基于多种机器学习的模式PM_(2.5)订正预报方法  

PM_(2.5) Forecast Correction Based on Multi-Machine Learning

作　　者：刘超 宫宇[1] 张碧辉 柯华兵 LIU Chao;GONG Yu;ZHANG Bihui;KE Huabing(National Meteorological Center,Beijing 100081,China;Key Laboratory of High Impact Weather(Special),China Meteorological Administration,Changsha 410073,China;Institute of Artificial Intelligence for Meteorology,Chinese Academy of Meteorological Sciences,Beijing 100081,China)

机构地区：[1]国家气象中心,北京100081 [2]中国气象局高影响天气(专项)重点开放实验室,湖南长沙410073 [3]中国气象科学研究院人工智能气象应用研究所,北京100081

出　　处：《热带气象学报》2024年第6期896-905,共10页Journal of Tropical Meteorology

基　　金：国家重点研发计划(2022YFC3701205);中国气象局高影响天气(专项)重点开放实验室以及中国气象局第二批揭榜挂帅项目(CMAJBGS202308)共同资助。

摘　　要：大气细颗粒物污染深刻影响着人体健康、大气能见度以及气候变化等诸多方面,对PM_(2.5)浓度进行精细化预报至关重要。因此,本研究基于随机森林(Random Forest,RF)、LightGBM(Light Gradient Boosting Machine)和XGBoost(Extreme Gradient Boosting Machine)等多种机器学习方法,分别构建了基于中国气象局雾-霾数值预报系统(CMA Unified Atmospheric Chemistry Environment,CUACE)V3.0版本的京津冀地区PM_(2.5)订正预报模型,并对比分析各个机器学习模型的预报效果和性能差异。研究结果表明:三种机器学习算法误差(ME)和平均绝对误差(MAE)均较CUACE雾霾模式明显降低,而且在不同预报时效下的ME和MAE变化幅度更小,反映出基于机器学习算法得到的PM_(2.5)订正预报稳定性较好。此外,在三种机器学习算法中,RF算法预报性能最佳,ME和MAE分别为-3.0μg·m^(-3)和23.6μg·m^(-3),RF平均绝对误差的改善幅度最为突出,达到11.5%,而且对区域内“正订正”的站点比例达到97.7%,明显优于LightGBM和XGBoost算法。另外,对2024年3月9日至12日雾霾天气过程进行预报检验评估,RF算法的TS评分最高,其轻度污染、中度污染以及重度污染及以上TS评分分别为0.43、0.19和0.03。由此可以看出,RF算法的预报效果更为突出,研究结果在实际业务预报中具有一定参考意义。Fine particulate matter(PM_(2.5))pollution in the atmosphere profoundly affects human health,atmospheric visibility,and climate change;therefore,accurate forecasting of PM_(2.5)concentration is essential.This study developed PM_(2.5)forecast correction models for the Beijing-Tianjin-Hebei region using the China Meteorological Administration Unified Atmospheric Chemistry Environment for Haze V3.0(CUACE-Haze 3.0)model and various machine learning methods,including random forest(RF),light gradient boosting machine(LightGBM),and extreme gradient boosting machine(XGBoost).The forecast performance and differences among these machine learning models were then compared and analyzed.The results show that the mean error(ME)and mean absolute error(MAE)for the three machine learning algorithms were significantly lower than those of the CUACE model.The variation ranges of ME and MAE under different forecast time intervals were smaller,indicating better stability of the PM_(2.5)forecasts obtained based on machine learning algorithms.Furthermore,among the three machine learning algorithms,RF exhibited the best forecast performance,with ME and MAE of-3.0μg m^(3)and 23.6μg m^(3),respectively.The improvement in MAE for RF was the most prominent,reaching 11.5%,and the proportion of stations with positive correction was 97.7%in this region,significantly better than those of LightGBM and XGBoost.Additionally,during the verification and evaluation of forecast for the haze from March 9 to 12,2024,RF achieved the highest threat score(TS),with TS scores of 0.43,0.19,and 0.03 for light pollution,moderate pollution,and severe pollution or above,respectively.This demonstrates that the forecast performance of the RF algorithm is superior,and the research results provide valuable references for operational forecasting.

关 键 词：机器学习 PM_(2.5) 中国气象局雾-霾数值预报系统 订正 

分 类 号：P45[天文地球—大气科学及气象学] P457