-
近地面高浓度臭氧(O3)会增强大气氧化性, 加重城市环境空气污染, 长期处于高浓度臭氧环境下会诱发心血管和呼吸系统疾病[1-2]. 准确预测臭氧浓度能够为臭氧防控治理提供重要支持, 及时污染预警可为居民出行决策提供建议, 降低健康影响. 臭氧浓度与前体物排放、气象、地形等因素密切相关, 具有高度复杂性和非线性变化特征[3], 存在着显著的时空关联特征[4]. 如何有效学习臭氧浓度分布的时空关联特征, 并用于臭氧浓度预测已成为关注的焦点.
目前大气臭氧浓度预测的方法主要有如下3种方法:(1)基于物理化学反应机制的空气质量模式[5-7], 该模式基于污染源排放清单、气象条件和大气边界条件, 模拟污染物在大气环境中的物理化学变化过程获得预测结果, 但该方法计算量大, 特征提取困难且运行成本高[8]. (2)基于统计学理论的预测模型[9-12], 该类统计方法对时间序列数据特征提取能力有限[13], 导致预测精度偏低且仅能实现较短步长预测. (3)基于机器学习算法的数值预测方法, 目前支持向量机算法及其改进算法被广泛应用于大气污染物浓度预测中[14-17]. 随机森林算法可以处理高维度数据并且可以得到变量重要性[18-20], 在大气污染物浓度预测领域取得了一定的成果. 深度学习算法[21-26]是机器学习领域最新的发展成果, 可深层次提取数据特征, 较好捕捉数据间的非线性关系. 深度学习中, 长短期记忆神经网络(long short-term memory neural network, LSTM) [27]能够提取时间序列数据的变化特征, 不受传统循环神经网络(recurrent neural network, RNN)梯度消失的影响[28], 同时具有序列到序列的多步预测能力. 有研究[29-31]使用LSTM模型预测臭氧浓度, 但LSTM模型无法考虑站点的空间关联影响, 导致模型预测准确度不高. 将 LSTM与卷积神经网络(convolutional neural network, CNN)[32-33], 耦合可处理空间信息, 从而更准确预测臭氧浓度.CNN适用于处理欧式空间数据, 在处理臭氧浓度分布这类非欧式空间数据时表现较差;而图卷积神经网络(graph convolutional neural network, GCN)[34]基于图傅里叶变换及拉普拉斯矩阵, 能够更好提取臭氧浓度分布这类非欧式空间数据特征. 因此, 将LSTM与GCN耦合能够捕获臭氧浓度的时空依赖关系, 相较于单独使用一种模型, 耦合模型预测准确性更高.
本研究建立了LSTM与GCN耦合的臭氧小时浓度预测模型, 并应用该模型预测北京市未来72 h臭氧浓度, 为臭氧预测预报提供了一种新的方法.
基于深度学习的城市臭氧小时浓度预测模型
Prediction model of urban ozone hourly concentration based on deep learning
-
摘要: 近地面高浓度臭氧(O3)对城市环境空气质量、植物生长和人体健康等均有较大影响. 因此,精准预报臭氧浓度对城市环境管理部门臭氧污染防治、居民出行决策建议、降低健康影响等具有重要意义. 深度学习模型对于非线性关系具有较强捕捉和学习能力,因此本研究提出一种基于深度学习算法的混合模型,利用图卷积神经网络(GCN)及长短期记忆神经网络(LSTM)分别捕捉臭氧浓度空间和时间变化特征,耦合气象因子,构建基于时空关联的臭氧小时浓度预测模型GCN-LSTM,并以北京市为例开展应用研究. 结果显示,GCN-LSTM模型可较好预测北京市未来72 h臭氧浓度,预测值与观测值决定系数为0.86;预测未来24、48、72 h臭氧浓度平均相对偏差分别为18.2%、19.2%和22.9%,RMSE值分别为17.3、23.7、25.4 μg·m−3,对于48 —72 h的长时预测准确度优于已有机器学习模型;当臭氧观测浓度介于0—80 μg·m−3、80—160 μg·m−3和160—200 μg·m−3时(共占总数据量的96.3%),预测平均相对偏差分别为20.1%、6.9%和16.4%;预测不同类型站点浓度时发现,城市清洁对照点、城市环境评价点、区域背景传输点和交通污染监控点的平均相对偏差分别为7.9%、13.2%、24.4%和29.3%,RMSE值分别为10.8、14.9、20.1、31.4 μg·m−3,模型对城市清洁对照点和城市环境评价点的预测准确度较高. 使用本模型对城市大气臭氧小时浓度预测,将较好助力城市大气臭氧污染防治工作.
-
关键词:
- O3 /
- 小时浓度预测 /
- 深度学习 /
- 图卷积神经网络 /
- 长短期记忆神经网络.
Abstract: High near-surface ozone concentrations (O3) have a significant impact on urban ambient air quality, plant growth and human health. Therefore, accurate forecasting of ozone concentrations is important for urban environmental management departments to prevent and control ozone pollution, advise residents on travel decisions, and reduce health impacts. Deep learning models have strong capturing and learning ability for nonlinear relationships. Therefore, this study proposes a hybrid model based on deep learning algorithm, using graph convolutional neural network (GCN) and long short-term memory neural network (LSTM) to capture the spatial and temporal variation of O3 concentration features respectively, coupled with meteorological factors, to build a GCN-LSTM based on spatio-temporal correlation of O3 hourly concentration prediction model, and to conduct an application study in Beijing. The results show that GCN-LSTM model can predict the future O3 concentration in Beijing for 72 hours with a correlation coefficient of 0.86 between the predicted and observed values; the mean relative bias (MRB) of the predicted future O3 concentrations for 24, 48 and 72 hours are 18.2%, 19.2% and 22.9%, with root mean square error (RMSE) values of 17.3, 23.7 and 25.4 μg·m−3 respectively; the prediction accuracy was better than that of the existing machine learning models for the long time of 48—72 h; when the observed O3 concentrations ranged from 0—80 μg·m−3, 80—160 μg·m−3 and 160—200 μg·m−3 (a total of 96.3% of the total data volume), the MRB were 20.1%, 6.9% and 16.4%, the RMSE were 10.8, 14.9, 20.1, 31.4 μg·m−3 respectively. The MRB of the predicted concentrations at different types of sites were found to be 7.9%, 13.2%, 24.4% and 29.3% for the urban clean control site, urban environmental assessment site, regional background transmission site and traffic pollution monitoring site, respectively, and model predicts urban clean control points and urban environmental assessment points with high accuracy. Using this model to predict the hourly O3 concentration in urban air will better help to prevent and control O3 pollution in urban air. -
-
[1] WANG M Y, YIM S H L, DONG G H, et al. Mapping ozone source-receptor relationship and apportioning the health impact in the Pearl River Delta region using adjoint sensitivity analysis [J]. Atmospheric Environment, 2020, 222: 117026. doi: 10.1016/j.atmosenv.2019.117026 [2] GUAN Y, XIAO Y, WANG F Y, et al. Health impacts attributable to ambient PM2.5 and ozone pollution in major Chinese cities at seasonal-level [J]. Journal of Cleaner Production, 2021, 311: 127510. doi: 10.1016/j.jclepro.2021.127510 [3] 董红召, 王乐恒, 唐伟, 等. 融合时空特征的PCA-PSO-SVM臭氧(O3)预测方法研究 [J]. 中国环境科学, 2021, 41(2): 596-605. DONG H Z, WANG L H, TANG W, et al. Research on PCA-PSO-SVM ozone prediction considering spatial-temporal features [J]. China Environmental Science, 2021, 41(2): 596-605(in Chinese).
[4] 李子凌. 基于时空数据的臭氧特征分析及其预测算法研究[D]. 北京: 北京交通大学, 2020. LI Z L. Research on ozone feature analysis and prediction algorithm based on spatio-temporal data[D]. Beijing: Beijing Jiaotong University, 2020(in Chinese).
[5] 肖德林, 邓仕槐, 邓小函, 等. 达州市城区环境空气质量变化趋势及CMAQ模型预报分析 [J]. 中国环境监测, 2021, 37(4): 92-103. XIAO D L, DENG S H, DENG X H, et al. Analysis of ambient air quality variation trend and CMAQ model forecast system in urban areas of Dazhou City [J]. Environmental Monitoring in China, 2021, 37(4): 92-103(in Chinese).
[6] RYU Y H, HODZIC A, DESCOMBES G, et al. Toward a better regional ozone forecast over CONUS using rapid data assimilation of clouds and meteorology in WRF-chem [J]. Journal of Geophysical Research:Atmospheres, 2019, 124(23): 13576-13592. doi: 10.1029/2019JD031232 [7] 周广强, 瞿元昊, 余钟奇. 长江三角洲城市臭氧数值预报与释用 [J]. 中国环境科学, 2021, 41(1): 28-36. ZHOU G Q, QU Y H, YU Z Q. Numerical forecast and improvement of ozone over YRD cities [J]. China Environmental Science, 2021, 41(1): 28-36(in Chinese).
[8] 邹国建. 基于时空特征学习的区域空气污染物扩散趋势预测研究[D]. 上海: 上海师范大学, 2020. ZOU G J. Study on prediction of regional air pollutant diffusion trend based on spatiotemporal feature learning[D]. Shanghai: Shanghai Normal University, 2020(in Chinese).
[9] 丁愫, 陈报章, 王瑾, 等. 基于决策树的统计预报模型在臭氧浓度时空分布预测中的应用研究 [J]. 环境科学学报, 2018, 38(8): 3229-3242. DING S, CHEN B Z, WANG J, et al. An applied research of decision-tree based statistical model in forecasting the spatial-temporal distribution of O3 [J]. Acta Scientiae Circumstantiae, 2018, 38(8): 3229-3242(in Chinese).
[10] 梁炜, 李雅箐, 黄喜寿, 等. 基于ARMA-GARCH模型的南宁市O3浓度预测研究 [J]. 广西科学, 2020, 27(1): 91-97. LIANG W, LI Y Q, HUANG X S, et al. Research on atmospheric ozone concentration prediction based on ARMA-GARCH model in Nanning [J]. Guangxi Sciences, 2020, 27(1): 91-97(in Chinese).
[11] 蔡旺华. 运用机器学习方法预测空气中臭氧浓度 [J]. 中国环境管理, 2018, 10(2): 78-84. CAI W H. Using machine learning method for predicting the concentration of ozone in the air [J]. Chinese Journal of Environmental Management, 2018, 10(2): 78-84(in Chinese).
[12] 彭岩, 冯婷婷, 王洁. 基于集成学习的O3的质量浓度预测模型 [J]. 山东大学学报(工学版), 2020, 50(4): 1-7. PENG Y, FENG T T, WANG J. An integrated learning approach for O3 mass concentration prediction model [J]. Journal of Shandong University (Engineering Science), 2020, 50(4): 1-7(in Chinese).
[13] 王舒扬, 姜金荣, 迟学斌, 等. 模式预报数据的深度学习PM2.5浓度预测模型[J]. 数值计算与计算机应用, 2022, 43(2): 142-153. WANG S Y, JIANG J R, CHI X B, et al. A deep learning model for forecasting PM2.5 combined with numerical model Data[J/OL]. Journal on Numerical Methods and Computer Application, 2022, 43(2): 142-153(in Chinese).
[14] SUN W, SUN J Y. Daily PM2.5 concentration prediction based on principal component analysis and LSSVM optimized by cuckoo search algorithm [J]. Journal of Environmental Management, 2017, 188: 144-152. [15] 宋国君, 国潇丹, 杨啸, 等. 沈阳市PM2.5浓度ARIMA-SVM组合预测研究 [J]. 中国环境科学, 2018, 38(11): 4031-4039. doi: 10.3969/j.issn.1000-6923.2018.11.005 SONG G J, GUO X D, YANG X, et al. ARIMA-SVM combination prediction of PM2.5 concentration in Shenyang [J]. China Environmental Science, 2018, 38(11): 4031-4039(in Chinese). doi: 10.3969/j.issn.1000-6923.2018.11.005
[16] 李建新, 刘小生, 刘静, 等. 基于MRMR-HK-SVM模型的PM2.5浓度预测 [J]. 中国环境科学, 2019, 39(6): 2304-2310. doi: 10.3969/j.issn.1000-6923.2019.06.009 LI J X, LIU X S, LIU J, et al. Prediction of PM2.5 concentration based on MRMR-HK-SVM model [J]. China Environmental Science, 2019, 39(6): 2304-2310(in Chinese). doi: 10.3969/j.issn.1000-6923.2019.06.009
[17] 康俊锋, 黄烈星, 张春艳, 等. 多机器学习模型下逐小时PM2.5预测及对比分析 [J]. 中国环境科学, 2020, 40(5): 1895-1905. doi: 10.3969/j.issn.1000-6923.2020.05.005 KANG J F, HUANG L X, ZHANG C Y, et al. Hourly PM2.5 prediction and its comparative analysis under multi-machine learning model [J]. China Environmental Science, 2020, 40(5): 1895-1905(in Chinese). doi: 10.3969/j.issn.1000-6923.2020.05.005
[18] ZENG Z L, WANG Z M, GUI K, et al. Daily global solar radiation in China estimated from high-density meteorological observations: A random forest model framework [J]. Earth and Space Science, 2020, 7(2): e2019EA001058. [19] 侯俊雄, 李琦, 朱亚杰, 等. 基于随机森林的PM2.5实时预报系统 [J]. 测绘科学, 2017, 42(1): 1-6. HOU J X, LI Q, ZHU Y J, et al. Real-time forecasting system of PM2.5 concentration based on spark framework and random forest model [J]. Science of Surveying and Mapping, 2017, 42(1): 1-6(in Chinese).
[20] HUANG K Y, XIAO Q Y, MENG X, et al. Predicting monthly high-resolution PM2.5 concentrations with random forest model in the North China Plain [J]. Environmental Pollution, 2018, 242: 675-683. doi: 10.1016/j.envpol.2018.07.016 [21] KAPADIA D, JARIWALA N. Prediction of tropospheric ozone using artificial neural network (ANN) and feature selection techniques [J]. Modeling Earth Systems and Environment, 2022, 8(2): 2183-2192. doi: 10.1007/s40808-021-01220-6 [22] KUMAR N, MIDDEY A, RAO P S. Prediction and examination of seasonal variation of ozone with meteorological parameter through artificial neural network at NEERI, Nagpur, India [J]. Urban Climate, 2017, 20: 148-167. doi: 10.1016/j.uclim.2017.04.003 [23] SAYEED A, CHOI Y, ESLAMI E, et al. Using a deep convolutional neural network to predict 2017 ozone concentrations, 24 hours in advance [J]. Neural Networks, 2020, 121: 396-408. doi: 10.1016/j.neunet.2019.09.033 [24] WANG H W, LI X B, WANG D S, et al. Regional prediction of ground-level ozone using a hybrid sequence-to-sequence deep learning approach [J]. Journal of Cleaner Production, 2020, 253: 119841. doi: 10.1016/j.jclepro.2019.119841 [25] 贾鹏程. 基于深度学习的长三角地区臭氧临近预报技术研究[D]. 南京: 南京信息工程大学, 2021. JIA P C. Deep learning based ozone prediction technique in Yangtze River Delta region[D]. Nanjing: Nanjing University of Information Science & Technology, 2021(in Chinese).
[26] 万显烈, 杨凤林, 王慧卿. 利用人工神经网络对空气中O3浓度进行预测 [J]. 中国环境科学, 2003, 23(1): 110-112. doi: 10.3321/j.issn:1000-6923.2003.01.025 WAN X L, YANG F L, WANG H Q. The approach of artificial neural network applied in ambient ozone forecast [J]. China Environmental Science, 2003, 23(1): 110-112(in Chinese). doi: 10.3321/j.issn:1000-6923.2003.01.025
[27] HOCHREITER S, SCHMIDHUBER J. Long short-term memory [J]. Neural Computation, 1997, 9(8): 1735-1780. doi: 10.1162/neco.1997.9.8.1735 [28] 周永生. 基于LSTM神经网络的PM2.5预测[D]. 长沙∶湖南大学, 2018. ZHOU Y S. PM2.5 Prediction based on LSTM neural network [D]. Changsha: Hunan University, 2018(in Chinese).
[29] AL-JANABI S, MOHAMMAD M, AL-SULTAN A. A new method for prediction of air pollution based on intelligent computation [J]. Soft Computing, 2020, 24(1): 661-680. doi: 10.1007/s00500-019-04495-1 [30] FREEMAN B S, TAYLOR G, GHARABAGHI B, et al. Forecasting air quality time series using deep learning [J]. Journal of the Air & Waste Management Association, 2018, 68(8): 866-886. [31] JIA P C, CAO N W, YANG S B. Real-time hourly ozone prediction system for Yangtze River Delta area using attention based on a sequence to sequence model [J]. Atmospheric Environment, 2021, 244: 117917. doi: 10.1016/j.atmosenv.2020.117917 [32] PAK U, KIM C, RYU U, et al. A hybrid model based on convolutional neural networks and long short-term memory for ozone concentration prediction [J]. Air Quality, Atmosphere & Health, 2018, 11(8): 883-895. [33] 方韬. 基于神经网络的近地面臭氧估算和预测研究[D]. 上海: 上海师范大学, 2020. FANG T. Study on estimation and prediction of near-surface ozone based on neural network[D]. Shanghai: Shanghai Normal University, 2020(in Chinese).
[34] ZHOU J, CUI G Q, HU S D, et al. Graph neural networks: A review of methods and applications [J]. AI Open, 2020, 1: 57-81. doi: 10.1016/j.aiopen.2021.01.001 [35] RESHEF D N, RESHEF Y A, FINUCANE H K, et al. Detecting novel associations in large data sets [J]. Science, 2011, 334(6062): 1518-1524. doi: 10.1126/science.1205438 [36] 高婵娟, 赵啟超, 丁若男, 等. 2018年吉林市大气污染物浓度变化及其与气象因素的相关性分析 [J]. 环境工程, 2021, 39(5): 71-79. GAO C J, ZHAO Q C, DING R N, et al. Variations of atmospheric pollutants concentrations and their correlation with meteorological factor in Jilin City in 2018 [J]. Environmental Engineering, 2021, 39(5): 71-79(in Chinese).
[37] ZHENG Y, YI X W, LI M, et al. Forecasting fine-grained air quality based on big data[C]//Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Sydney NSW Australia. New York, NY, USA: ACM, 2015: 2267-2276.