多机器学习模型下逐小时PM2.5预测及对比分析

为了能及时、准确的估算出PM_2.5浓度及污染等级，分别构建了K最邻近模型（KNN）、BP神经网络模型（BPNN）、支持向量机回归模型（SVR）、高斯过程回归模型（GPR）、XGBoost模型和随机森林模型（RF）6个PM_2.5浓度预测模型，选取江西省赣州市为实验区域，采用2017~2018年逐小时气象站数据、PM_2.5浓度数据和Merra-2再分析数据开展PM_2.5预测实验.结果表明，缺少污染物观测数据时，利用能见度和气象因子等数据也能较好的预测PM_2.5浓度.在PM_2.5浓度预测精度方面，XGBoost模型最高，随机森林模型次之，高斯过程回归模型最差.6个模型的预测精度总体呈现冬季最高，秋季和春季次之，夏季最低.XGBoost模型的PM_2.5污染等级预测准确率高于其他模型，综合准确率达87.6%，并且XGBoost模型具有训练时间短，占用内存小等优点.XGBoost模型的变量重要性结果表明，能见度变量的重要性最高，相对湿度和时间变量次之.本研究可为环境部门准确预测、预报PM_2.5浓度提供参考.

Hourly PM2.5 prediction and its comparative analysis under multi-machine learning model

Six models were built for timely and accurate estimation of PM_2.5 concentration and pollution levels, namely K Nearest Neighbor (KNN) model, BP Neural Network (BPNN) model, Support Vector Machine (SVM) regression model, Gaussian Process Regression (GPR) model, XGBoost model and Random forest(RF) model. Ganzhou City of Jiangxi Province was selected as the study area. The hourly ground-based meteorological data, PM_2.5 concentration data and Merra-2reanalysis data from 2017 to 2018 were used for modelling. The results show that PM_2.5 concentration can also be predicted using visibility and meteorological data when pollutant observation data are missing. In terms of the prediction accuracy of PM_2.5 concentration, the XGBoost model performs best, followed by the RF model, and the GPR model is the worst. The prediction accuracy of the six models was generally highest in winter, followed by autumn and spring, and lowest in summer. Compared with other models, the XGBoost model exhibits a more accurate prediction performance for PM_2.5 pollution level prediction with the comprehensive accuracy rate of 87.6%. Moreover, XGBoost model has the advantages of short training and small memory consumption. Visibility (followed by the relative humidity and time variable) play a key factor in the XGBoost models for PM_2.5 concentration prediction. This study can provide a reference for environmental departments to accurately predict and forecast PM_2.5 concentration.