摘要
细颗粒物(PM_(2.5))对大气污染和人体健康具有显著影响。为了提高PM_(2.5)质量浓度预报准确率,提出一种将先知(Prophet)时间序列分解算法和极限梯度提升树(Extreme Gradient Boosting,XGBoost)机器学习模型相结合的多变量混合预测模型(Prophet-XGBoost)。利用Prophet算法对时间序列可分解的特性,将PM_(2.5)高维质量浓度序列分解成若干低维时序特征分量,并与污染物和气象因素数据集成构建XGBoost预测模型,以得到PM_(2.5)质量浓度的预测值。试验中以南京市PM_(2.5)质量浓度历史数据为例进行实证分析。结果表明,结合Prophet时间序列分解的预测模型,PM_(2.5)质量浓度预测结果的决定系数R^(2)提升至0.658 4。由此可见,Prophet-XGBoost多变量混合预测模型较传统长短期记忆神经网络(Long Short-Term Memory,LSTM)、XGBoost模型能够更好地预测PM_(2.5)日均质量浓度的变化趋势。
The mass concentration of fine particulate matter(PM_(2.5)) has a significant impact on air pollution and human health.To improve the accuracy of PM_(2.5) mass concentration prediction,in light of the shortcomings of traditional machine learning models in time series forecasting,this paper proposes a multivariate hybrid prediction model(Prophet-XGBoost) based on the Prophet time series decomposition algorithm and the XGBoost machine learning model.Using the Prophet algorithm to decompose the time series,the PM_(2.5) high-dimensional mass concentration series is decomposed into several low-dimensional time series feature components such as trend items and periodic items,and the predicted value and confidence interval of PM_(2.5) mass concentration based on the Prophet model are obtained.After extracting and matching historical pollutant concentration data and meteorological factors data in time,they are converted into daily average data.And the pollutant and meteorological data are screened by algorithms such as Lasso regression,neural network,and random forest to determine the range of feature sets.Then integrated with low-dimensional time feature components such as trend terms and periodic terms obtained by Prophet decomposition,we build combination input variables of the model,input into the Prophet-XGBoost model for prediction,and finally obtain the predicted value of PM_(2.5) daily mass concentration.In the experiment,the historical data of the daily average mass concentration of PM_(2.5) in Nanjing is used as an example for empirical analysis.The results show that compared with the traditional Prophet model,LSTM model,and XGBoost model,the coefficient of determination(R^(2)) of the Prophet-XGBoost combined prediction model is increased from 0.555 2 to 0.658 4.The mean absolute error(E_(MAE)) and the root mean square error(E_(RMSE)) decreased by 37.33% and 43.07%,respectively.The prediction results are improved significantly.Thus,it can be seen that after adding time series features such as trend items and pe
作者
杨长春
聂倩倩
YANG Changchun;NIE Qianqian(School of Computer Science and Artificial Intelligence,Changzhou University,Changzhou 213164,Jiangsu,China;Business School,Changzhou University,Changzhou 213164,Jiangsu,China)
出处
《安全与环境学报》
CAS
CSCD
北大核心
2023年第12期4600-4608,共9页
Journal of Safety and Environment
