融合图卷积神经网络和注意力机制的PM2.5小时浓度多步预测

PM_2.5小时浓度多为单步预测。为实现PM_2.5小时浓度的多步预测，基于“编码器-解码器”的序列-序列预测（Seq2Seq）模型，集合图卷积神经网络提取非欧式空间数据特征的能力以及注意力机制自适应关注特征的能力，提出了融合图卷积神经网络和注意力机制的PM_2.5小时浓度多步预测（GCN_Attention_Seq2Seq）模型。并与Seq2Seq模型和使用了图卷积神经网络、未使用注意力机制的GCN_Seq2Seq模型进行了对照，以2015—2016年北京市22个空气质量监测站点的空气质量数据为样本进行实例验证，结果表明，Seq2Seq模型和图卷积神经网络（GCN）可对PM_2.5小时浓度数据的时空依赖进行有效建模，注意力机制有助于减缓多步预测中的预测精度衰减，提升PM_2.5小时浓度多步预测的精度。GCN_Attention_Seq2Seq模型可有效应用于多种长度的PM_2.5浓度预测窗口。

Multi-step prediction of PM2.5 hourly concentration by fusing graph convolution neural network and attention mechanism

The current studies about PM_2.5 hourly concentration prediction are mostly on single-step prediction.In order to achieve accurate prediction of PM_2.5 hourly concentration at multiple moments in a single prediction task,this article proposes a multi-step prediction model of PM_2.5 hourly concentration based on graph convolution neural network and attention mechanism,which is named GCN_Attention_Seq2Seq.The model based on Seq2Seq is able to extract the features of non-euclidean spatial data meantime pays attention to features adaptively.We take air quality data of 22 monitoring stations in Beijing from January 1st,2015 to December 29th,2016 as samples and compare GCN_Attention_Seq2Seq with GCN_Seq2Seq and Seq2Seq model. Results show that Seq2Seq and GCN can model spatio-temporal dependence effectively and the attention mechanism is helpful to improve the prediction accuracy and slow down the prediction accuracy decline in multi-step prediction,it indicates that the GCN_Attention_Seq2Seq model can be effectively applied to multi-step prediction of PM_2.5 concentration.