Accurate and uncertainty-aware spatiotemporal modeling of wind speed fields is essential for optimizing wind energy operations and identifying suitable turbine deployment sites. This study proposes a data-driven generative framework based on a 3D-CNN combined with a \(\beta\) -Variational Autoencoder ( \(\beta\) -VAE) to produce probabilistic 1-hour-ahead wind field forecasts over the Korean Peninsula. The five-year ERA5 dataset is split into three years for training and two subsequent years for validation and testing. Three-dimensional convolutions are used to capture both spatial and temporal dependencies, and the parameter \(\beta\) controls the trade-off between reconstruction fidelity and latent space regularization. A sensitivity analysis indicates that \(\beta = 0.01\) provides a favorable balance for forecasting. Using Root Mean Square Error, Continuous Ranked Probability Score, Variogram Score, and Probability Integral Transform-based diagnostics, we evaluated the trained model and found that it reproduces the ensemble statistics and spatial dependence structures of the observed wind fields. The model also yields reasonably well-calibrated probabilistic forecasts. Compared to the ConvLSTM \(\beta\) -VAE and 2D-CNN \(\beta\) -VAE baselines, the 3D-CNN \(\beta\) -VAE provides comparable skill at a 1-hour lead time and noticeably better probabilistic forecast performance and spatial consistency at longer lead times. These results suggest that the 3D-CNN \(\beta\) -VAE could serve as a scalable tool for offshore wind energy resource assessment and short-term turbine operation planning.
