This study proposes a hybrid forecasting framework that integrates Seasonal-Trend decomposition using LOESS (the following abbreviations are referred to as STL) with two time series models, TimesNet and TimeXer. To capture the complex periodic characteristics of the PM2.5 series, the original data are first decomposed into trend, seasonal and residual components via STL. The trend and seasonal components are then predicted using TimesNet, which maps one-dimensional time series into a two-dimensional representation to better model multi-scale periodicities and temporal dependencies. In parallel, the residual component is forecast using TimeXer, which incorporates exogenous variables to improve the modeling of endogenous dynamics. The final PM2.5 prediction is obtained by aggregating the forecasts of the three components. Experimental results demonstrate that the proposed STL-TimesNet-TimeXer model achieves high predictive accuracy, with an R2 of 0.969, MAE of 2.834, MSE of 19.063, and MAPE of 6.435. Comparative analyses against single-model baselines further confirm that STL-based decomposition significantly enhances forecasting performance, indicating that STL provides an effective and interpretable approach for modeling PM2.5 time series.
Xiong et al. (Fri,) studied this question.