随着风电功率预测对时效和更新频率的要求不断提高，基于数值天气预报进行预测的方法局限性愈发明显，人工智能气象模型的发展为风电功率预测提供了高效的气象预报来源。选取广东阳江一典型海上风电场作为服务对象，应用盘古模型和数值天气预报模式CMA-GD 进行2023 年和2024 年共17 个月的模拟计算，设计了4 组功率预测试验，对比分析了不同气象数据预报源、不同功率预测方法以及进行风速订正对于功率预测的改进效果。试验结果显示，在风电功率预测中盘古模型具备同等甚至超越数值天气预报模型的精准度。应用生成的场站功率曲线做功率预测时，基于盘古模式气象预报的短期功率预测准确率超过基于CMA-GD 的准确率，甚至在2024 年个别月份的预测效果与应用机器学习算法预测的功率相当。根据2023 年11 月的试验结果，基于订正的盘古模式预报和应用ResMLP 算法的功率预测模型，功率预测效果最佳，24 h 预测准确率为0.714，96 h 预测准确率为0.654，将该模型应用到2024 年1~6 月，效果稳定，可以达到电网的考核标准。研究为在风电功率预测业务中应用盘古模型或其他人工智能气象模型提供了一个可以借鉴的方案。

Abstract：With increasing demands for timeliness and update frequency in wind power prediction, the limitations of methods based on Numerical Weather Prediction (NWP) have become increasingly evident. The development of artificial intelligence meteorological models has provided an efficient source of meteorological forecasts for wind power prediction. Selecting an offshore wind farm in Yangjiang, Guangdong as the service object, the Pangu AI meteorological model and the numerical weather prediction model CMA-GD were applied to conduct simulation calculations for a total of 17 months in 2023 and 2024. Four sets of power prediction experiments were designed, which used different meteorological data forecast sources, different power prediction methods, and a wind speed correction method, to compare and analyze the improvement of wind power prediction. The experimental results show that the Pangu model achieves accuracy comparable to or even surpassing that of numerical weather prediction models. When using the generated site power curve for power prediction, the short-term power prediction accuracy based on the Pangu model meteorological forecast exceeds that based on CMA-GD, and even in some months of 2024, the prediction performance is comparable to that of using machine learning algorithms for power prediction. According to the November 2023 experimental results, the power prediction model based on corrected Pangu model forecasts and the ResMLP algorithm achieved the best power prediction performance, with a 24-hour prediction accuracy of 0.714 and a 96-hour prediction accuracy of 0.654. Applying this model to the period from January to June 2024 yielded stable results that met the grid assessment standards. This study provides a referable solution for applying the Pangu or other artificial intelligence meteorological models in wind power prediction operations.