Buoy trajectory prediction method based on stationary wavelet and wind-flow-position interaction

Abstract

The application relates to the field of ocean monitoring and forecasting, and discloses a buoy trajectory prediction method based on stationary wavelet and wind-flow-position interaction, which comprises the following steps: acquiring a trajectory sequence of a buoy in a historical period and corresponding ocean environment field data and performing standardization processing; decomposing the normalized trajectory sequence of the buoy through stationary wavelet transformation to obtain a plurality of sub-sequences of different scales, and then obtaining a plurality of segmented token sequences through dynamic segmentation; performing feature extraction on each segmented token sequence after feature mapping through a wind-flow-position interaction feature extraction mechanism; reconstructing the interaction features output by the wind-flow-position interaction feature extraction mechanism through stationary wavelet inverse transformation, and finally outputting a trajectory coordinate prediction value of the buoy in a future period after linear layer mapping. Through the introduction of wavelet multi-scale decomposition and dynamic segmentation, the model can explicitly capture the trend and details in the trajectory, and effectively cope with non-stationarity.

Description

Technical Field

[0001] This application belongs to the field of marine monitoring and forecasting, and specifically relates to a buoy trajectory prediction method based on stationary wavelets and wind-current-position interaction. Background Technology

[0002] Surface drifting buoys play a crucial role in marine environmental monitoring, climate research, and disaster response. Their trajectory prediction is essential for optimizing observation network deployment and ensuring the safety of maritime operations. However, buoy motion is subject to nonlinear coupling effects from multiple factors such as wind and currents, and their trajectories exhibit significant non-stationarity and multi-scale temporal dynamic characteristics, making accurate prediction extremely challenging.

[0003] Existing prediction methods suffer from the following limitations: Numerical simulation methods heavily rely on precise input flow field data, incur high computational costs, and struggle to adapt to complex nonlinear dynamic processes. Traditional statistical methods (such as SARIMA and Kalman filtering) are based on linear assumptions and steady-state distributions, making it difficult to effectively capture complex nonlinear interactions and multi-scale features in the marine environment. Early deep learning methods (such as CNN-LSTM and CNN-BiGRU-Attention) can handle nonlinear relationships, but their recurrent neural network structures suffer from gradient vanishing and error accumulation problems. Furthermore, they typically employ a single paradigm of "channel dependence" or "channel independence," failing to collaboratively model long-term time-series dependencies and complex inter-variable interactions, and particularly lacking the ability to explicitly represent the multi-scale characteristics of trajectory data.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address or at least mitigate one or more of the above problems, a buoy trajectory prediction method based on stationary wavelets and wind-current-position interaction is provided.

[0006] To achieve the above objectives, according to the first aspect of this application, a buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction is provided, comprising the following steps: The trajectory sequence of buoys and the corresponding marine environmental field data within a historical period are acquired and standardized; wherein, the trajectory sequence of the buoys includes the location, i.e., longitude and latitude, and the marine environmental field data includes wind speed and current speed; Multi-scale wavelet context dynamic segmentation: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain multiple subsequences of different scales, and then multiple segmented token sequences are obtained by dynamic segmentation. Wind-flow-location interaction feature extraction: For each segmented token sequence after feature mapping, features are extracted using a wind-current-location interaction feature extraction mechanism; The wind-flow-location interaction feature extraction mechanism includes at least a temporal branch, a variable branch, and an interaction branch. For the output after dynamic segmentation, a query vector Q, a key vector K, and a value vector V are first generated, and then the output is split into a temporal branch and a variable branch for inference calculation. The temporal branch calculates attention in the time dimension and introduces a learnable span mask. The variable branch calculates the interaction correlation weights between wind speed, flow speed, and location. The interaction branch weighted and fused the outputs of the temporal branch and the variable branch. Multi-module fusion and trajectory coordinate decoding: The interactive features output by the wind-current-position interaction feature extraction mechanism are reconstructed using stationary wavelet inverse transform, and finally output as predicted values ​​of the buoy's trajectory coordinates for future time periods after linear layer mapping.

[0007] To achieve the above objectives, according to a second aspect of this application, a buoy trajectory prediction system based on stationary wavelet and wind-current-position interaction is provided, the buoy trajectory prediction system comprising: Data acquisition and preprocessing module: Acquires the trajectory sequence of buoys and the corresponding marine environmental field data within a historical period and performs standardization processing; wherein, the trajectory sequence of the buoys includes the location, i.e., longitude and latitude, and the marine environmental field data includes wind speed and current speed; Multi-scale wavelet context dynamic segmentation module: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain multiple subsequences of different scales, and then multiple segmented token sequences are obtained through dynamic segmentation. Wind-flow-position interaction feature extraction module: Wind-flow-position interaction feature extraction: For each segmented token sequence after feature mapping, features are extracted using a wind-current-location interaction feature extraction mechanism; The wind-flow-location interaction feature extraction mechanism includes at least a temporal attention branch, a variable attention branch, and an interaction branch; wherein, the temporal attention branch calculates attention in the time dimension and introduces a learnable span mask, the variable attention branch calculates the interaction correlation weights between wind speed, flow speed, and location, and the interaction branch weightedly fuses the outputs of the temporal attention branch and the variable attention branch; Multi-module fusion and output module: used for multi-module fusion and trajectory coordinate decoding: the interaction features output by the wind-current-position interaction feature extraction mechanism are reconstructed using stationary wavelet inverse transform, and finally output as the predicted trajectory coordinates of the buoy in the future time period after linear layer mapping.

[0008] To achieve the above objectives, in accordance with a third aspect of this application, a computer-readable storage medium is provided storing a computer program, which, when executed by a processor, is used to implement the buoy trajectory prediction method based on stationary wavelets and wind-current-position interaction as described above.

[0009] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art: In this application, prediction accuracy is significantly improved: by introducing wavelet multi-scale decomposition and dynamic segmentation, the model can explicitly capture trends and details in the trajectory, effectively addressing non-stationarity. Modeling capabilities are enhanced: the innovative wind-current-position interactive attention mechanism breaks through the limitations of a single modeling paradigm, enabling simultaneous and effective learning of long-term time-series dependencies and complex coupling relationships between wind, current, and position variables.

[0010] This application demonstrates superior robustness and generalization: each core module makes a key contribution to performance improvement, especially the combination of wavelet transform and dynamic segmentation, which enables the model to exhibit stable high performance across different sea states and prediction tasks. It also boasts high computational efficiency: thanks to the model's structured design, this application maintains high accuracy while achieving inference speeds far exceeding traditional numerical methods and some deep learning models, making it more suitable for real-time or near-real-time operational prediction scenarios.

[0011] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0012] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application. The illustrative embodiments and descriptions of the application are used to explain the application, but do not constitute an undue limitation of the application. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0013] In the attached diagram: Figure 1 This is a flowchart illustrating the buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction in this specific embodiment; Figure 2 This is a schematic diagram of the wind-flow-position interactive attention mechanism in this specific embodiment. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0015] Please see Figure 1 This application provides a buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction, including the following steps: Acquire the trajectory sequence of buoys and the corresponding marine environmental field data within a historical period; wherein, the trajectory sequence of the buoys includes the location, i.e., longitude and latitude, and the marine environmental field data includes wind speed and current speed; Multi-scale wavelet context dynamic segmentation: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain multiple subsequences of different scales, and then multiple segmented token sequences are obtained by dynamic segmentation. Wind-flow-location interaction feature extraction: For each segmented token sequence after feature mapping, features are extracted using a wind-current-location interaction feature extraction mechanism; The wind-flow-location interaction feature extraction mechanism includes at least a temporal attention branch, a variable attention branch, and an interaction branch; wherein, the temporal attention branch calculates attention in the time dimension and introduces a learnable span mask, the variable attention branch calculates the interaction correlation weights between wind speed, flow speed, and location, and the interaction branch performs a weighted fusion of the outputs of the temporal attention branch and the variable attention branch; Multi-module fusion and trajectory coordinate decoding: The interaction features output by the multiple wind-current-position interaction feature extraction mechanisms are reconstructed using stationary wavelet inverse transform, and finally output as predicted values ​​of the buoy's trajectory coordinates for future time periods after linear layer mapping.

[0016] It should be noted that the execution subject of the buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction in this embodiment is a buoy trajectory prediction system based on stationary wavelet and wind-current-position interaction. This system can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, vehicle-mounted electronic devices, wearable devices, etc., and non-mobile electronic devices can be servers and personal computers, etc., which are not specifically limited in this application. The following description uses a server as the execution subject to illustrate the buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction in this embodiment.

[0017] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] Multi-scale wavelet context dynamic segmentation (MSWCDS): Stationary wavelet transform submodule: like Figure 1 As shown, the buoy trajectory sequence is represented as follows: First, the buoy trajectory sequence (B,T,C) is input into the stationary wavelet transform (SWT) submodule. The SWT decouples the multivariate buoy trajectory data into components at different frequency scales, achieving a multi-scale representation of the time-series signal. Unlike the traditional discrete wavelet transform (DWT), SWT achieves equal-scale convolution by inserting holes in the convolution kernel, ensuring the output sequence length matches the input and exhibiting translation invariance. To better adapt to the statistical characteristics of the trajectory data, the wavelet filter is designed to be optimized during training. The decomposed signal includes approximate components and detail components, corresponding to low-frequency trend information and high-frequency local perturbation features, respectively, providing multi-scale feature input for the subsequent multi-path inference module. In this paper, the approximate component is represented as (B,T,C), and the detail components are represented as detail coefficient 1 (B,T,C) and detail coefficient 2 (B,T,C), etc.

[0019] The trajectory sequence of the buoy is represented as The input is fed into the stationary wavelet transform (SWT) submodule to obtain a multi-scale representation. Here, B represents the batch size, T represents the sequence length, and C indicates the number of channels (number of variables). The formal definition of the stationary wavelet transform (SWT) operation is as follows: (1); in, It is a learnable filter with a kernel size of k, and S is the decomposition level.

[0020] SWT generates a sequence of tokens. These token sequences contain time and frequency information, independently capturing information for each channel at different time scales.

[0021] The advantage of SWT lies in its ability to provide time-invariant decomposition while preserving the original temporal structure. This is achieved by avoiding downsampling at each decomposition level, thereby maintaining the upsampled length / size T. SWT also possesses translation invariance, enabling it to effectively capture local events at multiple scales.

[0022] The core of SWT is the mother wavelet function. and scaling function The discrete wavelet family can be represented as: (2a); (2b); In this design, parameter s controls the wavelet scale (expansion level), changing the width of the wavelet function through exponential scaling; parameter k determines the translation position, controlling the positioning of the wavelet function on the time axis. This parameterized design enables joint analysis of the signal in the time and frequency domains.

[0023] Multi-scale wavelet coefficient acquisition process:

[0024] Normalized time series of surface drift buoy trajectories (B,T,C) is decomposed using stationary wavelet transform, producing approximate coefficients at each level s. (B,T,C) and detail coefficients (B, T, C). In this paper, the drifting buoy trajectory data is a multivariate sequence. To clearly demonstrate the specific processing details of obtaining multi-scale wavelet coefficients using stationary wavelet transform, the following only shows the processing procedure for stationary wavelet transform of a univariate sequence; the processing for other variable sequences is the same as in this paper.

[0025] SWT uses two core filters for decomposition: a low-pass filter h derived from the scaling function. , used to extract the low-frequency trend components of the signal; high-pass filter g: derived from wavelet function ψ(t), used to capture the high-frequency detail features of the signal.

[0026] The mathematical expressions for low-pass and high-pass filters are as follows: (3a); (3b); Multi-level decomposition: From initial conditions Initially, the decomposition calculation of level s is achieved through the following iterative process: (4a); (4b); in, and It is an upsampled version of the original filters h and g, achieved by inserting between each pair of original filter coefficients. This is achieved by using zeros. This upsampling operation preserves the signal length, ensuring the time invariance of the transformation.

[0027] Coefficient reconstruction and output: The final iterative decomposition produces the ultimate approximation coefficients. and a set of wavelet coefficients The time-frequency characteristics of the signal are captured across different scales at each time point t. The complete set of wavelet coefficients can be expressed as: ; Here, S(t) represents the complete set of wavelet coefficients obtained through multi-scale decomposition. This set contains the time-frequency domain feature representations of the signal at different time scales, providing a multi-resolution analysis basis for subsequent trajectory prediction. It should be noted that: Representing the approximation coefficient, it mainly captures the low-frequency trend components of the signal, corresponding to the long-term evolution pattern and macroscopic laws in the time series. and These are all detail coefficients, specifically designed to extract high-frequency detail features of the signal, corresponding to short-term fluctuations and instantaneous change patterns in the sequence.

[0028] Spatial Context Dynamic Segmentation (SCDS) submodule: After wavelet decomposition, multiple feature inference paths are constructed based on the decomposition components at different scales, such as the approximation coefficients, detail coefficient 1, and detail coefficient 2 obtained above. Each branch consists of a wind-current-position interaction feature extraction module designed for trajectory data features. Specifically, the approximation coefficients and detail coefficients at each level generated by wavelet decomposition correspond to different frequency scales. The model establishes an independent feature inference branch for each scale to achieve hierarchical modeling of multi-scale dynamics. Before inputting the wind-current-position interaction feature extraction module, this paper proposes to introduce a spatial context dynamic segmentation mechanism. That is, when the scale is coarser, the signal changes relatively smoothly, so fewer segments are needed; while at finer scales, the signal fluctuations are more intense, and more segments are used accordingly to refine the local modeling. Finally, after the spatial context dynamic segmentation submodule, the segmented approximation coefficients (B,N1,C1'), segmented detail coefficient 1 (B,N2,C2'), and segmented detail coefficient 2 (B,N3,C3') are obtained.

[0029] The lack of semantic relevance in time series data is a key criticism of Transformer-based prediction methods by linear models. Therefore, we will use an output signal S with shape (B, T, C) i (t) (wherein) The approximation coefficients, detail coefficients 1 and 2 are segmented into segments of different lengths. This method not only reduces computational complexity when inputting into subsequent attention-based feature extraction modules, but also preserves semantic relationships along the time dimension.

[0030] A key contribution of this method is the use of a dynamic segmentation strategy instead of a fixed-length strategy. Specifically, the segment length is adjusted according to the scale: when the scale is coarser, the signal changes relatively smoothly, so fewer segments are needed; while at finer scales, the signal fluctuations are more dramatic, so more segments are used to refine the local modeling, ensuring that each segment contains approximately the same amount of information, thereby promoting more effective feature extraction.

[0031] Mathematical modeling and algorithm implementation: Let the first The output signal of wavelet decomposition is Its shape is Break it down into If there are segments, then the length of each segment is: (5); in, T and These represent the length of each individual segment, the total duration, and the output signal, respectively. The total number of segments.

[0032] Segmented representation and feature reshaping: The k-th segment is defined as: (6); Finally, each segment is flattened and converted into a token, thus transforming the signal shape into... ,in: (7); Therefore, the approximation coefficient, detail coefficient 1, and detail coefficient 2 are respectively , and .

[0033] Final output: The segmented token sequence after dynamic segmentation is as follows:

[0034] This dynamic segmentation strategy offers three advantages over traditional methods: first, it achieves information density equalization by adapting to local signal characteristics; second, it optimizes computational efficiency while preserving semantic relationships; and finally, it provides more representative token inputs for subsequent attention mechanisms. Particularly when processing ocean buoy trajectory data, this design effectively addresses the common non-stationary characteristics in trajectory sequences, laying a solid foundation for accurate prediction.

[0035] The attention module used in the wind-flow-location interaction feature extraction step has a dual-branch structure, including: The wind-flow-position interaction feature extraction module comprises three core sub-modules: a wind-flow-position interaction attention sub-module and two segmented embedding sub-modules located at its input and output ends, respectively. This sandwich-like structural design ensures effective feature transformation and dimensional alignment at different processing stages.

[0036] Seg-Embedding submodule (input): Segmented signal The input is fed into the segment embedding submodule, which consists of two linear layers and a GELU activation function. The two linear layers are located at opposite ends of the GELU function, as shown below. Figure 1 As shown. It should be noted that the parameters of the segment embedding submodules in the three paths are configured independently according to the characteristics of their respective segment embedding features. (Independent Seg-Embedding submodule parameters are configured for the three data sources: wind field features, flow field features, and location features).

[0037] Wind-Flow-Position Interactive Attention (WCP) submodule: The wind-flow-position interaction attention submodule is mainly used for learning association relationships. Its detailed structure is as follows: Figure 2 As shown, it employs computational logic similar to traditional attention mechanisms, targeting the input... First, the query vector Q, key vector K, and value V matrix are generated through a fully connected layer. Then, the data is split into two parallel branches, namely the time-series branch and the variable branch, for inference calculations.

[0038] Timing branches: The operation is performed directly on the time dimension by calculating the query vector Q and the key vector K (of shape...). Scaling dot product attention between ) ( (corresponding to the dimension of the attention head) to capture temporal dependencies, where , Let be the number of attention heads and the number of embedding dimensions, respectively. The formula is expressed as: (8); Wherein, scaling factor (Corresponding to the dimension of the attention head) Normalize the dot product, span mask. This is used to filter the attention span, while the final weighted sum with the value vector V captures the temporal dependencies of the local span. The span mask is composed of learnable parameters. This is introduced to limit the computational scope of the attention weights. The mask is defined as a piecewise function, where the soft mask component takes the form of a sigmoid-like function, specifically expressed as follows: (9); in, This represents the distance between position r in the sequence and the current position t, and S is the adaptive span of attention.

[0039] Variable branch: First, transpose Q and K to get the input. Transpose The shape of the variables is used to calculate the correlation between them using the following formula: (10); in, and ; V is then aggregated based on the variable weights: (11); Where Q′ and K′ represent the transposed query and key (shaped as follows) ), The scaling factor represents the variable dimension (e.g., feature dimension), while the correlation after softmax normalization is used to weight the value matrix V to emphasize the relationship between variables.

[0040] The outputs of the time branch and the variable branch are multiplied by the learnable weights, respectively. and Then, by summing and combining, the calculation formula can be expressed as: ; Finally, the feature tensor undergoes an output transformation through a linear layer, maintaining the same shape as the input tensor. , where d is the embedding dimension. This design ensures that the dimensional structure of features remains compatible with subsequent processing modules after being enhanced by the attention mechanism.

[0041] Seg-Embedding submodule (output): The shape tensor of the above output The data is passed to a second segment embedding submodule, which has the same architectural design as the first. Through processing by this submodule, the dimension output is mapped to... The last dimension has a size of 2, corresponding to two target variables along the longitude and latitude directions, respectively. This dimensional mapping process realizes the transformation from a high-dimensional feature space to specific geographic coordinates, providing a direct output interface for trajectory prediction. The separation of longitude and latitude helps the model independently learn the motion patterns in both directions, improving prediction accuracy.

[0042] This dual-branch design effectively decouples temporal modeling from variable relationship learning: the temporal branch focuses on capturing the temporal evolution patterns in the sequence, while the variable branch focuses on exploring the interactions between variables. The outputs of the two branches are adaptively fused through learnable weights to ultimately form a comprehensive feature representation, providing a richer feature foundation for ocean buoy trajectory prediction.

[0043] Multimodal fusion: The outputs from the three-branch multi-level wavelet decomposition are first integrated through a stacking operation to generate an output tensor of shape (B, N, lev, 2). This tensor is then reconstructed by the SWT reconstruction submodule into a shape of (B, T, 2). This reconstruction process completes the transformation from multi-scale feature representation to the final trajectory coordinates: Stacking operations: effectively integrate features from different branches and scales; since each scale branch uses the same computational process, all branch outputs have a uniform shape. , as well as Stack them according to the path dimension to get Where lev = s + 1, that is: ; Inverse wavelet transform: The stationary inverse wavelet transform is used to restore the multi-scale coefficient representation to time-domain coordinates. Let the decomposed and processed information tensor be... Then it is necessary to start from arrive Reverse reconstruction, targeting the first Layer, using a reconstruction approximation filter dual to the decomposition filter. and detail filter The information coefficients are reconstructed in reverse, as shown in the following formula: (12); in, This represents the convolution operation. These are the approximate coefficients obtained after filtering at the lev-th layer. These are the detail coefficients obtained after filtering at the lev layer.

[0044] Iterate in sequence The final output of the layer is: (13); Through the above steps, the segmented representation is reconstructed into a complete temporal trajectory.

[0045] Based on the same inventive concept, this application also provides a buoy trajectory prediction system based on stationary wavelet and wind-current-position interaction, the buoy trajectory prediction system comprising: Data acquisition and preprocessing module: Acquires the trajectory sequence of buoys and the corresponding marine environmental field data within a historical period and performs standardization processing; wherein, the trajectory sequence of the buoys includes the location, i.e., longitude and latitude, and the marine environmental field data includes wind speed and current speed; Multi-scale wavelet context dynamic segmentation module: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain multiple subsequences of different scales, and then multiple segmented token sequences are obtained through dynamic segmentation. Wind-flow-position interaction feature extraction module: Wind-flow-position interaction feature extraction: For each segmented token sequence after feature mapping, features are extracted using a wind-current-location interaction feature extraction mechanism; The wind-flow-location interaction feature extraction mechanism includes at least a temporal attention branch, a variable attention branch, and an interaction branch; wherein, the temporal attention branch calculates attention in the time dimension and introduces a learnable span mask, the variable attention branch calculates the interaction correlation weights between wind speed, flow speed, and location, and the interaction branch performs a weighted fusion of the outputs of the temporal attention branch and the variable attention branch; Multi-module fusion and output module: used for multi-module fusion and trajectory coordinate decoding: the interaction features output by the wind-current-position interaction feature extraction mechanism are reconstructed using stationary wavelet inverse transform, and finally output as the predicted trajectory coordinates of the buoy in the future time period after linear layer mapping.

[0046] Based on the same inventive concept, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the buoy trajectory prediction method based on stationary wavelets and wind-current-position interaction as described above.

[0047] The program product of this application for implementing the above method may employ a portable compact disk read-only memory and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, system, or device.

[0048] It should be noted that a computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, system, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0049] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction, characterized in that, Includes the following steps: The trajectory sequence of buoys and the corresponding marine environmental field data within a historical period are acquired and standardized; wherein, the trajectory sequence of the buoys includes the location, i.e., longitude and latitude, and the marine environmental field data includes wind speed and current speed; Multi-scale wavelet context dynamic segmentation: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain multiple subsequences of different scales, and then multiple segmented token sequences are obtained by dynamic segmentation. Wind-flow-location interaction feature extraction: For each segmented token sequence after feature mapping, features are extracted using a wind-current-location interaction feature extraction mechanism; The wind-flow-location interaction feature extraction mechanism includes at least a temporal branch, a variable branch, and an interaction branch. For the output after dynamic segmentation, a query vector Q, a key vector K, and a value vector V are first generated, and then the output is split into a temporal branch and a variable branch for inference calculation. The temporal branch calculates attention in the time dimension and introduces a learnable span mask. The variable branch calculates the interaction correlation weights between wind speed, flow speed, and location. The interaction branch weighted and fused the outputs of the temporal branch and the variable branch. Multi-module fusion and trajectory coordinate decoding: The interactive features output by the wind-current-position interaction feature extraction mechanism are reconstructed using stationary wavelet inverse transform, and finally output as predicted values ​​of the buoy's trajectory coordinates for future time periods after linear layer mapping.

2. The method according to claim 1, characterized in that, The normalized buoy trajectory sequence is decomposed using a stationary wavelet transform to obtain multiple subsequences of different scales. Then, multiple segmented token sequences are obtained through dynamic segmentation, including: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain a low-frequency approximate coefficient sequence representing the trend and multiple high-frequency detail coefficient sequences representing the detailed fluctuations.

3. The method according to claim 2, characterized in that, The dynamic segmentation includes: The approximation coefficient sequences and detail coefficient sequences at each scale are segmented into token sequences of different lengths. The segment length is dynamically adjusted based on the local characteristics of the sequence: increasing the length in smooth regions and decreasing the length in fluctuating regions. Specifically, this includes: Let the first The output signal of wavelet decomposition is , shape Where B represents the batch size, T represents the time series length of the complete trajectory, and C is the number of channels for the variable. Split into If there are segments, then the length of each segment is: (5)； in, T and These represent the length of each individual segment, the total duration, and the output signal, respectively. The total number of segments.

4. The method according to claim 3, characterized in that, The feature extraction of each segmented token sequence after feature mapping, through the wind-current-location interaction feature extraction mechanism, includes: The segmented token sequence is segmented and embedded using a segmented embedding mechanism; wherein the segmented embedding mechanism includes at least two linear layers and a GELU activation function. Feature extraction is performed using a wind-current-position interaction feature extraction mechanism; The extracted features are then processed through the segmented embedding mechanism, where the last dimension has a size of 2, corresponding to two target variables along the longitude and latitude directions, respectively.

5. The method according to claim 4, characterized in that, The temporal branch computes attention along the time dimension, introducing a learnable span mask, including: Calculate the scaled dot product attention between the query vector Q and the key vector K; Introducing span mask The attention weights are limited by multiplying the attention weights by the scaled dot product; the span mask It is defined as a piecewise function, in which the mask component takes the form of a Sigmoid function; After applying Softmax, the result is multiplied by the value vector V to ultimately capture the time dependency of the local span.

6. The method according to claim 5, characterized in that, The variable branch calculation includes the interaction correlation weights between wind speed, current velocity, and location, including: Perform a transpose operation on the query vector Q and the key vector K. Input Transpose The shape, in which These represent the batch size, the number of heads in the multi-head attention mechanism, the time step, and the feature dimension, respectively. Transpose is the process of transposing the input... The last two dimensions Transpose This information is then input into formula (10) to calculate the correlation between variables. (10)； in, and , A scaling factor that represents a specific dimension of the variable; The value vector V is then aggregated based on the variable weights: （11）； Where Q′ and K′ represent the transposed query and key, with the shape as follows: The correlation after softmax normalization is used to weight the value vector V to emphasize the relationship between variables.

7. The method according to claim 6, characterized in that, The interaction branch weighted and fused the outputs of the temporal attention branch and the variable attention branch, including: The outputs of the time branch and the variable branch are multiplied by the learnable weights, respectively. and Then, by summing and combining, it can be expressed as: ； in, Indicates the output of the interactive branch. This represents the output of the timing branch. Indicates the output of the variable branch; Finally, the output tensor of the interaction branch undergoes an output transformation through a linear layer, maintaining the same shape as the input tensor. Where B represents the batch size and d is the embedding dimension.

8. The method according to claim 3, characterized in that, The step of reconstructing the interaction features output by the multiple wind-current-position interaction feature extraction mechanisms using stationary wavelet inverse transform, and finally outputting the predicted trajectory coordinates of the buoy for future time periods after linear layer mapping includes: The interaction features output from multiple wind-flow-position interaction feature extraction mechanisms are first stacked along the channel dimension to generate an integrated output tensor, which retains the shape (B, N, lev, 2), where B represents the batch size, N represents the number of trajectory segments, lev represents the wavelet decomposition level, and 2 represents the two-dimensional coordinates, i.e., longitude and latitude. The output tensor is then transformed from (B, N, lev, 2) to (B, T, 2) by a stationary wavelet inverse transform, where T represents the temporal length of the complete trajectory. The segmented trajectories are then continuously reconstructed in the temporal domain to obtain complete and smooth temporal trajectory data.

9. A buoy trajectory prediction system based on stationary wavelet and wind-current-position interaction, characterized in that, The buoy trajectory prediction system includes: Data acquisition and preprocessing module: Acquires the trajectory sequence of buoys and the corresponding marine environmental field data within a historical period and performs standardization processing; wherein, the trajectory sequence of the buoys includes the location, i.e., longitude and latitude, and the marine environmental field data includes wind speed and current speed; Multi-scale wavelet context dynamic segmentation module: The normalized buoy trajectory sequence is decomposed by stationary wavelet transform to obtain multiple subsequences of different scales, and then multiple segmented token sequences are obtained through dynamic segmentation. Wind-flow-position interaction feature extraction module: Wind-flow-position interaction feature extraction: For each segmented token sequence after feature mapping, features are extracted using a wind-current-location interaction feature extraction mechanism; The wind-flow-location interaction feature extraction mechanism includes at least a temporal attention branch, a variable attention branch, and an interaction branch; wherein, the temporal attention branch calculates attention in the time dimension and introduces a learnable span mask, the variable attention branch calculates the interaction correlation weights between wind speed, flow speed, and location, and the interaction branch weightedly fuses the outputs of the temporal attention branch and the variable attention branch; Multi-module fusion and output module: used for multi-module fusion and trajectory coordinate decoding: the interaction features output by the wind-current-position interaction feature extraction mechanism are reconstructed using stationary wavelet inverse transform, and finally output as the predicted trajectory coordinates of the buoy in the future time period after linear layer mapping.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it is used to implement the buoy trajectory prediction method based on stationary wavelet and wind-current-position interaction as described in any one of claims 1-8.

Images