This study proposes an improved hybrid approach for traffic prediction of autonomous underwater vehicles (AUVs) to address the challenges posed by complex deep-sea conditions and obstacles. Traditional models often suffer from prediction lag and reduced accuracy due to inadequate integration of environmental factors. To overcome these limitations, we introduce a novel G-GRU multistep prediction network that combines an enhanced polynomial fitting algorithm with a gated recurrent unit (GRU) neural network. For sin-gle-step prediction, the improved polynomial fitting algorithm adaptively se-lects terms and selection points to enhance precision, whereas for multistep prediction, the GRU network, with its reduced parameter set and shorter training duration, effectively captures long-term dependencies. This method-ology demonstrates superior prediction accuracy and reduced lag under both raster and latitude–longitude maps. The experimental results show that the G-GRU network significantly outperforms the traditional GRU model in terms of metrics such as the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), providing robust sup-port for AUV short-term positioning and long-term planning, thus advancing marine resource utilization and safeguarding national maritime interests.