Temporally boosting neural network for improving dynamic prediction of PM2.5 concentration with changing and unbalanced distribution

Increasing medical research evidence suggests that even low PM2.5 concentrations may trigger significant health issues. Hence, an accurate prediction of PM2.5 holds immense significance in securing public health safety. However, current data-drive predictive methods exhibit seasonal model performance decline and difficulties in predicting extremely high values. Those issues may stem from neglecting two crucial features in PM2.5 data streams, i.e., concept drift and imbalanced distribution. In this study, we validate this hypothesis by conducting an in-depth analysis of the characteristics of the PM2.5 data stream and the prediction errors of three mainstream models trained on this PM2.5 data stream, i.e., random forest, convolutional neural network and transformer. Based on the identified types of concept drift and the patterns of imbalanced distribution, we introduce the Temporally boosting neural network (Temp-boost), a novel ensemble learning method designed to enhance predictive accuracy by integrating static and dynamic models. Static models, which are trained on balanced historical datasets, typically receive infrequent updates. Conversely, dynamic models are trained on newly arrived data and undergo more frequent updates. We evaluated the performance of Temp-boost and the three mentioned models in predicting gridded PM2.5 concentrations across the North China Plain in 2019. Compared to the three models, the Temp-boost shows improved prediction accuracy for different seasons, with notable enhancements in high-pollution levels. Specifically, for pollution levels above lightly polluted, the Temp-boost effectively reduces the average MAE by 13.22 mu g m-3, RMSE by 13.32 mu g m-3 , with reductions peaking MAE at 26.45 mu g m-3,RMSE at 25.76 mu g m-3 in more severe case.