基于传统和深度学习技术的黄渤海域大风预报方法研究

Research on the gale prediction methods in the Yellow Sea and Bohai Sea based on traditional and deep learning technologies

基于黄渤海域站点风速观测资料以及TIGGE资料,选取欧洲数值预报中心(EC)、中国(CMA)、美国(NCEP)、加拿大(ECCC)4家集合预报产品,在综合评估各家性能的基础上,构建、优化和对比了海上大风集成平均(EM)、动态权重(WEM)、变权偏差订正(BCWEM)3类传统集成方法和长短期记忆神经网络(LSTM)方法。结果表明:LSTM在大风集成预报中性能最优。对于黄渤海域10 m风速预报,EC综合表现最好,NCEP在6级及以上大风段优势明显。各家预报误差均具有显著日变化特征,夜间预报能力弱于白天。优化训练期长度和去除表现较差成员可显著改善WEM和BCWEM的大风预报能力。相对EM的预报结果,WEM无明显改进,BCWEM和LSTM则有显著提升,后两者在全风速段和大风风速段上的预报误差均下降10%以上,且在夜间时段更为明显。BCWEM有效订正了EM和WEM方法对弱风速的预报偏差,LSTM则进一步减小了对强风速的预报误差,并提高了对大风站次的命中数和ETS评分。大风个例分析也表明,LSTM有效弥补了传统方法对低涡东移型大风漏报的问题,提升了对冷高压型大风的预报能力,优势明显。

Based on the wind speed observation data of stations in the Yellow Sea and Bohai Sea and the ensemble forecast products of the European Centre for Medium-Range Weather Forecasts(EC),China(CMA),the United States(NCEP)and Canada(ECCC)in the THORPEX Interactive Grand Global Ensemble(TIGGE)data,three traditional integration methods,including ensemble mean(EM),dynamic weight ensemble mean(WEM),bias correction weighted ensemble mean(BCWEM),and the long short-term memory neural network(LSTM)methods are constructed,optimized,and compared on the basis of comprehensive evaluation of the performance of the ensemble forecast products.The results show that LSTM has the best performance in sea gale integrated prediction.For the 10-m wind speed forecast in the Yellow Sea and Bohai Sea,EC has the best comprehensive performance,while NCEP has significant advantages in the gale prediction that is equal or greater than level 6.The diurnal variations of the forecast errors in four products are significant,and the prediction abilities of all products at nighttime are weaker than that at daytime.In the traditional methods,the gale prediction ability of WEM and BCWEM can be significantly improved by optimizing the length of training period and removing the members with poor performance.Compared with EM,WEM shows no significant improvement,while BCWEM and LSTM shows a significant improvement with a decrease in forecast error by more than 10% for both full wind speed and strong wind speed,which is more remarkable at nighttime.BCWEM effectively corrects the prediction bias of the EM and WEM methods for moderate and weak wind speed,while LSTM further reduces the prediction error for strong wind speed and improved the hit number of gale stations and ETS score.The cases analysis of gale also shows that LSTM effectively compensates for the missing report problem of gale in low vortex eastward type by traditional methods,and improves the prediction ability of gale in cold and high pressure type with significant advantage.