摘要
近年来,空气污染问题备受关注,近地面臭氧逐渐成为我国部分城市的首要污染物,因此对臭氧浓度的精准预测尤为重要.为了进一步提高臭氧浓度预测的精度,提出一种融合空间特征和统计特征的卷积神经网络和门控循环单元(Convolutional Neural Networks and Gate Recurrent Unit,CNN-GRU)臭氧浓度组合预测模型.首先,通过对时空因素以及其他大气污染物与臭氧浓度进行相关性分析,利用基于统计域的方法和克里金插值法对臭氧浓度时序数据进行预处理来提取臭氧浓度数据的时空特征,采用并联杂交CNN和GRU结构的组合预测模型得到最终的臭氧浓度预测结果.实验结果表明,CNN-GRU组合预测模型预测未来一小时的臭氧浓度可决系数、均方根误差和均方误差的值分别为0.9598,11.9508和8.2753,未来两小时的臭氧浓度可决系数、均方根误差和均方误差的值分别为0.8985,18.5373和13.0045,优于独立的CNN、长短期记忆(LongShortTermMemory,LSTM)网络、GRU、卷积-长短期记忆网络(Convolutional LSTM Network,ConvLSTM)、CNN-LSTM和CNN-GRU预测模型,这是由于CNN-GRU组合预测模型融合了空间和统计特征,可以多角度提取特征并采用并联杂交的网络结构,所以预测精度较高,且具备较好的鲁棒性.
In recent years,air pollution has attracted much attention and near ground ozone gradually becomes the primary pollutant in some cities in China.Therefore,accurate prediction of ozone concentration is particularly important.In order to improve the prediction accuracy of ozone concentration,In this paper,we propose a combined prediction model of ozone concentration based on CNN-GRU(Convolutional Neural Network and Gate Recurrent Unit)by fusing spatial and statistical features.Firstly,by analyzing the correlation between temporal and spatial factors as well as other atmospheric pollutants and ozone concentration,the statistical domain based method and Kriging interpolation method are used to preprocess the ozone concentration time series data to extract the spatial-temporal characteristics of the ozone concentration data.The combined prediction model of parallel hybrid CNN(Convolutional Neural Networks)and GRU(Gate Recurrent Unit)structure is used to obtain the final ozone concentration prediction results.The results show that the combined prediction model of ozone concentration based on CNN-GRU has the highest accuracy and the best prediction effect,the values of R2,RMSE(Root Mean Squared Error)and MAE(Mean Absolute Error)of ozone concentration in the next hour are 0.9598,11.9508 and 8.2753,in the next two hours are 0.8985,18.5373 and 13.0045,outperform the independent CNN,LSTM(Long Short-Term Memory),GRU,ConvLSTM(Convolutional LSTM Network),CNN-LSTM and CNN-GRU prediction models.Because the prediction model combines spatial and statistical features,multi angle feature extraction and parallel hybrid network structure are used to obtain high prediction accuracy and good robustness.
作者
杨雨佳
肖庆来
陈健
曾松伟
Yujia Yang;Qinglai Xiao;Jian Chen;Songwei Zeng(College of Mathematics and Computer Science,Zhejiang A&F University,Hangzhou,311300,China;Songyang County Natural Resources and Planning Bureau,Lishui,323400,China;State Key Laboratory of Subtropical Silviculture,Zhejiang A&F University,Hangzhou,311300,China;College of Forestry and Biotechnology,Zhejiang A&F University,Hangzhou,311300,China;College of Optical Mechanical and Electrical Engineering,Zhejiang A&F University,Hangzhou,311300,China)
出处
《南京大学学报（自然科学版）》
CAS
CSCD
北大核心
2023年第2期322-332,共11页
Journal of Nanjing University（Natural Science）
基金
国家自然科学基金(41471442)。
关键词
臭氧浓度预测
卷积神经网络
门控循环单元
空间特征
统计域
ozone concentration prediction
Convolutional Neural Network
Gated Recurrent Unit
spatial feature
statistical domain
