Air-sea interface collaborative observation and short-term prediction method driven by multi-source heterogeneous data

Abstract

The invention belongs to the technical field of sea surface forecasting, and particularly relates to a multi-source heterogeneous data-driven air-sea interface collaborative observation and short-term prediction method, which is characterized in that a self-attention mechanism in an Informer model is utilized to establish correlation modeling between an input sequence and an output sequence, obtaining a global feature matrix; according to the method, a BiGRU model is used for capturing local time sequence dependence features in sequence data to obtain a local feature matrix, and the method has the advantages that a parallel prediction model based on Informer and BiGRU-GlobalAttention is used for training and reasoning fused data, and short-time prediction of key marine meteorological elements in a period of time in the future is achieved.

Description

Technical Field

[0001] This application belongs to the field of sea surface forecasting technology, specifically involving a multi-source heterogeneous data-driven method for collaborative observation and short-term forecasting of the air-sea interface. Background Technology

[0002] Currently, air-sea interface observation buoys mostly operate independently, measuring only a single parameter (such as waves or temperature), making it difficult to simultaneously acquire multi-dimensional elements such as ocean currents and meteorological data, and thus hindering the formation of an effective space-based collaborative observation network. Furthermore, the diverse sources of observational data (including buoy sensors, satellite remote sensing, and shore-based weather stations) exhibit significant heterogeneity in data type, sampling frequency, and data format, making efficient fusion and in-depth utilization difficult using traditional methods. Existing technical solutions typically place the computational burden on the buoy itself, making it difficult to support complex prediction models and meet the operational needs for high-precision, short-term forecasts of the marine meteorological environment.

[0003] In recent years, with the development of sensor and communication technologies, buoys integrating multiple sensors have emerged, and attempts have begun to utilize data fusion technology, combining external data sources such as satellite remote sensing and reanalysis data, to improve the comprehensiveness of observations. Existing similar technical solutions mostly focus on data processing from a single buoy or simple multi-source data assimilation; for example, using machine learning algorithms on the buoy or a single server to perform offline predictions based on historical data. These solutions lack the design of a collaborative network observation mechanism for spatially distributed multi-buoy systems, and have failed to construct a complete system for real-time fusion and short-term prediction of multi-source heterogeneous data based on powerful shore-based computing power. They exhibit significant bottlenecks in data processing efficiency and prediction timeliness. Summary of the Invention

[0004] To address the aforementioned problems, this application provides a method for air-sea interface data fusion and short-term forecasting based on multi-buoy collaborative observation and shore-based centralized processing, aiming to achieve more comprehensive, accurate, and timely perception and forecasting of the marine atmospheric environment. The technical solution is as follows: A multi-source heterogeneous data-driven method for collaborative observation and short-term forecasting of the air-sea interface includes the following steps: S1. Obtain the raw data, normalize and scale each element, and transform the original discrete observations into a regularized feature matrix, thereby constructing a unified multi-source heterogeneous dataset. S2. The self-attention mechanism in the Informer model is used to establish the correlation model between the input and output sequences. The multi-head attention mechanism can capture important relationships in long sequences and obtain the global feature matrix. The BiGRU model is used to capture local temporal dependency features in the sequence data and obtain the local feature matrix. S3. The prediction matrix Y obtained by concatenating the global feature matrix and the local feature matrix and mapping it through the fully connected layer has the same dimension as the short-term prediction target and directly represents the environmental quantity of the air-sea interface in the future.

[0005] Preferably, in step S1, for data from different sources such as buoys, satellites, and shore base stations, time unification alignment is first performed to ensure the consistency of sampling step size and time scale; then, spatial registration is completed through linear or spline interpolation to eliminate the offset caused by differences in observation location; then, each element is normalized and scaled to weaken the impact of differences in dimensions and excessive numerical span; the original discrete observations are transformed into regularized feature matrices to form a fusion sample set with an T×F structure, where T represents the time step size and F represents the feature dimension, thereby constructing a unified multi-source heterogeneous dataset.

[0006] Preferably, the Informer model includes an encoder and a decoder, wherein the decoder includes a feedforward layer, a multi-head attention layer, and a masked multi-head ProbSparse self-attention layer; the ProbSparse self-attention mechanism is used to establish the correlation model between the input sequence and the output sequence, wherein the multi-head attention mechanism can capture important relationships in long sequences and connect the self-attention outputs of each group to form the final prediction result.

[0007] Preferably, the global feature matrix is ​​obtained using the Informer model. The specific steps are as follows: S2-110. Construct a unified multi-source heterogeneous dataset matrix from the observed samples. Through three independent linear mappings, where The weight matrix yields the query, key, and value. For each query vector Calculate the sparsity metric: ; in L This indicates the sequence length, i.e., the number of key vectors. This refers to the dimensions of each head; This indicates the number of keys (Keys) of the j-th key. S2-111. After calculating the sparsity metric for all query vectors, select the one with the highest score. u Key query vectors u It is a constant derived from empirical constants c and sequence length L The determined parameters are given by the following formula: ; S2-112. Sparse attention approximation, in selecting... uAfter obtaining the key query vectors, a sparse attention matrix is ​​obtained. : ; in This indicates sparse attention output. Is it a selection The key vector matrix and V represent the value matrix, i.e., the input feature mapping; S2-113. For the query position selected by Top-u, directly... The vector corresponding to the row in the output is filled in; for the remaining unselected query positions, the average of all Value vectors is used as an approximation to form the final context representation Z; S2-114. Multi-head parallelism: Repeat steps S2-110–S2-114 for h heads and summarize them by a linear mapping: ; in W represents the output of the i-th attention head. O This represents the output weight matrix; S2-115. Fully Connected and Feedforward Networks: The decoded result is input into a fully connected layer to obtain the global temporal feature output. : ; Where Z represents the multi-head attention output, W1 and W2 represent the weight matrices, and b1 and b2 represent the bias vectors. This represents the final output feature of the Informer branch, with a size of B×T×d, where B is the number of channels.

[0008] Preferably, the data processing of BiGRU involves simultaneous forward and backward information processing at each time step. The forward GRU processes the information flow from the beginning of the sequence to the current time step, while the backward GRU processes the information flow in reverse from the end of the sequence to the current time step. The GRU in each direction controls the flow of information through update gates and forget gates, and updates the hidden state through candidate activation functions. Finally, the forward and backward hidden states are merged to form a representation that can simultaneously consider the past and future information of the sequence, thereby providing support for sequence prediction or classification tasks.

[0009] Preferably, GRU introduces update gates and reset gates to control the degree of retention and updating of the hidden state information from the previous time step; update gate The calculation formula is: ; Reset door The calculation formula is: ; Final hidden state The updated formula is: ; in, , The weight parameter matrix, For learnable parameter matrix, , For bias terms, Select the hidden state later.

[0010] Compared with the prior art, the beneficial effects of this application are as follows: 1) Enhanced observation capabilities: Through multi-buoy collaborative networking, synchronous observation of a wider range and more elements in the target sea area is achieved, overcoming the limitations of the observation capabilities of a single buoy.

[0011] 2) High-efficiency data processing capability: The Informer + BiGRU-GlobalAttention parallel prediction architecture adopted in this invention reduces the computational complexity from the traditional O(n log n) to O(n log n) through the ProbSparse self-attention mechanism. The computational cost is reduced to O(LlogL), significantly improving the processing efficiency of long-sequence data. Simultaneously, the bidirectional information flow design of BiGRU can consider both past and future information, further optimizing data processing speed. This achieves efficient data processing, enabling rapid response to changes in the marine environment and meeting real-time or near-real-time monitoring and forecasting needs.

[0012] 3) Multi-source collaboration: This invention integrates data from different sensors (such as ocean drifting buoys, satellite remote sensing, and weather stations) through multimodal data fusion, overcoming the limitations of a single data source. Simultaneously, multi-head self-attention and global attention mechanisms capture global temporal dependencies, while local attention mechanisms focus on key parts of the sequence, extracting the features most influential on prediction. The parallel architecture processes observation variables with different sampling rates and different physical meanings separately; after AWF fusion, system bias is significantly reduced, improving prediction accuracy. Attached Figure Description

[0013] Figure 1 Data flow diagram of a parallel prediction model based on Informer and BiGRU models; Figure 2 Here is a detailed structural diagram of the Informer branch; Figure 3 It is a bidirectional gated loop unit; Figure 4 This is the temperature prediction result. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] A multi-source heterogeneous data-driven method for collaborative observation and short-term forecasting of the air-sea interface includes the following steps: S1. For data from different sources such as buoys, satellites, and shore base stations, time unification and alignment are first performed to ensure the consistency of sampling step size and time scale. Then, spatial registration is completed through linear or spline interpolation to eliminate the offset caused by differences in observation location. Next, each element is normalized and scaled to reduce the impact of differences in dimensions and excessive numerical span. The original discrete observations are transformed into regularized feature matrices to form a fusion sample set with an T×F structure, where T represents the time step size and F represents the feature dimension, thereby constructing a unified multi-source heterogeneous dataset.

[0016] The raw data is obtained, and each element is normalized and scaled to transform the original discrete observations into a regularized feature matrix, thereby constructing a unified multi-source heterogeneous dataset. S2. The self-attention mechanism in the Informer model is used to establish the correlation model between the input and output sequences. The multi-head attention mechanism can capture important relationships in long sequences and obtain the global feature matrix. The BiGRU model is used to capture local temporal dependency features in the sequence data and obtain the local feature matrix. The Informer model consists of an encoder and a decoder. The decoder includes a feedforward layer, a multi-head attention layer, and a masked multi-head ProbSparse self-attention layer. The ProbSparse self-attention mechanism is used to establish the relationship model between the input sequence and the output sequence. The multi-head attention mechanism can capture important relationships in long sequences and connect the self-attention outputs of each group to form the final prediction result.

[0017] The global feature matrix is ​​obtained using the Informer model. The specific steps are as follows: S2-110. Construct a unified multi-source heterogeneous dataset matrix from the observed samples. Through three independent linear mappings, where The weight matrix yields the query, key, and value. For each query vector Calculate the sparsity metric: ; in L This indicates the sequence length, i.e., the number of key vectors. This refers to the dimensions of each head; This indicates the number of keys (Keys) of the j-th key. S2-111. After calculating the sparsity metric for all query vectors, select the one with the highest score. u Key query vectors u It is a constant derived from empirical constants c and sequence length L The determined parameters are given by the following formula: ; S2-112. Sparse attention approximation, in selecting... u After obtaining the key query vectors, a sparse attention matrix is ​​obtained. : ; in This indicates sparse attention output. Is it a selection The key vector matrix and V represent the value matrix, i.e., the input feature mapping; S2-113. For the query position selected by Top-u, directly... The vector corresponding to the row in the output is filled in; for the remaining unselected query positions, the average of all Value vectors is used as an approximation to form the final context representation Z; S2-114. Multi-head parallelism: Repeat steps S2-110–S2-114 for h heads and summarize them by a linear mapping: ; in W represents the output of the i-th attention head. O This represents the output weight matrix; S2-115. Fully Connected and Feedforward Networks: The decoded result is input into a fully connected layer to obtain the global temporal feature output. : ; Where Z represents the multi-head attention output, W1 and W2 represent the weight matrices, and b1 and b2 represent the bias vectors. This represents the final output feature of the Informer branch, with a size of B×T×d, where B is the number of channels.

[0018] The data processing of BiGRU involves simultaneous forward and backward information processing at each time step. The forward GRU processes the information flow from the beginning of the sequence to the current time step, while the backward GRU processes the information flow in reverse from the end of the sequence to the current time step. Each direction of the GRU controls the flow of information through update gates and forget gates, and updates the hidden state through candidate activation functions. Finally, the forward and backward hidden states are merged to form a representation that can simultaneously consider the past and future information of the sequence, thereby supporting sequence prediction or classification tasks.

[0019] GRU introduces update gates and reset gates to control the degree to which hidden state information from the previous time step is retained and updated; update gate The calculation formula is: ; Reset door The calculation formula is: ; Final hidden state The updated formula is: ; in, , The weight parameter matrix, For learnable parameter matrix, , For bias terms, Select the hidden state later.

[0020] S3. The prediction matrix Y obtained by concatenating the global feature matrix and the local feature matrix and mapping it through the fully connected layer has the same dimension as the short-term prediction target and directly represents the environmental quantity of the air-sea interface in the future.

[0021] To achieve the above objectives, the technical solution of the present invention is generally divided into two main parts: a collaborative observation layer and a shore-based processing center.

[0022] The collaborative observation layer consists of multiple observation node buoys deployed in the target sea area. Each buoy is equipped with various marine and meteorological sensors, which are responsible for collecting raw data such as sea surface temperature, salinity, ocean current speed, wave height, air temperature, air pressure, humidity, wind speed and wind direction, and transmitting the data to the shore-based processing center in real time through a wireless communication network.

[0023] The shore-based processing center is the core of the system, responsible for receiving and storing data transmitted from all observation nodes' buoys, and integrating external data sources such as satellite remote sensing and shore-based meteorological stations. For data from different sources such as buoys, satellites, and shore-based stations, the system first performs time alignment to ensure consistency between the sampling step size and time scale. Then, spatial registration is completed through linear or spline interpolation to eliminate offsets caused by differences in observation locations. Next, each element is normalized and scaled to mitigate the impact of dimensional differences and excessively large numerical spans. After these steps, the original discrete observations are transformed into regularized feature matrices, forming a fused sample set with an T×F structure (where T represents the time step and F represents the feature dimension), thus constructing a unified multi-source heterogeneous dataset, providing standardized input for subsequent feature extraction, model training, and short-term prediction. Subsequently, the center uses an adaptive weighted fusion algorithm (AWF) to preprocess and fuse this dataset to eliminate system bias and heterogeneity. Finally, a parallel prediction model based on Informer and BiGRU-GlobalAttention is used to train and infer the fused data, enabling short-term predictions of key marine meteorological elements for a future period. The model data flow diagram is shown below. Figure 1 As shown.

[0024] The Informer branch consists of two parts: an encoder and a decoder. The decoder includes a feedforward layer, a multi-head attention layer, and a masked multi-head ProbSparse self-attention layer, as shown in the specific structure below. Figure 2 As shown. Its core idea is to use the ProbSparse self-attention mechanism to establish the relationship model between the input sequence and the output sequence. The multi-head attention mechanism can capture important relationships in long sequences and connect the self-attention outputs of each group to form the final prediction result.

[0025] ProbSparse Self-Attention is a sparse attention mechanism proposed by Informer models for long sequences. Through a three-step process of "measuring, filtering, and approximation," it transforms the dense attention of Transformer into sparse attention, making long sequence training feasible while minimizing performance loss. The core goal is to reduce the computational complexity of traditional Transformer models from O(L²) to O(LlogL) with almost no loss of accuracy. The original discrete observations are transformed into regularized feature matrices, forming an fused sample set with an T×F structure (where T represents the time step and F represents the feature dimension), thus constructing a unified multi-source heterogeneous dataset.

[0026] The input to the BiGRU-GlobalAttention branch is the same as that of the Informer branch, both constructing a unified multi-source heterogeneous dataset matrix from the observed samples. However, focusing on different feature dimensions, this branch emphasizes the fluctuation characteristics of elements within a local time window. The Bidirectional Gated Recurrent Unit (BiGRU) is used to capture local temporal dependency features in sequence data, and its structure is as follows: Figure 3 As shown.

[0027] This invention overcomes the limitations of single data sources by integrating data from different sensors (such as ocean drifting buoys, satellite remote sensing, and weather stations) through multimodal data fusion. Simultaneously, multi-head self-attention and global attention mechanisms capture global temporal dependencies, while local attention mechanisms focus on key parts of the sequence, extracting the features most influential on prediction. The parallel architecture processes observation variables with different sampling rates and physical meanings separately; after AWF fusion, system bias is significantly reduced, improving prediction accuracy.

Claims

1. A multi-source heterogeneous data-driven method for collaborative observation and short-term forecasting of the air-sea interface, characterized in that, Includes the following steps: S1. Obtain the raw data, normalize and scale each element, and transform the original discrete observations into a regularized feature matrix, thereby constructing a unified multi-source heterogeneous dataset. S2. The self-attention mechanism in the Informer model is used to establish the relationship model between the input sequence and the output sequence. The multi-head attention mechanism can capture important relationships in long sequences and obtain the global feature matrix. The BiGRU model is used to capture local temporal dependency features in sequence data, and the local feature matrix is ​​obtained. S3. The prediction matrix Y obtained by concatenating the global feature matrix and the local feature matrix and mapping it through the fully connected layer has the same dimension as the short-term prediction target and directly represents the environmental quantity of the air-sea interface in the future.

2. The method for collaborative observation and short-term prediction of the air-sea interface driven by multi-source heterogeneous data according to claim 1, characterized in that, Step S1: For data from different sources such as buoys, satellites, and shore base stations, time unification and alignment are first performed to ensure consistency between the sampling step size and the time scale. Then, spatial registration is completed through linear or spline interpolation to eliminate the offset caused by differences in observation locations. Next, each element is normalized and scaled to reduce the impact of differences in dimensions and excessive numerical spans. The original discrete observations are transformed into regularized feature matrices to form a fusion sample set with an T×F structure, where T represents the time step size and F represents the feature dimension, thereby constructing a unified multi-source heterogeneous dataset.

3. The method for collaborative observation and short-term prediction of the air-sea interface driven by multi-source heterogeneous data according to claim 1, characterized in that, The Informer model consists of an encoder and a decoder. The decoder includes a feedforward layer, a multi-head attention layer, and a masked multi-head ProbSparse self-attention layer. The ProbSparse self-attention mechanism is used to establish the relationship model between the input sequence and the output sequence. The multi-head attention mechanism can capture important relationships in long sequences and connect the self-attention outputs of each group to form the final prediction result.

4. The method for collaborative observation and short-term prediction of the air-sea interface driven by multi-source heterogeneous data according to claim 1, characterized in that, The global feature matrix is ​​obtained using the Informer model. The specific steps are as follows: S2-110. Construct a unified multi-source heterogeneous dataset matrix from the observed samples. Through three independent linear mappings, where The weight matrix yields the query, key, and value. For each query vector Calculate the sparsity metric: ； in L This indicates the sequence length, i.e., the number of key vectors. This refers to the dimensions of each head; This represents the number of keys (Keys) of the j-th key. S2-111. After calculating the sparsity metric for all query vectors, select the one with the highest score. u Key query vectors, u It is a constant derived from empirical constants c and sequence length L The determined parameters are given by the following formula: ； S2-112. Sparse attention approximation, in selecting... u After obtaining the key query vectors, a sparse attention matrix is ​​obtained. : ； in This indicates sparse attention output. Is it a selection The key vector matrix and V represent the value matrix, i.e., the input feature mapping; S2-113. For the query position selected by Top-u, directly... Fill the corresponding row of the vector into the output; For the remaining unselected query positions, the average of all Value vectors is used as an approximation to form the final context representation Z; S2-114. Multi-head parallelism: Repeat steps S2-110–S2-114 for h heads and summarize them by a linear mapping: ； ； in W represents the output of the i-th attention head. O This represents the output weight matrix; S2-115. Fully Connected and Feedforward Networks: The decoded result is input into a fully connected layer to obtain the global temporal feature output. : ； ； ； Where Z represents the multi-head attention output, W1 and W2 represent the weight matrices, and b1 and b2 represent the bias vectors. This represents the final output characteristic of the Informer branch. The output feature size is B×T×d, where B is the number of channels.

5. The method for collaborative observation and short-term prediction of the air-sea interface driven by multi-source heterogeneous data according to claim 1, characterized in that, The data processing of BiGRU involves simultaneous forward and backward information processing at each time step. The forward GRU processes the information flow from the beginning of the sequence to the current time step, while the backward GRU processes the information flow in reverse from the end of the sequence to the current time step. Each direction of the GRU controls the flow of information through update gates and forget gates, and updates the hidden state through candidate activation functions. Finally, the forward and backward hidden states are merged to form a representation that can simultaneously consider the past and future information of the sequence, thereby supporting sequence prediction or classification tasks.

6. The method for collaborative observation and short-term prediction of the air-sea interface driven by multi-source heterogeneous data according to claim 5, characterized in that, GRU introduces update gates and reset gates to control the degree to which hidden state information from the previous time step is retained and updated; update gate The calculation formula is: ； Reset door The calculation formula is: ； Final hidden state The updated formula is: ； in, , The weight parameter matrix, For learnable parameter matrix, , For bias terms, Select the hidden state later.

Images