A parallel convolution attention and temporal sequence attention neural network approach for ocean current prediction incorporating spatial-temporal coupling mechanism

Long Bai, Lan Qiu, Jia Zheng, Yulin Zhang, Xiaohong Chen, Yuanxi Sun

Abstract: Accurate prediction of ocean currents is crucial for long-term planning and decision-making in the context of climate change and ocean navigation. Traditional machine learning and deep learning methods struggle to comprehensively extract spatial–temporal features of ocean current vectors and validate multi-step prediction performance. This paper proposes a parallel convolution attention and temporal sequence attention neural network (PCNN-TSA) approach for ocean current prediction incorporating a spatial–temporal coupling mechanism. It incorporates different attention mechanisms in each module to comprehensively extract key information. Experimental results demonstrate that the PCNN-TSA model outperforms state-of-the-art temporal and image class prediction network models, with RMSE errors of 0.0014 m2/s2<math><mrow is="true"><msup is="true"><mi is="true">m</mi><mn is="true">2</mn></msup><mo stretchy="true" is="true">/</mo><msup is="true"><mi is="true">s</mi><mn is="true">2</mn></msup></mrow></math> and 0.0045 m2/s2<math><mrow is="true"><msup is="true"><mi is="true">m</mi><mn is="true">2</mn></msup><mo stretchy="true" is="true">/</mo><msup is="true"><mi is="true">s</mi><mn is="true">2</mn></msup></mrow></math> for single-step and multi-step current prediction, respectively. The model effectively extracts spatial and temporal characteristics of the current vector and accurately captures dominant currents within the next 8 days. It also exhibits the ability of migration training, making it applicable to other prediction models characterized by spatial–temporal information.