摘要
为了提升关联区域内VOCs浓度预测精度,基于深度学习理论构造了K-CNN-BiLSTM时空关联预测模型。同时,为了实现VOCs精细化治理,首先对研究区域进行了网格划分,采用IDW进行空间插值,计算整理得到VOCs的网格数据集。其次使用KNN算法计算空间相关性筛选得到空间相关矩阵,按照时序排列拼接成时空类图。然后将时空类图输入CNN模型中提取局部时空特征,最后将提取的时空特征送入双向LSTM中进行全局预测。以西安市某区为例,对VOCs浓度进行预测,并将预测结果进行时空分布可视化。结果表明:模型具备单步预测和多步预测能力,同时与CNN-BiLSTM、CNN-LSTM和LSTM相比考虑了VOCs浓度数据的时空关联性,预测精度更高;平均均方根误差(RMSE)、平均绝对值误差(MAE)和平均绝对百分比误差(MAPE)分别为6.352、5.442和10.252%,均优于对比模型。
To improve the prediction accuracy of VOCs concentration and realize the prediction of VOCs concentration in relevant regions,a spatio-temporal correlation prediction model of VOCs concentration based on K-CNN-BiLSTM is proposed based on depth learning theory.Firstly,the study area is divided into grids with a resolution of 1 km×1 km,and the inverse distance weighting method is used for spatial interpolation.The information of VOCs concentration data is collected from the whole region to achieve the fine control of VOCs.Secondly,the spatial correlation of the grid is calculated by the k-nearest neighbor algorithm,and the spatial correlation matrix of the target grid is obtained.The class diagram of spatio-temporal is obtained by splicing time series.A convolutional neural network is used to extract local spatio-temporal characteristics from spatio-temporal class diagrams.The spatio-temporal characteristics are passed to a bidirectional long and short-term memory neural network for global spatio-temporal prediction to obtain the predicted VOCs concentration.The prediction of VOC concentrations is carried out for an area of Xi city and the spatio-temporal distribution of the predicted results is visualized.The results show that the K-CNN-BILSTM model has single-step prediction ability and multi-step prediction ability,and can predict the changes of VOCs concentration 3h in advance.Meanwhile,compared with CNN-BiLSTM,CNN-LSTM and LSTM models,the prediction accuracy is higher by considering the spatio-temporal correlation of VOCs concentration data.The root mean-square error is 6.352,which is 1.979,3.089 and 3.972 lower than those of the three models respectively.The mean absolute value error is 5.442,which is 0.793,1.890 and 3.470 lower than those of the three models.The mean absolute percentage error is 10.252%,which is 2.09%,3.77%and 6.97%lower than those of the three models,respectively.Therefore,the K-CNN-BILSTM spatio-temporal correlation prediction model has good prediction ability and can master the changing
作者
黄光球
王瑞泽
陆秋琴
HUANG Guang-qiu;WANG Rui-ze;LU Qiu-qin(School of Management,Xi'an University of Architecture and Technology,Xi'an 710055,China)
出处
《安全与环境学报》
CAS
CSCD
北大核心
2023年第4期1336-1348,共13页
Journal of Safety and Environment
基金
国家自然科学基金项目(71874134)
陕西省自然科学基础研究计划项目(2019JZ-30)。
