基于时空优化的多尺度卷积神经网络空气质量预测模型  

Multi-scale Convolutional Neural Network Air Quality Prediction Model Based on Spatio-Temporal Optimization

作　　者：周杰 罗云芳[1] 雷耀建 李文敬 封宇[1] ZHOU Jie;LUO Yun-fang;LEI Yao-jian;LI Wen-jing;FENG Yu(College of Mechanical,Electrical and Information Engineering,Guangxi Vocational&Technical College,Nanning 530226,China;Guangxi College for Preshcool Education,Nanning 530022,China;School of Logistics Management and Engineering,Nanning Normal University,Nanning 530001,China)

机构地区：[1]广西职业技术学院机电与信息工程学院,南宁530226 [2]广西幼儿师范高等专科学校,南宁530022 [3]南宁师范大学物流管理与工程学院,南宁530001

出　　处：《计算机科学》2020年第S02期535-540,共6页Computer Science

基　　金：国家自然科学基金项目(61866006);广西教育厅自然科学基金项目(2019ITA01002,2018KY0951,2019KY1220,2017KY0980);广西职业技术学院自然科学类课题(桂职院(2019)176号191202)。

摘　　要：目前针对空气质量数值预测多采用单一站点时间序列特征进行浓度预测,没有考虑空气质量数值的变化受空间特征的影响。针对该问题提出一种基于时空优化的多尺度神经网络(MSCNN-GALSTM)模型用于空气质量预测,利用一维多尺度卷积核(MSCNN)提取空气质量数据中的局部时间关系和空间特征关系,并进行线性拼接融合,得出多站点多特征的相互时空特征关系,结合长短记忆网络(LSTM)处理时间序列的优势,并引入遗传算法(GA)对LSTM网络的参数集进行全局寻优,把多站点多特征的相互时空关系输入至LSTM网络中,进而输出多站点多特征的长期特征依赖关系。最后将MSCNN-GALSTM模型与单一LSTM基准模型和单尺度卷积神经网络模型进行对比,均方根误差(RMSE)下降约11%,平均预测准确率提升约20%。实验结果表明,MSCNN-GALSTM预测模型在空气质量数据预测中特征提取更加全面、层次更深,预测精度更高,并且表现出了更好的泛化能力。At present,the air quality prediction is mainly based on the time series of a single station,without considering the influence of the spatial characteristics on the air quality.To solve this problem,a multi-scale neural network(MSCNN-GALSTM)model based on spatiotemporal optimization is proposed for air quality prediction.one-dimensional multi-scale convolution kernel(MSCNN)is used to extract the local temporal and spatial characteristic relations in air quality data,the LINEAR SPLICING and fusion are carried out to obtain the space-time characteristic relation of multi-sites and multi-features,combine the advantage of long-short memory network(LSTM)to process time series,and introduce genetic algorithm(Ga)to optimize the parameter set of LSTM network globally,the time-space relationship of multi-site and multi-feature is input into the LSTM network,and then the long-term feature dependence of multi-site and multi-feature is output.Finally,the MSCNN-GALSTM model was compared with the single LSTM reference model and the single scale convolutional neural network model.The root mean square error(RMSE)decreased by about 11%and the average prediction accuracy increased by about 20%.The results show that the MSCNN-GALSTM model has more comprehensive feature extraction,deeper level,higher prediction accuracy and better generalization ability.

关 键 词：多尺度卷积 空间特征 时空优化 LSTM 预测模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]