A Model for Predicting Ship Emission Pollutants Based on MASTGCN Using AIS Information
-
摘要: 船舶排放的二氧化硫（SO2）是导致大气污染和海洋酸化的主要因素之一，其时空分异性显著且不均，当前船舶污染物预测模型在时空依赖性建模方面存在局限性，难以有效捕捉船舶SO2排放中的复杂时空关联特征。针对该问题，基于船舶自动识别系统（automatic identification system，AIS）数据及中国船舶基础信息数据，采用动力学方法结合排放因子量化船舶航行过程中的SO2排放量，为后续预测提供了数据支持。在预测模型构建方面，研究了融合多头自注意力机制的时空图卷积网络（multi-head attention spatial-temporal graph convolutional network，MASTGCN）预测模型。该模型以时空图卷积网络（spatial-temporal graph convolutional network，STGCN）为基础架构，在空间和时间维度中引入多头自注意力机制，通过动态权重分配强化对不同区域间空间关联性以及不同时段间时间关联性的建模能力，实现对船舶SO2排放的时空预测。实验结果表明，在注意力头数为5时，模型的平均绝对误差（mean absolute error，MAE）、均方误差（mean squared error，MSE）、均方根误差（root mean squared error，RMSE）以及浮点运算数（floating point operations，FLOPs）分别为0.057 5、0.120 6、0.347 3、3 030 M，模型准确度和计算复杂度的综合性能优于其他头数配置及STGCN模型。相较于STGCN模型，MAE、MSE、RMSE和FLOPs指标分别提高了27.6%、6.0%和1.3%。研究结果表明，多头注意力机制可以通过动态权重分配有效捕获船舶SO2排放的空间特征，5个注意力头的MASTGCN模型在预测精度上表现优秀，同时在计算复杂度方面保持相对合理。Abstract: Sulfur dioxide (SO2) emissions from ships are a major contributor to air pollution and ocean acidification, exhibiting significant spatial and temporal heterogeneity. Current prediction models for shipborne pollutants have limitations in modeling spatiotemporal dependencies, making it difficult to effectively capture the complex spatiotemporal correlation characteristics in SO2 emissions. To address this issue, based on automatic identification system (AIS) data and Chinese ship registry data, a dynamics-based method combined with emission factor approaches is used to quantify shipborne SO2 emissions during navigation, thereby providing a solid data foundation for subsequent prediction. In terms of model construction, a multi-head attention spatial-temporal graph convolutional network (MASTGCN) is proposed. Based on the spatial-temporal graph convolutional network (STGCN) architecture, MASTGCN incorporates multi-head self-attention mechanisms in both spatial and temporal dimensions. By dynamically allocating weights, it enhances the modeling capability to learn spatial dependencies across different regions and temporal dependencies across time intervals, thus improving the accuracy of spatiotemporal predictions for shipborne SO2 emissions. Experimental results show that when the number of attention heads is set to five, the model achieves a mean absolute error (MAE) of 0.057 5, mean squared error (MSE) of 0.120 6, root mean squared error (RMSE) of 0.347 3, and floating point operations (FLOPs) of 3 030 M. These results demonstrate superior overall performance in both accuracy and efficiency compared to other configurations and the baseline STGCN model. Specifically, MASTGCN with five attention heads outperforms STGCN by improving MAE by 27.6%, MSE by 6.0%, and RMSE by 1.3%. The findings indicate that the incorporation of multi-head attention mechanisms enables the model to effectively capture the spatial characteristics of SO2 emissions through dynamic weighting. The five-head MASTGCN model achieves excellent predictive accuracy while maintaining a relatively reasonable computational complexity.
-
图 5 MASTGCN模型整体结构示意图
Figure 5. Schematic of the overall structure of the MASTGCN model
图 8 头数为5时训练集与验证集上损失函数
Figure 8. Loss function on training and validation sets for head count of 5
图 10 不同注意力头数的RMSE误差对比
Figure 10. Comparison of RMSE error for different number of attention heads
表 1 设备排放因子
Table 1. Emission factors for equipment
设备 含硫量/ % $ \mathrm{SO}_{2} $排放因子$ /(\mathrm{g} /(\mathrm{kW} \cdot \mathrm{h})) $ 燃油消耗量$ /(\mathrm{g} /(\mathrm{kW} \cdot \mathrm{h})) $ 主机 0.50 2 203 辅机 0.50 2.1 217 锅炉 0.50 2.8 290 
下载: 导出CSV
表 2 实验环境
Table 2. Experimental environment
名称 型号版本 中央处理器（CPU） Intel（R）Xeon（R）Platinum 8358 CPU 图形处理器（GPU） NVIDIA Tesla-A40-48G 内存/GB 128 系统盘/数据盘GB 20/50 操作系统 Ubuntu 20.04 Python 3.8 PyTorch 2.0.0 CUDA 12.2
下载: 导出CSV
表 3 网络超参数设置
Table 3. Network hyperparameter settings
超参数 设置值 学习率 0.001 迭代次数 96 优化器 Adam 批量大小 32 时间步长 12
下载: 导出CSV
表 4 不同模型性能对比分析表
Table 4. Comparative analysis of the performance of
模型 MAE MSE RMSE ARIMA 0.6949 1.9240 0.9003 DT 0.0757 0.2793 0.5285 LR 0.0621 0.1490 0.3860 GRU 0.0654 0.1276 0.3572 LSTM 0.0612 0.1375 0.3709 Transformer 0.0998 0.0628 0.2506 CNN+LSTM 0.0608 0.1413 1.2441 STGCN 0.0794 0.1283 0.3520 MASTGCN 0.0575 0.1206 0.3473
下载: 导出CSV
表 5 不同头数性能对比分析表
Table 5. Comparative analysis of the performance of different head counts
模型 HEAD MAE MSE RMSE STGCN 0.0794 0.1283 0.3520 1 0.0590 0.1218 0.3490 2 0.0620 0.1222 0.3496 MASTGCN 3 0.0611 0.1202 0.3467 4 0.0642 0.1188 0.3446 5 0.0575 0.1206 0.3473 6 0.0598 0.1174 0.3427
下载: 导出CSV
表 6 不同头数模型复杂度对比分析表
Table 6. Comparison and Analysis of Model Complexity with Different Numbers of Heads
模型 HEAD Parameters FLOPs/M STGCN 57096 866 1 66120 1298 2 75064 1731 MASTGCN 3 84008 2164 4 92952 2597 5 101896 3030 6 110840 3463
下载: 导出CSV
-
[1] STRIEBIG B, SMITTS E, MORTON S. Impact of transportation on carbon dioxide emissions from locally vs. non-locally sourced food[J]. Emerging Science Journal, 2019, 3(4): 222-234. doi: 10.28991/esj-2019-01184 [2] 李沁宇, 刘鑫, 张金池. 长三角区域酸雨类型转变趋势研究[J]. 南京林业大学学报(自然科学版), 2021, 45(1): 168-174.LI Q Y, LIU X, ZHANG J C. Changing trends of acid rain types in the Yangtze River Delta region[J]. Journal of Nanjing Forestry University(Natural Sciences Edition), 2021, 45 (1): 168-174. (in Chinese) [3] CAO Y, ZHANG W, WANG X, et al. The relationship be tween atmospheric pollutant emissions and fuel qualities of in land vessels in Jiangsu province, China[J]. Journal of the Air & Waste Management Association, 2019, 69(3): 305-312. [4] 万霖, 何凌燕, 黄晓锋. 船舶大气污染排放的研究进展[J]. 环境科学与技术, 2013, 36(5): 57-62.WAN L, HE L Y, HUANG X F. Progress in research of air pollution emissions from ships[J]. Environmental Science & Technology, 2013, 36(5): 57-62. (in Chinese) [5] 中华人民共和国生态环境部. 中国移动源环境管理年报(2023年)[J]. 环境保护, 2024, 52(2): 48-62.Ministry of Ecology and Environment of the People's Republic of China. China mobile source environmental management annual report in 2023[J]. Environmental Protection, 2024, 52 (2): 48-62(. in Chinese) [6] CIMORELLI A J, PERRY S G, VENKATRAM A, et al. AERMOD: a dispersion model for industrial source applications. Part Ⅰ: general model formulation and boundary layer characterization[J]. American Meteorological Society, 2005, 44(5): 682-693. [7] 陈晓春, 李建晖, 陈少博, 等. 基于CALPUFF模型的山东省独立焦化行业大气污染物分析与预测[J]. 工程科学学报, 2024, 46(7): 1311-1323.CHEN X C, LI J H, CHEN S B, et al. Analysis and prediction of air pollutants in the independent coking industry in Shandong province based on the CALPUFF model[J]. Chinese Journal of Engineering, 2024, 46(7): 1311-1323. (in Chinese) [8] 燕鸥, 王体健, 阮兆元, 等. 基于ADMS模型的工业园区VOCs排放源强反演[J]. 环境科学, 2025, 46(3): 1340-1349.YAN O, WANG T J, RUAN Z Y, et al. VOCs emission retrieval in industrial parks based on the ADMS mode[J]. Environmental Science, 2025, 46(3): 1340-1349. (in Chinese) [9] GARBATOV Y, GEORGIEV P. Stochastic air quality dispersion model for defining queuing ships seaport location[J]. Journal of Marine Science and Engineering, 2022, 10(2): 140-140. doi: 10.3390/jmse10020140 [10] ABRUTYTĖ E, ŽUKAUSKAITĖ A, MICKEVIČIENĖ R, et al. Evaluation of NOx emission and dispersion from marine ships in Klaipeda sea port[J]. Journal of Environmental Engineering and Landscape Management, 2014, 22(4): 264-273. doi: 10.3846/16486897.2014.892009 [11] TSENG Y L, CHENG W H, YUAN C S, et al. Impacts of ship emissions and sea-land breeze on urban air quality using chemical characterization, source contribution and dispersion model simulation of PM2.5 at Asian seaport[J]. Environmental Pollution, 2024, 347: 123663. doi: 10.1016/j.envpol.2024.123663 [12] HONG H, CHOI I H, JEON H, et al. An air pollutants prediction method integrating numerical models and artificial intelligence models targeting the area around Busan port in Korea[J]. Atmosphere, 2022, 13(9): 1462-1462. doi: 10.3390/atmos13091462 [13] XU H, WU L, XIONG S, et al. An improved CNN-LSTM model-based state-of-health estimation approach for lithiumion batteries[J]. Energy, 2023, 276: 127585. doi: 10.1016/j.energy.2023.127585 [14] 黄怡容, 熊秋林, 熊正坤, 等. 基于CNN-LSTM的鄱阳湖生态经济区大气污染物时空预测[J]. 生态环境学报, 2024, 33 (12): 1891-1901.HUANG Y R, XIONG Q L, XIONG Z K, et al. The spatiotemporal prediction of air pollutants in the Poyang Lake ecological economic zone based on CNN-LSTM[J]. Ecology and Environmental Sciences, 2024, 33(12): 1891-1901. (in Chinese) [15] QI Y, LI Q, KARIMIAN H, et al. A hybrid model for spatiotemporal forecasting of PM2.5 based on graph convolutional neural network and long short-term memory[J]. Science of the Total Environment, 2019, 664: 1-10 doi: 10.1016/j.scitotenv.2019.01.333 [16] 马园. 基于深度学习的空气质量预测研究[D]. 济南: 山东大学, 2023.MA Y. Research on air quality prediction based on deep learning[D]. Jinan: Shandong University, 2023. (in Chinese) [17] 黄敏, 杨亚东, 吴新鹏. 基于时空注意力机制的STA-GRU船舶轨迹预测方法[J]. 交通信息与安全, 2023, 41(6): 82-89. doi: 10.3963/j.jssn.1674-4861.2023.06.009HUANG M, YANG Y D, WU X P. Ship trajectory prediction method of gated recurrent unit based on spatial-temporal attention mechanism[J]. Journal of Transport Information and Safety, 2023, 41(6): 82-89. doi: 10.3963/j.jssn.1674-4861.2023.06.009 [18] 王征, 张卫, 彭传圣, 等. 中国近周边海域船舶排放清单及排放特征研究[J]. 交通节能与环保, 2018, 14(1): 11-15.WANG Z, ZHANG W, PENG C S, et al. Study on characteristics of ship emissions from surrounding China seas[J]. Transport Energy Conservation & Environmental Protection, 2018, 14(1): 11-15. (in Chinese) [19] 朱倩茹, 廖程浩, 王龙, 等. 基于AIS数据的精细化船舶排放清单方法[J]. 中国环境科学, 2017, 37(12): 4493-4500.ZHU Q R, LIAO C H, WANG L, et al. Application of fine vessel emission inventory compilation method based on AIS data[J]. China Environmental Science, 2017, 37(12): 4493-4500. (in Chinese) [20] 叶斯琪. 珠江三角洲地区船舶排放特征及对区域空气质量影响的研究[D]. 广州: 华南理工大学, 2014.YE S Q. Study of characteristics of marine vessel emission and its impact on regional air quality of the Pearl River Delta region[D]. Guangzhou: South China University of Technology, 2014. (in Chinese) [21] AGRAWAL H, EDEN R, ZHANG X, et al. Primary particulate matter from ocean-going engines in the southern California air basin[J]. Environmental Science & Technology, 2009, 43(14): 5398-5402. -
点击查看大图
图(10) / 表(6)
计量
- 文章访问数: 26
- HTML全文浏览量: 16
- PDF下载量: 0
- 被引次数: 0