基于随机森林算法的长江三角洲地区PM_(2.5) 浓度模拟研究

Prediction of PM_(2.5) Concentration in Yangtze River Delta Based on Random Forest Algorithm

PM_(2.5)影响人们的生活,危害城市居民的健康,因而在大范围、连续空间上精准预测PM_(2.5)的浓度对于降低居民暴露在大气污染环境中的风险意义重大。基于空气中PM_(2.5)浓度对气象因子、社会经济因子和下垫面条件因子的响应关系,利用皮尔逊相关分析、PCA分析和随机森林模型RF构建了长江三角洲地区浓度的预测模型。研究发现:①PM_(2.5)浓度大小分布与夜间灯光指数NLI、国内生产总值GDP、降雨量PRE、温度TEP、相对湿度RH、植被指数EVI以及土地利用覆被LUCC呈显著相关(P<0.05),其中NLI和GDP与PM_(2.5)浓度呈正相关,PRE、TEP、RH与PM_(2.5)浓度呈负相关。②主成分数量为4时,方差累积贡献率达到86.7%,PRE、RH、GDP、NLI和EVI是影响长三角地区PM_(2.5)浓度空间变化的最重要的5个因子。③PCA-RF模型对于PM_(2.5)浓度的预测具有较好的表现且在长三角中西部的城市预测效果好于东南部沿海城市。相对于逐步回归SR模型,PCA-RF模型验证数据集上的均方根误差RMSE降低12.7%,决定系数R 2提高12.1%,平均绝对误差MAE降低8.3%。

PM_(2.5) affects people’s lives and endangers the health of urban residents.Therefore,accurate prediction of PM_(2.5) concentration in macroscopic and continuous spatial scale is of great significance for reducing the risk of urban residents exposed to air pollution.Based on the response relationship of PM_(2.5) concentration to meteorological factors,social economic factors and underlying surface condition factors,a prediction model of PM_(2.5) concentration in the Yangtze River Delta region was constructed by using correlation analysis,PCA and Random Forest model.The results showed that PM_(2.5) concentration distribution was significantly correlated with NLI,GDP,PRE,TEP,RH,EVI and LUCC(P<0.05).NLI and GDP were positively correlated with PM_(2.5) concentration,while PRE,TEP and RH were negatively correlated with PM_(2.5) concentration.When the number of principal components was 4,the cumulative contribution rate of variance reached 86.7%.PRE,RH,GDP,NLI and EVI were the top five factors affecting the spatial variation of PM_(2.5) concentration in the Yangtze River Delta.The PCA-RF model had a good performance on the prediction of PM_(2.5) concentration,and the prediction effect in the central and western cities of the Yangtze River Delta was better than that in the southeast coastal cities.The RMSE decreased by 12.7%,R 2 increased by 12.1%,and MAE decreased by 8.3%on validation dataset for PCA-RF.