Environmental Forecasting of Particulate Matter (PM 2.5 ) Using Seasonal-Trend Decomposition and iTransformer: A Multivariate Learning Framework

Accurate PM 2.5 forecasting is crucial for timely countermeasures and policy formulation. However, existing univariate PM 2.5 forecasting methods often exhibit limited accuracy. This study proposes a multivariate forecasting approach based on the iTransformer framework, enhanced with Seasonal-Trend Decomposition using Loess (STL). STL decomposes time series into seasonal and trend components, effectively capturing periodic fluctuations and long-term patterns while improving model interpretability. The iTransformer employs a multiattention mechanism to capture complex dependencies among variables, modeling their interactions while preserving temporal characteristics. We evaluated the proposed method using 11 years of pollutant data (PM 2.5 , PM 10 , and SO 2 ) from Yantai City and Beijing, comparing its performance with established methods including Transformer and DLinear. The proposed method achieved a mean absolute error of 0.306, mean square error of 0.234, and R 2 of 0.787, demonstrating superior accuracy. These results highlight its potential for issuing public health alerts, evaluating emission control policies, and supporting data-driven environmental management decisions.