Abstract
Accurate prediction of three-dimensional (3-D) ocean flow fields is essential for understanding small-scale dynamic processes, improving environmental monitoring, and enhancing the performance of underwater observation and operation sys-tems. However, predictive modeling using coastal acoustic to-mography (CAT) data remains relatively unexplored. This study proposes a novel physics-constrained deep learning framework, termed PC-LSTM-MLP, to achieve the short-term prediction of 3-D ocean flow fields from CAT observations. The framework consists of three integrated components: a long short-term memory (LSTM) network for capturing temporal dependencies in range-averaged currents (RACs), a multilayer perceptron (MLP) for nonlinear mapping to spatial Fourier coefficients, and a physics-constrained (PC) module that imposes a horizontal incompressibility constraint to enhance the accuracy of the predicted 3-D flow fields. Experimental results on real CAT data demonstrate that the proposed PC-LSTM-MLP framework effec-tively captures complex spatiotemporal dynamics and provides accurate predictions up to 25 minutes ahead. The model achieves an average mean absolute error (MAE) of 0.0241 m/s and an average root mean square error (RMSE) of 0.0269 m/s in 3-D flow field prediction. This research shifts CAT analysis from traditional inversion methods to intelligent, prediction-oriented modeling, underscoring the significant potential of deep learning in ocean flow field prediction.
Funding Information
  • National Natural Science Foundation of China (52071293)
  • Qiushi Eagle Program of Zhejiang University
  • Citrus Research and Development Foundation (RDF-24-02-085)