基于时空深度学习模型的PM_(2.5)预测

PM_(2.5)Prediction Based on Spatio-Temporal Deep Learning Model

伴随着快速城市化进程,空气污染尤其是PM_(2.5)严重影响着人们的身体健康,精准的空气质量预测能够为空气污染防治以及政府决策提供有力支撑。针对当前空气质量预测研究中存在的问题,包括缺失数据填充,时空特征信息提取等,提出一种基于三维卷积神经网络和长短时记忆神经网络构建的时空混合深度学习模型C3D-LSTM。模型通过三维卷积模块对时空维度上的特征信息进行联合提取,并利用长短时记忆网络学习长时间序列数据的能力,预测目标站点的PM_(2.5)的浓度。基于北京市22个站点的真实数据集进行实验,结果表明,所提模型在平均绝对误差、均方误差和拟合系数三种指标方面均优于其它基准空气质量预测模型。

With the rapid urbanization process,air pollution especially PM_(2.5)severely affects people's health.Predicting air quality accurately can provide substantial support for air pollution prevention and governments'policymaking.Aiming at the problems in current air quality prediction research,including missing data imputation,spatiotemporal feature extraction,a spatiotemporal deep learning model named C3D-LSTM is proposed based on three-dimensional convolutional neural network and long and short-time memory network,which extracts the features in the temporal and spatial dimensions using three-dimensional convolution module,learns the long-term temporal dependency using long and short time memory network and then predicts the PM_(2.5)concentration of the target station.The experiments are conducted based on the real dataset from 22 stations of Beijing city and the results show that the proposed model is superior over other baseline models on the metrics including mean absolute error,mean square error,and the fitting coefficient.