顾及地理环境各向异性的PM2.5浓度时空预测模型
王军杰(1996— ),男,河南周口人,硕士生,主要从事时空数据挖掘研究。E-mail: junjie.wang@csu.edu.cn |
Editor: 蒋树芳
收稿日期: 2024-04-26
修回日期: 2024-07-01
网络出版日期: 2024-09-10
基金资助
国家自然科学基金面上项目(42271485)
湖南省自然科学基金项目(2022JJ40585)
湖南省自然科学基金项目(2024JJ8343)
中南大学中央高校基本科研业务费专项资金资助(2024ZZTS0638)
A Spatiotemporal Prediction Model of PM2.5 Concentration Incorporating Geographic Environmental Anisotropy
Received date: 2024-04-26
Revised date: 2024-07-01
Online published: 2024-09-10
Supported by
National Natural Science Foundation of China(42271485)
Hunan Provincial Natural Science Foundation of China(2022JJ40585)
Hunan Provincial Natural Science Foundation of China(2024JJ8343)
Fundamental Research Funds for the Central Universities of Central South University(2024ZZTS0638)
PM2.5浓度预测对于空气污染的防控和治理具有重要意义。传统图卷积神经网络(Graph Convolutional Network, GCN)等时空预测模型主要通过监测站点间的欧氏距离来度量PM2.5分布的空间相关性,未顾及地形和风向等因素对大气污染物传输过程的各向异性影响,导致地形复杂区域内的预测结果精度偏低。为此,本文提出了一种顾及地理环境各向异性的PM2.5浓度时空图卷积网络预测模型。首先,考虑地理环境的各向异性特征,利用地形和风向等对站点间PM2.5传播的各向异性影响构建GCN的边,将站点PM2.5浓度、土地利用和其他气象因子建模为GCN的节点特征。其次,通过GCN提取站点PM2.5浓度的空间特征。最后,通过门控循环单元(Gate Recurrent Unit, GRU)建模站点PM2.5浓度的时间特征并进行预测。本文以山地型省份贵州省2017年逐小时PM2.5浓度记录进行实验,并采用一系列时空预测基线模型(GTWR、STSVR和GCN+GRU)与本文模型进行对比。实验结果表明: ① 本文模型的PM2.5浓度预测结果的RMSE、MAE别为10.047、6.848,相比基线模型平均下降了11.29%和12.16%;R2为0.883,比基线平均提升了3.72%; ② 通过不同地形对站点PM2.5浓度相关性影响分析,论证了山脉沟谷等地形会显著影响站点间的PM2.5浓度的相关性,进而影响站点PM2.5浓度预测结果; ③ 充分考虑地形和风向对PM2.5传播产生的各向异性影响,能够显著提升存在山脉和沟谷地形区域内PM2.5预测精度。
王军杰 , 王达 , 丁晨 , 唐建波 , 石岩 , 杨学习 . 顾及地理环境各向异性的PM2.5浓度时空预测模型[J]. 地球信息科学学报, 2024 , 26(9) : 2106 -2122 . DOI: 10.12082/dqxxkx.2024.240238
PM2.5 concentration prediction plays a pivotal role in the prevention and control of air pollution. Traditional forecasting models, such as the Graph Convolutional Network (GCN) and other spatial-temporal prediction models, measure the spatial correlation of PM2.5 distribution primarily by monitoring the Euclidean distance between monitoring stations. However, these models often fail to account for the anisotropy effects of terrain, wind direction, and other factors that significantly influence the transport process of air pollutants. This oversight can result in lower accuracy of prediction results, especially in areas with complex terrain. This paper proposes a novel spatiotemporal convolutional network prediction model for PM2.5 concentration that takes into account the anisotropy of the geographical environment. The model first constructs the edges of the GCN by incorporating the anisotropic effects of terrain and wind direction on PM2.5 propagation between stations. It then models the station PM2.5 concentration, land use, and other meteorological factors as node characteristics of the GCN. The model employs GCN to extract the spatial characteristics of PM2.5 concentration and subsequently uses a Gate Recurrent Unit (GRU) to model and predict the temporal characteristics of PM2.5 concentration at the station. The model's performance was evaluated using hourly PM2.5 concentration records from the mountainous Guizhou province in 2017. It was compared against several spatiotemporal prediction baseline models, including Geographically and Temporally Weighted Regression (GTWR), Spatio-Temporal Support Vector Regression (STSVR), and a combined GCN+GRU model. The experimental results demonstrate that the model proposed in this paper significantly outperforms the baseline models. The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) ratios of PM2.5 concentration predicted by the model are 10.047 and 6.848, respectively, which represent decreases of 11.29% and 12.16% compared to the baseline models. The R-squared value of 0.883 indicates an average improvement of 3.72% from the baseline. Furthermore, the analysis of the influence of different terrains on the correlation of PM2.5 concentration at different stations reveals that mountainous, gully, and other terrain features can significantly affect the correlation of PM2.5 concentration between stations. This, in turn, impacts the prediction results of PM2.5 concentration at different stations. The study concludes that fully considering the anisotropic effects of terrain and wind direction on PM2.5 propagation can substantially enhance the prediction accuracy of PM2.5 in mountain and gully terrain areas. By integrating the anisotropic characteristics of the geographical environment into the prediction model, this research contributes to the development of more accurate forecasting tools for PM2.5 concentrations. This development is expected to ensure accurate prediction of PM2.5 concentrations in areas with complex terrain and support the development of effective air pollution control strategies.
表1 本文实验数据及其特征描述Tab. 1 The experimental data and their characteristics |
数据类型 | 数据来源 | 数据年份 | 取值范围 | 时间分辨率 | 空间分辨率 | 均值 | 标准差 |
---|---|---|---|---|---|---|---|
PM2.5 /(µg/m³ ) | 中国环境监测总站 | 2017 | [1.00, 863] | 小时 | 站点 | 31.13 | 593.94 |
气温/℃ | ECMWF | 2017 | [-4.26, 39.92] | 小时 | 0.1°×0.1° | 15.42 | 55.15 |
气压/KPa | ECMWF | 2017 | [79.81, 97.99] | 小时 | 0.1°×0.1° | 87.38 | 13.74 |
U型风分量/(m/s) | ECMWF | 2017 | [-6.91, 7.20] | 小时 | 0.1°×0.1° | -0.67 | 1.47 |
V型风分量/(m/s) | ECMWF | 2017 | [-7.97, 10.23] | 小时 | 0.1°×0.1° | 0.30 | 6.03 |
降雨量/(mm/h) | ECMWF | 2017 | [0.0, 12.66] | 小时 | 0.1°×0.1° | 0.17 | 0.28 |
土地利用数据(NA) | 中国土地覆盖数据集 | 2017 | [1, 9] | - | 30 m | - | - |
DSM/m | ALOS | 2017 | [254, 2 606] | - | 30 m | 1 114.69 | 445.78 |
注:U型风分量和V型风分量分别为纬向风速和经向风速。 |
表2 不同方法在测试集上的性能比较Tab. 2 Performance comparison of different methods on a test set |
方法 | 输入特征 | RMSE | MAE | R2 |
---|---|---|---|---|
GTWR | PM2.5浓度、欧氏距离、气象因素、土地利用 | 19.125 0 | 15.599 5 | 0.683 8 |
STSVR | PM2.5浓度、欧氏距离、气象因素、土地利用 | 14.265 6 | 10.315 7 | 0.763 8 |
GCN+GRU | PM2.5浓度、欧氏距离、气象因素、土地利用 | 11.325 3 | 7.796 0 | 0.851 1 |
本文方法 | PM2.5浓度、地形距离、气象因素、土地利用 | 10.900 0 | 7.398 4 | 0.862 1 |
本文方法 | PM2.5浓度、(风向约束)地形距离、气象因素、土地利用 | 10.046 6 | 6.848 0 | 0.882 8 |
提升/% | - | +11.29 | -12.16 | +3.72 |
表3 不同站点PM2.5浓度预测结果RMSE对比分析表Tab. 3 RMSE comparison of the PM2.5 concentration prediction results at different stations |
站点 | RMSE | ||||||
---|---|---|---|---|---|---|---|
不利气象条件 | 有利气象条件 | ||||||
本文方法(r1) | GCN+GRU(r2) | (r1-r2)/r2×100% | 本文方法(r1) | GCN+GRU(r2) | (r1-r2)/r2×100% | ||
1441A | 7.217 7 | 7.547 3 | 4 | 11.491 3 | 14.768 2 | 22.19 | |
1914A | 3.668 7 | 3.764 5 | 2 | 14.697 0 | 15.994 2 | 8.11 | |
2579A | 16.014 0 | 16.308 8 | 1 | 17.190 0 | 19.263 0 | 10.76 | |
2583A | 8.240 7 | 8.312 6 | 0.8 | 8.580 2 | 9.973 3 | 13.97 |
注: r1和r2分别为本文方法和GCN+GRU方法的RMSE。 |
表4 不同分辨率DSM下的模型预测结果Tab. 4 Model prediction results under different resolution DSM |
不同分辨率/m | RMSE | MAE | R2 |
---|---|---|---|
30 | 10.900 0 | 7.398 4 | 0.862 1 |
300 | 11.390 2 | 7.684 8 | 0.849 4 |
3 000 | 11.685 8 | 8.180 7 | 0.841 4 |
30 000 | 11.737 3 | 8.047 7 | 0.840 0 |
表5 不同土地利用类型上站点的PM2.5浓度预测结果Tab. 5 Prediction results of PM2.5 concentration on different land use types |
预测结果 | 站点 | ||
---|---|---|---|
1911A | 2578A | 2581A | |
RMSE | 9.076 1 | 15.687 9 | 8.521 4 |
MAE | 6.340 1 | 11.370 1 | 6.084 6 |
R2 | 0.884 2 | 0.827 6 | 0.927 9 |
[1] |
|
[2] |
贺灿飞, 张腾, 杨晟朗. 环境规制效果与中国城市空气污染[J]. 自然资源学报, 2013, 28(10):1651-1663.
[
|
[3] |
中华人民共和国生态环境部. 2020中国生态环境状况公报[R]. 2021.
[Ministry of Ecology and Environment of the People's Republic of China. 2020 report on the state of the ecology and environment in China[R]. 2021.]
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
邓涛, 吴兑, 邓雪娇, 等. 珠三角空气质量暨光化学烟雾数值预报系统[J]. 环境科学与技术, 2013, 36(4):62-68.
[
|
[13] |
|
[14] |
王德冬, 秦聪. 区域PM2.5时空回归建模与预测[J]. 中国环境监测, 2019, 35(5):107-113.
[
|
[15] |
|
[16] |
|
[17] |
杨亚莉, 李智伟, 钟卫军. 基于二向注意力循环神经网络的PM2.5浓度预测[J]. 空军工程大学学报(自然科学版), 2020, 21(6):101-106.
[
|
[18] |
李龙, 马磊, 贺建峰, 等. 基于特征向量的最小二乘支持向量机PM2.5浓度预测模型[J]. 计算机应用, 2014, 34(8):2212-2216.
[
|
[19] |
戴李杰, 张长江, 马雷鸣. 基于机器学习的PM2.5短期浓度动态预报模型[J]. 计算机应用, 2017, 37(11):3057-3063.
[
|
[20] |
李祥, 彭玲, 邵静, 等. 基于小波分解和ARMA模型的空气污染预报研究[J]. 环境工程, 2016, 34(8):110-113,134.
[
|
[21] |
秦东明, 丁志军, 金玉鹏, 等. 基于自编码网络的空气污染物浓度预测[J]. 同济大学学报(自然科学版), 2019, 47(5):681-687.
[
|
[22] |
|
[23] |
白盛楠, 申晓留. 基于LSTM循环神经网络的PM2.5预测[J]. 计算机应用与软件, 2019, 36(1):67-70,104.
[
|
[24] |
|
[25] |
黄婕, 张丰, 杜震洪, 等. 基于RNN-CNN集成深度学习模型的PM2.5小时浓度预测[J]. 浙江大学学报(理学版), 2019, 46(3):370-379.
[
|
[26] |
马俊文, 严京海, 孙瑞雯, 等. 基于LSTM-GCN的PM2.5浓度预测模型[J]. 中国环境监测, 2022, 38(5):153-160.
[
|
[27] |
张云伟, 王晴茹, 陈嘉, 等. 城市街谷内PM2.5浓度时空变化及影响因素分析[J]. 中国环境科学, 2016, 36(10):2944-2949.
[
|
[28] |
贺祥, 林振山. 基于GAM模型分析影响因素交互作用对PM2.5浓度变化的影响[J]. 环境科学, 2017, 38(1):22-32.
[
|
[29] |
|
[30] |
许珊, 邹滨, 蒲强, 等. 土地利用/覆盖的空气污染效应分析[J]. 地球信息科学学报, 2015, 17(3):290-299.
[
|
[31] |
|
[32] |
黄小刚, 赵景波, 孙从建, 等. 汾渭平原PM2.5空间分布的地形效应[J]. 环境科学, 2021, 42(10):4582-4592.
[
|
[33] |
|
[34] |
|
[35] |
徐冰冰, 岑科廷, 黄俊杰, 等. 图卷积神经网络综述[J]. 计算机学报, 2020, 43(5):755-780.
[
|
[36] |
|
[37] |
黎煜昭, 刘启亮, 邓敏, 等. 基于物理约束GRU神经网络的河流水质预测模型[J]. 地球信息科学学报, 2023, 25(1):102-114.
[
|
[38] |
|
[39] |
|
[40] |
毛文婧, 王卫林, 焦利民, 等. 基于深度学习的中国连续空间覆盖PM2.5浓度预报[J]. 测绘学报, 2022, 51(3):361-372.
[
|
[41] |
|
[42] |
|
[43] |
钟雨桐, 韦晶, 郑月明, 等. 基于静风期污染物的PM2.5排放清单空间精细化方法[J]. 地球信息科学学报, 2021, 23(11):1971-1983.
[
|
[44] |
邹雨轩, 吴志峰, 曹峥. 耦合土地利用回归与人口加权模型的PM2.5暴露风险评估[J]. 地球信息科学学报, 2019, 21(7):1018-1028.
[
|
/
〈 |
|
〉 |