【作者】 刘微；

【导师】 陈福集；

【作者基本信息】 福州大学 ， 信息管理与信息系统， 2021， 博士

【摘要】 环境污染问题已成为全球面临的重大挑战,特别是大气污染问题严重影响社会发展和人类身体健康。大气颗粒物极具危害性,即使在浓度相对比较低的情况下,依然能够对人类健康和生态系统造成严重的破坏。PM2.5作为大气主要污染物,已经引起全球的广泛关注,所以掌握PM2.5浓度的变化规律,并对其进行高精度的预测,具有重要的现实意义。作为信息科学与环境科学的交叉领域,利用智能信息处理技术预测PM2.5浓度已成为前沿研究热点,吸引了国内外科学家的广泛注意。但是作为一个新兴课题,目前的研究还处于初级阶段,根据PM2.5浓度数据特点进行针对性的模型构建方面还有待深入研究。本文的主要贡献包括提出了计算每日气象距平与最有利于PM2.5积累的气象模式之间的余弦相似性,利用其提取的“气象模式”辅助PM2.5浓度分类,研究了考虑气象因素的PM2.5浓度等级分类;进一步将主成分分析与聚类算法结合,提取关键污染物因素和“污染物模式”作为预测模型的输入变量,研究了考虑气象因素和污染物因素的PM2.5浓度预测;最后,对PM2.5浓度时间序列的特点进行了研究,将混沌理论与智能信息处理技术相结合,提出了EEMD-PSR-CS-SVR组合预测模型,实现了未来24小时PM2.5浓度预测。本文的研究工作主要包括以下几个方面:（1）基于气象模式提取的PM2.5浓度等级分类研究大气中污染物浓度很大程度上受气象条件及其相互作用的影响,尤其是大气中颗粒物的沉积和扩散受气象因素的影响很大。在某些情况下,气象条件对PM2.5颗粒浓度的影响甚至大于其排放量本身。研究表明,与有利天气条件下较高的污染物排放相比,不利天气条件下较低的污染物排放反而会导致较高的PM2.5颗粒的沉积和更差的空气质量等级。我国大多数城市已经建立了设备齐全的气象观测网络,能够方便地提供日常气象数据,所以将气象模式分析应用于辅助空气质量预测,可以节约大量的资源和成本。本文提出了每日气象距平（daily meteorological anomaly）、最有利污染物积累的气象模式（the most favorable pattern for pollutant accumulation）及余弦相似性（cosine similarity）相结合的新型概念,并创新性地应用于PM2.5浓度的等级分类中,即通过计算每日气象距平与最有利PM2.5积累的气象模式之间的余弦相似性,并以此为依据选择气象变量构成新的“气象模式”（meteorological pattern）作为预测模型的主要输入变量以协助提高PM2.5浓度等级分类的精确度。所提出的方法使用的变量更少,数据获取更加方便,且分类效率和精度有了大幅度的提高。（2）考虑气象因素和污染物因素的PM2.5浓度预测研究本文充分考虑了影响PM2.5形成和传输等各方面因素,包括气象因素（大气压力、相对湿度、气温、风速、风向、累积降水量等）和污染物因素（PM10、SO₂、NO₂、O₃、CO等）,并分析了这些变量与PM2.5浓度之间的关系。为了更好地预测PM2.5浓度,我们提出将主成分分析和聚类算法相结合（PCA-clustering）,选择关键污染物变量并提取“污染物模式”（pollutant pattern）,作为PM2.5浓度预测的重要变量;同时采用不同优化算法对所提出的预测模型参数寻优,从而提高PM2.5浓度预测的精度和效率。（3）基于集合经验模态分解和相空间重构技术的PM2.5浓度时间序列的预测研究影响PM2.5浓度的因素较多,它们之间的关系较为复杂,很难进行数学建模,特别对其进行未来一段时间的较为精确的预测难度较大。针对这一难题,我们研究了PM2.5浓度时间序列的特点,创新性地将基于集合经验模态分解和相空间重构技术组合（EEMD-PSR）,并应用到预测模型中,充分发挥EEMD降低原始时间序列复杂性,PSR可利用有限数据重构原动力系统模型、以及参数优化的预测模型快速收敛到全局最优以解决非线性问题的优点,获得了比传统时间序列预测模型精度更高的未来24小时PM2.5浓度预测结果,同时,实验证明,EEMD-PSR策略对于时间序列的预测具有较好的普适性。以上研究探索了影响PM2.5浓度的气象因素、污染物因素、时间序列,通过提取气象模式、污染物模式、降低原序列的复杂度、预测模型的参数寻优等,大幅度提高了PM2.5等级分类和浓度预测的效率和精度,促进了智能信息处理技术在PM2.5浓度预测领域的应用,并可为政府在大气污染治理方面提供决策支持。更多还原

【Abstract】 Environmental pollution has become a major issue for the whole world,especially in terms of air pollution.Haze has a serious impact on social development and human health.In particular,the atmospheric Particulate Matter（PM）is extremely harmful,which can seriously damage human health and ecosystem even if the concentration is relatively low.As a major atmospheric pollutant,PM2.5 has attracted extensive attention all over the world.It is of great practical significance to master the variation law of PM2.5 concentration and predict it with high precision.As a cross field of information science and environmental science,using intelligent information processing technology to predict PM2.5 concentration has become a cutting-edge research hotspot and attracted extensive attention.However,as an emerging topic,the current research is still in its infancy.The targeted model construction based on the characteristics of PM2.5 concentration data needs to be further studied.This paper has made a series of innovative achievements.Firstly,this paper calculated the cosine similarity between the daily meteorological anomaly and the most favorable meteorological pattern for pollutant accumulation,which was used to extract“meteorological model”to assist the classification of PM2.5 concentration.And the prediction of PM2.5 concentration grade considering meteorological factors was accordingly studied.Furthermore,principal component analysis and clustering methods were combined to extract the key pollutant factors and patterns as the input variables of the prediction model,and the PM2.5 concentration prediction considering both meteorological and pollutant factors was investigated.Finally,the characteristics of PM2.5 concentration time series were studied.By combining chaos theory with intelligent information processing technology,an EEMD-PSR-CS-SVR combined prediction mode has been established and applied to predict PM2.5 concentration in the next 24 hours.The specific research work mainly includes the following aspects:（1）Prediction of PM2.5 grades based on meteorological pattern extractionThe concentration of pollutants in the atmosphere is greatly affected by the meteorological conditions and their interaction.Especially,the deposition and diffusion of the atmospheric particulate matter is highly affected by meteorological factors.In some cases,the affect of meteorological conditions on PM2.5 concentration is even greater than the particle emission itself.It has been shown that compared with the higher pollutant emission under good weather,higher PM2.5 particle deposition and worse air quality grade can be resulted from lower pollutant emission but adverse weather conditions.Most cities in China have established well-equipped meteorological observing networks,so the daily meteorological data can be easiliy aquired.Therefore,the application of meteorological pattern analysis to assist air quality prediction could save a lot of resources and costs.This paper puts forward a new concept in combination of daily meteorological anomaly,the most favorable pattern for pollutant accumulation and cosine similarity,which is further applied to the grade prediction of PM2.5 concentration.By calculating the cosine similarity between the daily meteorological anomaly and the most favorable meteorological model for PM2.5 accumulation,meteorological variables are extracted to form a new meteorological pattern as the main input variable of the prediction model,which can improve the accuracy of PM2.5 concentration grade prediction.The obtained results indicate the validity of meteorological pattern analysis for higher efficiency and accuracy prediction using easily accessible and fewer variables.（2）PM2.5 concentration prediction considering both pollutant factors and meteorological factorsWe fully consider various factors that affect the formation and transmission of PM2.5,including pollutants（PM10,SO₂,NO₂,O₃,CO）and meteorological factors（atmospheric pressure,relative humidity,air temperature,wind speed,wind direction,cumulative precipitation）,and comprehensively study the relationship between these variables and PM2.5 concentration.In order to better predict the concentration of PM2.5,we combine principal component analysis（PCA）and clustering methods to extract pollutant variables and patterns as the important PM2.5 concentration predictors.At the same time,different optimization algorithms are applied to optimize the parameters of the proposed prediction model to further increase the efficiency and accuracy of PM2.5 concentration prediction.（3）Prediction of PM2.5 concentration time series based on the technology of ensemble empirical mode decomposition and phase space reconstructionThere are many factors affecting PM2.5 concentration,and the relationship between them is complex,making the change of PM2.5 concentration nonlinear and unstable.Therefore,it is difficult to carry out mathematical modeling to predict the time series of PM2.5 concentration.To solve this problem,an EEMD-PSR-CS-SVR prediction model has been established by combining the technology of ensemble empirical mode decomposition（EEMD）and phase space reconstruction（PSR）with the support vector machine model optimized by cuckoo algorithm（CS-SVR）.Thereinto,EEMD reduces the complexity of the original time series,PSR reconstructs the prime mover model with limited data,and CS-SVR quickly converges to the global optimization to solve the nonlinear problem.Taking these advantages,the prediction results of PM2.5 concentration in the next 24 hours are obtained with higher accuracy than the traditional time series prediction model.Meantime,the experiments confirm the good universality of the EEMD-PSR strategy for time series prediction.These studies comprehensively investigated the influence of meteorological factors,pollutant factors and time series on PM2.5 concentration.By extracting meteorological model and pollutant model,reducing the complexity of the original series and optimizing the parameters of the prediction model,it greatly increases the efficiency and accuracy of PM2.5 grade and concentration prediction.The obtained results have promoted the application of intelligent information processing technology in the field of PM2.5 concentration prediction,and can provide decision support for the government in air pollution control.更多还原