基于后向传播神经网络的PM2.5和臭氧预测研究

Prediction of PM2.5 and ozone based on back propagation neural network

大气环境污染物对人体健康和生态环境有着重要影响,而准确预测PM2.5和臭氧(O3)的浓度对空气质量管控有着重要意义。基于后向传播(BP)神经网络,以杭州市2014年5月-2015年9月和2017年1月-2018年12月监测的国控点污染物浓度和气象条件为实验数据,搭建了1、2、3小时预测模型,对主要污染物PM2.5和O3进行预测,并提出了一种BP神经网络模型优化方案,对BP神经网络的各个重要参数进行优化,同时采用周期循环的动态学习率算法对模型进行训练,结果表明优化方案对模型表现提升明显,可进一步丰富和完善BP神经网络模型及其他人工神经网络模型优化的研究方案。而预测结果中,优化后的1小时预测模型的R^2范围为0.936~0.965,2小时预测模型的R^2范围为0.773~0.885,3小时预测模型的R^2范围为0.624~0.813,预测精度较高。

Ambient air pollutants have a direct influence on the human body and the environment,and accurate prediction of PM2.5 and ozone(O3)concentrations are very important for air quality control.Based on the back propagation(BP)neural network,in this paper,a prediction model to predict the concentrations of PM2.5 and O3.is built.The prediction range is 1~3 hours.The data included air pollutants concentrations and meteorological conditions in Hangzhou,and the air pollutants concentrations were monitored at national monitor stations.The time range of the data was from May 2014 to September 2015 and from January 2017 to December 2018,respectively.Meanwhile,this study proposed a BP neural network model optimization method,which could optimize the important parameters of the BP neural network.In addition,the model was trained by using a dynamic learning rate algorithm of cyclical cycle.The optimization results show that the optimization method significantly improves the performance of the BP neural network models.The optimization method enriches the research of BP neural network models and other artificial neural network models.Analyzing the prediction results of optimized models,the R^2 of the 1 h prediction models is 0.936~0.965,the R^2 of the 2 h prediction models is 0.773^-0.885,and the R^2 of the 3 h prediction model is 0.624~0.813,which shows the high accuracy of models.