基于多源数据特征驱动和多尺度分析的PM_(2.5)混合预测研究

A Hybrid PM_(2.5) Prediction Model Based on Multi-Source Data Features and Multi-Scale Analysis

精准把握PM_(2.5)污染的动态演变规律对政府和企业的大气污染防治决策至关重要.因此,文章提出了基于多源数据特征驱动及多尺度分析的混合预测建模框架,以提高PM_(2.5)预测精度.预测建模框架分为:1)多源数据分析,有效融合与PM_(2.5)污染相关的气象、污染、舆情等多源数据;2)多尺度分析,通过多元经验模态分解技术(MEMD)将多源数据分解成不同模态下的预测特征;3)混合预测分析,有序结合计量和机器学习模型,集成各模态预测值为最终结果.文章以北京市PM_(2.5)为研究案例结果表明:1)文章提出的混合模型的预测精度优于所有的基准模型;2)微博个数和情感能够叠加提升PM_(2.5)预测精度,且优于单因素预测结果;3)引入MEMD分解的模型精度显著高于基准模型.

Understanding the dynamic evolution of PM_(2.5)concentration and making accurate prediction can provide effective decision-making support for the government and enterprises to control air pollution.Therefore,this study aims to develop a hybrid prediction framework based on multi-source data features and multi-scale analysis,to improve the prediction accuracy of PM_(2.5).In particular,the prediction framework include:Multi-source data analysis,multi-source data(e.g.,meteorological data,environmental pollution data and public opinion)are introduced into model;multi-scale analysis,MEMD is used to decompose raw multi-source data into informative features;PM_(2.5)prediction,the independent prediction values of each mode are integrated into the final prediction result by orderly combining the traditional statistical models and machine learning models.The empirical study focuses on Beijing and shows that:1)The proposed model outperforms all the benchmarking models in accuracy;2)the numbers and sentiment of microblog can effectively improve the prediction accuracy of PM_(2.5);3)the accuracy of MEMD-based models is superior to that of the benchmark models.