太原市PM2.5预报统计修正模型及其应用检验

为提高太原市PM_2.5预报准确率，更好地服务于空气质量预报预警工作，在华北区域BREMPS（环境气象数值预报系统）预报结果的基础上，结合MR（多元线性回归）、BP（BP神经网络）和MLR（多层递阶）建立10 d的滚动修正模型，并对太原市2017年1月15日—4月15日ρ（PM_2.5）进行了修正.结果表明:3种修正模型对BREMPS预报的ρ（PM_2.5）小时值和日均值均有不同程度的改善，尤其是MLR修正结果在多项评价指标上明显优于MR和BP，其小时值的RMSE（均方根误差）由原来的42.46 μg/m3降至26.74 μg/m3，重污染和非重污染时段日均值的RMSE分别由未修正前的63.78、43.68 μg/m3降至28.52、21.27 μg/m3，日均值修正结果的基础评分从0.65升至0.88，预报准确率由原来的66.18%升至86.74%.从3种修正模型的构建来看，MR和BP方法对系统平稳状态的修正具有一定的优势，而对系统大幅变化的识别能力较弱，所以在天气变化时临界状态的修正结果误差较大，模型的稳定性较差.研究显示，MLR方法本身具有一定的自适应能力，稳定性和修正结果的整体趋势明显优于MR和BP方法，对太原市空气质量预报改进、重污染天气预警和大气污染防治等方面具有较大的应用价值. 

Implementation of Model Output Statistics on PM2.5 Forecast in Taiyuan City

In order to improve the accuracy of PM_2.5 forecast in Taiyuan City and better advance the efficiency of air quality forecast and early warning. Based on the forecasting products of BREMPS, multiple linear regression (MR), BP neural network (BP) and multi-level recursive method (MLR) were used to correct the PM_2.5 concentration in Taiyuan City from January 15th to April 15th, 2017. The results indicate that the hourly and daily PM_2.5 concentration of BREMPS forecast are optimized by three correction methods in varying degrees. In particular, the results of MLR correction are obviously superior to multiple linear regression and BP neural networks on multiple evaluation indicators, which the root mean square error of hourly PM_2.5 concentration from 42.46 μg/m3 down to 26.74 μg/m3, and respectively the root mean square error of daily PM_2.5 concentration in heavy and non-heavy pollution periods from 63.78, 43.68 μg/m3 narrowed to 28.52, 21.27 μg/m3. The basic score of the modified daily PM_2.5 concentration has increased from 0.65 to 0.88, and the accuracy of the forecast is increased from 66.18% to 86.74%. From the construction of the three correction schemes, the MR and BP methods have certain advantages in correcting the stationary state of the system, but the ability to identify the strenuous changing system is weak. Therefore, the correction results of the critical state during the weather systems change and the stability of the MR and BP methods are weak. The MLR itself has a certain self-adaptive ability, which the overall trend and magnitude of MLR correction results are obviously superior to the MR and BP. Therefore, it is of great application value to the improvement of air quality forecast in Taiyuan City, the weather warning of heavy pollution and the prevention and treatment of air pollution.