What question did this study set out to answer?

The main aim is to develop a high-precision model for predicting ocean sound speed profiles using a residual LSTM architecture.

April 28, 2026Intelligent Marine Technology and Systems0 citationsOpen Access

Sound speed profile prediction using a multi-parameter, single-layer residual LSTM approach

ZYZetao YangChang'an University TXTianhe XuChina Academy of Space Technology JWJunting WangChina Academy of Space Technology

Key Points

The main aim is to develop a high-precision model for predicting ocean sound speed profiles using a residual LSTM architecture.
Utilized a multi-parameter single-layer residual LSTM for sound speed prediction.
Compared the performance of the Res-LSTM with GRNN, RBF, and standard LSTM models.
Evaluated using Copernicus Marine reanalysis data.
Res-LSTM achieved RMSE improvements of 62.4% over GRNN, 46.0% over RBF, and 32.9% over LSTM.
Res-LSTM consistently showed the lowest RMSE across various time scales for sound speed prediction.
Ablation studies confirmed that residual connections significantly enhance model performance.

Abstract

Abstract Temporal variation errors in ocean sound speed constitute a critical factor affecting underwater navigation and positioning accuracy. However, real-time and large-scale measurement of ocean sound speed remains challenging. Therefore, the inversion of the sound speed profile (SSP) is essential for obtaining ocean sound speed field data at high spatiotemporal resolution. Because the ocean SSP inversion function is relatively complex, neural network-based inversion models demonstrate clear advantages over conventional SSP inversion methods. This study presents a multi-parameter, single-layer residual long short-term memory network (Res-LSTM) for SSP prediction. The proposed model enhances prediction accuracy and generalization capability in complex marine environments. It introduces residual connections within the traditional long short-term memory (LSTM) hidden layer, utilizing skip connections to form identity mappings. This architecture learns residuals between the inputs and outputs while integrating key influencing factors, such as temperature, salinity, and depth. The approach achieves high-precision SSP prediction through these technical innovations. To evaluate the performance of the Res-LSTM model, this study utilizes Copernicus Marine reanalysis data. The proposed model is systematically compared against benchmark methods, including generalized regression neural networks (GRNN), radial basis function networks (RBF), and standard LSTM architectures. Experimental results indicate that at four representative points, the root mean square error (RMSE) of the Res-LSTM decreased by 62.4%, 46.0%, and 32.9% compared with GRNN, RBF, and LSTM, respectively. Furthermore, in tests conducted across different time scales, the RMSE of Res-LSTM consistently remained the lowest. Ablation experiments further indicate that the introduction of residual connections is a key factor in the model’s performance improvement. Therefore, the Res-LSTM is an effective method for predicting SSP and can further improve prediction accuracy and stability.

问 AI

Bookmark

View Full Paper