基于机器学习的长株潭城市群PM_(2.5)重污染预报

Prediction of Heavy PM_(2.5) Pollution in Chang-Zhu-Tan Urban Agglomeration Based on Machine Learning

为提高PM_(2.5)重污染的预报准确率,融合气象和环境资料、前期观测和后期数值天气预报数据、地面和高空预报因子,建立预报时效较长且准确度较高的机器学习模型库。以长株潭城市群的PM_(2.5)重污染天气预报为例,将数据预处理、特征工程、算法优选、超参数调优等技术方法运用于模型中,建立的重污染预报机器学习模型库可预报PM_(2.5)浓度和等级,预警4 d内的PM_(2.5)重污染。为增强模型的透明度,对其进行可解释性研究。事前可解释性分析表明,PM_(2.5)浓度预报模型存在事前三特性:前期要素比后期要素重要,环境要素比气象要素重要,地面要素比高空要素重要;事后可解释性分析表明,常德2022年1月18日的重污染天气过程受上游传输和本地污染累积的共同影响,其中传输的作用稍大。

A machine learning model library with long prediction time and high accuracy was established based on meteorological and environmental data,early observation and later numerical weather forecast data,ground and high-altitude forecast factors to improve the prediction accuracy of PM_(2.5)heavy pollution.Taking heavy PM_(2.5)pollution forecast in Chang-Zhu-Tan urban agglomeration as an example,using data preprocessing,feature engineering,algorithm optimization and hyperparameter tuning and other technologies,this model library could predict the concentration and grade of PM_(2.5),and warn heavy PM_(2.5)pollution within 4 days.Interpretability of the model was studied to enhance its transparency.Ex ante interpretability analysis showed that PM_(2.5)concentration prediction model had three ex ante characteristics:preceding factors were more import ant than late factors,environmental factors were more important than meteorological factors,and ground factors were more important than high-altitude factors.Post interpretability analysis showed that the heavy pollution weather process on January 18,2022 in Changde was influenced by upstream transmission and local pollution accumulation,in which transmission played a larger role.