基于多尺度时空优化的空气质量预测方法

Multi-Scale Spatial-Temporal Optimization Method for Air Quality Prediction

本文提出一种基于多尺度时空优化的空气质量预测方法(multi-scale spatial-temporal network for air quality prediction,MSSTN-AQP),结合空气质量系统中存在的长短期时间依赖关系和动态空间依赖性,提高长期空气质量预测的准确性。首先,通过构建多尺度时空特征提取模块,从多源异构数据中提取时空特征。其次,构建动态空间特征提取模块。通过将图卷积网络与注意力机制进行有效结合,捕捉空气质量网络中的全局空间特征,用于对多种空间依赖关系的联合建模。最后,构建时间特征提取模块,对Transformer模型进行改进与优化。自适应时间Transformer模块主要用于模拟跨多个时间步长的双向时间依赖关系。此外,将上述时空特征提取模块进行有效集成化,构建端到端的空气质量预测模型。为了验证模型的有效性,在两个真实数据集中进行实验验证。实验结果表明,MSSTN-AQP在预测精度上更具优势,尤其是在长期的空气质量预测任务中优势更加明显。

In this article we propose an air quality prediction method based on Multi-Scale Spatial-Temporal Network for air quality prediction(MSSTN-AQP)to improve the accuracy of long-term air quality prediction by combining the long-and short-term time dependence and dynamic spatial dependence in the air quality system.First,the spatial-temporal features are extracted from multi-source heterogeneous data by constructing a multi-scale spatial-temporal feature extraction module.Second,the dynamic spatial feature extraction module is constructed.With an effective combination of the graph convolutional network with the attention mechanism,the global spatial features in the air quality network are captured and used for the joint modeling of multiple spatial dependencies.Finally,the temporal feature extraction module is constructed,which is to improve and optimize the Transformer model.The adaptive time Transformer module is mainly used to simulate bidirectional time dependencies across multiple time steps.Moreover,the above temporal feature extraction module is effectively integrated to construct an end-to-end air quality prediction model.To verify the effectiveness of MSSTN-AQP,extensive experiments were conducted on two real data sets.The experimental results showed that MSSTN-AQP was more advantageous in prediction accuracy,especially in long-term air quality prediction tasks.