A combined prediction model for depth-averaged current velocities of hybrid underwater profilers based on time series analysis

In persistent monitoring missions of hybrid underwater profilers (HUPs), accurate prediction of depth-averaged current velocities (DACVs) is essential for station-keeping and navigation. However, only a limited number of historical profiling cycles are available during early HUP operations, posing challenges for reliable DACV prediction. To address this, a combined prediction model based on wavelet transform and long short-term memory networks (WT-LSTM) is proposed for short-term DACV forecasting. DACVs observed by HUPs are formulated as a nonlinear autoregressive time series, then decomposed into multi-frequency sub-sequences using wavelet transform to capture long-term trends and high-frequency variations. Each sub-sequence is independently predicted using an LSTM network, and the final DACV prediction is obtained via wavelet reconstruction. The WT–LSTM model is evaluated using publicly available datasets and real sea-trial data from the South China Sea. Results show that, compared with baseline models, WT-LSTM achieves the highest R² values (0.9901/0.9815 for simulation, 0.9859/0.8810 for sea-trial), the narrowest confidence intervals, and the smallest RMSE (0.0124/0.0054 for simulation, 0.0091/0.0121 for sea-trial) and SD values. WT–LSTM consistently outperforms linear interpolation, RBFNN, and standalone LSTM models, with pronounced advantages when DACV samples are limited. These findings indicate that WT–LSTM provides a robust, practical solution for accurate DACV prediction in localized underwater observation, especially during early mission stages.