融合深度神经网络特征的ARIMAX雾霾PM2.5预测  

Haze PM2.5 Concentration Prediction Based on ARIMAX Model with Deep Convolutional Neural Network Features

作　　者：张钰洁 王钰 杨杏丽[1] ZHANG Yu-jie;WANG Yu;YANG Xing-li(School of Mathematical Sciences,Shanxi University,Taiyuan 030006,China;School of Modern Educational Technology,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学数学科学学院,山西太原030006 [2]山西大学现代教育技术学院,山西太原030006

出　　处：《计算机技术与发展》2023年第2期167-172,共6页Computer Technology and Development

基　　金：山西省应用基础研究计划项目(201901D111034);国家自然科学基金项目(62076156);统计与数据科学前沿理论及应用教育部重点实验室开放基金(KLATASDS2007)。

摘　　要：目前,雾霾污染问题是关乎国计民生的重大问题,它已经对人们的生产、生活、身体健康,以及生态环境和气候变化都产生了很大的影响。这样,如何通过监测雾霾变化获取的雾霾相关信息去准确预测雾霾污染物的浓度,以防治和减轻雾霾造成的严重后果变得尤为重要。因此,通过在简单有效的传统ARIMAX模型基础上融入深度神经网络语义特征,提出了一种新的雾霾PM2.5浓度预测框架。首先,把对雾霾预测有显著影响的气象因子温度、压力、相对湿度数据转换为图像数据;然后,运用ResNet-50(Residual Network-50)卷积神经网络模型提取深度语义特征,进而运用主成分分析(Principal Component Analysis,PCA)技术处理高维特征,得到最佳深度神经网络特征组合;最后,用ARIMAX技术建立雾霾PM2.5浓度预测模型。在收集的山西省2015~2019年PM2.5浓度和气象因子数据集上验证了该预测框架在皮尔逊相关系数(Pearson’s Correlation Coefficient,PCC)、均方误差(Mean Square Error,MSE)、均方根误差(Root Mean Squared Error,RMSE)和平均绝对误差(Mean Absolute Error,MAE)度量下,对于1、3、5和7天长短期预测,都始终优于传统的简单差分自回归滑动平均(Autoregressive Integrated Moving Average,ARIMA)模型、三因素ARIMAX模型、多元回归模型、ResNet-多元回归模型、长短期记忆网络(Long and Short-Term Memory,LSTM)模型和支持向量机(Support Vector Machine,SVM)模型。At present,the problem of haze pollution is a major issue related to the national economy and people's livelihood,which has already had a great impact on people's production,life,health,and ecological environment and climate change.In this way,how to accurately predict the concentration of haze pollution by monitoring the haze-related information obtained by monitoring the haze changes in order to prevent and reduce the serious consequences of haze has become particularly important.Therefore,by incorporating the semantic features of deep neural networks on the basis of the ARIMAX model,we propose a new haze PM2.5 concentration prediction framework.Firstly,the numerical meteorological data(temperature,pressure and relative humidity that have a significant impact on haze prediction)is transformed into image data.Secondly,the ResNet-50(Residual Network-50)model is used to extract deep semantic features and PCA(Principal Component Analysis)is to process high-dimensional features for the best combination of deep neural network features.Finally,ARIMAX technology is used to establish a haze PM2.5 concentration prediction model.Furthermore,on the collected data set of PM2.5 concentration and meteorological factors in Shanxi Province from 2015 to 2019,under the Pearson’s Correlation Coefficient(PCC),Mean Square Error(MSE),Root Mean Squared Error(RMSE)and Mean Absolute Error(MAE)measurements,the experimental results demonstrate the proposed prediction frameworks are always superior to the traditional Autoregressive Integrated Moving Average(ARIMA),three-factor ARIMAX,multiple regression,ResNet-multiple regression,Long and Short-Term Memory(LSTM)and Support Vector Machine(SVM),for 1,3,5 and 7-day long and short-term prediction.

关 键 词：PM2.5预测 ARIMAX模型 ResNet神经网络 主成分分析技术 深度语义特征 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP181[自动化与计算机技术—控制科学与工程]