Multi-Source PM2.5 Prediction Model Based on Fusion of Graph Attention Networks and Multiple Time Periods

Aiming at the problem that the traditional time series prediction model only considers a single node (region), does not take into account the spatial interactivity among multiple nodes and the cycle characteristics embedded in the time series data, and has low accuracy in the task of predicting the spatio-temporal sequences of multiple sources, this study proposes a feature extraction prediction model GMC (GAT-MULCYCLE). The model is designed to cope with the accuracy of complex prediction problems characterized by both spatial correlation and temporal periodicity (e.g., multi-site PM2.5 prediction). In this study, spatial correlation is first extracted using GAT to dynamically focus on the contribution of different neighboring nodes. Then, focusing on the multiple cycles present in the time series, the extracted features are fused for final prediction. Comparison tests with 10 other related models in the PM2.5 prediction task in three cities, Beijing, Shenyang and Qingdao, show that compared with the baseline model with the best prediction results, our proposed method reduces the average of the two evaluation metrics (Mean Squared Error MSE and Mean Absolute Error MAE) by (9.50% and 8.87%). It shows that GMC has smaller error and accurate prediction among the same type of models, which can extract the spatio-temporal features of sequence data more accurately and is more suitable for the prediction task of multi-source time series data.