基于机器学习算法的典型工业区环境空气VOCs浓度预测研究  

Research on Predicting VOCs Concentration in Air of Typical Industrial Zones Based on Machine Learning Algorithm

作　　者：车祥 CHE Xiang(Shanghai Environmental Monitoring Center,Shanghai 200235,China)

机构地区：[1]上海市环境监测中心,上海200235

出　　处：《环境科学研究》2024年第8期1694-1702,共9页Research of Environmental Sciences

基　　金：国家重点研发计划项目(No.2022YFC3700104,2022YFC3703501);上海市生态环境局青年科研项目(No.202404)。

摘　　要：挥发性有机物(VOCs)是重要的臭氧前体物,实现VOCs浓度预测可为区域大气污染预警与精准管控提供支撑。基于2022年上海市某典型工业区的区域观测VOCs小时体积分数、同期中尺度天气模型(WRF)分层气象因子预报数据、重点企业用电量数据,利用长短期记忆网络(LSTM)算法构建了区域VOCs预测模型,对2023年1-4月区域VOCs小时体积分数进行预测,预报时效为24 h。基于区域污染溯源结果,构建了正常排放情景数据集,将预测值、观测值、正常排放情景数据集进行比较和分析,评估了预测值的准确性。结果表明:①区域VOCs体积分数预测值变化趋势与观测值基本一致,总体略低于观测值,在小时及日均两个级别均具有良好的线性相关关系。VOCs小时体积分数预测值的均方根误差(RMSE)为23.71×10^(-9),标准化平均偏差(NMB)为-10.4%,相关系数(r)为0.53,平均绝对误差(MAE)为12.73×10^(-9),两倍因子百分比(FAC2)为86.9%,VOCs日均体积分数预测值的RMSE、r、MAE、FAC2分别为11.33×10^(-9)、0.81、7.06×10^(-9)、96.7%。②模型可准确预测正常排放情景下的区域VOCs体积分数变化,突发异常排放导致的短时VOCs高值是VOCs小时体积分数预测值误差的主要来源,且贡献率一般超过30%。不同气象条件下异常排放对区域VOCs体积分数的影响程度不同,扩散条件愈不利,异常排放对区域VOCs浓度的影响愈大。③预测模型误差的来源分析表明,更多种类高分辨率活动量表征数据及更精细的气象预测数据有助于进一步提升预测模型的准确性。研究显示,LSTM算法可实现工业区环境空气VOCs浓度预测,进而识别气象条件及异常排放对区域VOCs浓度的影响。Volatile organic compounds(VOCs)are important precursors of ozone.Realizing VOC concentration prediction can provide support for VOC early warning and precise control in complex areas.Based on the monitoring of hourly volume fraction of VOCs in a typical area of Shanghai in 2022,a regional VOCs prediction model was constructed using the Long Short-Term Memory Network(LSTM)algorithm to predict mesoscale layered meteorological factors from the Weather Research and Forecasting(WRF)during the same period,and the electricity consumption data of key enterprises in the region.The regional hourly volume fraction of VOCs from January to April 2023 was predicted,with a period validity of 24 hours.A normal emission scenario dataset was constructed based on the regional pollution tracking results.The accuracy of the predicted values was evaluated by analyzing them with observed values and dataset under normal emission scenarios.The results indicate that:(1)The predicted values of regional volume fraction of VOCs have a similar tend to the observed values,with good linear correlation at both hourly and daily levels,and overall,slightly lower than the observed values.For the prediction of hourly volume fraction of VOCs,the root mean square error(RMSE)is 23.71×10^(-9),the normalized mean bias(NMB)is-10.4%,the correlation coefficient(r)is 0.53,the mean absolute error(MAE)is 12.73×10^(-9),and FAC2 is 86.9%.The RMSE,r,MAE and FAC2 of the predicted daily average volume fraction of VOCs are 11.33×10^(-9),0.81,7.06×10^(-9)and 96.7%,respectively.(2)The prediction model can accurately predict the volume fraction of VOCs in the region under normal emission scenario without any unexpected situation.It was found that high short-term VOCs caused by sudden abnormal emissions were the main source of error in predicting hourly VOCs volume fraction,with a contribution rate typically exceeding 30%.Under different meteorological conditions,the impact of abnormal emissions on the volume fraction of regional VOCs varies.The more unfavorable th

关 键 词：机器学习 长短期记忆网络 挥发性有机物(VOCs) 预测 环境空气 

分 类 号：X831[环境科学与工程—环境工程]