Development of a 2-D deep learning regional wave field forecast model based on convolutional neural network and the application in South China Sea

Abstract

Currently, the methods of wave prediction based on deep learning theory primarily focus on single-point wave prediction; however, two-dimensional (2-D) wave field prediction can help understand the overall wave situation in a certain area, which has practical value. Given the current situation, in which numerical wave forecasting requires vast computing resources and huge time cost, a 2-D deep learning regional wave field forecast model based on a convolutional neural network (CNN) is proposed to forecast the significant wave height (SWH) in the South China Sea. In this study, the random search algorithm was used to optimize the hyper-parameters of the CNN model with the SWH, 10 m u-component of wind (U10), and 10 m v-component of wind (V10) as input parameters. The 2-D correlation coefficient (R2) was used to evaluate the correlation between the wave field and the wind field, and a sensitivity analysis of 56 different working conditions with the optimal forecast model was performed to obtain the best input scheme. Five evaluation indicators were used to evaluate the accuracy and stability of the model. Three typical field positions were selected. Month-averaged and year-averaged wave field forecasts were studied to comprehensively evaluate the model forecast results. The results indicate that the existing models can not only accurately forecast the change in wave height along the timeline, but also provide a good estimation of the spatial wave height distribution in the 2-D wave field. SWH forecasts for lead time periods of 12 h, 24 h, 48 h, 72 h were performed using the optimal input scheme and the optimal model. The mean absolute percentage errors (MAPE) for these lead time periods were 8.55%, 12.95%, 16.85%, and 19.48%, respectively, which demonstrates the ability of the model to perform long-term forecasts.