A Novel PM2.5 Prediction Model Combining Adaptive Variational Mode Decomposition And Multivariate Temporal Graph Neural Network Architecture

PM2.5 refers to the inhalable airborne pollutant particles with diameters less than or equal to2.5μm.These particles are capable of harming the human respiratory system and cardiovascular system,as well as for spreading viruses.The above potential hazards have been prevented through PM2.5 forecasting.The PM2.5 prediction is of great significance to public health.Affected by factory operation and traffic travel factors,PM2.5 is nonlinear and non-stationary,resulting in low PM2.5 prediction accuracy.Most of the existing PM2.5 prediction models do not consider the extraction of correlation features in multivariate inputs such as PM10 and CO related to PM2.5,and it is difficult to guarantee the PM2.5 prediction effect under multiple time steps.A PM2.5prediction model based on Adaptive Variational Mode Decomposition(AVMD)and Multivariate temporal Time Graph Neural Network(Mtem GNN)is proposed to solve above problems.The main contents are as follows:(1)Due to the nonlinearity and nonstationarity of PM2.5,it is difficult for ordinary models to extract the internal features of PM2.5.An AVMD algorithm is proposed to automatically decompose PM2.5 into a set of stationary sub-modalities.AVMD automatically finds the optimal decomposition parameter combination of Variational Mode Decomposition(VMD)according to data characteristics,which perfectly solves the problem that the number of modes K and the strength of bandwidth constraint α are difficult to determine in VMD decomposition.AVMD decomposes PM2.5 into a set of sub-modes with stable and obvious frequency domain characteristics,which greatly reduces the difficulty of PM2.5 prediction.(2)Considering that the traditional Euclidean distance cannot represent the correlation characteristics among PM2.5 and PM10,CO and other related pollutants,a graph neural network is introduced to construct a correlation graph structure of non-Euclidean distance.A self-attention mechanism is used to automatically calculate the similarity among PM2.5 and other air pollutants and build a correlation graph structure based on this,and the graph convolutional network is used to further learn the graph structure features.The use of graph neural network effectively extracts the correlation features among multivariate input variables,which is beneficial to the improvement of PM2.5 prediction accuracy.(3)Considering the long-term correlation of PM2.5,the long-term PM2.5 data in the past affects the future prediction results,so the Gated Recurrent Unit(GRU)is used to learn the longterm memory characteristics of PM2.5.GRU extracts long-term and short-term time series features of time series,solves the problems of gradient disappearance and gradient explosion in long-term series training,and reduces unnecessary computation.The long-range correlation of PM2.5extracted by GRU further improves the prediction accuracy of PM2.5,making the model more suitable for predicting PM2.5 with multiple time steps.The PM2.5 prediction model proposed in this thesis solves the problem that PM2.5 is difficult to accurately predict.Experiments show that the prediction model proposed in this paper is far better than the existing baseline models on the real Beijing dataset.The proposed model predicts the PM2.5 in the next 12 hours with high accuracy,which can not only provide good air pollution protection advice for travelers but also help the government in air pollution control.