Hybrid framework for improved PM2.5 prediction based on seasonal-trend decomposition and tailored component processing.

Precise prediction of PM2.5 concentrations is crucial for effective environmental management and the safeguarding of public health. However, accurately forecasting PM2.5 levels presents significant challenges due to the complex, non-linear, and non-stationary characteristics inherent in the PM2.5 concentration data, which are influenced by diverse meteorological conditions, emission sources, and atmospheric transport dynamics. Existing prediction methods often struggle to adequately capture these multifaceted patterns simultaneously; single models may fail to address both long-term trends and seasonal variations as well as short-term stochastic fluctuations, while many hybrid decomposition approaches may not optimally utilize techniques suited to the distinct nature of each decomposed component, thereby limiting potential accuracy improvements. To overcome these limitations, this paper proposes HISTCP, a novel hybrid framework. HISTCP leverages Seasonal-Trend decomposition using LOESS (STL) to initially separate the PM2.5 series into trend, seasonal, and residual components. Then, specific processing techniques are applied based on the informational characteristics of the different components. The framework's superior performance and robustness were demonstrated through rigorous experiments on PM2.5 datasets from five diverse Chinese cities, outperforming several baseline and state-of-the-art models across MSE, MAPE, MAE and R