Physics-informed temporal convolutional network with auto-regressive residual for significant wave height prediction

Abstract

Significant wave height (SWH) serves as a critical metric for assessing marine conditions. While current neural network models have demonstrated considerable success in SWH prediction, their efficacy is frequently limited by the need for more data and the lack of interpretability. Addressing these challenges, we propose a physics-informed temporal convolutional network with auto-regressive residual (PITCN-AR) for SWH prediction, which effectively integrates spatiotemporal dependency information into the training process. First, the PITCN-AR framework utilizes temporal convolutional networks (TCN) as the backbone for the physics-informed neural network. This design adeptively captures the spatiotemporal dependencies of wave height data. To enhance training efficiency and introduce additional dependencies, we integrate an auto-regressive model into the residual block of TCN. Furthermore, TCN is embedded within a hybrid data-physics loss function, and a mini-batch learning strategy is implemented to optimize the performance of PITCN-AR. Finally, the efficacy and practicality of PITCN-AR are rigorously validated using both synthetic and real-world data. Comparative analysis reveals that our model outperforms several state-of-the-art methods, achieving a lower root mean square error and a higher coefficient of determination, thereby underscoring its better predictive accuracy and reliability.