作者:朱菊香1, 谷卫2, 罗丹悦2, 潘斐2, 张赵良1
作者单位:1. 无锡学院轨道交通,江苏 无锡 214105;
2. 南京信息工程大学自动化学院,江苏 南京 210000
关键词:卷积神经网络;改进长短期记忆网络;PM2.5浓度预测;注意力机制
摘要:
为提高PM2.5的预测精度,提出一种基于卷积神经网络(CNN)、改进长短期记忆网络(ILSTM)和注意力机制(attention)组合的预测模型。ILSTM删除LSTM中的输出门,改进其输入门和遗忘门,并引入转换信息模块(CIM),以防止学习过程中的过饱和。该模型将一维卷积神经网络的特征提取和改进长短期记忆网络学习序列依赖性的能力相结合,得到过去不同时间特征状态对未来PM2.5浓度的影响,可以有效模拟PM2.5在时间和空间上的依赖性,并通过注意力机制自动权衡过去的特征状态,进一步提升PM2.5预测的准确度。实验结果表明:CNN-ILSTM-attention模型的拟合度达到98.5%,与LSTM模型、CNN-LSTM模型和CNN-ILSTM模型相比,分别提高26%、9.2%和6.2%,具有较高的预测精度和应用价值。
Prediction modeling of PM2.5 in subway station based on attention mechanism and CNN-ILSTM
ZHU Juxiang1, GU Wei2, LUO Danyue2, PAN Fei2, ZHANG Zhaoliang1
1. Rail Transit, Wuxi University, Wuxi 214105, China;
2. School of Automation, Nanjing University of Information Technology, Nanjing 210000, China
Abstract: In order to improve the prediction accuracy of PM2.5, a combined prediction model based on convolutional neural network (CNN), improved long short-term memory network (ILSTM) and attention mechanism was proposed. ILSTM removes the output gate in LSTM, improves its input gate and forget gate, and introduces a transition information module (CIM) to prevent oversaturation during learning. One-dimensional convolutional neural network models of feature extraction and the ability to improve both short-term and long-term memory network learning sequence dependent state of different time of the past, the combination of characteristics of PM2.5 concentrations in the future, can effectively simulate the PM2.5 dependent on time and space, and through the attention mechanism automatic weighing the characteristics of the state in the past, to further improve the accuracy of PM2.5 prediction. The experimental results show that the fitting degree of CNN-ILSTM-attention model reaches 98.5%, which is improved by 26%, 9.2% and 6.2%, respectively, compared with LSTM model, CNN-LSTM model and CNN-ILSTM model. It has high prediction accuracy and application value.
Keywords: convolutional neural network;improved long short-term memory network;PM2.5 concentration prediction;attention mechanism
2024, 50(7):53-62 收稿日期: 2022-07-05;收到修改稿日期: 2022-09-11
基金项目: 国家自然科学青年基金(51206082);江苏省自然科学基金(be2015692);江苏省高等学校自然科学研究项目(21KJB460005)
作者简介: 朱菊香(1979-),女,江苏常州市人,副教授,硕士生导师,研究方向为自动化及控制技术、检测技术。
参考文献
[1] XI G, LI W D. A graph-based LSTM model for PM2. 5 forecasting[J]. Atmospheric Pollution Research, 2021 (12): 101150.
[2] 杨国亮, 余华声, 黄聪. 基于优化组合模型的PM2.5浓度预测[J]. 计算机工程与设计, 2023, 44(10): 3132-3137.
YANG G L, YU H S , HUANG C. Prediction of PM2.5 concentration based on optimized combination model[J]. Computer Engineering and Design, 2023, 44(10): 3132-3137.
[3] 邢红涛, 郭江龙, 刘书安, 等. 基于CNN-LSTM混合神经网络模型的NOx排放预测[J]. 电子测量技术, 2022, 45(2): 98-103.
XING H T, GUO J L, LIU S A, et al. NOx emission prediction based on CNN-LSTM hybrid neural network model[J]. Electronic Measurement Technology, 2022, 45(2): 98-103.
[4] LOY-BENITEZ J, VILELA P, LI Q, et al. Sequential prediction of quantitative health risk assessment for the fine particulate matter in an underground facility using deep recurrent neural networks[J]. Ecotoxicology and Environmental Safety, 2019(169): 316-324.
[5] NOURI A, MEHDI G L. Prediction of PM2. 5 concentrations using principal component analysis and artificial neural network techniques: A case study: Urmia, Iran[J]. Environmental Engineering Science, 2021 (38): 89-98.
[6] LI M. Prediction of PM2. 5 concentration based on the similarity in air quality monitoring network[J]. Building and Environment, 2018(137): 11-17.
[7] 刘拥民, 罗皓懿, 谢铁强. 基于XGBoost-ARIMA方法的PM2.5质量浓度预测模型的研究及应用[J]. 安全与环境学报, 2023, 23 (1): 211-221.
LIU Y M, LUO H Y, XIE T Q. Research and application of PM2.5 mass concentration prediction model based on XGBoost ARIMA method [J]. Journal of Safety and Environment, 2023, 23 (1): 211-221.
[8] FENG R. Recurrent neural network and random forest for analysis and accurate forecast of atmospheric pollutants[J]. Journal of Cleaner Production, 2019 (231): 1005-1015.
[9] 于伸庭, 刘萍. 基于长短期记忆网络-卷积神经网络(LSTM-CNN)的北京市PM2.5浓度预测[J]. 环境工程, 2020, 38(6): 176-180.
YU S T, LIU P. Prediction of PM2.5 concentration in Beijing based on long short termmemory network convolutional neural network (LSTM-CNN) [J]. Environmental Engineering, 2020, 38(6): 176-180.
[10] ONG B, KOJI Zl. Dynamically pre-trained deep recurrent neural networks using environmental monitoring data for predicting PM2. 5[J]. Neural Computing and Applications, 2016(27): 1553-1566.
[11] 朱菊香, 谷卫, 罗丹悦, 等. 基于PSO优化BP神经网络的多传感器数据融合[J]. 中国测试, 2022, 48(8): 94-100.
ZHU J X, GU W, LUO D Y, et al. Multi sensor data fusion based on PSO optimized BP neural network[J]. China Measurment & Test, 2022, 48(8): 94-100.
[12] ALBIRMIZY S, BANAZ Q. Multivariate time series forecasting with lstm for madrid, spain pollution[C]//2019 International Conference on Computing and Information Science and Technology and Their Applications (ICCISTA), 2019.
[13] ABIRAMI S. Hybrid spatio-temporal deep learning framework for particulate matter (PM2.5) concentration forecasting[C]//2020 International Conference on Innovative Trends in Information Technology (ICITIIT), 2020.
[14] 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.
BAI S N, SHEN X L. PM2.5 prediction based on LSTM recurrent neural network[J]. Computer Applications and Software, 2019, 36(1): 67-70.
[15] 赵彦明. 基于时空相关性的LSTM算法及PM2.5浓度预测应用 [J]. 计算机应用与软件, 2021, 38 (6): 249-255.
ZHAO Y M. LSTM algorithm based on spatiotemporal correlation and its application in predicting PM2.5 concentration [J]. Computer Applications and Software, 2021, 38 (6): 249-255.