Research On Prediction Method Of Ocean Observation Data Based On Attention Mechanism

Ocean observation data prediction plays an important role in analyzing the marine environment and predicting major marine disasters.However,issues such as low quality,outliers,and missing values in ocean observation data can impact the accuracy of downstream applications.Traditional methods fail to meet the expected goals when dealing with ocean observation data.Attention mechanism has been widely used in various deep learning models to improve prediction accuracy and efficiency.Therefore,a deep learning algorithm based on attention mechanism can effectively mine valuable information from ocean observation data,and is crucial for marine environment analysis and major marine disaster warning.The research contents of this article are as follows:(1)This thesis proposes a single-channel ocean observation data prediction method based on attention mechanism and a real-time prediction network based on Transformer.The basic structure of the Transformer model was modified by omitting the embedding layer and Softmax layer used in natural language processing and classification tasks,respectively.The time encoding method was adopted to incorporate time information into the model.A sliding window was introduced in the Transformer model to efficiently predict ocean observation data by moving the window.The proposed algorithm can flexibly capture the correlation between different time positions of the input sequence and dynamically focus on the input sequence that has a strong correlation with the prediction vector.The experiment results show that the proposed model has accurate prediction ability for long-term dependent data,is less affected by variable-length historical input data,and has strong robustness to different data types.(2)This thesis also proposes a missing value prediction method for multi-channel ocean observation time-series scalar data based on a three-stage attention recurrent neural network.In the first stage,a convolutional attention module is used to refine the input sequence to enhance the model’s representational capacity for input sequences.In the second stage,a local spatial attention module is used to selectively capture the dynamic spatial correlation between different input sequences.In the third stage,a time attention module is used to adaptively capture the dynamic temporal correlation between different time intervals in the input sequence.This method overcomes the limitation of relying only on single-channel data to fill in missing values.(3)This thesis proposes a spatiotemporal fusion modeling and prediction method for multi-channel ocean observation data based on attention mechanism,and constructs a recurrent neural network model with multiple attention mechanisms as a red tide disaster warning model.The proposed model enhances the representation capability and memory of the input sequence,efficiently handles long-term dependent sequences,and accurately captures the dynamic spatiotemporal evolution mechanisms of various observation data.The experimental results demonstrate that the proposed method can train a prediction model from ocean observation data,accurately capture the features and evolution mechanisms of ocean observation data,and is of crucial significance for red tide disaster early warning.