An Ensemble Model with Adaptive Variational Mode Decomposition and Multivariate Temporal Graph Neural Network for PM2.5 Concentration Forecasting

Abstract

Accurate prediction of PM2.5 concentration for half a day can provide valuable guidance for urban air pollution prevention and daily travel planning. In this paper, combining adaptive variational mode decomposition (AVMD) and multivariate temporal graph neural network (MtemGNN), a novel PM2.5 prediction model named PMNet is proposed. Some studies consider using VMD to stabilize time series but ignore the problem that VMD parameters are difficult to select, so AVMD is proposed to solve the appealing problem. Effective correlation extraction between multivariate time series affects model prediction accuracy, so MtemGNN is used to extract complex non-Euclidean distance relationships between multivariate time series automatically. The outputs of AVMD and MtemGNN are integrated and fed to the gate recurrent unit (GRU) to learn the long-term and short-term dependence of time series. Compared to several baseline models—long short-term memory (LSTM), GRU, and StemGNN—PMNet has the best prediction performance. Ablation experiments show that the Mean Absolute Error (MAE) is reduced by 90.141%, 73.674%, and 40.556%, respectively, after adding AVMD, GRU, and MtemGNN to the next 12-h prediction.