| 时间尺度重构EEMD-GRNN改进模型预测PM2.5的研究 |
| |
| 引用本文： | 符海月,张祎婷.时间尺度重构EEMD-GRNN改进模型预测PM2.5的研究[J].地球信息科学,2019,21(7):1132-1142. |
| |
| 作者姓名： | 符海月 张祎婷 |
| |
| 作者单位： | 南京农业大学 土地管理学院, 南京 210095 |
| |
| 基金项目： | 国家自然科学基金项目(41871319) |
| |
| 摘 要： | 合理构建PM2.5浓度预测模型是科学、准确地预测PM2.5浓度变化的关键。传统PM2.5预测EEMD-GRNN模型具有较好的预测精度,但是存在过于关注研究数据本身而忽略其物理意义的不足。本研究基于南京市2014-2017年PM2.5浓度时间序列数据,分析PM2.5浓度多尺度变化特征及其对气象因子和大气污染因子的尺度响应,基于时间尺度重构进行EEMD-GRNN模型的改进与实证研究。南京市样本数据PM2.5浓度变化表现为明显的天际尺度和月际尺度,从重构尺度（天际、月际）构建GRNN模型更具有现实意义;同时,PM2.5对PM10、NO2、O3、RH、MinT等因子存在多尺度响应效应,以其作为GRNN模型中的输入变量更具有时间序列上的解释意义。改进后的EEMD-GRNN模型具有更高的PM2.5浓度预测精度,MAE、MAPE、RMSE和R2分别为6.17、18.41%、8.32和0.95,而传统EEMD-GRNN模型的模型有效性检验结果分别为8.37、27.56%、11.56、0.91。对于高浓度天（PM2.5浓度大于100 μg/m3）的预测,改进模型更是全面优于传统EEMD-GRNN模型,MAPE为12.02%,相较于传统模型提高了9.03%。
|
| 关 键 词： | 南京市 PM2.5 本征模函数 时间尺度重构 多尺度响应 集合经验模态分解 广义回归神经网络 |
| 收稿时间： | 2018-11-22 |
Improving the EEMD-GRNN Model for PM2.5 Prediction based on Time Scale Reconstruction |
| |
| Haiyue FU,Yiting ZHANG.Improving the EEMD-GRNN Model for PM2.5 Prediction based on Time Scale Reconstruction[J].Geo-information Science,2019,21(7):1132-1142. |
| |
| Authors: | Haiyue FU Yiting ZHANG |
| |
| Institution: | College of Land Management, Nanjing Agricultural University, Nanjing 210095, China |
| |
| Abstract: | Reasonable construction of PM2.5 concentration prediction models is the key to scientifically and accurately predict PM2.5 concentration dynamics. The traditional EEMD-GRNN model predicts PM2.5 concentration with good prediction accuracy, but focuses more on research data and less on its physical meaning. Based on the time series data of PM2.5 concentration in Nanjing duing 2014-2017, this study analyzed the multi-scale variations of PM2.5 concentration and its scale response to meteorological and atmospheric pollution factors. Additionally, the EEMD-GRNN model was reconstructed and validated based on time scale reconstruction. Results show that: ① based on scale reconstruction, the improved EEMD-GRNN model was scientific in predicting the PM2.5 concentration as it considered the variation of PM2.5 concentration and its multi-scale responses to atmospheric pollutants and meteorological factors. The PM2.5 concentration of Nanjing’s sample data had obvious inter-daily and inter-monthly variations. Thus, it is more practical to construct GRNN modeling at the reconstruction scales (i.e., daily and monthly). Meanwhile, PM2.5 was sensitive to factors such as PM10, NO2, O3, RH, and MinT. Thus, choosing these factors as input variables in the GRNN model can be more explanatory for time-series. ② the improved EEMD-GRNN model was more accurate in predicting PM2.5 concentration. The MAE, MAPE, RMSE, and R2 of the improved models are 6.17, 18.41%, 8.32, and 0.95, respectively, which are superior to the validity test results of the traditional EEMD-GRNN model. Furthermore, for the prediction of high-concentration days (i.e., PM2.5 concentration greater than 100 μg/m3), the improved model was more comprehensive than the traditional EEMD-GRNN model. With the new EEMD-GRNN model, the MAPE is 12.02%, 9.03% higher than the traditional model. Our findings indicate that the improved EEMD-GRNN based on scale reconstruction is an efficient method to scientifically and accurately predict PM2.5 concentration, especially in days with high PM2.5 concentration. |
| |
| Keywords: | Nanjing PM2 5 intrinsic mode function(IMF) time scale reconstruction multi-scale response ensemble empirical mode decomposition generalized regression neural network |
|
| 点击此处可从《地球信息科学》浏览原始摘要信息 |
| 点击此处可从《地球信息科学》下载免费的PDF全文 |
|
