基于DLENN模型的沈阳地区PM2.5浓度预测

受到重工业发展规模、北温带季风气候、秋冬季燃煤取暖、机动车拥堵状况以及微观气象条件等各种因素影响,沈阳地区PM2.5浓度变化具有趋势性、周期性及随机性特征。针对上述三种特征,论文构建了一种集成双向长短期记忆网络的神经网络预测模型DLENN(Double-LSTM Ensemble Neural Network),内含的两个方向LSTM分别刻画PM2.5浓度变化趋势性和周期性,然后采用线性回归杂合神经网络来捕捉PM2.5浓度变化的随机性。基于沈阳地区11个监测站2016至2017年空气质量和气象条件数据,本文将DLENN模型分别与自回归移动平均ARIMA模型、支持向量机SVM模型、随机森林RF模型和梯度提升树GBDT集成学习方法进行对比实验,结果表明DLENN预测模型稳定优于其他方法,其预测误差RMSE相对于ARIMA、SVM和集成模型分别下降了9.23%、3.83%、5.49%.

DLENN-based Prediction of PM Concentration in Shenyang

Influencedby thecharacteristics and scale of heavy industry,the monsoonal climate in the north temperate zone,coal heating in autumn & winter, traffic congestion and motor emissions,and microscopic meteorological conditions,the PM2.5 concentrationsin Shenyang can be decomposed into trend,periodic and random fluctuation component.This paper pertinently proposes a model,called DLENN(Dual LSTM Ensemble Neural Network),to predict the PM2.5concentrations reasonably.The DLENN uses two LSTM models to characterize the trend and periodicity of PM2.5concentrations respectively, and adopt a linear part plus neural networkto describe the fluctuation.Basedoncollected data on air quality and meteorological conditions fromeleven monitoring stations in Shenyang since 2016 through 2017,this paper compares the proposed DLENN model with other mainstream methods,including ARIMA model,Support Vector Machine(SVM),ensemble model of Random Forest(RF) and Gradient Descent Boosting Tree(GBDT).Experimental results show that the prediction error RMSE of DLENN is lower than SVM,ensemble model of RF and GBDT,and ARIMA by 3.83%, 5.49%, 9.23% respectively.