Accurate offshore wind speed forecasting plays a crucial role in site selection, turbine layout, and energy yield estimation for wind farms. This study presents an enhanced data-driven prediction framework based on a Backpropagation (BP) neural network optimized by Particle Swarm Optimization (PSO), aiming to improve short-term wind speed prediction accuracy under data-limited conditions. The model is trained using hourly wind speed data from 2020 to 2022, collected at a coastal meteorological station in northeastern China. Two modeling strategies are implemented: (1) a multi-year unified training approach capturing long-term temporal dependencies, and (2) a seasonal decomposition strategy in which spring, summer, autumn, and winter data are modeled independently using dedicated BP-PSO models.

To evaluate forecast performance, model predictions are compared with measured wind speed from January to July 2023. The multi-year model achieves superior performance with RMSE = 1.235 and MAE = 0.924, indicating strong generalization across different seasonal conditions. Seasonal models demonstrate varying accuracy: spring (RMSE = 1.243), summer (RMSE = 1.324), and combined seasonal (RMSE = 1.255). These results suggest that although season-specific training may enhance interpretability, it does not necessarily outperform global training due to limited seasonal data and lack of hyperparameter adaptation.

In conclusion, the proposed BP-PSO model offers a robust and low-cost solution for wind speed forecasting in offshore applications. The multi-year framework demonstrates better generalization, while seasonal modeling provides insight into intra-annual wind variations. These findings support the use of hybrid optimization algorithms in enhancing wind resource assessments under real-world operational constraints.