• 关键词：自适应偏最小二乘回归法  CUACE模式  PM_2.5偏差订正  改进方法

• 基金项目：国家重点研发计划课题（No.2016YFC0203301）；国家基金委重点研究项目（No.91644223）；大气重污染成因与治理攻关项目（No.DQGG0104）；环保公益性行业（气象）科研专项（No.201509001）；中国气象科学研究院基本科研业务费专项（No.2016Y005）；北京市基金重点项目（No.8171002）；中国气象局预报员专项（No.CMAYBY2016-091）

• 吕梦瑶
• 中国气象局国家气象中心, 北京 100081

• 程兴宏
• 中国气象局大气化学重点开放实验室, 北京 100081

• 张恒德
• 中国气象局国家气象中心, 北京 100081

• 刁志刚
• 上海点明信息技术有限公司, 上海 200940

• 谢超
• 中国气象局国家气象中心, 北京 100081

• 刘超
• 中国气象局国家气象中心, 北京 100081

• 江琪
• 中国气象局国家气象中心, 北京 100081

• 摘要：针对GRAPES-CUACE模式预报的6种常规污染物浓度，采用非线性动力统计-订正方法——自适应偏最小二乘回归法，建立了中国不同地区的CUACE模式预报偏差订正模型，采用多种敏感性试验优选了不同季节各区域的最优自变量组合方案，并对2016年1-3月、11-12月全国342个城市PM_2.5浓度预报值进行了滚动订正检验，分析了订正前后PM_2.5浓度的时空变化特征，重点分析了该方法在京津冀、长三角、珠三角、川渝地区等关键区域的适用性及其改进效果.结果表明：①CUACE模式预报PM_2.5浓度普遍低于观测浓度，且与实测值的相关系数较低；CUACE 15 km分辨率模式PM_2.5浓度预报效果优于54 km分辨率模式，其中长三角地区改进最显著，珠三角和京津冀次之，川渝地区预报效果较差.②订正后的PM_2.5浓度更接近于实测值，订正后误差明显减小，相关系数明显提高，而且订正值与实测值的散点集中分布于对角线附近.③长三角地区PM_2.5浓度订正效果最好，准确率可达72.3%；珠三角地区次之，准确率为66.3%；京津冀和川渝地区订正效果稍差，但准确率亦可达63.6%和62.6%.④订正后污染日和非污染日的准确率、相关系数分别提高了57.5%和25.9%、304.8%和15.2%；绝对平均偏差、均方根误差分别减小了38.9%和18.7%、21.8%和8.5%.⑤针对北京、上海、广州、乐山的不同重污染过程，订正后的平均绝对误差分别减小了12.07%、46.63%、36.66%、17.71%，相关系数分别提升了25.86%、22.22%、16.92%、162.5%，说明该订正方法适用于不同地区的不同重污染过程的预报.

• Abstract：In this study, concentrations of six kinds of conventional air pollutants predicted by the GRAPES-CUACE model are corrected using the dynamical-statistical method based on the adapting partial least square regression technique (APLSR),and an improved correction method for predictions of the CUACE model is established in different regions of China. Furthermore, an optimal independent variable combination scheme for different regions and seasons is optimized by a variety of sensitivity tests in 2015, and the correction test for PM_2.5 concentration is carried out in 342 cities in Jan., Feb., Mar., Nov., Dec., 2016 using the selected combination scheme. Temporal and spatial variation characteristics of PM_2.5 concentrations before and after the correction are analyzed focusing on the applicability and improvement effect of the dynamical-statistical method in five key regions. It is shown that:① the averages of PM_2.5 concentrations are underestimated by CUACE model in most parts of China, and the correlation coefficient between the predicted values and the observations is low; the prediction effect of the CUACE-15 km model is better than the CUACE-54 km model. ② The forecasted PM_2.5 concentrations corrected by APLSR and observations is obviously better than that of the CUACE-15 km model, which means the corrected values are closer to the observations, and the correlation coefficients between corrected PM_2.5 concentrations and observations increase remarkably, while scattered plots are centrally distributed near the diagonal line. ③ The accuracy rates of corrected PM_2.5 concentrations by APLSR in Yangtze River Delta, Pearl River Delta, Beijing-Tianjin-Hebei region, Sichuan and Chongqing provinces are 72.3%, 66.3%, 63.6% and 62.6%. ④ After corrections, the accuracy rates and correlation coefficients of corrected PM_2.5 concentrations in polluted and clean days have been improved 57.5% and 25.9%, 304.8% and 15.2% respectively; the absolute mean deviations and the root mean square error have been reduced by 38.9% and 18.7%, 21.8% and 8.5%.⑤After the corrections, the absolute mean deviations at Beijing, Shanghai, Guangzhou, Leshan have been reduced by 12.07%,46.63%,36.66%,17.71%, and the correlation coefficients at those cities have been improved by 25.86%,22.22%,16.92%,162.5% in different heavy pollution period. It demonstrates that the correction method is applicable to predict the PM_2.5 concentrations in heavy pollution processes in different regions.

• Key words：the adapting partial least square regression technique  the CUACE model  deviation correction for PM_2.5  improved method

• 
  摘要点击次数： 1002 全文下载次数： 1568