Deep learning is rapidly becoming a key tool for wave forecasting. As a classic model in image processing, U-Net offers powerful capabilities for extracting deep information and thus holds great potential for intelligent, high-precision regional wave forecasting. This study focuses on winter waves in the North China Sea. We develop an intelligent regional wave-forecasting model based on the U-Net architecture, and this model can also be applied to ice-covered waters. Using this model, we design 54 experimental schemes with three forecast lead times (6, 12 and 24 h). These schemes combine three input variables: significant wave height, 10 m wind speed, and sea ice concentration. After training and verification, the U-Net model achieves mean absolute errors (MAE) and correlation coefficients (CC) of 0.08 m and 0.98 (for a 6 h lead time), 0.11 m and 0.96 (for a 12 h lead time), and 0.17 m and 0.96 (for a 24 h lead time). When sea ice is present, adding ice concentration information improves forecast accuracy. Taking Liaodong Bay as an example, the inclusion of ice information reduces the forecast mean significant wave height by up to 0.60 m, leading to a 67.7% increase in forecast accuracy. We also compare the U-Net model with a convolutional neural network (CNN). Over 90% of the North China Sea shows lower MAE values for U-Net, with average reductions of 0.05 m in the Bohai Sea and 0.07 m in the Yellow Sea. However, the U-Net model tends to underestimate extreme wave heights above 3 m. Therefore, future work should focus on improving forecasts for such extreme events.