基于大数据的城市空气质量时空预测模型

Research on Spatial-temporal Forecasting Model for Urban Air Quality Monitoring Based on Large Data Analysis

针对常用的基于数据驱动的空气质量预测方法只考虑当地站点时序特性的缺陷问题,提出一种时空特性的空气质量预测算法,通过长短期记忆网络(Long Short-term Memory,LSTM)构成SequencetoSequence范式处理时间序列的变长输入与输出,获取本地PM2.5时间序列规律;通过卷积神经网络(CNN)获取站点间的空间相关性及特征的进一步抽象,结合两类网络产生从结构上反映时间与空间相关性的预测结果。对比实验的结果表明,LSTM-CNN混合模型在公开数据集上,与神经网络、回归树以及简单的LSTM系列模型相比,取得了更好的预测效果,证实所提算法的优越性能。

Commonly used data-driven air quality prediction method only considers the temporal features of the local station.This paper proposes an air quality prediction algorithm with spatial and temporal characteristics.The Sequence to Sequence paradigm is used to obtain the temporal feature of the local PM2.5.It is composed of LSTM,and it can handle variable-length input and output of time series.Convolutional neural network(CNN)is used to obtain the spatial correlation between stations and the high-dimensional representation of features.The hybrid network is a combination of two types of networks.Its prediction essentially reflects the temporal and spatial correlation.A comparative experiment is conducted on public datasets.Compared with neural networks,regression trees,and simple LSTM series models,the LSTM-CNN hybrid model achieves better prediction results.