基于机器学习的兰州市空气质量预报方法研究
发布时间：2019-02-14 20:49
【摘要】：空气污染物(特别是PM2.5)严重危害人体健康。本文利用2001-2011、2013-2015年兰州市空气污染逐日监测资料,在分析了兰州市2001-2011年3种主要污染物SO2、NO2、PM10,2013-2015年6种主要污染物PM10、PM2.5、NO2、SO2、CO和O3的污染特征的基础上、以2014-2015年欧洲中期天气预报中心(ECMWF)与T639预报产品为预报因子,并引入小波分解方法,分别建立了基于BP神经网络、最小二乘法支持向量机(LS-SVM)和Elman神经网络的兰州市6种主要空气污染物浓度未来2日预报模型,并对预报结果进行检验,分析了各建模方法的优劣;最后结合以上模型以支持向量回归方法(SVR)建立了6种主要污染物集合预报模型,并进行仿真业务化预报检验。结果表明:(1)2013-2015年间PM10依然兰州市的主要污染物,是造成春季重污染天气的首要原因;2013-2015年SO2年平均浓度相比2001-2011年下降明显,并且从2013年起低于同期NO2的年平均浓度;O3的年平均浓度逐年增加,在2015年作为首要污染物的天数大幅度增加,成为夏季的最重要的污染物之一。(2)以LS-SVM建立的6种污染物24h和48h预报模型的评价指标整体好于BP神经网络和Elman神经网络;以BP神经网络建立的模型的稳定性相对较差,48h的预报精度衰减幅度最高。(3)以ECMWF建立的预报模型对未来2d的PM10、PM2.5、NO2、SO2和CO的日均质量浓度的预报效果好于T639,而T639对预报O3有一定优势。(4)通过小波分解方法对污染物资料进行预处理后,LS-SVM的24h和48h预报模型的预测精度得到有效改善。(5)集合预报模型对6种主要污染物的日均质量浓度的24h和48h预报精度比未进行集合处理的模型高;集合预报模型预测的AQI与实际AQI相比,24h和48h预报的平均误差为9.874和12.315,平均相对误差为12.4%和15.1%,均方根误差为14.033和17.095;空气质量指数等级的24h预报率为76.7%,漏报率为9.1%,误报率为14.2%;48h的预报率为71.5%,漏报率为11.0%,误报率为17.5%;集合模型对首要污染物的24h预报率为76.3%,48h预报率为70.0%。本文研究结果对提高兰州空气质量业务预报能力具有一定参考价值。
[Abstract]:Air pollutants (especially PM2.5) are seriously harmful to human health. Based on the daily monitoring data of air pollution in Lanzhou from 2001 to 2011 to 2015, the PM10,PM2.5,NO2,SO2, of six major pollutants in Lanzhou from 2001 to 2011 was analyzed in this paper. Based on the pollution characteristics of CO and O3, the (ECMWF) and T639 forecasting products of the European Center for Medium-Term Weather Forecast (ECWFC) in 2014-2015 are taken as forecasting factors, and the wavelet decomposition method is introduced to establish the BP neural network, respectively. LS-SVM and Elman neural network are used to predict the concentration of six major air pollutants in Lanzhou in the next 2 days. The results of prediction are tested and the advantages and disadvantages of each modeling method are analyzed. Finally, six major pollutant ensemble prediction models are established by using the support vector regression (SVR) method combined with the above models, and the simulated operational prediction tests are carried out. The results are as follows: (1) PM10 is still the main pollutant in Lanzhou from 2013 to 2015, which is the main cause of heavy pollution weather in spring; The annual average concentration of SO2 in 2013-2015 was significantly lower than that in 2001-2011, and was lower than the annual average concentration of NO2 in the same period from 2013 to 2013. The annual average concentration of O3 increased year by year, and the number of days as the primary pollutant increased significantly in 2015. It has become one of the most important pollutants in summer. (2) the evaluation indexes of 6 kinds of pollutants prediction models based on LS-SVM are better than those of BP neural network and Elman neural network on the whole; The stability of the model based on BP neural network is relatively poor, and the attenuation range of prediction accuracy is the highest at 48 h. (3) the prediction model based on ECMWF is better than T639 in predicting the daily average concentration of PM10,PM2.5,NO2,SO2 and CO in the next 2 days. However, T639 has some advantages in predicting O3. (4) after pretreatment of pollutant data by wavelet decomposition, The prediction accuracy of 24 h and 48 h prediction models of LS-SVM is improved effectively. (5) the accuracy of 24 h and 48 h prediction of daily average concentration of six major pollutants by the ensemble prediction model is higher than that of the model without collective treatment. Compared with the actual AQI, the AQI predicted by the ensemble prediction model has an average error of 9.874 and 12.315 at 24 h and 48 h, an average relative error of 12.4% and 15.1%, and a root mean square error of 14.033 and 17.095. The 24-hour forecast rate of air quality index grade was 76.7, the false alarm rate was 9.1, the false alarm rate was 14.2and 48h prediction rate was 71.5, the false alarm rate was 11.0 and the false alarm rate was 17.5; The 24 hour forecast rate of the main pollutants by the ensemble model is 76. 3% and 48 h forecast rate is 70. 0%. The results of this paper have certain reference value for improving the operational forecast ability of Lanzhou air quality.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X51
本文编号：2422594
[Abstract]:Air pollutants (especially PM2.5) are seriously harmful to human health. Based on the daily monitoring data of air pollution in Lanzhou from 2001 to 2011 to 2015, the PM10,PM2.5,NO2,SO2, of six major pollutants in Lanzhou from 2001 to 2011 was analyzed in this paper. Based on the pollution characteristics of CO and O3, the (ECMWF) and T639 forecasting products of the European Center for Medium-Term Weather Forecast (ECWFC) in 2014-2015 are taken as forecasting factors, and the wavelet decomposition method is introduced to establish the BP neural network, respectively. LS-SVM and Elman neural network are used to predict the concentration of six major air pollutants in Lanzhou in the next 2 days. The results of prediction are tested and the advantages and disadvantages of each modeling method are analyzed. Finally, six major pollutant ensemble prediction models are established by using the support vector regression (SVR) method combined with the above models, and the simulated operational prediction tests are carried out. The results are as follows: (1) PM10 is still the main pollutant in Lanzhou from 2013 to 2015, which is the main cause of heavy pollution weather in spring; The annual average concentration of SO2 in 2013-2015 was significantly lower than that in 2001-2011, and was lower than the annual average concentration of NO2 in the same period from 2013 to 2013. The annual average concentration of O3 increased year by year, and the number of days as the primary pollutant increased significantly in 2015. It has become one of the most important pollutants in summer. (2) the evaluation indexes of 6 kinds of pollutants prediction models based on LS-SVM are better than those of BP neural network and Elman neural network on the whole; The stability of the model based on BP neural network is relatively poor, and the attenuation range of prediction accuracy is the highest at 48 h. (3) the prediction model based on ECMWF is better than T639 in predicting the daily average concentration of PM10,PM2.5,NO2,SO2 and CO in the next 2 days. However, T639 has some advantages in predicting O3. (4) after pretreatment of pollutant data by wavelet decomposition, The prediction accuracy of 24 h and 48 h prediction models of LS-SVM is improved effectively. (5) the accuracy of 24 h and 48 h prediction of daily average concentration of six major pollutants by the ensemble prediction model is higher than that of the model without collective treatment. Compared with the actual AQI, the AQI predicted by the ensemble prediction model has an average error of 9.874 and 12.315 at 24 h and 48 h, an average relative error of 12.4% and 15.1%, and a root mean square error of 14.033 and 17.095. The 24-hour forecast rate of air quality index grade was 76.7, the false alarm rate was 9.1, the false alarm rate was 14.2and 48h prediction rate was 71.5, the false alarm rate was 11.0 and the false alarm rate was 17.5; The 24 hour forecast rate of the main pollutants by the ensemble model is 76. 3% and 48 h forecast rate is 70. 0%. The results of this paper have certain reference value for improving the operational forecast ability of Lanzhou air quality.
【学位授予单位】：兰州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X51
【参考文献】
相关期刊论文 前10条
1 何建军;余晔;刘娜;赵素平;陈晋北;;基于WRF模式的兰州秋冬季大气污染预报模型研究[J];气象;2013年10期
2 ;环境空气质量指数(AQI)技术规定(试行)[J];中国环境管理干部学院学报;2012年01期
3 王自发;庞成明;朱江;安俊岭;韩志伟;廖宏;;大气环境数值模拟研究新进展[J];大气科学;2008年04期
4 佟彦超;;中国重点城市空气污染预报及其进展[J];中国环境监测;2006年02期
5 吴小红,康海燕,任德官;基于神经网络中小城市空气污染指数预估器的设计[J];数学的实践与认识;2005年02期
6 杨民,丁瑞强,王式功,尚可政;兰州市大气气溶胶的特征及其对呼吸道疾病的影响[J];干旱气象;2005年01期
7 郎君,苏小红,周秀杰;基于有机灰色神经网络模型的空气污染指数预测[J];哈尔滨工业大学学报;2004年12期
8 杨民,王式功,李文莉,刘治国,尚景文;沙尘暴天气对兰州市环境影响的个例分析[J];气象;2004年04期
9 金龙,况雪源,黄海洪,覃志年,王业宏;人工神经网络预报模型的过拟合研究[J];气象学报;2004年01期
10 刘宇,胡非,王式功,邹捍,杨德保,尚可政;兰州市城区稳定边界层变化规律的初步研究[J];中国科学院研究生院学报;2003年04期
相关会议论文 前1条
1 姜金华;胡非;陈玉春;彭新东;;兰州市冬季二氧化硫浓度的数值模拟[A];庆祝中国力学学会成立50周年暨中国力学学会学术大会’2007论文摘要集（下）[C];2007年
，本文编号：2422594
本文链接：https://www.wllwen.com/shengtaihuanjingbaohulunwen/2422594.html
