Model and data-driven hydrodynamic identification and prediction for underwater gliders

Accurate real-time estimation and prediction of hydrodynamic parameters are critical for precise control of underwater gliders (UGs) in ocean observation missions. However, the time-varying oceanic environment introduces significant uncertainties, making it difficult for existing measurement methods to capture the inherent nonlinearities and spatiotemporal dependencies in hydrodynamic data. This study presents a comprehensive hydrodynamic identification, preprocessing, and prediction framework driven by model and data-based approaches. Initially, the hydrodynamic data are identified by integrating the UG motion model with sea trial data using an internal penalty function. Next, the identified hydrodynamic data are decomposed with an optimized variational mode decomposition method, and the resulting intrinsic mode functions are clustered via hierarchical analysis to improve interpretability. Subsequently, a hybrid neural network combining convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and LSTM is developed to establish an offline-enhanced real-time hydrodynamic prediction model. Testing results show that the CNN-LSTM model enhanced with CNN-BiLSTM feature extraction significantly improves error control and noise robustness, achieving reductions of 26.25% in root mean square error and 27.74% in mean absolute error, compared to the baseline CNN-LSTM model. Finally, the trained enhanced CNN-LSTM model is applied to real-time hydrodynamic prediction during the sea trial, achieving a coefficient of determination of 0.9756 and 0.9839 for drag and lift data, respectively, while also demonstrating excellent performance across all other evaluation metrics. Briefly, this study contributes valuable insights into the processing and measurement of complex UG data in dynamic marine environments.