基于交互差分时空LSTM的网格化臭氧浓度预测  

Prediction of gridded ozone concentration based on interaction differential spatiotemporal LSTM

作　　者：刘恩海 任晓康 张智 李妍 赵娜 张军 LIU Enhai;REN Xiaokang;ZHANG Zhi;LI Yan;ZHAO Na;ZHANG Jun(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;Hebei Meteorological Disaster Prevention and Environmental Meteorology Center,Shijiazhuang Hebei 050000,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401 [2]河北省气象局灾害防御和环境气象中心,河北石家庄050000

出　　处：《河北工业大学学报》2023年第3期36-43,共8页Journal of Hebei University of Technology

基　　金：河北省高等学校科学研究项目(ZD2021311);河北省气象局科研项目(20ky08);河北省自然科学基金(F2020202008)。

摘　　要：臭氧浓度的预测对于大气环境治理、空气质量改善等起到了重要的作用。本文提出了一种交互差分时空LSTM网络预测模型(ST-IDN)来挖掘臭氧浓度历史数据的时间相关性和空间相关性,并成功将其应用到网格化臭氧浓度数据预测上。在该模型中,首先交互模块(IC)可以通过一系列的卷积操作来捕捉短期上下文信息,其次层融合模块(LF)可以融合不同层的空间信息来获得上一时刻丰富的空间信息,最后差分时空LSTM模块(DSTM)将捕捉到的时间信息和空间信息进行统一建模实现臭氧浓度预测。所构建模型分别与卷积LSTM网络(ConvLSTM)、预测循环神经网络(PredRNN)以及Memory in Memory网络(MIM)模型在河北省气象局提供的臭氧浓度数据上进行了对比分析,ST-IDN模型的平均绝对误差分别降低了19.836%、12.924%、7.506%。实验结果表明,所提出的模型能够提高臭氧浓度的预测精度。The prediction of ozone concentration plays an important role in controlling atmospheric environment and im-proving air quality.In this paper,an interactive difference spatiotemporal LSTM network prediction model(ST-IDN)is proposed to mine the temporal and spatial correlation of the historical data of ozone concentration,and is successfully ap-plied to the prediction of gridded ozone concentration data.In this model,firstly,the interactive module(IC)can capture short-term context information through a series of convolution operations,and secondly,the layer fusion module(LF)can fuse spatial information of different layers to obtain abundant spatial information at the last moment.Finally,the differen-tial space-time LSTM module(DSTM)will use the captured time and space information for unified modeling to predict ozone concentration.The constructed models were compared with convolution LSTM network(ConvLSTM),predictive cy-clic neural network(PredRNN)and Memory in Memory network(MIM)models respectively on the ozone concentration data provided by Hebei Meteorological Bureau.The average absolute error of ST-IDN model is reduced by 19.836%,12.924%and 7.506%,respectively.The experimental results show that the proposed model can improve the prediction accuracy of ozone concentration.

关 键 词：交互差分时空LSTM预测网络 网格化臭氧浓度数据 臭氧浓度预测 时间信息 空间信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]