TransWaveNet: Cross-structural time series prediction of wave height based on a simplified transformer

Accurate prediction of wave attenuation behind breakwaters is critical for coastal protection and offshore aquaculture operations. This study proposes TransWaveNet, a simplified model based on transformers, which aims to forecast future wave heights (12 h) behind breakwaters using historical wave height data (12 h) measured in front of breakwaters. The data are derived from synchronous wave monitoring data collected from a floating breakwater in Lianjiang County, China (15 months, hourly intervals), used to construct the TransWaveNet model. A systematic evaluation indicates that TransWaveNet outperforms LSTM (long short-term memory), ANN (artificial neural network), and Seq2Seq (sequence to sequence) baseline models in predicting significant wave height (Hs) over a 12-h period, achieving a prediction accuracy of R2 (R-squared)= 0.73, MSE (mean squared error) = 0.02 m2, and RMSE (Root Mean Square Error) = 0.142 m. The model revealed two key findings: first, the 12-h input sequence achieves an optimal balance between accuracy and computational efficiency; second, while effectively capturing the Hs decay pattern, the framework exhibits minor limitations at maximum wave height (Hmax). The framework provides a lightweight, data-driven solution for cross-breakwater wave forecasting in practical applications.