沈阳市PM2.5浓度ARIMA-SVM组合预测研究

首先利用回归树分类方法，对采暖期与非采暖期各日进行气象类型划分，识别出易造成重污染天气的气象类型.其次分别在各气象类型内，以污染源排放量为自变量，利用差分自回归滑动平均与支持向量机（ARIMA+SVM）组合方法建立起PM_2.5浓度日均值预测模型，并选取2013年01月~2017年06月间，沈阳市区内9个环境监测点PM_2.5浓度日均值进行实证分析.结果表明，使用气象分类下的ARIMA+SVM组合模型对PM_2.5浓度日均值进行预测，相比于不划分气象类型时的普通机器学习模型，其模型预测值与实测值趋势的吻合度更高，且对峰-谷值的识别能力更强.在采暖期与非采暖期，组合模型均具有平均绝对误差更低、预测正确率更高的优点.

ARIMA-SVM combination prediction of PM2.5 concentration in Shenyang

Firstly, meteorological types of heating period and non-heating period were classified using the method of regression tree classification, and meteorological types which are likely to cause severe pollution were identified. Secondly, the daily mean value prediction model of PM_2.5 concentration of different meteorological types was established using the combination of Autoregressive Integrated Moving Average Model and Support Vector Machine (ARIMA+SVM), which takes the emission of pollution sources as independent variables. In this paper daily mean PM_2.5 concentration of 9environmental monitoring points with continuous data in Shenyang during Jan 2013 to June 2017 was analysed. The results show that, compared with ordinary machine learning model without weather classification, the prediction of daily mean PM_2.5 concentration using ARIMA+SVM combined model based on meteorological classification has a better agreement with actual value, and its ability to identify the peak and valley values is much stronger. In heating and non-heating period, this combined model has the advantages of lower average error and higher prediction accuracy.