基于GA/PSO BP神经网络的石家庄VOCs环境浓度预测模型研究  

Prediction model of VOCs in Shijiazhuang based on GA/PSO BP neural network

作　　者：王欣 郭婧涵 耿雅娴 王树桥 葛宇轩 袁京周 张丁超 韩梦非 WANG Xin;GUO Jinghan;GENG Yaxian;WANG Shuqiao;GE Yuxuan;YUAN Jingzhou;ZHANG Dingchao;HAN Mengfei(College of Environmental Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050018,China;National and Local Joint Engineering Research Center for Volatile Organic Compounds and Fetid Pollution Control Technology,Shijiazhuang 050018,China)

机构地区：[1]河北科技大学环境科学与工程学院,石家庄050018 [2]挥发性有机物与恶臭污染防治技术国家地方联合工程研究中心,石家庄050018

出　　处：《安全与环境学报》2024年第4期1560-1568,共9页Journal of Safety and Environment

基　　金：国家自然科学基金项目(51804096);河北省高等学校科学技术研究重点项目(ZD2020345);河北省自然科学基金面上项目(B2020208023);河北省臭氧及PM2.5污染物多源监测与预警体系优化融合研究及示范项目(21373903D)。

摘　　要：为了提升挥发性有机物(Volatile Organic Components,VOCs)的预测精度,在反向传播(Back Propagation,BP)网络结构的基础上使用优化算法分别为遗传算法(Genetic Algorithms,GA)优化BP神经网络(GA BP)和粒子群算法(Particle Swarm Optimization,PSO)优化BP神经网络(PSO BP)对VOCs质量浓度进行预测。首先,对污染物及气象因子进行筛选。采用相关性分析法及逐步回归法进行分析筛选,并筛选出合适的输入变量。其次,建立BP神经网络结构。利用BP、GA BP、PSO BP神经网络,以石家庄市2022年夏季污染数据为样本对VOCs质量浓度进行预测。结果显示,经相关性分析及逐步回归法筛选,将PM_(2.5)质量浓度、O_(3)质量浓度、NO_(2)质量浓度、温度、相对湿度作为输入变量。经预测结果对比,PSO BP神经网络模型的预测精度较高,烷烃、烯烃、芳香烃和含氧烃实测值与预测值之间的拟合程度(R^(2))分别为0.80、0.55、0.78、0.67。研究结果可为日后VOCs污染预报预警提供理论参考。To improve the accuracy of Volatile Organic Compounds(VOCs)prediction in Shijiazhuang region,based on the BP neural network algorithm,this study optimizes it with Genetic Algorithm(GA)and Particle Swarm Optimization(PSO).This study utilizes data from the summer of 2022 and preprocesses it by categorizing it into Alkanes,Alkenes,Alkynes,Aromatic hydrocarbons,Halogenated hydrocarbons,and Oxygenated Volatile Organic Compounds(OVOCs).The emission inventory identifies several VOCs components highly affected by OFP,including Alkanes,Alkenes,Aromatic hydrocarbons,and OVOCs.Missing values are handled using two methods:numerical interpolation for datasets with fewer missing values,and modeling interpolation for datasets with more missing values.Significant variables that influence VOCs including PM_(2.5),NO_(2),O_(3),temperature,and Relative Humidity(RH),are filtered through both correlation analysis and stepwise regression modeling.These influential factors are used as input variables,applying three neural network structures:BP,GA BP,and PSO BP to predict the 24-hour concentration of different types of VOCs on a particular day,allowing for a comparison with the actual values.By observing the curves between the predicted and actual values,along with comparing the errors and goodness of fit among the three networks,the results show that PSO BP reduces the mean square errors(I MSE)by 69.40%,37.84%,57.38%,and 49.88%compared to BP,and 63.95%,34.29%,42.29%,and 46.72%compared to GA-BP when predicting Alkanes,Alkenes,Aromatic hydrocarbon,and OVOCs,respectively.Compared to BP,GA BP reduces the mean square errors(I MSE)by 15.11%,5.40%,26.16%,and 5.93%,respectively.Through analysis,the optimized network structure overcomes the limitation of poor global search capability found in a single BP network,and its advantage over the GA algorithm lies in the memory function of the PSO algorithm.As a result,PSO-BP holds promising prospects for predicting VOCs during the summer season.

关 键 词：环境工程学 挥发性有机物(VOCs) 神经网络 智能优化算法 遗传算法 粒子群算法 

分 类 号：X511[环境科学与工程—环境工程]