利用2017-2021年的ERA5再分析资料和京津冀国家站地面资料，结合多种机器学习方法建立预报模型，开展轻雾、大雾客观预报。探讨了再分析资料、地形因素的影响，并结合多模型集成、统计消空进一步优化模型。结果表明:(1)XGBoost(eXtreme Gradient Boosting)、LightGBM(Light Gradient Boosting Machine)、随机森林等集成学习方法的预报效果均优于决策树方法;(2)在引入ERA5再分析资料、地形建模后，XGBoost、LightGBM模型的预报性能显著提高。相比仅使用地面要素建模，大雾预报的TS(Threat Score)提升了 30％、32％，达到0。52、0。49，命中率分别为0。62、0。87。此外，经过多模型集成后，轻雾、大雾预报的TS提升到了 0。51、0。54;(3)2022年秋季一次大雾过程中，本方法提前72 h准确预报了京津冀地区的大雾，其中以LightGBM模型表现最好。0～72 h轻雾预报和0～36 h逐小时大雾预报的TS均达到0。3，预报准确率、时效性均优于ECMWF(European Center for Medium Weather Forecasting)模式。
Application of multiple machine learning methods in low-visibility weather forecasting in Beijing-Tianjin-Hebei region
Based on observation data from the Beijing-Tianjin-Hebei National Weather Stations between 2017 and 2021,and ERA5 reanalysis data,a forecasting model for light fog and fog was developed by using a variety of machine learning algorithms.The study also investigated the influence of reanalysis and topographic factors on model performance,and utilized a method combining multi-model integration and statistical voiding to optimize the model.The main findings are as follows:(1)ensemble learning methods such as XGBoost(eXtreme Gradient Boosting),LightGBM(Light Gradient Boosting Machine),and random forest outperform the decision tree method in terms of low-visibility weather forecast ability;(2)the performance of the XGBoost and LightGBM models is significantly improved when introducing ERA5 reanalysis and topographic factors.Specifically,the TS(Threat Score)of fog forecast is 30％and 32％higher than that built on surface elements only,reaching 0.52 and 0.49,and the POD(Probability of Detection)is 0.62 and 0.87,respectively.In addition,the TS of light fog and fog forecast increase to 0.51 and 0.54 after stacking two models;(3)during a regional fog event in the fall of 2022,our methods accurately predict fog 72 h in advance.In particular,the LightGBM model performs best,with 0-36 h fog forecast TS and 0-72 h light fog forecast TS reaching 0.3,which is better than ECMWF(European Center for Medium Weather Forecasting)in accuracy and timeliness.
国家科技期刊平台
NSTL
万方数据
