| 引用本文: | 杨留明，高帅鹏，黄飞，岑路娟.基于神经网络模型对城市空气质量预报方法的优化研究[J].环境监控与预警,2023,15(2):33-39 |
| YANG Liu-ming, GAO Shuai-peng, HUANG Fei, CEN Lu-juan.Optimization of Urban Air Quality Prediction Method Based on Neural Network Model[J].Environmental Monitoring and Forewarning,2023,15(2):33-39 |
|
| 摘要: |
| 基于郑州市2017年1月1日—2022年2月28日环境空气细颗粒物（PM2.5）逐日质量浓度监测数据和同期气象数据，利用反向传播（BP）神经网络构建了环境空气PM2.5质量浓度预报模型，实现了对郑州市后1日环境空气PM2.5质量浓度日均值进行预报。构建了考虑大气氧化性因素（情景一）和不考虑大气氧化性因素（情景二）这2种情景，并对2种情景下的预报效果进行评价。结果显示，在情景一下，各季节PM2.5预报质量浓度与实况质量浓度的标准化平均偏差（NMB）和均方根误差（RMSE）均处于较低水平，表明预报效果均具有较好的稳定性；各季节PM2.5实况质量浓度与预报质量浓度之间的相关系数（r）、一致性指数（IA）、准确率（Q）和级别预报准确率（G）均处于较高水平，其中Q值均＞79%，G值均＞80%，表明各季节PM2.5实况质量浓度与预报质量浓度趋势的吻合程度较高。情景一各季节PM2.5预报质量浓度与实况质量浓度的NMB和RMSE均低于情景二，降幅分别为0.8%~2.3%和2.3~4.2 μg/m3 ；r值、IA、Q值和G值均高于情景二，增幅分别为3.5%~36.1%，2.2%~14.6%，64%~9.4%和3.5%~9.1%，表明考虑大气氧化性因素能够优化该模型。 |
| 关键词: 细颗粒物预报 反向传播神经网络 统计模型 大气氧化性 MATLAB软件 |
| DOI： |
| 分类号:X51；X823 |
| 基金项目:安徽省重点研发计划项目（2022107020025） |
|
| Optimization of Urban Air Quality Prediction Method Based on Neural Network Model |
|
YANG Liu-ming, GAO Shuai-peng, HUANG Fei, CEN Lu-juan1,2
|
|
1.Henan GRG Metrology & Test Co. Ltd., Zhengzhou,Henan 450001, China;2.Key Laboratory of Southern Farmland Pollution Prevention and Control, Ministry of Agriculture, Changsha, Hunan 410128, China
|
| Abstract: |
| Based on the daily concentration monitoring data of ambient air quality in Zhengzhou from January 1, 2017 to February 28, 2022 and the meteorological data of the same period, BP neural network was used to construct the prediction model of ambient air PM2.5 concentration based on whether to add atmospheric oxidizing factors, so as to predict the daily average value of ambient air PM2.5 concentration in Zhengzhou in the next day, and evaluate the prediction effect under the two scenarios. The results showed that under the scenario of considering atmospheric oxidative factors, the standardized mean deviation and root mean square error between the predicted concentration and the actual concentration in each season were at a low level, indicating that the prediction effect of each season had good stability. The correlation coefficient, consistency index, accuracy and level prediction accuracy between the predicted concentration and the actual concentration in each season were at a relatively high level, of which the accuracies of the four seasons were higher than 79%, and the accuracies of the level forecast were higher than 80%, indicating that the actual concentration in four seasons were in good agreement with the predicted concentration trend. And in this scenario, the standardized mean deviation and root mean square error between the forecast concentration and the actual concentration in each season were lower than those in the other scenario, and the reduction ranges were 0.8%～2.3% and 2.3～4.2 μg/m3, respectively; the correlation coefficient, consistency index, accuracy and level prediction accuracy were all higher than those of the other scenario, with increases ranging from 3.5% to 36.1%, 2.2% to 14.6%, 6.4% to 9.4%, and 3.5% to 9.1%, respectively, indicating that atmospheric oxidation was an important parameter for model optimization. |
| Key words: PM2.5 forecast BP neural network Statistical model Atmospheric oxidation MATLAB |
