• 关键词：决策树  统计预报模型  臭氧预报

• 基金项目：国家自然科学基金（No.4177114，41271116）；江苏省"双创团队"项目（No.0014）

• 丁愫
• 中国矿业大学环境与测绘学院, 徐州 221116

• 陈报章
• 中国矿业大学环境与测绘学院, 徐州 221116

• 王瑾
• 中国矿业大学环境与测绘学院, 徐州 221116

• 陈龙
• 中国矿业大学环境与测绘学院, 徐州 221116

• 张晨雷
• 中国矿业大学环境与测绘学院, 徐州 221116

• 孙少波
• 天津大学表层地球系统科学研究院, 天津 300072

• 黄丛吾
• 中国矿业大学环境与测绘学院, 徐州 221116

• 摘要：近地面层臭氧是光化学污染的主要污染物之一.臭氧污染不仅严重影响着空气质量并且危害人类健康与动、植物生长.本研究以徐州市为研究区，基于环境监测站连续监测数据分别采用分类回归树（CART）、随机森林（RF）和M5模型树方法建立臭氧浓度统计预报模型，选取1、4、7、10等4个月作为季度代表进行区域臭氧浓度预测的研究.以2015整年逐小时徐州市国控大气监测站实时监测的臭氧浓度（因变量）和气象因子数据（自变量）为训练样本建立臭氧浓度统计预报模型.模型验证结果表明，总体上3种决策树模型能够较好的预测臭氧浓度动态变化特征，月尺度上预测值与观测值相关系数均值为0.68，平均绝对误差和均方根误差均值分别为21.63 μg·m^-3和27.42 μg·m^-3.在此基础上，基于站点观测所建立臭氧统计预报模型，以WRF模型模拟的气象场作为输入，预报区域网格化臭氧预报值，并发现臭氧浓度空间分布与站点观测特征总体一致性较好.经与观测值进行对比，结果表明两者相关系数均值为0.58，平均绝对误差及均方根误差分别为29.38 μg·m^-3和37.15 μg·m^-3，预报准确率均高于75%.同时利用周步长观测值与预测值建立的多元线性集合预报回归方程对3种决策树模型的预报值进行修正，在一定程度上提高了预报值的精度.

• Abstract：Ozone in surface layer is one of the main pollutants of photochemical pollution. It not only severely affects air quality, but also harms the heaths of human, flora and fauna. In this study, based on the CART (Classification and Regression Tree), random forest (RF), M5 model tree algorithms, and atmospheric data, we developed an ozone concentration statistical forecast model. Using the model, we predicted the ozone pollution in Xuzhou during January, April, July and October. The training data was from Xuzhou country atmospheric monitoring stations and included ozone concentration (dependent variable) and meteorological data (independent variable) during the periods from January 2015 to December 2015 in hour. The verifications indicated that overall the statistical forecast model reflected ozone concentration dynamic variations well, with the correlation coefficient (r) value of 0.68, and average of mean absolute (BIAS) and root-mean-square error (RMSE) values of 21.63 μg·m^-3 and 27.42 μg·m^-3 respectively. Based the regional weather forecast data from the Weather Research and Forecasting (WRF) model, we obtained grid-based regional ozone concentration distributions. The patterns of our modelling ozone concentration and observed data agreed well. The r value was about 0.58, and the BIAS and RMSE values were 29.38 μg·m^-3 and 37.15 μg·m^-3, respectively. The overall accuracy reached above 75%. Besides, we further used the multiple linear regression method to correct the model predictions at weekly scale, which significantly improved model results.

• Key words：decision tree  statistical forecast model  ozone forecast

• 
  摘要点击次数： 1177 全文下载次数： 1104