• 关键词：决策树  随机森林  机器学习  风场  WRF模式  长沙

• 基金项目：国家重点研发计划（No.2018YFC0214003 ，2016YFA0602004）；四川省教育厅项目（No.15ZA0191）

• 张琛
• 中国科学院大气物理研究所，创新转化基地，淮南 232000;中国科学院大学，北京 100049

• 廖婷婷
• 成都信息工程大学大气科学学院，高原大气与环境四川省重点实验室，成都 610225

• 孙扬
• 中国科学院大气物理研究所，创新转化基地，淮南 232000

• 孟祥来
• 中国科学院大气物理研究所，创新转化基地，淮南 232000;中国科学院大学，北京 100049

• 张成影
• 中国科学院大气物理研究所，创新转化基地，淮南 232000;成都信息工程大学大气科学学院，高原大气与环境四川省重点实验室，成都 610225

• 摘要：基于决策树和随机森林两种机器学习算法，以长沙市国控站气象观测数据和WRF模式模拟得到的风场数据构建数据集，对WRF预报长沙市城区风场数据进行优化.同时，利用树模型特征选择法，筛选对近地面风场数据有重要影响的气象要素，将其作为两种机器学习算法的输入，并利用网格搜索法进行调参建模.最后，将训练结果与地面观测数据进行对比评估模型的性能，开展长沙市城区风场短时预报应用研究.结果表明：相比于WRF模式，随机森林模型和决策树模型的均方根误差平均降低34%和17%，平均绝对误差平均降低33%和13%，相关系数平均提高26%和19%，风向预报准确率平均提高1%和17%；随机森林模型和决策树模型可成功模拟风场日变化特征，而且刻画的风场空间变化特征更为精细，能很好地描述城区内大气污染物随时间的动态变化特征.这表明构建的机器学习风场模型具有较好的模拟性能，可应用于空气质量改善及环境风险评估等方面的研究，显示出机器学习方法在提升风场预报方面的潜力.

• Abstract：To optimize the WRF forecast data of the wind field in the urban area of Changsha， this paper constructs a data set based on the two machine learning algorithms of decision tree and random forest， using the meteorological observation data from the national air environmental monitoring station in Changsha and the wind field data simulated by WRF model. The tree model is applied to identify the feature which significantly affect the near surface wind field， among all the available meteorological elements. The extracted feature set is used as input for each machine learning algorithm. The grid search method was used for parameter adjustment modeling. We evaluated the model performance via comparing training results with the observational data and applied the model to the application research on the short-term forecast of the wind field in the urban area of Changsha. The results showed that compared with the WRF model， the root mean square errors of the random forest model and the decision tree model were decreased by 34% and 17% on average， and the mean absolute errors were decreased by 33% and 13% on average， and the correlation coefficients were increased by 26% and 19% on average， the accuracy of wind direction forecasting were increased by 1% and 17% on average， respectively； It was found that the Decision Tree and Random Forest models can successfully simulate the diurnal variations of the wind field， describe the spatial variations of the wind field more finely， and reasonably reproduce the dynamic change characteristics of atmospheric pollutants in the urban area over time. The constructed machine learning wind field model has good performance，and can be applied to air quality improvement and environmental risk evaluation， showing the potential of machine learning methods in improving wind field forecasting.

• Key words：decision tree  random forest  machine learning  wind field  WRF model  Changsha

• 
  摘要点击次数： 311 全文下载次数： 452