基于一种样本卷积和交互网络模型的空气质量预测

Air quality forecasting based on a sample convolution and interaction network model

空气质量预测是治理和减轻空气污染的有效手段。为了提高预测精度,构建了一个新的空气质量预测模型,即样本卷积和交互网络(sample convolutional and interaction network,SCINet)模型。该模型由多个SCIBlock按照完全二叉树结构排列而成,通过翻转奇偶分裂重新排列生成一个新的序列,该结构能够更好地捕捉多变量大气污染物彼此间复杂的依赖关系和局部趋势。因为大气污染物监测数据具有季节性和随机性,所以使用两个SCINet进行叠加,既能扩大卷积运算的接受域,又能实现多分辨率分析。此外,通过模型深度及超参数调优,使其更加拟合空气质量时序数据特征,能够有助于提取目标变量的时间关系特征。最后,通过北京PM_(2.5)数据集和北京多站点空气质量数据集进行实证研究,结果表明,SCINet模型具有更高的预测精度,在短期预测中其均方根误差比对比模型中表现最佳的DAQFF模型减少了31.59%,在长期预测中减少了24.36%。

Air quality forecasting is an effective means of managing and mitigating air pollution.To enhance prediction accuracy,a new air quality prediction model,namely the sample convolutional and interaction network(SCINet),is introduced in this paper.The model is composed of multiple SCI-Blocks arranged in a complete binary tree structure.Whereafter,the time series is rearranged through flipping odd-even splits,and a new sequence is generated,which is able to capture the complex dependencies and local trends of multivariate atmospheric pollutants better.Given the seasonality and randomness of monitoring data for atmospheric pollutant,the paper employs two SCINets for stacking,which not only expands the receptive field of convolutional operations,but also enables multi-resolution analysis.Furthermore,through the optimization of model depth and hyperparameters,the model may fit the temporal characteristics of air quality time series data better,which is helpful to extract the temporal relationship features of the target variable.In the end,the Beijing PM_(2.5) dataset and the Beijing multi-site air quality dataset are utilized to evaluate the effectiveness of SCINet.The results show that SCINet has higher prediction accuracy,whose the root mean square error(δRMSE)is reduced by 31.59%in short-term prediction and 24.36% in long-term prediction compared with the best-performing DAQFF model.