Abstract
Introduction Ocean wave conditions forecasting is crucial for reducing wave-related disasters and enhancing prevention and mitigation capabilities in China's coastal regions. Methods This study develops a convolutional neural network (CNN) model optimized by a random search algorithm to predict significant wave height (Hs) and mean wave period (Tm). Using SWAN nearshore wave simulation data as input, the model analyzes the impact of different historical input step lengths and various forecast lead times on prediction performance. It is applied to the Yantai Fishing Zone, China. Results The optimal input step length for the model is 3 hours, achieving correlation coefficients (CC) of 0.9997 for Hs and 0.9969 for Tm. The mean absolute errors (MAE) are 0.0075 m and 0.0562 s, and the root mean square errors (RMSE) are 0.0149 m and 0.2014 s, respectively. Based on the 3-hour input step length, forecasts were conducted for lead times of 3, 6, 9, and 12 hours. As the forecast lead time increased, prediction accuracy declined, but the model still effectively captured the main trends of Hs and Tm. Discussion The model errors remain within an acceptable range, and its computational efficiency is significantly superior to traditional numerical methods. This demonstrates the model's good applicability in the study area, indicating its potential to effectively enhance fishery production efficiency and optimize fishing operation scheduling.