit to process the time series information, and output the feature dynamics that change over time and the hidden state representing the final moment of the time series information.

[0009] In a preferred embodiment, in step S401, a graph convolution operation is performed on the input data X: ; in, Represents the result after graph convolution operation, are learnable weights, is the normalized directed adjacency matrix, is a nonlinear activation function; Will Split along the last dimension into 3 parts: ; They correspond to the input receiving information of the update gate, reset gate and new memory unit in the gated recurrent unit respectively.

[0010] In a preferred embodiment, in step S402, in the improved gated recurrent unit, the hidden state of each moment to the previous moment is Perform graph convolution to obtain the impact of the previous state on the current state: ; in, are learnable weights, It is a nonlinear activation function. h2h represents the current state information received from the previous state. h2h is split into three parts along the last dimension: ; They correspond to the information of the previous state received by the update gate, reset gate and new memory unit respectively.

[0011] In a preferred embodiment, in step S403, the update gate z is first calculated. t : ; in, is the bias term of the update gate; Then calculate the reset gate r t : ; in, is the bias term for the reset gate; Then calculate the new candidate memory : ; in, represents element-wise product, is the computation result of the new memory unit in the graph convolution, Represents a new memory unit generated by the hidden state of the previous moment, which is calculated by graph convolution from the hidden state of the previous moment. is the bias term of the new memory unit; Calculate the hidden state at the current moment : ; in, It is the hidden state obtained by graph convolution processing the data at the previous moment.

[0012] In a preferred embodiment, in step S5, the FGGNN model adopts an encoder-decoder architecture, the encoder includes three layers of improved gated recurrent units, the decoder includes three layers of improved gated recurrent units, the three layers of improved gated recurrent units in the decoder correspond to the three layers of improved gated recurrent units in the encoder, and each improved gated recurrent unit in the decoder receives a hidden state generated by the corresponding improved gated recurrent unit in the encoder.

[0013] In a preferred embodiment, the encoder and decoder are combined into an end-to-end system. The encoder consists of multiple GCN and RNN layers to extract the spatiotemporal features of the input time series data X and generate a hidden state list : ; Among them, the number of final states is the number of GRU units in the encoder; The decoder accepts the hidden state list from the encoder as input, performs time series prediction, and generates the final predicted future time series. : .

[0014] In a preferred embodiment, in step S1, each buoy record is traversed, and the timestamp, longitude and latitude query wind speed and current data are merged with the buoy track longitude and latitude data into time series data; Calculate the latitude and longitude time step displacement of each buoy as the displacement feature: ; ; Among them, lng_diff represents the time step displacement in the longitude direction, and lat_diff represents the time step displacement in the latitude direction. represents the longitude at time t, represents the longitude at time t-1, represents the latitude at time t, Represents the latitude at time t-1.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: (1) Representation of buoy trajectory time series data based on graph structure: The present invention proposes a graph structure representation method, which uses the advantages of graph neural networks to map the buoy trajectory and the marine environmental elements (such as longitude and latitude, ocean currents, wind speed, etc.) at the buoy location into a graph structure, and efficiently represents them in the Encoder part. Each marine environmental element (such as ocean currents, wind speed, temperature, etc.) is a node in the graph, and the interaction between nodes is represented by a directed adjacency matrix. The connection relationship between these nodes reflects the mutual influence and effect between environmental elements (such as the relationship between wind speed and ocean currents). The construction of the adjacency matrix is ​​based on the physical background, such as the influence of ocean currents on wind speed or the coupling relationship between wind speed and buoy trajectory, so as to more accurately describe the interaction law of environmental elements.

[0016] On the one hand, this representation method effectively expresses the physical interaction relationship between different marine environmental elements by introducing a graph structure, which enhances the interpretability of the model; on the other hand, the graph structure can effectively reduce redundant connections between features, thereby significantly improving computational efficiency and reducing unnecessary computational overhead. Compared with the traditional data representation method based on a single feature, the graph structure can capture the complex interactions between environmental elements, allowing the model to more comprehensively and accurately reflect the dynamic changes of the marine environment, thereby improving the accuracy and reliability of the prediction.

[0017] (2) Learning the relationship between feature dimensions based on directed graph aggregation: The feature dimension relationship learning method based on directed graph aggregation of the present invention can effectively simulate the complex dynamic relationship between marine environmental elements and buoy trajectories. By normalizing the degree and adjusting the size of the adjacency matrix, the adjacency matrix can more accurately reflect the mutual influence between different marine environmental elements, and how these influences act on the changes in the buoy trajectory during the time series process. The adjusted adjacency matrix can not only show the relationship between marine environmental elements, but also effectively control the intensity of the influence of these elements on the prediction of buoy trajectories. During the model training process, through the dynamic adjustment of the learnable adjacency matrix weights, the model can flexibly optimize the influence path between different features, avoiding the overfitting problem that may occur in traditional methods, while enhancing the generalization ability of the model.

[0018] This feature dimension relationship learning method based on directed graph aggregation can better cope with the highly complex and nonlinear feature interactions in the marine environment, and effectively improve the prediction accuracy of the buoy trajectory. Especially when dealing with complex marine environments, the model has stronger adaptability and prediction stability.

[0019] (3) Improved gated recurrent unit: In order to solve the shortcomings of traditional gated recurrent units when processing graph-structured data, the present invention proposes an improved GRU structure. Traditional GRU uses a fully connected layer to learn feature dependencies in time series, but for graph-structured data, the traditional fully connected layer cannot fully consider the complex dependencies between nodes. Therefore, the present invention replaces the traditional fully connected layer by introducing a graph convolution operation in GRU. The graph convolution operation can capture the mutual dependencies between nodes when processing multi-feature graph-structured data at each time step, thereby more effectively extracting spatial and temporal features in graph-structured data.

[0020] This improvement enables the model to better model the correlation between different features in time series prediction tasks, and is particularly suitable for processing multimodal data (such as the correlation between buoy trajectory data and environmental factors such as ocean currents and wind speed). In addition, after the introduction of graph convolution, the model can reduce redundant connections, thereby reducing computational complexity and improving efficiency. In the buoy trajectory prediction task, the improved GRU can not only improve the model's ability to learn time series features, but also effectively enhance the accuracy of modeling the relationship between features, significantly improving the prediction effect.

[0021] (4) FGGNN model based on improved GRU and directed graph aggregation mechanism: The present invention proposes a new graph neural network model (FGGNN) based on improved GRU and directed graph aggregation mechanism. The model adopts an encoder-decoder architecture and realizes efficient temporal feature learning through a three-layer improved GRU structure. Compared with the traditional buoy trajectory prediction method, the FGGNN model of the present invention can simultaneously process the latitude and longitude data of the buoy and a variety of marine environmental factors (such as wind speed, ocean current, etc.). By integrating these multi-dimensional features, the model can capture more complex environmental change patterns, thereby improving the accuracy and adaptability of trajectory prediction.

[0022] The three-layer improved GRU structure in the model has a strong temporal learning ability and can accurately model the temporal dependence of the buoy trajectory. In addition, the FGGNN model also models the complex relationship between feature dimensions through a directed graph aggregation mechanism, further enhancing the adaptability of the model in complex marine environments. In the context of multimodal data fusion, the FGGNN of the present invention can not only process features from different sources, but also effectively learn the interactions between these features, providing more accurate buoy trajectory predictions.

[0023] Compared with traditional methods, the FGGNN model of the present invention can better cope with the dynamic and complex marine environment, while improving the accuracy of trajectory prediction. It has strong application value and is particularly suitable for complex time series prediction tasks that require multimodal feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The directed graph structure for characterizing the characteristics of multiple marine environment elements of the present invention; Figure 2 It is a structural diagram of the improved gated recurrent unit of the present invention; Figure 3 The overall encoder-decoder architecture diagram of the model of the present invention; Figure 4 It is a graph of the longitude variation of the drifting buoy; Figure 5 This is a graph showing the dimensional changes of the drifting buoy; Figure 6 This is a graph showing the distance a drifting buoy moves every hour; Figure 7 This is a comparison chart of actual longitude and predicted longitude; Figure 8 This is a comparison chart of the actual latitude value and the predicted value. DETAILED DESCRIPTION

[0025] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] Example 1

[0027] The following combination Figure 1-Figure 6 , the present invention is further described in detail. The specific implementation steps are as follows: Step S1. Obtain time series data from an ocean drifting buoy, and obtain the features in the time series data through interpolation calculation.

[0028] Time series data is obtained from drifting buoys in the ocean, including the latitude and longitude information of the buoys. At the same time, the ocean current speed and wind speed at the corresponding position are interpolated from the ocean environment field grid data according to the latitude and longitude coordinates of the buoy track points. The ocean environment field data can be obtained from the European Copernicus Ocean Data Sharing Center or through remote sensing reanalysis data.

[0029] Step S101. First, traverse each buoy record, interpolate the corresponding wind speed and current data from the Copernicus Ocean Environment Field grid data according to the timestamp, longitude and latitude of each record data, and merge the data with the buoy trajectory longitude and latitude data into time series data.

[0030] Step S102. Calculate the latitude and longitude time step displacement of each buoy as the displacement feature: ; ; Among them, lng_diff represents the time step displacement in the longitude direction, and lat_diff represents the time step displacement in the latitude direction. represents the longitude at time t, represents the longitude at time t-1, represents the latitude at time t, Represents the latitude at time t-1. After calculation, the time step displacement of longitude and latitude is also used as a feature in the time series data.

[0031] Step S2: Analyze the interaction relationship between features in the time series data, generate a directed adjacency matrix of the feature graph based on it, and perform normalization.

[0032] Step S201: Analyze the interaction relationship between features in the time series data, and generate a directed adjacency matrix through an edge set.

[0033] Analyze the interaction between features in time series data, such as Figure 1 The connection arrows between the nodes are shown. If an edge has only one arrow, it means a one-way relationship, and two arrows mean a two-way relationship. A set of two-dimensional arrays is used to represent this connection relationship, and the array shape is 2×L: ; Among them, L represents the number of nodes in each row, represents the starting node, represents the abort node, so, , , ,..., All are directed edges. Based on this two-dimensional array, you can create an adjacency matrix. Create an all-zero adjacency matrix of size: ; where n represents the number of features, then the position corresponding to each directed edge is set to 1, indicating that there is a directed connection relationship between the node pairs, such as Figure 1 As shown, wind_u represents the wind speed in the longitude direction, wind_v represents the wind speed in the latitude direction, longitude represents longitude, latitude represents latitude, current_u represents the current speed in the longitude direction, and current_v represents the current speed in the latitude direction.

[0034] Step S202: Add self-loops to the directed adjacency matrix so that each node is at least connected to itself.

[0035] In order to avoid that some nodes completely lose information due to the lack of information about neighboring nodes in subsequent graph convolution operations, self-loops are added to the directed adjacency matrix so that each node is at least connected to itself.

[0036] ; in is the identity matrix, is the original directed adjacency matrix, is the adjacency matrix after adding self-loops.

[0037] Step S203. By calculating the degree matrix , and obtain the directed adjacency matrix after degree normalization.

[0038] Its elements Representation Node The degree of the node The degree matrix is ​​a diagonal matrix whose diagonal elements are the degrees of the nodes.

[0039] To normalize the degree, the square root of the reciprocal of the degree is used for normalization. The formula for degree normalization is as follows: ; in, Represents the degree matrix The reciprocal square root of the degree of each node is the reciprocal square root of the degree of each node. This operation can reduce the impact of the degree on the adjacency matrix, so that nodes with high degrees will not have too much impact on the model. The directed adjacency matrix finally returned is It is a degree-normalized matrix, which will be used as input in the graph neural network to improve the robustness of the model to the graph.

[0040] Normalization is used to improve the representation of feature maps to prevent the degree from affecting the final model performance.

[0041] Step S3: Use the weighted adjacency matrix to perform a convolution operation on the feature graph structure to obtain time series data.

[0042] The structure of the graph is adjusted through a learnable weighted adjacency matrix, and then the graph convolution operation is applied, and finally the output is obtained through a fully connected layer (linear transformation).

[0043] Step S301: Perform a weighted operation on the normalized directed adjacency matrix using the adjacency matrix weights to obtain a weighted adjacency matrix.

[0044] First, through the learnable adjacency matrix weights To normalize the adjacency matrix Perform weighted operation to obtain the weighted adjacency matrix : ; in, Represents element-wise product.

[0045] in, is a , where n is the number of features, which adjusts the strength of the relationship between nodes in the adjacency matrix through the learning rate; is the normalized adjacency matrix (the degree-normalized matrix mentioned above). By weighting the adjacency matrix, the model can dynamically adjust the adjacency relationship of each node based on the graph structure information in the data.

[0046] Step S302: Based on the weighted adjacency matrix, perform a convolution operation on the structure of the feature graph and output the convolved feature data of each node graph.

[0047] After calculating the weighted adjacency matrix, the graph convolution operation is applied to operate the input data at a single moment to implement the convolution operation on the graph structure, thereby obtaining the graph convolution output of each node: ; in Represents the i-th feature data of the feature map after convolution, is the jth feature of the input data X, is an element in the weighted adjacency matrix, representing the connection weight between node i and node j.

[0048] Step S303. Use a trainable parameter to overall scale the feature data after graph convolution to obtain the intermediate representation T2 of the time series data: ; in, It is a learnable parameter that can scale the overall graph structure feature data after graph convolution. Then the ReLU activation function is used to introduce nonlinearity, thereby increasing the expressive power of the model.

[0049] Step S304: Perform batch normalization on the intermediate representation of the time series data to obtain the time series data after batch normalization. : ; In the formula, BatchNorm stands for batch normalization. Batch normalization normalizes the data of each small batch to have zero mean and unit variance, thereby accelerating the gradient descent process and preventing the gradient from disappearing or exploding. The result after graph convolution, fully connected layer and batch normalization is the output of this step and has the following shape: ; Among them, OutputShape represents the shape of the output after the above processing, batch_size represents the batch size of the data, num_nodes represents the number of nodes, that is, the number of features of the input data, and num_timesteps represents the number of time steps of the input data.

[0050] Step S4. Use the improved gated recurrent unit to extract the temporal features of the time series data.

[0051] The previous step is to use a graph structure to receive the buoy trajectory time series feature data. The time series task also requires the model to be able to extract time series features, so the gated recurrent unit is used next to extract the time series features and then predict the trajectory for a period of time in the future. In particular, the present invention improves the gated recurrent unit and uses a graph convolution operation instead of a full connection operation.

[0052] Step S401: Use graph convolution operation, i.e., GCNBlock, to replace the fully connected operation in the gated recurrent unit to form an improved gated recurrent unit GCNGRU, so as to further extract the relationship between time series data features while learning the time relationship. Figure 2 The figure shows the improved gated recurrent unit, where G represents the graph convolution operation, σ represents the sigma activation number, tanh represents the hyperbolic tangent activation function, X represents the Hadamard product, + represents matrix addition, H[t-1] represents the state output at the previous moment, H[t] represents the state output at the current moment, y[t] represents the output at the current moment, and x[t] represents the input at the current moment.

[0053] Assume that the output data shape of the above steps is B × N × S × C, where B is the batch size, N is the number of nodes, S is the number of time steps, and C is the number of features of each node. It is passed to the GCNGRU module, and the relationship between features is extracted again through the GCNBlock, that is, the input data is first subjected to graph convolution: ; in, (input-to-hidden) indicates the result of the current input X after being processed by GCN. X is the current input. are the learnable weights computed by GCN, is the normalized adjacency matrix, is a non-linear activation function.

[0054] Then Split along the last dimension into 3 parts: ; The corresponding inputs of the update gate, reset gate, and new memory unit in the gated recurrent unit receive information respectively. The update gate controls how much information in the current hidden state comes from the hidden state at the previous moment and how much information comes from the newly calculated state. The reset gate controls how much information in the hidden state at the previous moment needs to be forgotten in order to calculate the new memory. The new memory unit calculates a new candidate state based on the current input and the result of the reset gate.

[0055] Step S402. In the gated recurrent unit, the hidden state of each moment to the previous moment Perform graph convolution to obtain the impact of the previous state on the current state: ; in, is the hidden state at the previous moment, is the normalized adjacency matrix, are the learnable weights computed by GCN, is a nonlinear activation function, and h2h represents the current state information received from the previous state. Similarly, h2h is split into 3 parts along the last dimension: ; They correspond to the information of the previous state received by the update gate, reset gate and new memory unit respectively.

[0056] Step S403. Use GRU to process the time series information, output the feature dynamics that change over time and the hidden state that represents the final moment of the entire time series information.

[0057] With i2h and h2h, we can calculate the three core parts of GRU: update gate, reset gate, and calculate new candidate memory.

[0058] First, the update gate is calculated, which determines h t-1 To what extent: ; in, It is the update gate component of the current input after graph convolution transformation, which indicates the contribution of the input feature to the current update gate. It is the update gate component of the hidden state at the previous moment after graph convolution transformation, which indicates the contribution of past information to the current update gate. is the bias term of the update gate.

[0059] Then the calculation of the reset gate is performed: ; in, It is the reset gate related part of the current input after graph convolution transformation, indicating the contribution of the input feature to the current reset gate. It is the reset gate related part of the hidden state of the previous moment after the graph convolution transformation, which indicates the contribution of past information to the current reset gate. is the bias term for the reset gate.

[0060] Then calculate the new candidate memory: ; in, represents the element-wise product, also known as the Hadamard product ⊙, It is the calculation result of the new memory unit in the GRU network. It is obtained after the current input passes through the graph convolution (GCN) and is used to determine the updated part of the new memory at the current moment. It represents the new memory unit generated by the hidden state of the previous moment. It is calculated by the hidden state of the previous moment through graph convolution and is used to determine the memory update at the current moment. is the bias term for the new memory cell.

[0061] Then we can calculate the hidden state at the current moment: ; in, is the aforementioned update gate, It is the hidden state obtained by GRU processing the data at the previous moment. The GRU gated recurrent unit improved by graph convolution outputs two main parts. First, it outputs a sequence, which is the hidden state of the entire input sequence after multiple time steps. It is composed of the features that change over time. In addition, it also outputs the hidden state of the last time step, that is, the final hidden state, which represents the model state at the last moment of the entire sequence.

[0062] Step S5. Construct a FGGNN model to predict the trajectory of the ocean drifting buoy based on the time characteristics obtained in step S4.

[0063] The FGGNN model adopts an encoder-decoder architecture, in which the encoder contains three layers of improved gated recurrent units, and the decoder includes three layers of improved gated recurrent units. The three layers of improved gated recurrent units in the decoder correspond to the three layers of improved gated recurrent units in the encoder, and each gated recurrent unit in the decoder receives the hidden state generated by the corresponding gated recurrent unit in the encoder; a feature dimension relationship learning module based on graph aggregation is used to accept the buoy trajectory time series data.

[0064] The overall framework adopted by the present invention is the encoder-decoder framework, that is, the Encoder-Decoder framework, where Encoder represents the encoder and Decoder represents the decoder. The GRU unit in the Encoder will output two parts of content, one part is the hidden state list of the entire time sequence, which will be used as the input of the GCNBlock in other Encoders. The other part is the hidden state at the final moment, which will be input to the GRU unit in the Decoder framework.

[0065] Step S501. Build an encoder and perform graph convolution on the input features through GCNBlock: ; Among them, the specific operation of GCN is as mentioned above, which is to model the relationship between input features through the graph convolution module GCNBlock, use the weighted adjacency matrix to transfer the relationship between nodes, and obtain the graph aggregation operation output GCNOutput.

[0066] Step S502: The output GCNOutput of the graph aggregation operation is modeled through GRU for time series.

[0067] Step S503. Construct a decoder, such as Figure 3 As shown in the figure, there are 3 GRU units in the decoder. Each GRU unit in the decoder receives the hidden state output by the corresponding layer GRU unit in the encoder as the state input, and receives the output of the GCNBlock above it as the regular input. That is, each GRU unit receives two channels of information, one for state information and one for regular input information. In particular, there is no GCNBlock module above the first GRU unit at the top of the decoder, so it receives None as the regular input. There is a GCNBlock immediately below each GRU unit in the decoder, and the output of each GRU unit will be used as the input of the subsequent GCNBlock. Step S504. The GCNBlock module in the decoder receives the output of the GRU unit in the decoder, performs feature relationship modeling, and outputs the predicted buoy drift trajectory data.

[0068] The GCNBlock module in the decoder receives the output of the GRU unit in the decoder and models the relationship between features. The last GCNBlock module in the decoder outputs the predicted time series data as the output of the entire framework.

[0069] In the overall model, the Encoder and Decoder are combined into an end-to-end system. The Encoder consists of multiple GCN and RNN layers to extract the spatiotemporal features of the input time series data X and generate a hidden state list: ; The number of final states is the number of GRU units in the encoder. The encoder maps the input time series data to a new representation space, which usually contains rich representations of spatiotemporal features.

[0070] The decoder accepts the hidden states from the encoder as input. It makes time series predictions based on these states and generates the final output: ; in Represents the hidden state list output by Encoder, Represents the output of the decoder (that is, the entire overall model), that is, the predicted future time series.

[0071] Based on the processed data set, the present invention uses the FGGNN (Flow Graph Neural Network) model to train and test the ocean drifting buoy trajectory. The data set is divided into training set, validation set and test set in a ratio of 8:1:1, and the model is implemented using the Pytorch framework. In the main experiments, the architecture of the encoder and decoder are both 3 layers. The time interval of the data is set to 1 hour, and the appropriate time step is selected from 12 hours and 24 hours. Finally, the input historical trajectory data length T is defined as 24 hours, and the predicted time range is also 24 hours, considering that a longer prediction time may reduce the prediction accuracy.

[0072] During the training process, considering the GPU memory limitation, the batch size was set to 32. The Adam optimizer was used as the optimizer, the initial learning rate was set to 1e-4, and the learning rate was gradually adjusted during the training process through the learning rate scheduler. The total training rounds (epochs) were set to 5000 and adjusted in time according to the training speed.

[0073] Example 2 The present invention compares the prediction accuracy by calculating the mean square error (MSE) between the predicted longitude and latitude change value and the actual trajectory longitude and latitude change value as an evaluation index.

[0074] This paper chooses to compare FGGNN with five trajectory prediction methods, and tests the performance of the models of the comparison methods on the drifting buoy trajectory dataset, using the hyperparameters recommended in the original method, including the number of layers, batch size, etc., to approach the optimal performance of the model. The input length and output length of all models are the same.

[0075] The models used for comparison and their structures are as follows: (1) CNN-LSTM: This model combines convolutional neural networks (CNN) and long short-term memory networks (LSTM). The CNN part consists of three convolutional layers, a batch normalization (BatchNorm) layer, a dropout layer, an expansion layer, and a fully connected layer, which is responsible for extracting spatial features. The LSTM part contains two LSTM layers, which process the time series data output from the CNN layer and capture the time series features.

[0076] (2) BiLSTM: This model consists of two bidirectional LSTM (BiLSTM) layers, with a total of four LSTM units. The input of the model consists of multiple features, which are processed by the BiLSTM layer to obtain time series information. The output is processed by a fully connected layer to generate the final prediction result.

[0077] (3) TRFM-LS: This model is improved on the traditional Transformer architecture. The LSTM layer is introduced to preprocess the sequence data and the output of the LSTM layer is used as the input of the decoder, thereby enhancing the modeling ability of time series data and better capturing long-term dependencies.

[0078] (4) RGA: This model combines the residual network (ResNet) and the gated recurrent unit (GRU). First, the input data is processed by the ResNet layer to extract high-level features, then passes through two GRU layers to capture temporal dependencies, and finally the output is weighted by the attention mechanism. This structure can efficiently perform time series prediction and is particularly suitable for processing complex time series patterns.

[0079] (5) TrAISformer: This model uses a Transformer architecture similar to GPT. The Transformer consists of multiple stacked self-attention layers, each of which acts as an autoregressive model. Through the dot product multi-head attention mechanism, the model can process multiple subspaces in parallel, thereby improving the model's efficiency and prediction accuracy in time series prediction tasks.

[0080] In order to verify the effect of the FGGNN model in trajectory prediction, the model was trained and tested using actual ocean drifting buoy trajectory data in the northwest Pacific region.

[0081] By comparing the fit between the model prediction results and the actual trajectory, the closer the model is to the actual trajectory, the higher the prediction accuracy of the model. The experimental results show that FGGNN performs well in predicting trajectory sequences and has strong application value. By comparing the predicted value and the actual value in the same time period, the prediction effect can be observed intuitively. The predicted trajectory curve is an important indicator for evaluating the prediction accuracy of the model.

[0082] Figure 7 and Figure 8 The comparison results between the predicted trajectory and the true trajectory of the FGGNN model proposed in this invention are shown. As can be seen from the figure, FGGNN shows higher prediction accuracy than other comparison models. At the same time, Table 1 shows the prediction results of the mean square error (MSE) of FGGNN and other comparison models on the same data set.

[0083] Table 1. Errors of different models in predicting drifting buoy trajectories

[0084] The FGGNN proposed in this paper significantly improves the prediction accuracy of trajectory sequences. From the data in Table 1, it can be seen that the error value of FGGNN is lower than all other comparison models. The mean square error of the longitude displacement value of the predicted ocean drifting buoy trajectory is 0.000231, and the mean square error of the latitude displacement value is 0.000256. The experimental results show that FGGNN can effectively capture the relationship between the various features in the buoy trajectory data, and extract time-dependent features through time series data, thereby significantly improving the accuracy of trajectory prediction.

[0085] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. The embodiments should therefore be considered exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for predicting ocean drifting buoy trajectory based on FGGNN model, characterized in that: The steps include: Step S1, obtaining time series data from an ocean drifting buoy, and obtaining features in the time series data through interpolation calculation; Step S2, analyzing the interaction relationship between features in the time series data, generating a directed adjacency matrix of the feature graph, and performing normalization processing; Step S3, using the weighted adjacency matrix to perform a convolution operation on the feature graph to obtain time series data; Step S4, extracting the time features of the time series data using an improved gated recurrent unit; Step S5: construct an FGGNN model to predict the trajectory of the ocean drifting buoy based on the time characteristics obtained in step S4.

2. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 1, characterized in that: The step S2 comprises: Step S201, analyzing the interaction relationship between features in the time series data, and generating a directed adjacency matrix through an edge set; Step S202, adding self-loops to the directed adjacency matrix so that each node is at least connected to itself; ; in, is the identity matrix, is the original directed adjacency matrix, is the adjacency matrix after adding self-loops; Step S203: Calculate the degree matrix D to obtain the degree-normalized directed adjacency matrix : 。 3. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 2, characterized in that: The step S3 comprises: Step S301: Adjacency matrix weights The normalized adjacency matrix Perform weighted operation to obtain the weighted adjacency matrix : ； in, represents element-wise product; Step S302: Based on the weighted adjacency matrix, a convolution operation is performed on the structure of the feature graph, and feature data after convolution of each node graph is output; ； in Represents the i-th feature data of the feature map after convolution, is the jth feature of the input data X, is an element in the weighted adjacency matrix, representing the connection weight between node i and node j; Step S303: Overall scale the feature data after graph convolution to obtain the intermediate representation T2 of the time series data: ； in, is a learnable parameter, and ReLU is the activation function; Step S304: perform batch normalization on the intermediate representation of the time series data to obtain the time series data after batch normalization. : ； In the formula, BatchNorm means batch normalization.

4. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 3, characterized in that: The step S4 comprises: Step S401: Use graph convolution operation to replace the fully connected operation in the gated recurrent unit to form an improved gated recurrent unit, and extract the relationship between time series data features while learning the time relationship; Step S402: In the improved gated recurrent unit, at each moment, graph convolution is performed on the hidden state of the previous moment to obtain the influence of the state of the previous moment on the state of the current moment, thereby forming time series information; Step S403: Use the improved gated recurrent unit to process the time series information, and output the feature dynamics that change over time and the hidden state representing the final moment of the time series information.

5. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 4, characterized in that: In step S401, a graph convolution operation is performed on the input data X: ； in, Represents the result after graph convolution operation, are learnable weights, is the normalized directed adjacency matrix, is a nonlinear activation function; Will Split along the last dimension into 3 parts: ； They correspond to the input receiving information of the update gate, reset gate and new memory unit in the gated recurrent unit respectively.

6. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 5, characterized in that: In step S402, in the improved gated recurrent unit, each moment is a hidden state of the previous moment. Perform graph convolution to obtain the impact of the previous state on the current state: ； in, are learnable weights, It is a nonlinear activation function. h2h represents the current state information received from the previous state. h2h is split into three parts along the last dimension: ； They correspond to the information of the previous state received by the update gate, reset gate and new memory unit respectively.

7. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 6, characterized in that: In step S403, the update gate z is first calculated. t : ； in, is the bias term of the update gate; Then calculate the reset gate r t : ； in, is the bias term for the reset gate; Then calculate the new candidate memory : ； in, represents element-wise product, is the computation result of the new memory unit in the graph convolution, Represents a new memory unit generated by the hidden state of the previous moment, which is calculated by graph convolution from the hidden state of the previous moment. is the bias term of the new memory unit; Calculate the hidden state h at the current moment t : ； in, It is the hidden state obtained by graph convolution processing the data at the previous moment.

8. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 4, characterized in that: In step S5, the FGGNN model adopts an encoder-decoder architecture, the encoder includes three layers of improved gated recurrent units, the decoder includes three layers of improved gated recurrent units, the three layers of improved gated recurrent units in the decoder correspond to the three layers of improved gated recurrent units in the encoder, and each improved gated recurrent unit in the decoder receives a hidden state generated by the corresponding improved gated recurrent unit in the encoder.

9. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 5, characterized in that: The encoder and decoder are combined into an end-to-end system. The encoder consists of multiple GCN and RNN layers to extract the spatiotemporal features of the input time series data X and generate a hidden state list. : ； Among them, the number of final states is the number of GRU units in the encoder; The decoder accepts the hidden state list from the encoder as input, performs time series prediction, and generates the final predicted future time series. : 。 10. The method for predicting the trajectory of an ocean drifting buoy based on the FGGNN model according to claim 1, characterized in that: In the step S1, each buoy record is traversed, and the timestamp, longitude and latitude query wind speed and current data are merged with the buoy track longitude and latitude data into time series data; Calculate the latitude and longitude time step displacement of each buoy as the displacement feature: ； ； Among them, lng_diff represents the time step displacement in the longitude direction, and lat_diff represents the time step displacement in the latitude direction. represents the longitude at time t, represents the longitude at time t-1, represents the latitude at time t, Represents the latitude at time t-1.

Images